The next step is experimental. Knock each gene down in the relevant cell type and measure the transcriptional response. See whether the disease program shifts. Then decide which target deserves deeper validation. A virtual cell model moves part of that workflow into computation. Given a starting cell state and a perturbation, it predicts what the cell is likely to do next.

That is the useful claim. A virtual cell is not a whole human in software or a substitute for biology. It is a learned model of cellular state transitions, usually trained on single-cell atlases plus perturbation data, and used to rank experiments before the lab runs them.

For target discovery, that is enough. The field has spent decades asking whether a target matters. Virtual cells recast that question in intervention terms by asking what happens when a gene is reduced, a pathway is blocked, or a cell population sees a drug.

Why virtual cells matter in 2026

Target discovery already asks whether a target is linked to disease, expressed in the right tissue, and supported by network or multi-omics evidence. Virtual cells pick up at the next step. Once a target looks plausible on paper, the practical question is how a relevant cell state changes when the team perturbs it.

Three changes made that question computationally serious. Single-cell atlases are now large enough to serve as training corpora. Perturbation datasets have moved from boutique experiments to industrial scale. Public benchmarks are finally testing whether models generalize instead of rewarding polished demos.

First, single-cell scale is no longer hypothetical. Public atlases contain tens to hundreds of millions of cells, and platforms such as CZ CELLxGENE make those data usable as model-training corpora. Models such as Geneformer, scGPT, scFoundation, AIDO.Cell, Universal Cell Embeddings, and TranscriptFormer treat gene-expression profiles as structured biological objects that can be embedded, compared, and transferred across tissues or species.

Second, perturbation data now exist at sizes that can train models rather than merely illustrate them. Perturb-seq combines CRISPR-based perturbation with single-cell RNA sequencing, so each perturbed cell carries both the intervention label and the transcriptomic readout. Genome-scale Perturb-seq has already been run across millions of human cells.

Drug-perturbation datasets are larger still. Tahoe-100M contains 100 million single cells across 50 cancer models and 1,100 drug perturbations, and the 2026 Tahoe, Arc Institute, and Biohub partnership announced a plan to generate more than 120 million cells across 225,000 drug-patient interactions (Tahoe Therapeutics, 2025 and 2026).

Third, the field now faces real benchmark pressure. In 2025, the Arc Institute launched the Virtual Cell Challenge to test whether models can predict transcriptional responses to perturbations in a held-out cell type. The first round did not produce an AlphaFold-style breakaway win. On some metrics, many models struggled against naive baselines. That makes the benchmark more useful, not less. It marks where the field is ready for deployment and where it is still doing research.

Concept Translation: A virtual cell is a learned transition model for biology. The input is a cell state plus an intervention. The output is a predicted next state or readout. In reinforcement-learning terms, the cell is the state, the perturbation is the action, and the transcriptomic or phenotypic response is the transition. The catch is observability. RNA counts are a narrow sensor, not the whole cell.

That ambiguity matters because target discovery does not return a clean software object called target. It returns a probabilistic, context-dependent intervention hypothesis. Virtual cells do not remove that uncertainty. They make it easier to interrogate.

What a virtual cell actually predicts

The phrase “virtual cell” covers several model classes. In drug-discovery writing, it is often stretched beyond its useful meaning.

A virtual cell model predicts cellular behavior under conditions that were not directly measured. The condition might be a gene knockout, gene knockdown, CRISPR activation, small-molecule exposure, cytokine treatment, disease context, species transfer, or a combination. The output might be a transcriptome, a protein-expression profile, an image-derived morphology vector, a viability phenotype, a pathway score, or a cell-state label.

Near-term work still predicts RNA most of the time. That is a convenience rather than the endpoint biology cares about. Single-cell RNA-seq is high-throughput, broadly available, and already organized into public atlases. It also gives a rich readout of cell state, but the tradeoff is clear. Drug targets are usually proteins, safety phenotypes can be electrophysiological or immunological, and clinical endpoints sit far downstream. In practice, virtual-cell models operate as assay models and sit well upstream of clinical-response modeling.

In practice, the workflow is direct.

1. Start with disease-relevant cells, ideally patient-derived cells or a strong surrogate.
2. Encode the baseline cell state, including cell type, disease state, and any available molecular context.
3. Specify a perturbation, such as knocking down a candidate target or exposing the cell to a tool compound.
4. Predict the response, often as differential expression, pathway activation, or movement in a learned cell-state space.
5. Rank perturbations for wet-lab validation.
6. Feed the validation data back into the model or the next screen.

That loop is much narrower than “simulate biology.” It is active learning for target validation.

Three model layers show up repeatedly.

Cell-state encoders. These models map a single-cell profile into an embedding. The embedding can support cell-type annotation, disease-state classification, batch correction, cross-species transfer, and clustering. Geneformer, scGPT, scFoundation, Universal Cell Embeddings, AIDO.Cell, and TranscriptFormer mostly live here, although several also support perturbation tasks.

Perturbation response predictors. These models try to predict how gene expression changes after an intervention. CPA, GEARS, scGen, scGenePT, TxPert, STATE, and newer virtual-cell systems belong here. GEARS is a useful reference point because it predicts transcriptional outcomes of multi-gene perturbations using a graph of gene relationships, which makes the link to knowledge graphs explicit.

Multimodal virtual instruments. The 2024 Cell virtual-cell perspective uses the phrase “virtual instruments” for models that expose the cell through different readouts. One instrument may predict RNA, another protein localization, another imaging morphology, and another pathway response. That matters because target discovery rarely fails on RNA alone. It fails when the downstream phenotype does not follow.

The data substrate for virtual cells

Virtual cells are data-hungry because cells have many degrees of freedom. A mammalian cell has tens of thousands of transcripts, thousands of proteins, many possible states, and many interventions that can push it around. Scaling a language model mostly means collecting more text. Scaling a virtual cell means collecting more cells under controlled perturbations.

The data come from five main sources.

Reference atlases. CZ CELLxGENE, Tabula Sapiens, the Human Cell Atlas, GTEx-derived resources, organoid atlases, and disease-specific single-cell collections provide baseline cell states. They teach the model what a cell type looks like across tissues, donors, diseases, and species. Their best use is representation learning.

Genetic perturbation screens. Perturb-seq, CRISPRi, CRISPRa, and CROP-seq-style methods create labeled intervention data. Replogle and colleagues showed genome-scale Perturb-seq across more than 2.5 million human cells, using CRISPRi to target expressed genes and read out transcriptional phenotypes. This is the core signal for predicting what happens when a gene is reduced or activated.

Chemical perturbation screens. Drug perturbations connect target biology to pharmacology. They are messier than genetic perturbations because drugs have dose, exposure time, uptake, polypharmacology, metabolism, and off-target activity. They are also closer to the real question in drug discovery. A target perturbation asks what happens when the target changes. A drug perturbation asks what happens when a molecule enters a cell and hits whatever it actually hits.

Imaging and spatial data. RNA readouts lose morphology and tissue architecture. Image-based profiling, spatial transcriptomics, subcellular localization data, and multiplexed protein imaging put some of that structure back into the model. This is where virtual cells begin to connect with tissue specificity and on-target, off-tissue safety.

Clinical and patient-derived context. Patient-derived organoids, ex vivo tumor samples, disease tissue atlases, and electronic health record-linked molecular data are the route from cell models toward patient stratification. This layer is thin compared with public atlas data. It is also the layer that matters most for translation.

Scale alone is not enough. Coverage matters more than raw cell count. A hundred million cells from a narrow set of cancer cell lines will not teach the model how a macrophage behaves in fibrotic lung tissue. A perturbation-rich dataset can teach intervention response in one assay and still fail on another platform, donor population, or disease model.

This ambiguity is also a data problem. A cell-line perturbation is not the same object as a patient-tissue perturbation. A transcriptional response is not the same object as a clinical response. A virtual cell learns a conditional distribution over what it was shown. When the deployment setting changes, the model has to show that it transfers.

Concept Translation: Scaling a virtual cell is closer to scaling a robotics dataset than scaling a language model. Text-based scaling laws exploit the fact that text is nearly free to collect at internet scale. Cellular perturbation data has to be generated experimentally - every additional perturbation × cell-type × dose × timepoint combination is a wet-lab experiment. The cost curve, the coverage problem, and the data-quality concerns are closer to robotics or autonomous-driving data collection than to LLM pretraining. That asymmetry shapes which scaling strategies actually pay off here.

Model families behind virtual cells

Current models fall into four broad families.

Foundation models for cell representations

Geneformer brought the language-model analogy into a form many ML readers immediately recognize. It treats a cell as an ordered sequence of expressed genes and pretrains a transformer over millions of single-cell transcriptomes. The original corpus covered roughly 30 million human single-cell transcriptomes. Later scaling work expanded the corpus beyond 100 million transcriptomes and reported scaling behavior across model sizes.

scGPT follows a related transformer approach, training on more than 33 million cells and supporting tasks such as batch correction, cell-type annotation, multi-omics integration, gene-network inference, and perturbation prediction. scFoundation scales in a different architectural direction, using a large model trained on more than 50 million cells with whole-transcriptome input space. TranscriptFormer extends the idea across species, with CZI reporting training on 112 million cells from 12 species and downstream use for cross-species cell-type and disease-state tasks.

Concept Translation: Treating a cell as an ordered sequence of genes is the move that makes the rest of the foundation-model toolkit drop in. In NLP, a sentence is an ordered sequence of tokens; the transformer learns contextual embeddings by predicting masked tokens or next tokens given context. In Geneformer and scGPT, a cell is an ordered list of its expressed genes (typically ranked by expression level), and the transformer learns contextual gene embeddings by predicting masked genes given the rest of the cell. The architecture is the same. What changes is the vocabulary (genes instead of tokens), the corpus (single-cell transcriptomes instead of text), and the downstream tasks (cell-type classification, perturbation prediction). The same scaling arguments and pretraining intuitions apply, with the data-cost asymmetry above as the main caveat.

These models matter even when they fall short on perturbation prediction. Strong cell embeddings can surface rare disease states, align patient samples with reference atlases, and provide features for downstream target-prioritization models. That alone is already useful in target discovery.

Perturbation predictors

Perturbation models sit closest to the virtual-cell promise. CPA learns a latent representation that separates basal cell state from perturbation effects. GEARS uses graph neural networks over gene relationships to predict responses to novel multi-gene perturbations. scGenePT injects gene-level language embeddings into a single-cell perturbation architecture. Tahoe-x1 trains a billion-parameter-scale model on perturbation-rich single-cell data, with explicit attention to genes, cells, and drugs.

The output is usually a vector of expression changes. That makes evaluation awkward. One model may capture the broad cell-state shift while missing a few disease-relevant genes. Another may rank the differentially expressed genes well while giving poor absolute values. A third may do well on global averages by predicting something close to the mean cell. A benchmark has to say which failure matters for the use case.

Graph and knowledge-guided models

Virtual-cell prediction is also a graph problem. Genes already sit inside regulatory, signaling, and protein-interaction networks. Perturb one gene and the effect propagates through that graph. The knowledge-graph article made the same point at the level of target nomination. Here the same logic operates inside the cell.

GEARS is the clearest example for target-discovery readers because it uses a gene-gene knowledge graph to generalize from observed perturbations to unseen combinations. Graph-aware models such as GREmLN encode regulatory structure into attention mechanisms. Knowledge-guided world models such as VCWorld push further, trying to produce stepwise mechanistic predictions rather than only final output vectors.

The open question is how much explicit biological structure still helps once datasets become very large. In chemistry, learned representations and physics priors now coexist. Virtual cells are likely to settle into the same pattern. Scaling helps, but models that ignore known biology still throw away signal.

Generative and world-model approaches

Some of the most ambitious virtual-cell papers describe world models that simulate cellular trajectories under intervention. Flow matching, diffusion-style generation, state-space models, and latent dynamical systems all point in that direction. The aim is to represent a path through cell states rather than a static expression vector.

Concept Translation: “World model” is borrowed from reinforcement learning and robotics, where it means a learned dynamics model the agent can roll out in simulation to plan or imagine consequences before acting in the real environment. Used for cells, it means a model that predicts how cellular state evolves under a perturbation across time, not just what the endpoint looks like. The same questions that come up for RL world models - does it predict accurate trajectories, does it generalize to unseen actions, does it stay calibrated under longer rollouts - come up here too. The drug-discovery payoff is that the resulting model can be queried about dose timing, exposure duration, combination order, and resistance trajectories, not just final endpoints.

This matters for drug discovery because many therapeutic effects are temporal. A target may look safe at 6 hours and toxic at 6 days. A gene knockdown may shift a disease marker quickly, while the functional phenotype follows later. A virtual cell that predicts only one endpoint has limited reach. A virtual cell that predicts trajectories could guide dose, timing, combination order, and resistance hypotheses.

In 2026, “world model” often means “a model trained to predict perturbation-conditioned cell-state distributions.” It is not a full simulator of cellular physiology.

Benchmarks for virtual cells

Benchmarks are now the most useful part of the field because claims have outrun evidence.

The 2025 Virtual Cell Challenge framed the test in concrete terms. A computational model was asked to stand in for a Perturb-seq experiment. The challenge used approximately 300,000 single-cell RNA-seq profiles from H1 human embryonic stem cells with 300 CRISPRi perturbations. H1 cells created a distribution shift from common training data such as K562 and A375. That design forced context generalization rather than memorization.

Concept Translation: The H1 vs K562/A375 design is the standard out-of-distribution generalization test from ML, applied to cell biology. K562 and A375 are the workhorse cell lines that dominate Perturb-seq training data - they are the in-distribution majority. Holding out H1 forces the model to predict perturbation responses in a cell type it has never seen perturbed. A model that succeeds is doing real generalization across cell-type domain shift; a model that fails is doing memorization of the training distribution. This is the same evaluation discipline that distinguishes a vision model that generalizes from one that overfits to ImageNet’s specific photographic conventions.

The evaluation used three primary metrics.

Perturbation Discrimination Score (PDS) measured whether predicted perturbation effects were distinguishable from each other.

Differential Expression Score (DES) measured whether models identified the correct upregulated and downregulated genes.

Mean Absolute Error (MAE) measured gene-level prediction accuracy across the transcriptome.

The results were sobering in a productive way. Arc’s wrap-up reported that perturbation prediction models were not consistently outperforming naive baselines across all metrics, although they improved on perturbation discrimination and differentially expressed gene identification. Winning approaches mixed deep learning with classical statistical features. Pure end-to-end learning did not settle the problem.

A Nature Methods benchmark reached a similar caution from a different angle. Ahlmann-Eltze, Huber, and Anders compared five foundation models and two other deep learning models against simple baselines for predicting transcriptome changes after single or double perturbations. None outperformed the baselines in their setup. The result shows that perturbation prediction is harder than cell embedding, and model papers need baselines simple enough to embarrass weak claims.

Concept Translation: A naive baseline is the negative control for a model paper. If a transformer cannot beat “predict no change” or “add the two single perturbation effects,” it has not learned the biological interaction the paper claims. The same discipline applies in docking against decoys or in testing a classifier against a majority-class predictor.

For target discovery, a benchmark should answer practical questions.

• Prediction of perturbation effects in a new cell type.
• Prediction for a new perturbation target when related genes were seen during training.
• Prediction of combinations where the single perturbations were seen but the pair was not.
• Recovery of disease-relevant differentially expressed genes rather than only the largest global shift.
• Prediction of functional endpoints such as viability, cytokine release, differentiation, or drug sensitivity.
• Uncertainty estimates calibrated well enough to guide experimental triage.

A target-discovery benchmark that cannot answer those questions is a model-quality benchmark, not a decision benchmark.

That is why AlphaFold analogies still overreach. The field looks closer to CASP before the jump, when the community was still deciding what useful prediction meant.

How virtual cells change target discovery

Virtual cells add a new layer to the target-validation stack. They do not replace evidence frameworks, tissue-specificity analysis, druggability assessment, or knowledge graphs. They carry those pieces into a perturbation forecast.

Candidate-target triage

A knowledge graph can nominate a target-disease edge. Human genetics can make that edge more credible. Tissue specificity can warn that modulation may be unsafe. A virtual cell adds the intervention question by predicting response in a relevant cell context.

Suppose a target is upregulated in diseased fibroblasts and sits near a fibrosis pathway in a knowledge graph. A virtual-cell workflow can simulate knockdown in disease-like fibroblast states and rank whether the predicted response moves away from the fibrotic program. That is not proof. It is a cheaper way to decide which targets deserve Perturb-seq, organoid testing, or animal studies.

Screen design

Perturbation experiments are expensive because the search space is large. There are thousands of genes, many cell types, multiple disease contexts, and long lists of chemical interventions. Virtual cells can help choose the next experiment.

This is active learning in practice. The model proposes perturbations with high expected information gain or high expected therapeutic relevance, the lab runs the selected experiments, and the new data update the model. The aim is to cut wasted assays rather than chase perfect accuracy.

Biomarker, combination, and resistance hypotheses

A target only matters in the patients whose disease biology depends on it. Virtual cells can test whether a perturbation response is specific to a cell state, genotype, subtype, or disease stage. If predicted target knockdown reverses a disease program only in cells carrying a particular pathway signature, that signature becomes a responder hypothesis. The same logic extends to combination settings, where synthetic lethality and drug synergy turn target discovery into a pair problem. GEARS-style models were built for multi-gene perturbation prediction, and newer generative approaches can represent response distributions under combinations.

Evaluation is the hard part. Additive baselines are strong. If perturbation A and perturbation B each have known effects, summing them will often work surprisingly well. A useful combination model has to find the cases where additivity breaks and show whether the deviation reflects synergy, buffering, toxicity, or assay noise.

Tissue safety and cell-type specificity

On-target, off-tissue toxicity is already a target-selection problem. Virtual cells can sharpen that analysis. Instead of asking only where a target is expressed, ask how each cell type is predicted to respond when the target is perturbed.

That shift matters. Two cell types can express the same target and react very differently. One may tolerate inhibition because a backup pathway is active. Another may die or release inflammatory cytokines. A virtual cell that predicts response across cell types could turn expression atlases into functional safety screens.

This use case is still early, and RNA alone misses many safety phenotypes. Even so, the field is moving in the right direction, from maps of target location to predictions of cellular response after modulation.

What virtual cells cannot do yet

The useful caution is more specific than “virtual cells are overhyped.” Most models are trained on proxies for the thing drug discovery actually wants.

RNA is not phenotype, and cell lines are not tissues. A transcriptome can move in the right direction while protein activity, secretion, morphology, electrophysiology, or viability does not. This is the core mRNA-protein gap in target biology. RNA is easy to measure at scale, but drug effects live across more layers.

The model systems create another gap. Many perturbation datasets come from cell lines because they are scalable, yet cell lines lack immune context, vascular signaling, liver metabolism, microbiome effects, tissue stiffness, and developmental history. Patient-derived organoids and ex vivo systems help, but they are harder to scale and standardize.

Coverage is sparse, and technical variation is easy to learn. A model trained on a few thousand perturbations has seen a tiny slice of the intervention space. Even 225,000 drug-patient interactions cover only a small part of possible drugs, doses, combinations, cell states, and exposure times. Batch effects, donor effects, guide efficiency, multiplicity of infection, sequencing depth, dissociation stress, and dropout all create patterns a model can learn. A model that performs well inside one assay distribution can fail when the assay changes.

Causality and translation remain the hardest gaps. A perturbation experiment is closer to causal evidence than an observational atlas, but most models still see the intervention label and the outcome rather than the full causal path between them. A target-discovery decision needs mechanism in addition to prediction.

The clinical question is whether a drug helps a patient population. A virtual cell usually predicts a cellular readout in a model system, and many steps separate those two objects. The model can reduce uncertainty at one step. It cannot compress the whole pipeline into a software call.

These limits tell you how to use the models. They are strongest for ranking, triage, experiment design, and hypothesis generation. They are weakest when treated as proof that a target is validated.

Where the field is heading

The next few years will decide whether virtual cells become routine target-discovery infrastructure or remain a research theme with polished demos.

Four directions matter most.

Perturbation-rich data instead of larger atlases alone. Atlases teach a model what cells look like. Perturbations teach it what cells do when pushed. The Tahoe, Arc, and Biohub partnership matters because it treats perturbation diversity as the bottleneck.

Multimodal readouts. RNA alone will not carry the field. Imaging, protein abundance, chromatin accessibility, secreted factors, viability, electrophysiology, and spatial position each answer different biological questions. A virtual cell that predicts several readouts can support target decisions better than one that predicts expression alone.

Uncertainty and abstention. A virtual cell used for target selection should know when it is outside its training distribution. Abstaining is a useful behavior. A confident wrong prediction can send a team into a wasted screen.

Lab-in-the-loop learning. The best near-term systems will look less like static foundation models and more like experimental-design engines. They will propose the next perturbation, observe the result, update the model, and repeat. That is a plausible way for ML to shorten the target-validation cycle without pretending to replace it.

Biology does not expose a clean target_discovery(disease) API. The useful destination is a tighter loop between hypothesis, perturbation, prediction, validation, and revision.

In 2026, the answer is mixed. Virtual-cell models already help organize single-cell data, align disease states, prioritize perturbation experiments, and generate target hypotheses. Benchmarks also show that perturbation prediction remains hard and that simple baselines still win in some settings. That is the right kind of progress.

A target hypothesis becomes more useful once it can be stress-tested as a perturbation forecast, and a virtual cell gives a team one more place to do that before the next round of validation spending.

Further reading

• Bunne, C., Roohani, Y., Rosen, Y., Gupta, A. et al. (2024). How to build the virtual cell with artificial intelligence: priorities and opportunities. Cell 187, 7045-7063. The main field-defining perspective for AI virtual cells.

• Roohani, Y. H., Hua, T. J., Tung, P. Y. et al. (2025). Virtual Cell Challenge: toward a Turing test for the virtual cell. Cell 188, 3370-3374. The benchmark framing for perturbation-response prediction.

• Ahlmann-Eltze, C., Huber, W. and Anders, S. (2025). Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods 22, 1657-1661. The cautionary benchmark paper every virtual-cell model claim should be read against.

• Roohani, Y. H., Huang, K. and Leskovec, J. (2024). Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nature Biotechnology 42, 927-935. A graph-learning reference point for perturbation prediction.

• Replogle, J. M. et al. (2022). Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell 185, 2559-2575.e28. A key data-generation paper for single-cell perturbation maps.

• Cui, H. et al. (2024). scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods 21, 1470-1480. A widely used transformer model for single-cell biology.

• Theodoris, C. V. et al. (2023). Transfer learning enables predictions in network biology. Nature 618, 616-624. The Geneformer paper and a useful example of transfer learning for gene-network dynamics.

• Hao, M. et al. (2024). Large-scale foundation model on single-cell transcriptomics. Nature Methods 21, 1481-1491. The scFoundation paper.

• CZI Virtual Cells Platform. A public platform for models, datasets, benchmarks, and workflows related to virtual cell modeling.

• Arc Institute Virtual Cell Challenge 2025 Wrap-Up. A clear public summary of the first challenge’s results, including the role of baselines and hybrid methods.

Concept Translation: Synthetic lethality is a context-conditional prediction problem. The label “lethal” is not a property of either gene alone - it’s a property of the pair, conditioned on a genetic background. The ML framing is closer to a logical AND on two perturbations than to a continuous dose-response curve. The search space, the label structure, and the evaluation protocols look different from the rest of target discovery: you’re not asking “is this gene a good target?” but “is this gene a good target given that gene B is already broken?”

Why synthetic lethality matters for drug discovery in 2026

Synthetic lethality has moved from biological observation to clinical strategy. The first-generation PARP inhibitors form a real class: olaparib, rucaparib, niraparib, and talazoparib were approved between 2014 and 2018, generated real revenue, and exposed two problems the field has spent the early 2020s trying to solve.

The first is acquired resistance. Tumors revert BRCA mutations. Some restore homologous recombination, the cell’s high-fidelity pathway for repairing double-strand DNA breaks. Others adapt replication-fork protection, the machinery that keeps DNA replication from collapsing under stress. Progression-free-survival gains attenuate over time. This is the resistance story most discussions of PARP inhibitors focus on.

The second is biomarker discipline. Mature randomized data through 2022 and 2023 made the point more sharply. In advanced epithelial ovarian cancer, four randomized controlled trials with long enough follow-up to read out overall survival (OS, time until death from any cause) showed that earlier progression-free-survival (PFS, time until the cancer grows or the patient dies) gains in less tightly selected populations did not translate to OS. In several non-BRCA-mutant arms, patients on PARP inhibitors died sooner than controls. Between late 2022 and late 2023, the FDA withdrew or narrowed six advanced epithelial ovarian cancer indications across olaparib, niraparib, and rucaparib; the surviving maintenance indications in recurrent ovarian cancer (use after an initial response to keep the disease under control) were restricted to patients with confirmed deleterious BRCA mutations, with manufacturers issuing “Dear Healthcare Provider” letters to clinicians (ASCO and SGO summaries of the FDA revisions, 2022–2023). The class works in the subgroup the synthetic-lethal argument identifies. Broaden the indication on the strength of a surrogate endpoint - an earlier readout used in place of survival - and mature OS data can reverse the story. That lesson now sits near the center of any serious synthetic-lethal drug program.

A second wave of DNA-damage-response inhibitors is aimed squarely at those residual problems. Two programs return to the same BRCA-deficient setting from different angles. Artios Pharma’s ART6043, a DNA polymerase θ (Polθ) inhibitor, received FDA Fast Track designation on February 23, 2026 for germline BRCA-mutant HER2-negative breast cancer (patients who inherit a BRCA mutation and whose tumors do not overexpress HER2), and is launching a global randomized Phase 2 in combination with olaparib (Artios Pharma press release, Feb 23, 2026; ongoing Phase 1/2a NCT05898399). KSQ Therapeutics’ USP1 inhibitor KSQ-4279 (also reported as RO7623066 / RG6614 under a Roche partnership announced July 2023) completed first-in-human Phase 1 testing with anemia as the principal dose-limiting signal (Yap et al., 2024 ASCO Annual Meeting, abstract 3005; NCT05240898). A competing USP1 program, TNG348, was terminated in May 2024 for liver toxicity, with Grade 3/4 liver-function abnormalities observed in patients who remained on study longer than eight weeks (Tango Therapeutics press release, May 23, 2024; Phase 1/2 NCT06065059). That is a reminder that correct biology does not guarantee a drug-like molecule.

A third program reaches a different genetic context. WRN (Werner syndrome RecQ helicase) is synthetic-lethal in tumors with microsatellite instability and defective mismatch repair (MSI-H/dMMR) - tumors that fail to correct copying mistakes in repetitive DNA and therefore accumulate unstable short repeats. That population includes a substantial fraction of colorectal, gastric, and endometrial cancers. WRN inhibitors entered the clinic over 2024–2025, and three programs have made the most progress: NDI-219216 (Nimbus Therapeutics), a non-covalent oral WRN inhibitor that completed Phase 1/2 dose-escalation roughly nine months ahead of schedule with no dose-limiting toxicities and no maximum tolerated dose reached; HRO761 (Novartis), a first-in-class oral non-covalent WRN inhibitor with an ongoing Phase 1/1b trial (NCT05838768) in MSI-H/dMMR solid tumors that presented favorable safety and encouraging antitumor activity at ESMO 2025; and NTX-452 (Nimbus), a follow-on candidate with preclinical complete-response data that began first-in-human studies in the first half of 2025 (Nimbus and Novartis disclosures; ESMO 2025 abstracts). WRN is the clearest post-BRCA-PARP program the field has produced so far: a different genetic vulnerability, a different biomarker, the same synthetic-lethal logic.

There is also a methods story underneath the clinical one. Synthetic-lethality target identification has become one of the most data-rich sub-problems in target discovery. CRISPR screens generate gigabytes of guide-RNA count data per campaign. Drug-combination screens evaluate thousands of pairs against hundreds of cell lines. The raw material for supervised classification, unsupervised community detection, and link prediction is already there. This post is about what that methods stack actually looks like.

What synthetic lethality means

The textbook definition is a pair of genes whose simultaneous inactivation kills the cell, but loss of either gene alone does not. In the usual diagram, gene A and gene B perform redundant or parallel functions. Lose A and B compensates. Lose B and A compensates. Lose both, and compensation fails.

In cancer, one of those two losses is often already present. The tumor has disabled gene B through somatic mutation, copy-number loss, or epigenetic silencing. A drug that inhibits gene A now hits a vulnerability found in the cancer cell but not in the patient’s healthy tissue. That selective window is why synthetic lethality matters as a target-discovery strategy rather than only as a combination-therapy strategy. It offers a way to drug targets that would otherwise be too toxic systemically, because the liability appears only in the mutant background. The PARP-inhibitor regulatory recalibration of 2022–2023 is the reminder that this logic holds only inside the genetic background it predicts. Drift outside that background, and benefit can turn to harm.

Drug synergy is a different problem: it asks whether two drugs together produce more effect than the sum of their individual effects. Synthetic lethality asks whether losing one intact gene is lethal when another is already broken. The former is a pharmacological question; the latter is a genetic one. They meet in the clinic. A PARP inhibitor in a BRCA-mutant tumor exploits a genetic synthetic-lethal relationship and shows single-agent efficacy. The methods used to find synergy and synthetic lethality differ, so I treat them separately below.

Why cancer is the natural hunting ground for synthetic lethality

Most clinically validated synthetic-lethal pairs sit in DNA damage repair: PARP1 with BRCA1/2, Polθ with BRCA1/2, USP1 with BRCA1/2, WEE1 with TP53-mutant backgrounds, ATR with ATM-deficient backgrounds, and WRN with MSI-H/dMMR backgrounds. This concentration is not accidental. Cancer cells accumulate mutations at far higher rates than normal tissue because they have already broken, at least in part, one or more of the pathways that repair damage. That breakage fuels the oncogenic mutational trajectory and creates dependencies on whatever parallel repair machinery remains intact. Hit the remaining machinery, and the tumor’s own mutability kills it.

Outside DNA repair, the synthetic-lethal network is less well charted. That gap is both a problem and an opportunity for ML.

The target–target problem

Most of this series asks “is this the right target?” That is a target–disease question. Synthetic lethality is different: the answer to “is target A worth developing?” depends on whether some other target B is already broken in the patient’s tumor. It is a target–target problem, and it reshapes what target discovery means.

Practically, this has three consequences.

The candidate pool is combinatorial. For ~20,000 protein-coding genes, there are ~200 million unordered pairs. Genome-scale screening of pairs remains expensive, especially compared with single-gene essentiality screens, which are now routine. That cost is why computational prioritization exists.

The biomarker is central, and the bar is rising. A PARP inhibitor in a BRCA-wildtype tumor does very little, and the FDA’s 2022–2023 narrowing of the ovarian-cancer indications across olaparib, niraparib, and rucaparib formalized the principle: when mature OS data did not support broader use, the indication snapped back to confirmed deleterious BRCA mutations. Any SL-derived target program carries a built-in patient-selection story. That is an advantage, because biomarker-stratified programs have higher likelihood of approval. It is also demanding - you have to develop and validate the assay alongside the molecule, and resist the regulatory or commercial pull to broaden the program before the OS data is in.

Novelty changes meaning. In the single-target paradigm, “novel” means the gene has never been drugged. In the SL paradigm, “novel” can also mean the pair has never been drugged. A well-studied gene becomes a first-in-class target the moment someone validates it as synthetic-lethal with a disease-relevant mutation. That changes what counts as a first-in-class program and how you would score it — the kind of question the novel-vs-repurposed framework takes up.

Concept Translation: Combinatorial scaling is the recurring problem. ~200 million unordered gene pairs is roughly the same order of magnitude as a fully-specified pairwise-feature-interaction matrix in a tabular ML problem with 20K features - and most ML practitioners don’t enumerate every pairwise interaction either. They use linear models with explicit pair selection, tree-based models that pick informative pairs implicitly, or low-rank embeddings that score pair similarity without storing every pair. SL prediction does the same thing: prune the search space with structural priors (network proximity, pathway membership, mutual exclusivity), then test the survivors experimentally.

Identifying SL pairs with method 1 through CRISPR screens

The experimental workhorse is the pooled CRISPR screen.

The classical screen

Design a library of guide RNAs covering the genes of interest, typically genome-wide, with ~60,000 guides and 3–4 guides per gene. Deliver the library into a cancer cell line, often one carrying a specific mutation such as BRCA1-null (the cell lacks functional BRCA1). Split the population: treat one half with the drug or condition under test, and leave the other as vehicle control (the same solvent without the active drug). Let the cells grow for a fixed number of doublings, then harvest. Copy the guide-RNA sequences out of genomic DNA with PCR (a method for copying specific DNA sequences), then sequence the pool and count the guides.

The readout is direct. A guide RNA targeting a gene whose loss sensitizes the cell to the drug becomes under-represented in the treated arm versus control because the cells carrying that guide died. A guide targeting a gene whose loss has no effect tracks the control. For synthetic-lethal discovery against a PARP inhibitor in BRCA1-null cells, the depleted guides point to genes that become essential when both BRCA1 and the PARP inhibitor are compromising the DNA-repair landscape.

Concept Translation: A pooled CRISPR screen is a massively parallel ablation study with barcoded readout. Each guide RNA is a barcode that identifies which gene was knocked out in which cell. The cell population goes through a treatment perturbation and the surviving pool is sequenced. The ratio of guide counts between arms is the per-gene ablation effect - the same quantity an ML practitioner running ablation experiments is computing, but at 60,000 ablations in parallel rather than one at a time.

MAGeCK and the statistical layer

Raw guide counts are noisy. Each gene has multiple guides with variable cutting efficiency. Library preparation and sequencing introduce batch effects, and cells under drug treatment drift. The standard analytical pipeline is MAGeCK (Model-based Analysis of Genome-wide CRISPR/Cas9 Knockout; Li et al., Genome Biology 15:554, 2014). It performs median-ratio normalization across samples, models guide-RNA read counts with a negative binomial distribution, tests each guide for differential abundance, and then aggregates guides to genes with a rank-based step that pools signal across guides targeting the same gene. The output is a per-gene score with a p-value and false-discovery rate, plus pathway-level enrichment.

For ML, MAGeCK is where CRISPR screens stop being raw counts and become a labeled dataset. The depleted-gene list becomes the positive class for any downstream classifier. The full screen, shared across labs, becomes training data for models that try to predict synthetic-lethal pairs without running a new screen every time.

CRISPRi, combination screens, and Perturb-seq

Classical CRISPR mutates the target gene. That triggers a DNA damage response, which is fine for most targets but confounding for a synthetic-lethality screen that is itself about DNA damage. CRISPRi uses a nuclease-dead Cas9 (dCas9) fused to the KRAB transcriptional repressor - a protein domain that shuts transcription down. The guide RNA brings the complex to the gene’s promoter, KRAB suppresses transcription, and the DNA is never cut. You get loss of function without the confounding DDR signal, and the effect is reversible, which lets you check whether a phenotype is guide-specific.

Concept Translation: CRISPR-Cas9 vs CRISPRi maps neatly onto hard vs soft ablation. CRISPR-Cas9 cuts the gene out (closer to removing a feature from the dataset entirely). CRISPRi turns transcription down without altering the DNA (closer to applying a learnable mask that scales the feature contribution toward zero). For a study where the cutting itself triggers downstream artifacts - DNA damage in this case, which confounds DNA-repair experiments - the soft-ablation version isolates the loss-of-function effect from the cutting-induced effect.

Combination CRISPR screens place two guides in a single vector and deliver them together. This is the direct experimental approach to finding synthetic-lethal pairs: knock out both genes and measure dropout. The combinatorial scaling is the obvious limit. Even a 1,000 × 1,000 pairwise screen is a million guide combinations, not something you run casually.

Perturb-seq combines a pooled CRISPR screen with single-cell RNA sequencing. Each cell receives a guide RNA; when you run the single-cell readout, you recover both the transcriptome and the identity of the knocked-out gene from the same cell. Instead of a single dropout phenotype, you get a transcriptional signature for every perturbation. That signal is far richer for ML. It lets you cluster perturbations by downstream effect, build models of pathway dependencies, and identify genes whose loss produces transcriptional states similar to known synthetic-lethal partners.

Concept Translation: Perturb-seq is what happens when you replace a scalar phenotype (alive / dead, fitness score) with a high-dimensional embedding (the transcriptional state of the cell after perturbation). Two perturbations that produce similar transcriptional states are likely hitting related biology, and the distance between perturbation embeddings becomes a useful feature for ML. The same intuition that drives representation-learning approaches in other domains - learn an informative embedding, then do downstream tasks in that space - applies here, with the added advantage that the perturbation labels come for free from the guide-RNA barcode.

Identifying SL pairs with method 2 through computational prediction

CRISPR screens are definitive but expensive, and they have reproducibility issues across labs, libraries, and cell lines. A parallel literature tries to predict synthetic-lethal pairs from existing data.

Mutual exclusivity in tumor genomics

The hypothesis is direct. If two genes are synthetic-lethal, a tumor carrying mutations in both would not survive. Across a large cohort of sequenced tumors, synthetic-lethal pairs should therefore show mutually exclusive mutation patterns. Any given tumor has a hit in gene A or gene B, but rarely both. The method reduces to a statistical test for co-occurrence versus the baseline expected from mutation frequencies alone. This is cheap, runs on existing cohort data such as The Cancer Genome Atlas (TCGA), and produces candidates. It also produces many false positives, because mutual exclusivity can arise for reasons unrelated to lethality - for example, two mutations producing the same downstream phenotype.

Concept Translation: Mutual exclusivity is a population-level signal that substitutes for an experiment you can’t run. You can’t ablate two genes in one human tumor and check whether it dies; instead you check whether two ablations ever appear together in the existing tumor population, on the logic that if the combination were lethal, evolution would have removed it. The signal is a survivorship-based negative test - observed-but-empty cells in a co-occurrence contingency table. Like all observational substitutes for experiments, it generates hypotheses cheaply and has a high false-positive rate that calls for downstream experimental validation.

Multi-omics layering

Mutation data alone is sparse. Adding epigenetic silencing (promoter methylation, chromatin state) and transcriptomic downregulation as proxies for functional loss makes the signal denser. A gene that is not mutated but is transcriptionally silenced is, for the cell, nearly as lost as a mutated one. Combining these layers under the rule “gene is functionally absent if mutated or silenced or strongly downregulated” expands the positive pool and gives a more reliable substrate for testing mutual-exclusivity-style hypotheses.

DAISY and related in-silico methods

DAISY (Data mining synthetic lethality identification pipeline; Jerby-Arnon et al., Cell 158:1199–1209, 2014) is a representative in-silico synthetic-lethality pipeline that integrates cancer-sample copy-number variation (gene amplifications and deletions), transcriptomics, and short-hairpin-RNA-interference dropout data - an older screen type that suppresses genes rather than cutting them. It constructs a cancer synthetic-lethal network where edges are candidate SL pairs supported by consistent evidence across data types. The network can then be mined for gene-gene and, because some genes are drug targets, gene-drug SL interactions. DAISY was the foundational in-silico SL method; multiple successor approaches have extended it with updated data modalities and ML backbones.

Yeast conservation

Core cellular machinery, including DNA repair, cell cycle control, and protein homeostasis, is deeply conserved across eukaryotes. Large-scale yeast double-knockout screens have been built out into broad genetic interaction maps, and conserved synthetic-lethal pairs in those maps are prior hypotheses for human pairs. The inference chain is “pair is lethal in yeast → pathway is conserved → pair may be lethal in mammalian cancer with the right mutational context,” and it works often enough to be useful as a prioritization filter.

SynLethDB

When you want to check whether any previous experimental or computational evidence supports a candidate pair, SynLethDB (Guo, Liu & Zheng, Nucleic Acids Res. 44:D1011–D1017, 2016; SynLethDB 2.0 in 2022) is the main aggregated resource - a synthetic-lethality knowledge graph curating supporting publications, experiments, and predictions across species. It is a reasonable starting point before running a new screen. It is also a live use case for the knowledge-graph methodology unpacked in the capstone on knowledge-graph-based target discovery.

Where ML actually enters

It helps to separate machine learning from statistical modeling and curation.

The classical SL prediction problem is binary classification. Given a gene pair, predict synthetic-lethal or not. Features are drawn from protein-protein interaction networks, gene co-expression, pathway membership, sequence or structure similarity, evolutionary conservation, shared regulators, and, when available, multi-omic co-occurrence patterns across tumor cohorts. Models range from support vector machines and random forests to gradient-boosted trees on tabular features and graph neural networks operating on learned representations of each pair.

Formulated as a graph problem, SL prediction is a link-prediction task on a heterogeneous biological network. Nodes are genes, edges are any of the thousands of known gene-gene relationships, and the model learns to predict a new edge type (“synthetic lethal with”) from the surrounding topology. This is the natural formulation for graph machine learning. It also sits close to the way knowledge-graph link prediction is used for target-disease scoring more generally — the subject of the knowledge-graph capstone.

The network framing also gives a useful intuition about where SL partners are likely to sit. Biological networks are scale-free - they have hubs with many connections and bridges with high betweenness (nodes that lie on many shortest paths between modules). Hubs tend to be essential. Inhibit them and the cell dies regardless of genetic background, so they make poor SL candidates. Bridges and mid-degree nodes that connect parallel modules are more interesting. Losing one is survivable if the parallel module is intact, and the synthetic-lethal relationship with the parallel module appears as a structural feature of the graph. In practice, graph neural networks trained on such networks learn much of this topology implicitly. You do not need to hand-engineer bridge-like features if you have enough labeled pairs.

The other problem of drug synergy prediction

SL prediction is about which two genes. Drug synergy prediction is about which two molecules. The two problems overlap. A drug that hits one SL-partner gene will often synergize with a drug that hits the other. Most drug-combination screens operate one level up, at the level of molecules with broad target profiles, so the ML problem changes accordingly.

A 2022 large-scale combination screen by Jaaks and colleagues (Nature 603:166–173, 2022) evaluated 2,025 pairs of clinical-stage compounds against 125 breast, colon, and pancreatic cancer cell lines (51 breast, 45 colon, 29 pancreatic) with deep molecular characterization. The key number for model builders: across 108,259 combination–cell-line pairs, only about 5.2% showed synergy, with the rate highest in pancreatic lines and lowest in breast. Only 27% of the synergistic pairs were synergistic at both high and low drug concentrations. Synergy is a narrow window, not a property of the pair in isolation (Jaaks et al. 2022 Nature).

A few implications for anyone building ML on drug-combo data.

It’s an imbalanced classification problem. With about 5% positives, a model predicting “no synergy” for everything is already 95% accurate. Training and evaluation therefore need to handle class imbalance explicitly through sample weighting and metrics such as AUROC, AUPRC, and top-K precision.

Context matters more than chemistry. The same pair is synergistic in one cell line and antagonistic in another. Any model that predicts synergy from molecular descriptors alone and ignores cell-line molecular context will overfit to the average and miss the biology. Models trained across many tissues are useful for screening prioritization, but the real clinical value lies in per-tissue, or better yet per-subtype, models that condition on molecular context.

Subtyping matters. Breast-cancer analyses typically stratify by the PAM50 intrinsic subtypes, an expression-based labeling scheme - luminal A, luminal B, HER2-enriched, basal-like, and normal-like. Colorectal work uses CRIS subtypes, a comparable expression-based subtype scheme. A pair may show synergy only in a single subtype - basal-like breast, microsatellite-stable or KRAS-mutant colon - and the Jaaks et al. screen documents several such subtype-specific combinations. A model that ignores subtype membership misses this; a model that encodes it as a feature or conditioning variable can capture it.

The pathway-level signal is real. Across the three cancer types in that screen, the EGFR signaling pathway was the one where synergistic pairs clustered most consistently, with most other pathways showing synergy in only one or two of the three. ML can surface that kind of signal directly by grouping pairs by shared pathway membership and testing for pathway-level enrichment in the synergistic set. That, in turn, points to where mechanistic explanations for synergy will be easiest to write.

Methodologically, the models used on this kind of data span the familiar ML spectrum. Support vector machines and random forests with chemical-structure descriptors and QSAR-style features (hand-engineered numerical summaries of molecular structure). Deep neural networks that take molecular fingerprints, target profiles, and cell-line omics and learn joint representations. Graph neural networks that represent drugs and cell lines as nodes in a drug-target-cell interaction graph and learn synergy as an edge-weight prediction. No single family wins consistently across benchmarks. Performance depends heavily on the dataset and the evaluation protocol, which is itself a marker of a young ML application area.

The worked example: BRCA, PARP, and the next wave

The BRCA-PARP story is still the best worked example because it makes the full pipeline visible: biological hypothesis, biomarker definition, drug development, resistance, and label revision.

BRCA1 and BRCA2 are tumor-suppressor genes whose products are essential for homologous-recombination repair, the high-fidelity pathway that fixes DNA double-strand breaks. Loss-of-function mutations in either are common in hereditary breast and ovarian cancer and appear sporadically in other tumor types. In BRCA-mutant cells, homologous recombination is compromised, so the cell relies on error-prone backup pathways to get through S phase, the DNA-replication phase of the cell cycle. PARP1 detects single-strand breaks and recruits the repair machinery that resolves them before replication turns them into lethal double-strand breaks.

In a BRCA-wildtype cell, PARP1 inhibition is tolerable because homologous recombination can still handle the double-strand breaks that leak through. In a BRCA-mutant cell, the same inhibition produces more double-strand breaks than the crippled repair machinery can process, and the cell dies. The first clinical validation arrived in December 2014, when olaparib received FDA approval for BRCA-mutant ovarian cancer; rucaparib followed in December 2016, niraparib in March 2017, and talazoparib in October 2018 (the latter for germline-BRCA-mutant HER2-negative metastatic breast cancer). PARP inhibitors are now a multi-drug class for BRCA-mutant and homologous-recombination-deficient (HR-deficient) breast, ovarian, prostate, and pancreatic cancers.

Clinical use exposed two problems. Resistance was expected: tumors revert BRCA mutations, restore homologous recombination, or adapt replication-fork protection. Biomarker drift was the harder lesson. As the first wave of PARP inhibitors moved into broader ovarian-cancer indications - including post-platinum-chemotherapy maintenance (the drug given after an initial response to keep the disease controlled), in unselected or HRD-positive populations rather than strict BRCA-mutant ones - early progression-free-survival gains read out but mature overall-survival data did not follow. Here HRD-positive means tumors with a broader homologous-recombination-deficiency signature that extends beyond BRCA mutation status alone. Across four randomized controlled trials in advanced ovarian cancer, longer follow-up showed an OS detriment in non-BRCA-mutant arms compared with controls. Between late 2022 and late 2023, the FDA narrowed or withdrew six advanced epithelial ovarian cancer indications across olaparib, niraparib, and rucaparib, with manufacturers issuing “Dear Healthcare Provider” letters and the surviving recurrent ovarian-cancer maintenance indications restricted to confirmed deleterious BRCA mutations (ASCO and SGO summaries of the FDA revisions, 2022–2023). The synthetic-lethal argument held. The indication broadening did not.

That history sets the bar for the next wave. The field is now pursuing targets that either layer on top of PARP in BRCA-mutant settings to address resistance, or extend synthetic-lethal logic to a different mutational background altogether.

Polθ (POLQ) is a DNA polymerase that drives theta-mediated end joining (TMEJ), a parallel error-prone repair pathway that BRCA-deficient tumors rely on in place of homologous recombination. Polθ was identified as synthetic-lethal with BRCA1/2 in 2015–2016 work. Small-molecule Polθ inhibitors entered the clinic in the early 2020s. Artios Pharma’s ART6043 received FDA Fast Track designation on February 23, 2026 in combination with olaparib for germline BRCA-mutant (gBRCAm) HER2-negative breast cancer in patients without prior PARP-inhibitor treatment, and Artios is launching a global randomized Phase 2 in that population (Artios press release, Feb 23, 2026; ongoing Phase 1/2a NCT05898399). Dana-Farber’s repurposed novobiocin program targets Polθ through a different binding mode and sits in early clinical testing.

USP1 is a deubiquitinase that regulates translesion synthesis and Fanconi-anemia pathway signaling - two DNA-damage-tolerance systems that help cells keep copying damaged DNA and coordinate repair when replication stalls. USP1-BRCA1/2 synthetic lethality was established through a combination of genome-scale CRISPR screens and targeted validation. It is a concrete example of the methods in this post generating a clinical program. KSQ-4279 (also reported as RO7623066 / RG6614 under a Roche partnership announced July 2023) emerged as a first-in-class USP1 inhibitor and completed first-in-human Phase 1 testing (Yap et al., 2024 ASCO Annual Meeting, abstract 3005; NCT05240898) with anemia (≈36% as a single agent, climbing to ≈87% in combination with olaparib) as the principal dose-limiting signal. A competing USP1 program, TNG348 (Tango Therapeutics), was terminated in May 2024 for liver toxicity, with Grade 3/4 liver-function abnormalities observed in patients who remained on study longer than eight weeks (Tango Therapeutics press release, May 23, 2024). The biology may be right and the molecule can still fail.

WRN is the clearest example of synthetic-lethality logic extending to a non-BRCA background. WRN is a RecQ-family helicase essential for genome maintenance during DNA replication. In tumors with microsatellite instability and defective mismatch repair (MSI-H/dMMR), a substantial fraction of colorectal, gastric, and endometrial cancers, loss of mismatch repair leaves cells dependent on WRN-mediated resolution of the replication stress generated by MSI. Functional-genomics screens (Project Score and Open Targets among others) identified WRN as a top SL target in MSI-H/dMMR backgrounds in the late 2010s, and three programs are now in or entering the clinic. NDI-219216 (Nimbus Therapeutics), a non-covalent oral WRN inhibitor, completed Phase 1/2 dose escalation roughly nine months ahead of schedule with no dose-limiting toxicities and no maximum tolerated dose reached (dose escalation never hit a safety ceiling during the study), and demonstrated WRN target engagement of >24 hours (evidence that the drug was hitting WRN for more than a day). HRO761 (Novartis) is a first-in-class oral non-covalent WRN inhibitor in a Phase 1/1b trial (NCT05838768) for MSI-H/dMMR solid tumors; ESMO 2025 interim results showed a favorable safety profile (manageable, low-grade gastrointestinal events) and encouraging signs of durable antitumor activity. NTX-452 (Nimbus) is a follow-on candidate showing tumor regression and complete responses at low oral doses in preclinical models refractory to immunotherapy and standard chemotherapy, and began first-in-human studies in the first half of 2025 (Nimbus and Novartis disclosures; ESMO 2025 abstracts).

The pattern is consistent. Polθ, USP1, and WRN were each identified computationally or through screening, validated experimentally, and translated into clinical programs anchored on a specific genetic biomarker. The biomarker is part of the program from day one - germline-BRCA-mutant for Polθ, BRCA1/2-deficient for USP1, MSI-H/dMMR for WRN. That is the structural lesson the PARP regulatory recalibration drove home. It is what the end-to-end SL pipeline looks like when it works.

Robotics, AI, and the feed-forward combo screen

Beyond finding SL pairs, ML also enters the screening process itself. Large-scale drug-combination screening is experimentally intensive. 2,025 pairs × 125 cell lines × multiple concentrations × biological replicates is a large number of wells. The industry trajectory has moved toward full laboratory automation and, beyond that, toward a feed-forward loop - a robotic platform runs part of a screen, feeds the results into an ML model that scores the expected value of every remaining experiment, replans the next batch to maximize information gain, and iterates.

The ML methods here are active learning and Bayesian optimization, familiar from other high-throughput-screening contexts. The setup is direct. A surrogate model of synergy, an acquisition function that trades off predicted hits against uncertain regions of the combination space, and an experimental loop that updates the surrogate as results arrive. The combinatorial explosion that makes exhaustive screening infeasible is exactly what makes active learning valuable. You do not need to evaluate every pair if the model can prioritize well.

Concept Translation: This is the same active-learning loop that ML practitioners build for any expensive labeling problem - annotation, simulation, hyperparameter search. The acquisition function is whatever balances exploitation (test pairs the surrogate already predicts as hits) against exploration (test pairs in regions where the surrogate is uncertain). Upper confidence bound, expected improvement, Thompson sampling - the standard library of acquisition functions transfers to combo screens with no modification. The experimental loop just costs more per query than a typical labeling task and produces noisier observations, which puts more weight on uncertainty-aware models.

The longer-term vision of patient-sample-derived combination screens run in hospital settings to inform individual treatment decisions remains aspirational in published work as of 2026, but it is the logical continuation of the current trajectory.

Where the field is heading, and where caution is warranted

Scalability of validation. Computational SL predictions substantially outpace experimental validation, and the failure rate from computational hit to experimentally confirmed SL pair is high. The methods literature tends to report validation on a handful of top-ranked candidates; the actual precision of a model deployed at scale is usually worse than the paper numbers suggest. This mirrors the broader target-discovery validation problem that the evidence-frameworks post examines.

Druggability of SL partners. Identifying a synthetic-lethal partner solves only half the problem. Many SL partners are scaffolding proteins, transcription factors, or complex assemblies with no obvious small-molecule binding pocket. Whether a given SL target is tractable depends on modality choice. A target that is intractable for small molecules may still be accessible to PROTACs, molecular glues, or antibody-drug conjugates. Polθ, USP1, and WRN turned out to be straightforward small-molecule targets; the next wave of SL candidates may not.

Toxicity is not predicted by SL screens. A synthetic-lethal pair in cancer cells says nothing about whether inhibiting target A will be tolerable in healthy tissue. The USP1 and Polθ programs both ran into clinical tolerability surprises: TNG348 was terminated in May 2024 for liver toxicity (Grade 3/4 LFT abnormalities at >8 weeks), and KSQ-4279 showed anemia signals and dose-limiting blood toxicity across the USP1 class. The PARP class illustrates the same issue one level up. Tolerability is a function of mechanism, but also of the population receiving the drug. The 2022–2023 OS-detriment story shows what happens when a drug that is tolerable and effective in one genetic background is pushed into a broader one without mature OS data to support it. Predicting tolerability from mechanism remains an open problem, and the SL framework provides only indirect traction on it.

Single-cell-conditioned SL maps. The bulk of the SL literature operates on cell-line-level data. Perturb-seq and related single-cell CRISPR-screening methods open the door to within-tumor SL maps that condition on cell state. Is a pair lethal in proliferating cells but not in quiescent, non-dividing ones? In one tumor microenvironment niche (a local cellular neighborhood) but not another? This is an area where ML models and experimental methods are evolving together, and where the 2026-onward literature is worth watching.

Two decades after its first clinical validation, synthetic lethality now looks like a mature target-discovery strategy rather than a research curiosity. The ML methods are not exotic. Most of the predictive work is binary classification, link prediction, and active learning, applied carefully to a domain with unusually clear labels, accumulating experimental data, and a clinical readout that closes the loop.

Further reading

• Bryant, H. E., Schultz, N., Thomas, H. D. et al. (2005). Specific killing of BRCA2-deficient tumours with inhibitors of poly(ADP-ribose) polymerase. Nature 434, 913–917. doi:10.1038/nature03443. One of the two foundational papers (with Farmer et al. 2005 below) establishing the PARP-BRCA synthetic-lethal relationship in cell-line studies; the direct primary reference for the “first clinical validation arrived in December 2014” framing in the post’s worked-example section. Readers who want to see the original evidence that anchored the entire next twenty years of synthetic-lethality drug development should start here.

• Farmer, H., McCabe, N., Lord, C. J. et al. (2005). Targeting the DNA repair defect in BRCA mutant cells as a therapeutic strategy. Nature 434, 917–921. doi:10.1038/nature03445. The companion paper to Bryant et al. 2005, published back-to-back in the same Nature issue, establishing the same PARP-BRCA relationship from an independent group. The two papers together are the canonical citation for the PARP-BRCA synthetic-lethality story - citing only one would be selectively citing the historical record.

• Li, W., Xu, H., Xiao, T. et al. (2014). MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens. Genome Biology 15, 554. DOI: 10.1186/s13059-014-0554-4. PMID: 25476604. The original MAGeCK methods paper; the authoritative reference for the statistical analysis pipeline described in the CRISPR-screens section. Readers who want to implement any of the draft’s CRISPR-screen analysis will end up here first. Tooling-currency note: MAGeCK has had multiple major releases since this 2014 paper; the core negative-binomial-plus-rank-aggregation methodology is unchanged, but the software ecosystem (MAGeCK-VISPR for visualization, MAGeCK-iNC for interactive comparisons, MAGeCK-MLE for maximum-likelihood estimation in complex designs) has expanded. Readers building a current-day pipeline should cite the 2014 paper for the methodology and the most recent major release’s documentation for the software.

• Guo, J., Liu, H., Zheng, J. (2016). SynLethDB: synthetic lethality database toward discovery of selective and sensitive anticancer drug targets. Nucleic Acids Research 44(D1), D1011–D1017. DOI: 10.1093/nar/gkv1108. The original SynLethDB methods paper; the primary-reference resource for anyone checking whether a candidate SL pair has prior evidence. Database-currency note: the database has been continuously updated since the original 2016 publication; SynLethDB 2.0 (Wang et al., Database 2022, baac030) added a knowledge-graph front end, CRISPR-derived SLs, and computational-prediction evidence types. Readers who use the database in practice should cite both the methods paper and the current release documentation.

• Society of Gynecologic Oncology - Revisions to FDA Approvals for PARP Inhibitors in the Management of Ovarian Cancer. Society of Gynecologic Oncology summary. The clinical-society summary of the 2022–2023 FDA narrowing of advanced-EOC indications across olaparib, niraparib, and rucaparib in response to mature overall-survival data. The primary-source-of-record is the FDA labeling change history for each individual product, but the SGO summary is the most accessible single document covering all six indications and the OS-vs-PFS discordance that drove the recalibration - useful background for the post’s “biomarker drift” framing in the worked-example section. See also: the JCO review by Tew et al. (2024), Overall Survival and the Evolving Benefit-Risk Assessment for Poly(ADP-ribose) Polymerase Inhibitors in Advanced Ovarian Cancer, doi:10.1200/JCO-24-02834, for the deeper review of the trial-level data.

• Chan, E. M., Shibue, T., McFarland, J. M. et al. (2019). WRN helicase is a synthetic lethal target in microsatellite unstable cancers. Nature 568, 551–556. doi:10.1038/s41586-019-1102-x. The foundational paper establishing WRN as a synthetic-lethal target in MSI-H/dMMR cancers - the third major branch of the worked-example section. Companion functional-genomics screens (Behan et al. 2019, Nature; Lieb et al. 2019, eLife; Kategaya et al. 2019, iScience) corroborated the finding from independent angles and are the historical anchor for the WRN clinical programs (HRO761, NDI-219216, NTX-452) discussed in the post’s next-wave section. Readers interested in the broader functional-genomics screening framework that generated WRN as a hit should start here.

• Jerby-Arnon, L., Pfetzer, N., Waldman, Y. Y. et al. (2014). Predicting cancer-specific vulnerability via data-driven detection of synthetic lethality. Cell 158, 1199–1209. DOI: 10.1016/j.cell.2014.07.027. PMID: 25171417. The original DAISY paper, foundational for in-silico SL prediction. The pipeline integrates copy-number variation, transcriptomics, and shRNA dropout data to construct a cancer synthetic-lethal network, and remains the methodological reference point for in-silico SL approaches. Most subsequent in-silico SL methods are extensions of DAISY’s three-evidence-type framework.

• Jaaks, P., Coker, E. A., Vis, D. J. et al. (2022). Effective drug combinations in breast, colon and pancreatic cancer cells. Nature 603, 166–173. DOI: 10.1038/s41586-022-04437-2. The 2,025-combination, 125-cell-line, 108,259-pair screen that anchors the post’s drug-synergy section: the 5.2% synergy rate, the cross-tissue synergy ranking (highest in pancreas, lowest in breast), the basal-like and microsatellite-stable subtype findings, and the TOP1+CHEK1 inhibitor lead combination (irinotecan + rabusertib in mouse xenograft). Readers building synergy-prediction models should treat this dataset as a benchmark.

We will cover contemporary, publicly accessible chemical data repositories to keep track of in accordance with your own personal projects and passions that might come to light while working through this book. Most of the databases we will look at contain existing compounds, though there are also large databases (up to billions) of virtual molecules that do not exist but could be synthesized. By no means do we intend to provide a comprehensive reference of benchmark datasets, as these are in flux and data sets for specific tasks are best found in primary literature.

This is Appendix B of my book, Build AI Drug Discovery Pipelines. The appendices are freely available through Manning, but most readers never check them. I keep an updated version of this one here rather than only in print, because as a catalog of datasets it is the part of the book most likely to go out of date. Last updated June 2026.

Bioactivity & Binding Affinity

ChEMBL & ChEBI

ChEMBL is an open access, literature derived (primarily from medicinal chemistry literature) data warehouse containing manually curated, high quality data with FAIR data principles (findable, accessible, interoperable, reusable) at its core. ChEMBL covers more than 40 years of published research and is maintained by the European Bioinformatics Institute (EBI). As of ChEMBL 35 (released 2024), ChEMBL constitutes ~1.6M assays, ~15.6K drug targets, ~2.5M compounds, and over 20M bioactivity measurements drawn from ~90K publications. ChEBI, also maintained by EBI, is a freely available dictionary of ~60K small-molecule chemical entities of biological interest (hence the initialism, ChEBI).

FAIR Data Principles

Adherence to FAIR principles enhances ChEMBL’s utility for ML applications in drug discovery. Data preprocessing—often the most time-consuming step in building ML models—becomes more standardized and reproducible. Feature engineering benefits from consistent representation of chemical structures and biological targets. Model validation gains reliability because test sets can be constructed with awareness of data provenance, avoiding leakage between training and test data. Importantly, FAIR principles facilitate model sharing and benchmarking across the research community, as models built on ChEMBL can be evaluated against common, well-understood datasets rather than proprietary or inconsistently processed data collections.

Each FAIR principle has a role:

• Findable. Humans and machines can locate the data through persistent identifiers and rich metadata. ML practitioners can programmatically access specific subsets of data—such as all measurements for a particular kinase family, enabling automated pipeline construction for model training.
• Accessible. ChEMBL data can be retrieved through standardized protocols like REST APIs, simplifying data collection workflows. Instead of scraping PDFs or navigating complex interfaces, researchers can write scripts that directly query and download precisely the bioactivity data needed for their models.
• Interoperable. ChEMBL uses standardized vocabularies, ontologies, and data formats that allow seamless integration with other databases and tools. Compound structures are represented using standard SMILES and InChI notations, making them directly compatible with cheminformatics tools like RDKit for feature generation. Biological targets are linked to UniProt identifiers, enabling automatic integration with protein sequence and structural databases. This standardization reduces the data preprocessing burden and target annotations can be reliably merged across datasets.
• Reusable. ChEMBL data comes with clear provenance information and detailed metadata about experimental conditions. This contextual information informs decisions about data filtering and quality control. For instance, knowing the exact assay type that generated a measurement allows practitioners to control for systematic biases by either removing certain assay types or explicitly modeling them as additional features.

BindingDB

BindingDB is a public repository of measured binding affinities—reported as Ki, Kd, IC50, EC50, and thermodynamic quantities (ΔG°, ΔH°, −TΔS°)—between protein targets and drug-like molecules. BindingDB curates affinities from both medicinal chemistry literature and issued US patents, a patent-derived slice that ChEMBL does not systematically cover. As of 2025, BindingDB holds roughly 3.29M binding datapoints across ~1.35M molecules and ~9,454 targets, including ~58,852 protein–small-molecule crystal structures for which an affinity measurement is also recorded. Content is released with monthly updates and is accessible via web interface, REST API, and bulk download. Although BindingDB overlaps with ChEMBL (roughly 55% of ChEMBL’s data is also in BindingDB), the patent-derived and mass-spectrometry-measured affinities support models that need breadth beyond literature-curated activities.

IUPHAR/BPS Guide to Pharmacology (GtoPdb)

The IUPHAR/BPS Guide to Pharmacology (GtoPdb) is an expert-curated database of drug targets, ligands, and their quantitative interactions, jointly maintained by the International Union of Basic and Clinical Pharmacology (IUPHAR) and the British Pharmacological Society. Its current release covers roughly 3,000 targets, 12,000 ligands, and 23,000 quantitative interaction records, with three releases per year. The World Health Organization draws on GtoPdb for its standardized drug nomenclature, and the database is widely treated as the reference for expert-curated target pharmacology. Content is available through a web interface, REST API, and bulk downloads. GtoPdb prioritizes carefully reviewed target family definitions and quantitative interaction data.

Additional bioactivity resources

Other resources for bioactivity and binding affinity:

• PubChem BioAssay is the NCBI bioassay component of PubChem, aggregating standardized high-throughput screening data deposited by NIH Molecular Libraries, academic centers, and industrial partners across hundreds of millions of tested datapoints.
• BindingMOAD contains ~41,000 high-quality protein–ligand complexes with manually curated binding data; largely static since 2020 but still widely used for benchmark construction.
• PDBbind is a curated subset of PDB entries with experimentally measured binding affinities, organized into general, refined, and core sets, and long used as a docking benchmark.

Chemical Structure & Property Databases

PubChem & ToxNet

PubChem contains data on 1114M compounds, 302M substances, and 304M bioactivities aggregated across 903 data sources. PubChem is federally sponsored and maintained by the National Library of Medicine (NLM) within the US National Institutes of Health (NIH). In addition to information on chemical structures, you can find related data on genes, proteins, pathways, and literature. PubChem has also demarcated a subset of their data that is related to COVID-19 and SARS-CoV-2. In 2019, content from ToxNet was also integrated into PubChem. The collection of ToxNet data within PubChem covers toxicology data, including but not limited to hazardous substances, genetic toxicology, and chemical carcinogenesis.

DrugBank & DrugCentral

DrugBank is a comprehensive database of structural (chemical, pharmacological) and biological target (sequence, structure, pathway) information covering approximately 15K drugs, including ~2,700 FDA-approved small molecules, ~1,500 FDA-approved biologics, 134 nutraceuticals, and ~6,700 experimental or discovery-phase drugs. The full DrugBank database requires a commercial license, which should be factored in early if you are building a pipeline you intend to publish or commercialize. DrugCentral, maintained by the University of New Mexico, is an openly licensed (CC-BY-SA) companion resource covering approximately 4,900 approved drugs with regulatory, pharmacology, and indication data. DrugCentral is commonly paired with DrugBank when a pipeline requires redistribution-friendly licensing; it overlaps substantially in coverage for marketed drugs while being more limited on experimental drugs.

Additional structure and property resources

Other resources covering chemical identity, physicochemical properties, and metabolite coverage:

• ChemSpider contains information on ~115M chemical structures aggregated from 276 data sources by the Royal Society of Chemistry (RSC); structures in RSC publications are added automatically.
• The Human Metabolome Database (HMDB) contains information on ~221K small-molecule metabolites found in the human body, plus ~8.6K linked protein sequences. HMDB is integrated with DrugBank, T3DB (toxins), SMPDB (pathway diagrams), and FooDB (food components).
• The NIST Chemistry WebBook contains thermochemical data on compounds and reactions as well as IR, mass, and UV/Vis spectroscopic and gas chromatography data.
• ChEBI is EMBL-EBI’s freely available dictionary of ~60K small-molecule chemical entities of biological interest; discussed together with ChEMBL in the preceding Bioactivity section because the two are co-maintained.
• SureChEMBL is EMBL-EBI’s repository of approximately 28M chemical structures mined from worldwide patent filings, with daily updates and open licensing. It complements ChEMBL for patent-space analyses in Chapters 2, 3, 7, and 10.

Structural Biology

Cryo-EM, nuclear magnetic resonance (NMR), and X-ray crystallography are useful tools in structure-based drug design. They aid in determining the three-dimensional crystal structure of biological macromolecules, such as a protein or nucleic acid structure, and complexes, such as protein-ligand complexes. The resultant information serves as a gold standard source on the atomic-level physical interactions between a drug and its protein target.

Protein Data Bank

The Protein Data Bank (PDB) is a repository of over 230,000 experimentally resolved 3D structures of large biological molecules (proteins, DNA, and RNA), including structures of proteins with bound small-molecule ligands. Note that PDB was one data source used in training AlphaFold. Other large experimentally derived repositories for organic and inorganic crystallographic information include the Cambridge Structural Database (CSD) and the Crystallography Open Database (COD).

AlphaFoldDB

AlphaFoldDB contains approximately 214M predicted protein 3D structures (AlphaFold DB v4) generated by AlphaFold2, effectively covering every sequence in UniProt. It is maintained jointly by DeepMind and EMBL-EBI and has fundamentally changed the scale at which structure-informed drug discovery can operate—from roughly 230K experimental structures to two orders of magnitude more, spanning nearly every sequenced organism. Each predicted structure ships with per-residue pLDDT confidence scores and predicted aligned error (PAE) maps, which should be inspected before committing to any structure-based modeling; the sidebar following this section discusses the trade-offs in more detail. Chapter 12 and Appendix D reference AlphaFoldDB repeatedly as a complement to experimentally resolved structures.

Additional structural resources

Additional structural resources worth knowing about:

• EMDB (Electron Microscopy Data Bank) hosts 48,000+ cryo-EM and cryo-ET maps (as of 2025); maintained at EMBL-EBI, it is the essential companion to PDB in the cryo-EM era.
• BMRB (Biological Magnetic Resonance Bank) provides NMR data for proteins, nucleic acids, and metabolites.
• UniProt is the reference protein sequence and functional-annotation database, linked to from essentially every structural and bioactivity resource covered in this appendix.
• InterPro, Pfam, CATH, and SCOP are integrated protein domain, family, and structural classification hierarchies; InterPro now subsumes Pfam.
• OPM (Orientations of Proteins in Membranes) provides spatial positioning of membrane proteins within lipid bilayers, relevant for membrane-target SBDD in Chapter 9.
• ASD (Allosteric Database) curates allosteric proteins, sites, and modulators, useful alongside Chapter 9’s discussion of allostery.

Experimentally Resolved Structures Versus Predicted Structures

Note that there is a critical distinction between experimentally resolved protein structures, such as those in the Protein Data Bank (PDB), and computationally predicted structures like those in AlphaFoldDB. Experimental structures, determined through X-ray crystallography, cryo-electron microscopy, or NMR spectroscopy, provide direct physical evidence of protein conformations, typically with resolution ranging from 1-3Å for high-quality structures. These experimental structures capture specific states of proteins and include crucial information about water molecules, metal ions, and binding site flexibility.

In contrast, AlphaFoldDB structures are generated through model predictions without experimental validation for each specific protein. While AlphaFold2 achieves remarkable accuracy for overall protein folding (especially for well-conserved domains), its predictions may be less reliable for other use cases. The model provides per-residue confidence scores (pLDDT) that should be evaluated when using these structures for drug design.

For structure-based drug discovery applications, experimentally resolved structures should be prioritized when available. AlphaFoldDB structures can serve as valuable alternatives when experimental data is lacking, but should be used with appropriate caution, particularly for novel target classes or when the predicted binding site shows low confidence scores. Ideally, computational predictions should be validated through orthogonal experimental approaches before committing significant resources to structure-based drug design campaigns.

Ultra-Large Virtual Libraries

ZINC22

ZINC22 is the successor to ZINC15/ZINC20 and the current staple for large-scale 3D virtual screening. It contains approximately 54.9B tangible 3D-ready molecules and incorporates ultra-large make-on-demand catalogs such as Enamine REAL alongside in-stock and historical vendor compounds. Every molecule is either in stock, available on synthesis-on-demand terms with reasonable success rates, or sourced from curated vendor catalogs, so docking hits can realistically be purchased and assayed. Access is free; tranches are downloadable and an API and web front-end support per-tranche retrieval. ZINC22 powers most contemporary ultra-large docking campaigns, like the ones we discussed in Chapters 2, 7, 9, and 10, and has replaced earlier ZINC releases as the default library for 3D virtual screening.

Enamine REAL Space

Enamine REAL Space is the largest enumerated make-on-demand virtual library, maintained by the commercial supplier Enamine. As of 2025, REAL Space enumerates approximately 81.8B molecules built from a small set of validated reactions and a curated building-block inventory, with the smaller REAL Database subset at ~6.75B molecules; Enamine reports roughly 80% synthesis success on actual orders. The library is freely downloadable for academic use (directly from Enamine or through ZINC22), and individual compounds can be purchased on 2–6 week timelines. Landmark ultra-large-scale docking studies such as Lyu et al. (2019) and Sadybekov et al. (2022, Nature) rely on REAL as their compound pool, and it is updated quarterly as new building blocks and reactions come online.

Additional virtual libraries

One further open enumerated library worth noting:

• GDB (Generated Database) provides exhaustive enumeration of organic molecules up to a specified atom count; GDB-17 covers approximately 166.4B molecules with up to 17 heavy atoms. Useful for probing the boundaries of synthesizable chemical space in Chapters 7 and 10.

ML Benchmark Suites

MoleculeNet

MoleculeNet is a collection of 17 curated datasets spanning quantum mechanics, physical chemistry, biophysics, and physiology, totaling more than 700K compounds. MoleculeNet (alongside TDC) remains a heavily cited benchmark suite for molecular property prediction, and many papers report results on at least a subset of its tasks. Readers should be aware of documented train/test leakage in several sub-datasets (HIV and BACE are the best-known cases, but documented quality issues are becoming pervasive) where scaffold splits alone do not prevent near-duplicates from appearing on both sides of the split. The Notes on Usage section later in this appendix discusses split strategies in detail, and Polaris Hub (below) addresses many of these issues head-on.

Therapeutics Data Commons (TDC / PyTDC)

The Therapeutics Data Commons (TDC) is an ML-ready benchmark platform maintained by an academic consortium. It exposes more than 66 datasets across 22 learning tasks—single-instance prediction, multi-instance prediction, and generation—covering target discovery, activity, efficacy, safety, and manufacturing. TDC-2 (released in 2024) extends coverage to multimodal single-cell data, clinical trial outcomes, and protein–peptide interactions. Installation is a single pip install PyTDC, and everything is released under the MIT license. In practice TDC has become the default benchmark suite for the models built in Chapters throughout this book, though recent analyses have brought forward quality issues.

Polaris Hub

Polaris Hub is a next-generation benchmarking platform backed by the Polaris consortium (Valence Labs / Recursion, the Chan Zuckerberg Initiative, and several major pharmaceutical labs). It was created specifically to address the quality issues that accumulated in MoleculeNet and TDC over the years: duplicated records, leakage between splits, overly optimistic random splits on data where temporal or structural splits are more realistic, and insufficiently grounded industrial datasets. Polaris emphasizes time-based splits, matched molecular pair analysis, and datasets donated by industrial partners, and it is designed to be consumed through the polaris-lib Python package. At the time of writing, Polaris is moving quickly to become the default benchmark platform for industrial-strength evaluation, and readers should check it alongside (or instead of) MoleculeNet and TDC.

PoseBusters

PoseBusters is the de-facto physical-validity benchmark for deep-learning protein–ligand docking and co-folding methods, produced by Buttenschoen and colleagues at Oxford. Rather than scoring poses on RMSD to a reference alone, PoseBusters checks whether predicted poses are chemically and geometrically sensible: correct stereochemistry, plausible bond lengths and angles, absence of severe clashes, and acceptable protein–ligand interaction geometry. The PoseBusters Benchmark Set consists of 308 curated protein–ligand complexes released from 2021 onward, specifically to avoid training-set contamination for models trained on older PDB snapshots. The package is MIT-licensed and installable as a Python library. It has become a standard reporting requirement for DiffDock, AlphaFold 3, RoseTTAFold All-Atom, and similar methods as those covered in Chapters 9, 11, 12, and Appendix D.

PLINDER

PLINDER (Protein–Ligand INteractions Dataset and Evaluation Resource; plinder.sh) is the largest annotated protein–ligand interaction benchmark, and it is the modern answer to the leakage problems that have dogged docking evaluation for a decade. Developed by a consortium that includes OpenFold, Isomorphic Labs, Genentech, and Stanford, PLINDER covers 449,383 protein–ligand complexes with rich annotations—pocket identity, ligand similarity, protein similarity, and assembly metadata—and, critically, ships with similarity-aware splits that control for leakage at the pocket, ligand, and protein levels simultaneously. Earlier benchmarks routinely allowed a docking model trained on pre-2021 PDB to encounter near-identical pockets or ligands at test time, inflating reported performance; PLINDER’s splits make that failure mode much harder. The full dataset and Python API are available under the MIT license through a Google Cloud Storage bulk download. PLINDER is rapidly becoming the reference benchmark for models like those covered in Chapters 9, 11, 12, and Appendix D.

CrossDocked2020

CrossDocked2020 is a training set of approximately 22.5M cross-docked protein–ligand poses across ~18,450 complexes, produced by the Koes Lab at the University of Pittsburgh. Each native PDB pose is supplemented with additional poses generated by cross-docking the ligand into related pockets, yielding a pose distribution that captures both correct and incorrect binding modes. This makes CrossDocked2020 the standard training set for structure-based generative models such as Pocket2Mol, TargetDiff, and DiffSBDD, as well as for learned scoring functions. The dataset is freely available.

Additional benchmark resources

Several additional benchmarks appear frequently in the ML drug-discovery literature:

• OGB (Open Graph Benchmark) is the standard GNN evaluation harness; molecular subsets include ogbg-molhiv and ogbg-molpcba. Relevant for Chapters 8 and 11.
• GuacaMol is BenevolentAI’s generative-model benchmark for distribution learning and goal-directed generation (Chapter 10).
• MOSES (Molecular Sets) is Insilico Medicine’s complementary generative benchmark with standardized metrics and baselines (Chapter 10).
• LIT-PCBA covers 15 targets curated from PubChem HTS campaigns; designed to avoid the ligand-bias issues of DUD-E and the preferred modern alternative for virtual-screening benchmarks.
• The Astex Diverse Set is 85 carefully curated protein–ligand complexes and a long-standing docking gold standard.
• DUD-E provides 102 targets with actives and property-matched decoys. Use cautiously: documented ligand-bias lets models discriminate actives from decoys on crude physicochemical features alone; treat as a historical baseline rather than a rigorous evaluation.
• CASP is the biennial blind assessment of protein structure prediction, and the venue in which AlphaFold2 was demonstrated. Relevant for Chapter 12 and Appendix D.
• CAMEO provides continuous weekly evaluation of structure-prediction methods, complementing CASP’s biennial cadence.

Toxicity & Safety

Tox21 / ToxCast / EPA CompTox Chemicals Dashboard

The Tox21 and ToxCast programs, together with the EPA’s CompTox Chemicals Dashboard, constitute the dominant source of high-throughput in vitro toxicity data for ML. Tox21 is a joint effort of the US EPA, NIH (NCATS, NTP/NIEHS), and FDA that screened roughly 10K compounds against ~60 nuclear receptor and stress response assays; ToxCast extends this to ~9K compounds across more than 1,800 assays; and the CompTox Dashboard aggregates these together with hazard, exposure, and toxicokinetic data on approximately 1.2M chemicals. All data are CC0-licensed and accessible through a REST API and bulk download, with quarterly updates. These datasets provide gold-standard training resource for ML models of toxicity endpoints, such as those trained in Chapter 5’s cytochrome P450 case study.

Additional toxicity and safety resources

Two complementary pharmacovigilance and side-effect resources:

• SIDER annotates ~1,400 marketed drugs with ~5,800 side-effect terms extracted from product labels. Caveat: last comprehensive update 2015; pair with FAERS for current signals.
• FAERS (FDA Adverse Event Reporting System) provides freely available quarterly data files of adverse-event reports.

Reaction & Synthesis

Open Reaction Database (ORD)

The Open Reaction Database (ORD) is an open, schema-standardized repository for organic reaction data, developed by a collaboration between Google Research, the Doyle Lab at Princeton, and the Coley Lab at MIT. It currently hosts more than 2M reactions contributed by academic and industrial groups, and it is the primary FAIR alternative to proprietary databases such as Reaxys and SciFinder for reaction-informed modeling. Reactions conform to a structured protobuf schema that captures reagents, conditions, outcomes, and provenance, and the full dataset is released under CC-BY-SA with a GitHub-hosted Python interface.

USPTO Reactions (Lowe dataset)

The USPTO Reactions dataset, compiled by Daniel Lowe during his PhD and hosted on Figshare under CC0, contains approximately 1.8M reactions extracted from US patent filings between 1976 and September 2016. Standardized subsets—USPTO-50K and USPTO-MIT—are the foundation training set for essentially every open-source retrosynthesis and forward-prediction ML model, from Molecular Transformer to Chemformer and beyond. The dataset is static at its 2016 cutoff, which is its main limitation for newer reaction chemistry; the commercial Pistachio database (NextMove Software) is a direct continuation that many industrial labs license. Pistachio, together with Reaxys (Elsevier) and CAS SciFinder, remain the dominant commercial reaction databases but are accessible only through institutional subscriptions and are not practical as open training resources.

Specialized Modalities

R-BIND

R-BIND (RNA-targeted BIoactive ligaNd Database) is the reference resource for small molecules that bind non-ribosomal RNA targets. Maintained by the Hargrove Lab at Duke, it provides manually curated bioactivity annotations along with a separate fragments subset, currently covering approximately 160+ small molecules with documented RNA-binding activity. The scale is small but the curation quality is high. Chapter 6’s case study on small-molecule binding to an RNA target illustrates how to use related data.

Additional specialized-modality resources

Resources for other emerging modalities covered in Part 2:

• PROTAC-DB v3.0 (2024) catalogs 6,100+ PROTACs, 500+ warheads, 220+ E3-ligase ligands, and 2,600+ linkers. Useful for targeted-protein-degradation research.
• COCONUT (Collection of Open Natural Products) aggregates 695,000+ natural products from 60+ sources under CC0.
• LOTUS integrates ~750K structure–organism pairs with Wikidata under CC0, complementary to COCONUT.
• RNAcentral aggregates ~36M non-coding RNA sequences across dozens of RNA-focused databases.
• Rfam provides ~4,100 curated RNA families with covariance models.
• NDB (Nucleic Acid Database) provides 3D structures of nucleic acids and nucleic acid–protein complexes, complementing PDB for RNA/DNA structural work in Chapter 6.

Target Discovery & Systems Biology

Open Targets Platform

Open Targets is the systematic target-identification and prioritization platform jointly developed by EMBL-EBI, the Wellcome Sanger Institute, and a 13-company pharmaceutical consortium (GSK, Pfizer, BMS, Sanofi, MSD, Genentech, Biogen, Novartis, Takeda, Merck KGaA, Calico, Roche, and EMBL). It integrates genetics, genomics, transcriptomics, drug, animal model, and literature evidence into a single target–disease association framework. Release 25.09 (September 2025) covers 63,226 targets, 28,327 diseases and phenotypes, and approximately 18M target–disease associations drawn from 22 data sources. Everything is CC0 and accessible through a web UI, GraphQL API, bulk downloads, and a Google BigQuery mirror, with quarterly releases.

Additional target-discovery and systems-biology resources

Further target-discovery and systems-biology resources, grouped roughly by data type:

• DisGeNET covers ~1.13M gene–disease associations across ~21K genes and ~30K diseases. Caveat: moved to a freemium model in 2023; academic access remains available.
• DepMap hosts Broad Institute CRISPR and RNAi loss-of-function screens across ~1,100 cancer cell lines with paired omics; CC-BY.
• Connectivity Map / LINCS L1000 (CLUE) contains 1M+ perturbational gene-expression signatures from small-molecule and genetic perturbations; central to the drug-repurposing methods covered in Chapter 7.
• Therapeutic Target Database (TTD) catalogs 4,000+ targets and 50,000+ drugs with therapeutic annotations.
• STRING integrates protein–protein interactions across ~12K organisms, combining experimental, predicted, and text-mining evidence.
• BioGRID provides manually curated genetic and protein interactions.
• Reactome curates human (and other) biological pathways under CC-BY.
• KEGG provides pathway, drug, and disease data. Free for academic use within rate limits; commercial use has required a license since 2022.
• GEO (Gene Expression Omnibus) is the NCBI public repository of bulk and single-cell expression datasets.
• GTEx provides tissue-specific expression across 54 human tissues from healthy donors.
• TCGA / GDC Data Portal hosts The Cancer Genome Atlas and the broader Genomic Data Commons multi-omics cancer datasets.

Clinical & Pharmacovigilance

Clinical and pharmacovigilance resources close the loop between preclinical modeling and observed human outcomes, where downstream clinical context shapes what a useful target or useful model looks like.

• ClinicalTrials.gov registers 490,000+ trials worldwide; API v2.0 and bulk download available. Primary source for trial-level clinical outcome data.
• FAERS (FDA Adverse Event Reporting System) is also covered in the Toxicity & Safety section above; it is the main source of US post-marketing adverse-event signals.
• PharmGKB is an NIH-funded pharmacogenomics knowledge base integrating gene–drug–phenotype evidence with CPIC clinical-guideline annotations; CC-BY-SA.
• CPIC (Clinical Pharmacogenetics Implementation Consortium) publishes clinical guidelines for gene-guided prescribing, co-developed with PharmGKB.

Notes on Usage

As mentioned at the beginning of this chapter, our models are subject to inherent limitations in the data they are trained on. We’ll unearth data quality issues while working through applied examples in the book, but we list several core data set properties you should verify prior to building a model.

Source Categorization

Primary data sources consist of the original experimental data gathered and submitted by researchers. Secondary data sources are built on top of primary data with additional analyses and interpretations. Primary data may be specific to the original researcher’s use case, which secondary data might supplement and generalize with information from other primary or secondary data sources. Additional curation during creation of secondary data can improve content of the primary data to enable new insights and even approximate information that is otherwise infeasible to collect via experimentation. Constructing a secondary data source is also cheaper relative to the expenses needed to generate primary data.

However, secondary data can introduce noise if using supplemental data of questionable authenticity or relevance, or from a biased source. In practice, many databases are a mix of primary and secondary sources and disentangling the two is not straightforward. To determine the reliability of a data source, we can investigate its data provenance. Data provenance is the record of how the data was generated, transmitted, and stored, from its origin to its entry into the database.

Data Exchange

Providers of data are, unsurprisingly, consumers of data. In the case of public databases, contents are freely accessible and often integrated across multiple other databases. For example, UniChem maintains a non-redundant database that cross-references identifiers and structures across 40 different databases, including ChEMBL, ZINC, PubChem, HMDB, ChEBI, and DrugBank. It is a good idea to be aware of any data exchange the database you source from is involved in and to understand how that database evaluates the verity of integrated data sources. Specifically, data sources might vary in how they represent the same molecule or units of measurement, and error from one data source is at risk of propagating across all linked data sources.

When integrating data from multiple sources, several preprocessing strategies can mitigate inconsistencies and prevent error propagation:

• Standardize units of measurement (e.g., converting all IC50 values to nM)
• Normalize molecular representations using canonical SMILES or InChI identifiers
• Implement rigorous duplicate detection and resolution protocols
• Establish clear rules for handling conflicting data points from different sources
• Document data provenance to track the origin of each data point
• Apply quality filters based on experimental confidence scores or statistical validation

In the case where we want to use ChEMBL, we might check their documentation on how they extract, curate, and annotate drug data. We’d learn that information on the literature reference, assay, target, organism, and structural and property information of the compound is manually extracted from full text articles via a service. The data then undergoes automated and manual curation steps to standardize activity types, standardize units, compute additional properties, and flag and correct incorrect or duplicate data.

When evaluating chemical databases, we must also consider recency and update frequency. ChEMBL, for instance, releases major updates approximately twice yearly, while PubChem receives daily updates as new compounds are deposited. Some specialized databases might update only annually or less frequently. This update cadence directly impacts model quality, particularly for rapidly evolving research areas involving emerging targets. Using a database that hasn’t been updated in 18 months might mean missing crucial compounds that could serve as valuable training examples.

Additionally, data corrections and standardization improvements often accompany database updates. Earlier versions might contain structural errors, incorrect activity values, or inconsistent representations that get rectified in subsequent releases. When selecting a database, examine its version history documentation to understand both update frequency and what specific improvements each update contained. Consider documenting the version and timestamp of each dataset used in your model training to ensure reproducibility and to help evaluate performance differences between model iterations.

Train/Test Splits

A random split assumes that training and test data are identically distributed, which is almost never true for drug-discovery data: chemical series cluster, assays drift over time, and literature coverage is biased toward well-studied targets. A scaffold split (Bemis–Murcko framework separation) removes obvious near-duplicates and is a useful minimum bar, but it does not prevent scaffolds from sharing substituents or near-identical analogs appearing on either side of the split. Time-based splits, i.e., training on data published before some cutoff date, testing on everything after, more closely simulate prospective deployment and often reveal that headline performance numbers are dramatically inflated. Similarity-aware splits, which enforce separation by pocket, ligand, and protein similarity simultaneously, are considered best practice for docking and structure-based ML; PLINDER’s splits (described in the ML Benchmark Suites section above) implement this and should be preferred over split strategies that ship with older benchmarks.

Benchmark Leakage and Contamination

Many published performance numbers on older benchmarks are misleading because the training data of modern models overlaps heavily with the test sets. E.g., in structure-based docking, models trained on any pre-2021 snapshot of PDB effectively have memorized large portions of the PDBbind and CASF test sets. PoseBusters mitigates this with a post-2021 benchmark set; PLINDER mitigates it with similarity-aware splits; but many published comparisons on older benchmarks do neither, and their reported improvements should be treated with caution. As another example, DeepChem’s MoleculeNet HIV and BACE datasets have documented near-duplicate pairs across scaffold splits, meaning that even scaffold-split accuracy can be overestimated. A third is historical and still cited: DUD-E’s actives and decoys are distinguishable by simple physicochemical descriptors, so models that appear to discriminate binders from non-binders may in fact be discriminating drug-like from non-drug-like chemistry (LIT-PCBA is one recommended replacement benchmark). When reading a new modeling paper, a sensible first question is: what was the split strategy, and is the reported benchmark known to leak?

Activity Cliffs and Assay-Merge Noise

Aggregated bioactivity databases (ChEMBL, BindingDB, PubChem BioAssay) merge data from tens of thousands of distinct assays run across decades in different labs. Even for the same compound against the same target, reported IC50 values can disagree by more than an order of magnitude because of assay format, cell line, reference compound, substrate concentration, and plate-to-plate variability. In the worst case, the same molecule is labeled “active” in one assay and “inactive” in another. Worse, the drug-discovery literature is populated with activity cliffs where a tiny chemical modification, such as adding or removing a methyl group or swapping a heteroatom, changes potency by 10–100×. Models trained on naively merged data can learn neither the real SAR signal nor a calibrated noise floor and tend to fail where cliffs have highest impact. Practical mitigations include filtering to a single assay type per target before modeling (at a cost in dataset size), modeling assay identity as an explicit categorical feature, reporting per-assay performance as well as global performance, and validating predictions on held-out series rather than held-out random samples.

Commercial vs. Free Licensing

Several widely cited resources have moved to or reinforced commercial terms, even since this appendix was first drafted! If you are building a model or product you intend to publish, distribute, or commercialize, audit the license terms of every resource in your pipeline early.

Garbage In, Garbage Out

ML facilitates drug discovery, but not without its shortcomings. Data with suspect quality eventuates as a faulty model, i.e., garbage in, garbage out. Lack of quality data limits the ability to model and search the vast space of drug-like molecules, leading to greater focus on well-explored chemical structures with more data and less complex interactions. The more limited the diversity of chemical structures in our data set, the more restricted our model’s applicability domain. The applicability domain scopes the reliability of our model when applied to molecules that are increasingly different in their structure or composition from the molecules in our training data set. Even the available data may only serve as an approximation of reality. Our ML models capture trends in the data – low quality data results in low quality models.

As an example, suppose we want to screen, or filter out, drug candidates that might cause drug-induced liver injury (DILI), which is the leading cause of drug withdrawal from the market [1]. Coincidentally, we happen to have a data set of molecules, each with “Yes” or “No” labels that indicate whether they are associated with DILI, that we can use to model and predict DILI on any molecule! We craft a proposal for our DILI model and send it to our employer’s pharmacovigilance department, which is responsible for drug safety with respect to clinical use.

Their response is lackluster and pinpoints an issue in our data set – the labels are conditional. Drug X might provoke DILI, but only in patients with a specific genotype or only in patients concomitantly taking drug Y or only when administered above a certain dose (even dihydrogen monoxide, also known as water, can be toxic at extreme doses). Once the adverse reaction occurs, we might observe differences in severity levels, frequency of occurrence across gender or age groups, and duration of the effect (i.e., short-term versus long-term). For this use case, a binary classifier is not adequate, and we need to reconsider our data selection and modeling strategy. Importantly, the further detached our data and model become from the full context under which the drug is taken, the less reliable improved model performance becomes as an indicator of improved safety and efficacy.

Continuing this example further, how might our DILI data set have been created? Experiments to directly measure DILI in a living organism (i.e., in vivo experiments) are difficult, expensive, low-throughput, and may have more confounding factors compared to measurements on microorganisms, cells, or molecules in a controlled environment, such as a test tube or petri dish, outside their normal biological environment (i.e., in vitro experiment). If we want to generate a DILI data set without in vivo experiments in an animal model or human, we could measure DILI via a proxy assay. One possible contributor to DILI is bile salt export pump (BSEP) inhibition, which we can measure with a BSEP inhibition assay on hepatocytes (a major type of liver cell that constitutes most of the total liver cell population).

Though proxy assays are a useful source data, we ultimately care about the phenotypic endpoint (DILI in this case). The further our data set moves away from in vivo context; the further our model’s utility moves away from clinical relevance. A silver-lining is that these limitations can be interpreted as challenges to overcome, and the field is ripe for contributions and advancements from readers such as yourself. We will formally cover the limitations of ML, and ways to navigate them, throughout the second half of the book.

Note. Even high quality data can produce poor results if the context of data collection is ignored or misunderstood! Always investigate the experimental conditions under which your training data was collected.

Where to Look Next

The resources catalogued here represent a snapshot. New databases appear, older ones change license terms, scales grow by orders of magnitude, and curation practices evolve. Community review articles (e.g., Nucleic Acids Research publishes an annual database issue each January) can help you keep current. Whatever resource you use, cite the specific version or release (the ChEMBL release number, the UniProt release date, the Open Targets quarterly tag) and archive the data snapshot you trained on. Models are only reproducible if the data they were trained on is.

I maintain this catalog beyond the book’s release. If a resource here has changed, or one is missing that belongs on the list, let me know.

References

[1] Babai, S., Auclert, L., and Le-Louët, H. (2018). Safety data and withdrawal of hepatotoxic drugs. Therapie. https://doi.org/10.1016/j.therap.2018.02.004

Adapted from Appendix B of Build AI Drug Discovery Pipelines. The book puts these datasets to work in applied case studies.

Why this matters in 2026

Most discussion of ML in drug discovery in 2026 still centers on chemistry. Generative models design novel small molecules. AlphaFold 3 and its open-weights successors - Boltz-2, Chai-1, Protenix - have made routine biomolecular complex prediction practical for most proteins, including in the presence of ligands, nucleic acids, and post-translational modifications. Virtual screening (computationally ranking large compound libraries for likely binders) runs at ten-thousand-compound scale in days.

This post focuses on the earlier step: deciding what a molecule should hit. When that decision is wrong, Phase II - the first stage where efficacy gets a real test in patients - fails regardless of how good the chemistry turns out to be. (The front-of-funnel decision walks through that economic argument in detail.)

Three developments since early 2024 have changed the shape of this space. Industry trackers now count more than 173 AI-discovered drug programs in clinical trials, with 15–20 expected to enter pivotal Phase III trials this year - the larger confirmatory studies usually meant to support approval. The field is no longer mostly preclinical. Recursion’s May 2025 discontinuation of REC-994, its lead pre-merger AI-discovered candidate for cerebral cavernous malformation, provided the first prominent counterexample after long-term Phase II data failed to confirm earlier efficacy trends. That matters because AI-designed compounds can fail in the clinic for the same reasons traditional ones do. A third shift is structural: AlphaFold-3-class systems now sit in an open-source field rather than a single closed one, which changes the cost calculus for building target-discovery platforms on top of predicted structures.

This post is the capstone of the series. The point is integration - how evidence-framework thinking (evidence frameworks), novelty scoring (novel vs repurposed targets), and the network-medicine view fit together inside a single end-to-end program, with synthetic lethality extending the framework to combination targets.

From biological networks to knowledge graphs

Network biology predates today’s knowledge-graph language. In much of computational biology, systems biology refers to the same basic idea. In drug discovery, this way of thinking often goes under the name network medicine - using networks to reason about how diseases cluster in pathways and how interventions propagate through them. The premise is direct. Protein targets do not act as isolated points. They sit inside interaction networks, and any attempt to modulate one protein propagates through its edges. A target-discovery view that ignores that propagation will misread biological effect.

The zoo of biological networks breaks into roughly five types. Protein–protein interaction networks use proteins as nodes and physical interactions as edges; proteins with many neighbors tend to be essential for cell survival. Gene regulatory networks capture transcription-factor-to-target-gene relationships. Gene co-expression networks connect genes whose expression covaries across conditions, on the assumption that they are co-regulated. Metabolic networks use metabolites as nodes and biochemical reactions as edges. Signaling networks represent information flow from receptors through kinase cascades to effectors. Beyond these sit drug–disease association networks, sequence-similarity networks, non-coding-RNA–protein networks, and a long tail of specialty graphs.

Three structural features matter most for target discovery. First, degree distributions follow a power law - these networks are scale-free, with a small number of highly connected hubs and a long tail of sparsely connected nodes. Such networks tolerate random node removal relatively well and fail badly when hubs are removed, which is why hub proteins can be attractive therapeutic targets or dangerous poison targets. Second, nodes with high betweenness centrality, often called bridging nodes, connect otherwise distinct subnetworks. Hitting a bridging node is usually less lethal than hitting a hub, which is one reason these nodes attract interest as safer targets. Third, biological networks are rich in motifs - small overrepresented subgraph patterns that often correspond to specific biological processes.

Concept Translation: Scale-free degree distributions and hub fragility are familiar from network science generally - the World Wide Web, citation networks, and protein interaction networks all share the same statistical signature. A useful intuition: random failures are absorbed because most nodes are sparsely connected, but targeted hub removal cascades because a small number of nodes carries most of the connectivity. That property does double duty in drug discovery: it tells you which proteins are valuable to hit (hubs near the disease) and which are lethal to hit (hubs essential everywhere). Bridging nodes - high betweenness centrality, low degree - sit between modules and tend to disrupt specific pathways without crashing the whole network. The same robustness math from social-network analysis transfers directly.

Knowledge graphs generalize this picture. A biomedical knowledge graph is a large heterogeneous graph in which both nodes and edges have types. Node types include genes, proteins, diseases, drugs, chemicals, pathways, tissues, cell types, phenotypes, and ontology terms. Edge types include activation, inhibition, binding, expression, disease-gene association, drug-target interaction, co-occurrence, and many others. A well-assembled biomedical KG draws from dozens of public resources - DrugBank for drug–target pairs, OMIM for Mendelian disease–gene links, STRING for protein–protein interaction data, Reactome for pathways, ChEMBL for compound–target bioactivity, the GWAS Catalog for genetic associations - then combines them with text-mined edges from the biomedical literature.

The scale is large and still growing. The Clinical Knowledge Graph published by Santos and colleagues in 2022 (Santos et al., Nat Biotechnol 40:692–702, DOI 10.1038/s41587-021-01145-6) reports close to 20 million nodes across 36 types and around 220 million relationships across 47 types, drawn from 26 biomedical databases and roughly 7 million abstracts. Hetionet, the open-source Himmelstein knowledge graph built around drug repositioning (Himmelstein et al., eLife 6:e26726, 2017), is smaller (47,031 nodes / 11 types, 2.25 million relationships / 24 types) but widely used in academic methods work. PrimeKG, released in 2023 and aimed at precision medicine (Chandak, Huang & Zitnik, Sci Data 10:67, 2023), integrates 20 high-quality resources to describe 17,080 diseases with around 4 million relationships. Commercial equivalents, including the ones used in the worked examples below, are typically larger, updated more often, and less transparent in how they are assembled.

The most important property of these graphs, for ML purposes, is incompleteness. We do not know most of the relationships we would need among every gene, drug, and disease. The edges we do have are enriched for what humans have already studied, so well-studied proteins collect more edges and less-studied ones fewer; that is the “ignorome” problem covered in detail in novel vs repurposed targets. In one sentence: knowledge-graph-based target discovery tries to predict the missing edges.

Text mining and where most new edges come from

No one can hand-curate a knowledge graph at PubMed scale. Several thousand new biomedical papers are indexed per day, and the useful subset - papers with new target–disease claims, new pathway assignments, new drug–target binding assays - is already larger than any curation team can track. Text mining is therefore how most new relationships enter the graph.

The text-mining pipeline that feeds knowledge graphs has three stages: information retrieval, named entity recognition, and relation extraction.

Information retrieval is the comparatively easy stage. Identify which documents in a corpus are relevant to a query. Industry has done that at scale for decades.

Named entity recognition (NER) locates predefined entity types in text - genes, proteins, diseases, chemicals, cell lines, mutations, and biological processes. For biomedical NER, the state of the art sits around 90% F1-score in aggregate, though performance varies meaningfully by entity type. Gene and cell-line names are relatively easier because they tend to be less ambiguous. Chemicals and disease names are harder because the same concept often appears under many surface forms, and informal writing is rarely tidy. For most practical purposes, that level of performance is usable. The remaining errors cluster in predictable places, and downstream graph-construction steps can often absorb them through redundancy.

Relation extraction (RE) is the hard part. Given two entities, determine whether a relationship exists and, if so, what kind. In biomedical text, that means triples like gene–activates–biological-process, chemical–ameliorates–disease, or disease–affects–gene. Approaches range from simple co-occurrence (where two entities appearing in the same abstract are treated as a hypothesis) to rule-based patterns and machine-learned classifiers. Aggregate RE performance sits around 50% accuracy. That figure makes text-mined edges hard to treat as stand-alone evidence. In practice, they are treated as weaker evidence than curated edges, and graph-scoring algorithms weight them accordingly.

Four problems keep biomedical RE difficult. Drug and disease names are heterogeneous and context-dependent - “Type 2 diabetes,” “T2D,” and “non-insulin-dependent diabetes mellitus” all name the same condition, but mapping surface form to canonical entity is still nontrivial. Labeled training data is scarce. Document styles vary sharply: journal articles, clinical trial registrations, and patents read like different languages even when they discuss the same target. And models still struggle to distinguish a paper’s new finding from background knowledge the paper is merely citing.

Text mining also reaches beyond PubMed. Patents capture industrial priorities before papers do. Grant applications show what researchers believe can attract funding. Clinical trial registries such as ClinicalTrials.gov show what compounds are being tested against what diseases. Electronic health records, where accessible, describe observed drug effects in real patient populations. Pulling these into one graph is mostly an engineering task rather than a scientific one, but it is the sort of engineering task that takes teams years to do well.

Graph machine learning for link prediction in target identification

Once you have a large, heterogeneous, incomplete knowledge graph, the ML problem that best matches “find new drug targets” is link prediction. Given the edges you have, predict the ones you are missing. For target identification, the mapping is direct: a missing edge between a disease node and a gene node is a hypothesis that the gene matters to the disease.

The standard approach is graph representation learning. Learn a vector embedding for each node - and often each edge type - so that structurally similar nodes sit near one another in embedding space. Three families dominate. Matrix-factorization-based methods decompose the adjacency matrix, or a weighted variant, into low-rank factors; this includes DeepWalk’s matrix view and related spectral approaches. Random-walk-based methods such as DeepWalk, node2vec, and metapath2vec generate sequences of nodes by walking the graph and then train a word2vec-style model on those sequences. The resulting embeddings capture local neighborhood structure and higher-order connectivity. Deep-learning-based methods use graph neural networks - graph convolutional networks (GCNs), graph attention networks (GATs), relational GCNs (R-GCNs), and heterogeneous graph transformers (HGTs) - to learn embeddings through message passing. In message passing, each node updates its representation by aggregating information from its neighbors. The heterogeneous variants are built for the multi-node-type setting biomedical KGs require.

Once nodes are embedded, link prediction becomes a scoring problem. A multi-layer perceptron, operating on concatenated or dot-producted embeddings of two candidate nodes, outputs the probability that an edge exists. Training uses known edges as positives and sampled non-edges as negatives. The same machinery handles related tasks:

• Drug–target interaction prediction. Will this compound bind this protein?
• Drug–drug interaction prediction. Do these two drugs interact adversely?
• Adverse-event prediction. What post-market adverse events should we expect from a novel target?
• Drug repurposing. What approved drugs might treat this disease via some existing but unrecognized edge?

Concept Translation: All four downstream tasks reduce to the same architecture - predict an edge of a specific type between two nodes - differing mainly in which node and edge types participate. From an ML engineering perspective, this is one of the rare cases where a single backbone (a heterogeneous graph encoder + a scoring head) genuinely covers a wide range of business problems with different tasks just being different decoder heads. That is why target discovery, drug repurposing, and adverse-event prediction sit so close together in the literature; they are sibling problems on the same graph.

What makes biomedical knowledge graphs harder than canonical social-network or citation-network benchmarks is heterogeneity. A general-purpose GCN treats all nodes and edges as the same kind of object. Biomedical graphs contain dozens of each, and edge semantics matter. A binds edge between a drug and a protein is not the same as an inhibits edge, and neither means the same thing as a text-mined co-occurs edge. Heterogeneous graph neural networks and relational GNNs address this by learning separate transformations for each edge type, though the added parameters create their own problem when training data is sparse.

Concept Translation: Heterogeneous GNNs run into a familiar parameter-sparsity problem - every additional edge type multiplies the learnable transformations, but the labeled examples per edge type don’t grow proportionally. Some edge types have millions of examples (drug–target, gene–pathway), and others have only a few hundred. The standard mitigations are familiar from low-resource ML: parameter sharing across related edge types, regularization, meta-learning across edge types, and inductive priors that bias the model toward biologically reasonable propagation patterns.

LLMs and the 2024+ shift

In the course material underlying this series, LLMs appeared as a short forward-looking section: biomedical NER, relation extraction, and chat-based graph exploration. That was roughly the 2024 picture. By 2026, the picture is broader, but the uses are similar.

In practice, LLMs do three things in the knowledge-graph-for-target-discovery workflow.

Biomedical NER, better than earlier systems. Fine-tuned BERT-family models (BioBERT, PubMedBERT, BioGPT) pushed NER benchmarks several points past pre-LLM approaches, and current instruction-tuned models push them further still. In practice, that means a scientist can point an off-the-shelf model at a corpus and get usable entity annotations without building a specialized pipeline. The remaining errors cluster around rare entity types and domain-specific abbreviations that are underrepresented in pretraining.

Relation extraction, still hard. Zero-shot RE with large models performs better than zero-shot RE with smaller ones, but it still underperforms specialized supervised pipelines in entity-dense biomedical settings. The more useful shift is in few-shot RE. With a small set of examples for a relation of interest, large models can produce reasonable extractions across the rest of a corpus. The bottleneck moves from building a 10,000-example labeled dataset to curating 20 good examples.

Concept Translation: Few-shot RE is the part of the LLM era that has changed biomedical KG construction the most. Annotating 10,000 examples of gene–activates–pathway triples used to require a domain-expert team and months of work. With a modern instruction-tuned model and 20 well-chosen examples, the same task can be set up in an afternoon. The annotation cost has not gone to zero - it has shifted from “label tens of thousands of examples” to “curate prompts and a small evaluation set.” The bottleneck is now in evaluation rather than labeling, which is a familiar shift from other LLM application domains.

Natural-language interfaces to knowledge graphs. On the practitioner side, modern target-discovery platforms increasingly expose a chat interface over a structured knowledge graph instead of a SPARQL endpoint or Neo4j Cypher box (both are formal graph-query interfaces). The goal is to let a scientist ask “what targets are upregulated in IPF patient tissue and also have a bridging-node role in known pulmonary fibrosis pathways?” without first translating the request into a formal query language. Behind the scenes, the LLM generates structured queries, executes them against the graph, and summarizes the results. This is where the strengths and weaknesses of LLMs meet most directly. Strong systems ground each claim in a specific node or edge and show provenance. Weak ones generate plausible relationships that the graph does not support.

The largest open question for LLM-augmented KG workflows is evaluation. Even a low fabrication rate is hard to manage when the errors look fluent. A formal query interface either succeeds or fails. A chat interface can fail quietly. Platforms that take this seriously include source citations with every claim surfaced to the user.

Worked example: the rentosertib target-discovery arc

Rentosertib, formerly known by the internal codes ISM001-055 and INS018_055, is the clearest public example in 2026 of an end-to-end AI-driven target-and-molecule program. It deserves a close read for two reasons. The public documentation is unusually thorough - two Nature-family papers in fifteen months and a Harvard Business School teaching case. And the program runs through every part of the methodology discussed above.

The program-level fact that stands out is speed. Project initiation to preclinical candidate nomination took about 12 to 18 months; project initiation to first dosing in humans took under 30 months. The traditional industry benchmark for the same span is 4.5 to 6 years. The open question is whether that compression holds across other programs or whether it depended on favorable conditions specific to TNIK and IPF, including strong network-medicine support for the target and relatively well-characterized fibrotic biology. Insilico has framed the result as a counterexample to Eroom’s Law - Moore’s Law spelled backwards, the observation that pharmaceutical R&D productivity has fallen despite rising technological capability. That is the strongest version of the speed claim.

Target identification - multi-omics to TNIK

The program began with idiopathic pulmonary fibrosis, a progressive and fatal lung disease with limited treatment options. The target-identification pipeline combined several streams: multi-omics analysis of lung tissue from IPF patients versus healthy controls to find differentially expressed genes; pathway activation inference to identify dysregulated fibrosis pathways; text mining across the IPF literature to collect known disease–gene associations; and knowledge-graph reasoning over protein–protein interactions and pathway data to propagate signal from known IPF targets to their network neighbors. (The structured evidence frameworks for weighing target-disease linkage across multi-omics streams are the subject of the evidence frameworks post.) TNIK, TRAF2- and NCK-interacting kinase, emerged as a novel candidate IPF regulator. The key point is novelty. TNIK was not an established IPF target at the start of the program. It emerged from computational analysis rather than from the pursuit of an already published target hypothesis. That distinguishes it from the many AI-supported programs in which ML optimizes a molecule against a preselected target. In the target-discovery literature, TNIK is described as a central regulator of fibrosis and inflammation because multi-omics signal and proximity in the graph to genes already implicated in IPF pathways converged on it, not because the literature had already settled on it.

Concept Translation: “Knowledge-graph reasoning to propagate signal from known targets to their neighbors” is graph diffusion in operational terms. Start with seed nodes that are known IPF-associated genes, then run a random walk or label-propagation algorithm across the graph and rank nodes by how much weight they accumulate. Nodes that aren’t in the seed set but accumulate high weight - TNIK, in this case - are the candidates. This is the same algorithm family used for personalized PageRank or graph-based semi-supervised learning, applied to the target-prioritization problem.

Target validation - parallel preclinical and chemistry

Validation ran in parallel with early chemistry, as it does in many modern programs. Wet-lab validation is expensive, and strict sequencing lengthens the cycle. Preclinical models confirmed that TNIK inhibition reduced fibrotic phenotypes in multiple systems. Generative chemistry tools produced a series of TNIK inhibitors; a small subset was synthesized and tested; and the compound that became rentosertib emerged from that iteration. The full discovery-to-Phase-1 history appeared in Nature Biotechnology in March 2024 (Ren et al., doi:10.1038/s41587-024-02143-0). That compressed validation-chemistry loop is what produced the 12-to-18-month preclinical-candidate-nomination timeline noted above. The Harvard Business School case study treats the parallelism, more than any single computational breakthrough, as the most reproducible lesson.

Phase I - safety and pharmacokinetics

Two Phase 1 trials, one in New Zealand and one in China, together with a Phase 0 microdose trial, established safety, tolerability, and a favorable pharmacokinetic profile in healthy volunteers by 2023. A microdose trial gives a very small, subtherapeutic dose to gather early human data, and pharmacokinetics refers to how the body absorbs, distributes, metabolizes, and clears the drug. This was standard Phase I work. What stands out is how quickly the program reached human dosing, not any unusual feature of Phase I design or conduct.

Phase IIa - the GENESIS-IPF trial

The Phase IIa trial (NCT05938920) was a multicenter, double-blind, placebo-controlled, randomized study of 71 IPF patients across 21 sites in China, with 12 weeks of dosing (Xu, Ren, Wang et al. 2025, Nature Medicine, PMID 40461817). The primary endpoint was safety, and rentosertib met it. The secondary endpoint was change in forced vital capacity from baseline at week 12. The 60 mg daily arm showed a +98.4 mL change in FVC, compared with a −20.3 mL decline in the placebo arm. The trial was not statistically powered to establish efficacy - it was a proof-of-concept study designed to look for an early signal rather than to settle the question definitively. Even so, the direction of effect is what Nature Medicine described as the first clinical proof of concept for an end-to-end AI-discovered drug when the results were published in June 2025 (Xu, Ren, Wang et al., doi:10.1038/s41591-025-03743-2; PubMed PMID 40461817).

Rentosertib was also the first AI-designed drug to receive a USAN (United States Adopted Name). That detail is small, but it suggests the regulatory apparatus is starting to treat these programs as part of ordinary drug development rather than as an exception.

Interpretation - what 71 patients does and doesn’t prove

The trial enrolled 71 patients for 12 weeks, so it is not a registrational result - it is not the kind of trial package used to support approval. It is hypothesis-generating. As of April 2026, no pivotal trial is running; Insilico is in regulatory discussions about a Phase IIb pivotal study, and a separate US Phase IIa (NCT05975983) is enrolling, with eight of a planned 60 patients reported to have completed the 12-week treatment as of mid-2025 and a primary completion date of February 2026. If a pivotal study reads out negative, rentosertib will join the long list of Phase IIa direction-of-effect signals - cases where the treated arm moved in the expected direction but the result did not hold up in larger testing - that did not replicate.

For target-discovery methodology, the more instructive part lies upstream of Phase IIa. TNIK had to be identified before any chemistry could begin, and that happened through a workflow that combined multi-omics, literature mining, and knowledge-graph reasoning. The Phase IIa result matters because it suggests the target call was directionally right. The harder question is generalization across diseases, and that will take a decade of readouts to answer. Recursion’s May 2025 discontinuation of REC-994 is a useful reminder that AI-discovered targets can fail for the same reasons traditionally discovered targets fail; the failure mode is familiar (the LOA economics that explain why single-program failures matter less than cumulative field readouts are the subject of the therapeutic-area economics post).

What’s next, and where to watch

Four developments are worth tracking over the next 12–24 months.

Rentosertib Phase III. The pivotal trial has not started as of April 2026; Insilico’s public statements indicate that regulatory discussions are ongoing, and late-2025 reporting pointed to Q4 2025 as the target for late-stage trial initiation. A successful Phase III would be the next major readout because it would remove the easiest objection to the story - one 71-patient Phase IIa study is too thin a basis for broad claims. It still would not settle the methodological question on its own. A failed Phase III would also leave the broader question open.

Generalization across modalities. Rentosertib is a small-molecule kinase inhibitor. The same methodology has not yet been publicly demonstrated at comparable depth for antibody targets, nucleic-acid therapeutics, or targeted-protein-degradation modalities such as PROTACs and molecular glues, let alone cell therapies. Published target-discovery platforms can, in principle, output candidates regardless of downstream modality. The empirical record of “we predicted this target, built a non-small-molecule against it, and it worked” is still much thinner. The druggability post covers why that matters.

Open-source vs proprietary knowledge graphs. The field is split. Open resources such as Open Targets, Hetionet, PrimeKG, and Clinical Knowledge Graph are large, documented, and freely available, but they update slowly. Proprietary platforms inside target-discovery companies are larger, updated continuously, and tied to proprietary ML models, but they are hard to validate from the outside. Whether a fully open target-discovery stack can match a well-resourced proprietary one remains unresolved. The answer matters. A yes would democratize the workflow; a no would concentrate capability inside a small number of well-funded platforms.

Regulatory path. The FDA published its draft guidance on AI in drug development in January 2025, with final guidance expected in Q2 2026. The draft sets out a risk-based credibility framework for sponsors who use AI in regulatory submissions. For knowledge-graph-based target identification, which sits several steps upstream of any direct regulatory decision, the practical implications are still unsettled. The most immediate effect is likely to be pressure for provenance - a clear trail from each claim back to its supporting source data and model version. Every claimed edge in a submission KG will need a traceable source, and every ML-predicted target will need documented training data and model versioning. Tooling for that level of provenance is not yet standardized.

Concept Translation: The provenance question for regulatory-grade knowledge graphs is closer to model cards, dataset cards, and ML lineage tooling than to anything traditional in computational biology. Each predicted target needs a traceable answer to “where did the supporting edges come from, when were they extracted, what model or curator extracted them, and what version of the graph produced this prediction?” The ML community has been working through similar questions for several years (think bias audits, dataset documentation, version-pinned model artifacts). Drug-discovery teams are about to inherit the same problem at higher stakes.

For most of the last decade, target discovery was the part of drug development least altered by ML. That is no longer true. The rentosertib Phase IIa result does not resolve whether knowledge-graph-based target identification can supplant traditional, hypothesis-driven discovery; that will require many more programs and much more time. What it does establish is that ML has reached the front of the funnel. The live questions now concern reliability, boundary conditions, and failure modes.

Those are better questions than the field was asking two years ago.

Further reading

• Xu, Z., Ren, F., Wang, P. et al. (2025). A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized Phase 2a trial. Nature Medicine. DOI 10.1038/s41591-025-03743-2; PubMed PMID 40461817. The Phase IIa results paper for the rentosertib program - the primary source for the +98.4 mL FVC change, the 71-patient GENESIS-IPF trial design, the 22-site China execution, and the biomarker-correlation analyses the post summarizes in the “Worked example” section. Readers who want to verify the specific trial endpoints, dose arms, or safety findings should go here. Important note on currency: the published paper is the Phase IIa result only; Phase 3 status changes on a quarterly cadence, and readers should cross-reference ClinicalTrials.gov for the current pivotal-trial status at the time they read this post.

• Ren, F., Aliper, A., Chen, J. et al. (2024). A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nature Biotechnology. DOI 10.1038/s41587-024-02143-0. The discovery-to-Phase-1 history paper - the methodological companion to the Phase IIa paper above, covering target identification (the multi-omics and knowledge-graph workflow that surfaced TNIK), generative chemistry (the molecule-design arc), preclinical validation, and Phase 1 safety/PK. This is the primary source for the “Target identification” and “Target validation” subsections of the rentosertib worked example. Readers who want to understand the methodology in more detail than the blog post provides should go here first; it is the denser technical reference of the two rentosertib papers.

• Open Targets Platform documentation (2026). Open Targets Platform: data sources and scoring. [URL: https://platform.opentargets.org/docs]. The canonical open-source reference public biomedical knowledge graph - free, well-documented, continuously updated. The documentation covers the evidence categories, data sources (ChEMBL, UniProt, GWAS Catalog, OMIM, Reactome, and roughly 20 others), and target-disease scoring methodology that a practitioner would need to adapt for building a similar system or integrating Open Targets into their own pipeline. This replaces what would otherwise be a proprietary-platform reference and is the cleanest methodology-to-implementation handoff in the Further Reading block. Aligns with the sponsor-neutrality audit: an open, free, well-documented platform is a better reader resource than a commercial vendor’s whitepaper. (For the operational scoring rationale and the human-genetics weighting in particular, see also the Annual Review of Biomedical Data Science methodology paper from the Open Targets team - discoverable via the platform documentation.)

• Himmelstein, D. S., Lizee, A., Hessler, C., Brueggeman, L., Chen, S. L., Hadley, D., Green, A., Khankhanian, P., Baranzini, S. E. (2017). Systematic integration of biomedical knowledge prioritizes drugs for repurposing. eLife 6:e26726. DOI 10.7554/eLife.26726. The canonical academic reference for a working biomedical knowledge graph (Hetionet) combined with a link-prediction-based drug repurposing workflow. Hetionet is open-access, an order of magnitude smaller than commercial KGs but rigorously documented, and the paper walks through the graph-representation-learning-to-link-prediction pipeline the “Graph machine learning” section describes. Recommended as the single best academic introduction to the methodology for a reader who wants to see a working biomedical KG plus link-prediction treatment in one primary source.

Concept Translation: A phase transition is the binary pass/fail event that advances (or kills) a drug program at the boundary between preclinical, Phase I, Phase II, Phase III, regulatory filing, and approval. Clinical development is a Markov chain over those states, with attrition at every step. The headline numbers in this post are the per-step success rates and the joint probabilities they produce.

Why likelihood of approval matters in 2026

The 2011–2020 BIO/Informa/QLS Clinical Development Success Rates report remains the default citation for industry success rates. In 2026, two updates matter.

First, the headline has moved. For the 2014–2023 ten-year cohort tracked by Citeline (formerly Informa Pharma Intelligence, now part of Norstella), the average Phase I likelihood of approval has fallen to about 6.7%, down from 10.4% for the 2014 cohort alone. Phase-by-phase, the latest transitions are roughly 47% out of Phase I, 28% out of Phase II, 55% out of Phase III, and 92% from filing to approval. Phase II remains the industry’s main loss point. Phase I success fell from over 75% in the 2006–2008 cohort to under 40% in the most recent one. Some failures now surface earlier because biomarker-based efficacy surrogates are measured in Phase I, so programs that once would have failed in Phase II now fail sooner. The target pool has also shifted toward harder, first-in-class biology. The headline 90% overall clinical-failure rate from Phase I to approval still holds. Roughly half of failures are attributable to lack of efficacy and another 30% to safety, with the balance split across biomarker misidentification, trial-design issues, adherence, and unfavorable commercial landscape.

Second, the structure of success remains uneven. Biomarker-stratified programs reach approval at roughly twice the rate of unstratified ones. Stratified means the trial prospectively selects or enriches patients using a measurable marker such as a mutation, expression level, or protein abundance. That finding was striking when the 2011–2020 report first presented it (15.9% vs 7.6% Phase I LOA), and it has held up across methodology refreshes. It remains the strongest program-level effect a team can influence. It is also a target-discovery decision as much as a clinical-development decision, because the choice to pursue a target whose disease biology supports a stratifying biomarker is made upstream. This 2× effect sits alongside a separate, partially overlapping 2× effect from human genetic evidence. Programs whose targets carry general genetic support (broad GWAS signals, common-variant associations) approve at roughly twice the unsupported base rate, and programs supported by high-confidence causal evidence (Mendelian traits, high-impact coding variants) approach 3×. The effects compound when a program’s stratifying biomarker derives directly from its causal-genetics rationale.

Concept Translation: Biomarker stratification is patient pre-selection by a measurable feature, run before enrollment rather than after the trial reads out. Mechanically it is the same move as filtering a training set to a labeled subpopulation where the target effect is expected to be detectable. The trial then asks a sharper question on a smaller cohort, and the design carries less noise per patient. The 2× LOA improvement is the field’s empirical reward for sharper-question trial design.

Two background shifts shape the picture. The GLP-1 receptor agonist wave, the diabetes-and-obesity drug class built around glucagon-like peptide-1 signaling, has rewritten big-pharma revenue rankings since 2023; Eli Lilly’s market capitalization briefly crossed $1 trillion in 2025, and Evaluate Pharma projects Lilly to be the #1 pharma by 2030 revenue at roughly $113B, driven by Mounjaro and Zepbound. What changed is the revenue map; the success-rate map did not move in the same way. Oncology is still the most crowded and lowest-yielding area, hematology is still the highest-yielding. Which disease areas are hard to drug remains a question of biology and trial design that the GLP-1 cycle did not affect.

Concept Translation: GLP-1 receptor agonists mimic the body’s own glucagon-like peptide-1 hormone, which signals satiety and slows gastric emptying. Semaglutide and tirzepatide hit GLP-1 receptors (and in tirzepatide’s case, also GIP receptors) in the gut, brain, and pancreas, blunting appetite and improving glucose control. The class is on track to become the largest in pharma history, which is what makes the cardiometabolic revenue map shift so quickly even though base success rates in chronic disease have not changed.

The other shift is regulatory output. The FDA approved 55 novel drugs in 2023, breaking the 2020 record of 53. The counts in 2024 and 2025 were 50 and 46. The ten-year rolling average through 2024 is about 46.5 novel approvals per year, a historical high-water mark. Of those, 16 of 55 (29.1%) in 2023 and 13 of 50 (26.0%) in 2024 were oncology indications, the largest single-therapeutic-area share of CDER’s novel approvals in both years. CDER is the FDA center that handles most standard drug approvals. Add CBER’s cellular-therapy approvals, lifileucel (Amtagvi) for advanced melanoma in February 2024 and afamitresgene autoleucel (Tecelra), the first TCR-T cell therapy for synovial sarcoma, in August 2024, and oncology’s regulatory share is larger still when measured by approved-product count rather than CDER novel-NME count alone. CBER is the FDA center that handles biologics such as cell and gene therapies, TCR-T means T-cell-receptor-engineered T cells, and NME means new molecular entity. The industry is approving more drugs per year on average than ever, oncology remains the largest concentration at the regulatory end, and Phase I LOA has fallen by about a third. Those facts describe the same productivity story from different angles.

What “likelihood of approval” actually is

The term has a specific operational definition. Likelihood of approval (LOA) is the joint probability that a compound currently in a given phase eventually reaches FDA (or equivalent) approval. It is computed as the product of the phase-to-phase transition success probabilities for every phase remaining, across a reference cohort of programs. A worked example. If a drug is currently in Phase II and you observed 50% success for Phase II → III, 50% for Phase III → filing, and 50% for filing → approval across a cohort of ten programs per transition, the LOA for that Phase II drug is 0.5 × 0.5 × 0.5 = 12.5% across n=30 observations. For readers new to clinical development, the shorthand is simple. Phase I asks mainly whether the drug is safe and how the body handles it, Phase II asks whether there is an efficacy signal in patients, Phase III is the larger confirmatory test, and filing is the submission step to the regulator (usually as an NDA for small molecules or a BLA for biologics).

Concept Translation: LOA is the path probability through a Markov chain that allows only forward transitions or termination. The “states” are clinical phases and the “absorbing state” is approval. The headline LOA number is the cohort-average probability of absorption when starting from Phase I, which is a base rate, not a per-program prediction, in the same way base accuracy on a benchmark is not a prediction for a specific input. Conditioning on biomarker stratification, genetic support, or therapeutic area shifts the relevant cohort and therefore the conditional base rate.

A few practical things follow.

• LOA is a cohort statistic, not a program-level probability. It is the rate at which a randomly drawn program from the reference cohort reached approval. Applied to your Phase II compound, it’s only as useful as the reference cohort is relevant.
• “LOA from Phase I” is the standard headline number (industry-wide, 6.7% in the 2014–2023 cohort). It requires four multiplicative transitions and is therefore the most volatile statistic as you slice across disease areas and modalities.
• Transition success rates at each phase matter most if you’re trying to understand where programs die. The 2011–2020 BIO data and the 2014–2023 Citeline data agree on the central point. Phase II is the attrition chokepoint. Phase I is mainly about safety and pharmacokinetics (PK), meaning how the body absorbs, distributes, and clears the drug. Phase III is about powered efficacy, and most programs that reach it have already shown some human proof of concept, an early sign that the mechanism produces the intended effect in patients. Filing-to-approval is high (~92%) because most selection has already happened. Phase II is where target and indication choices meet the biology. AbbVie’s $8.7B acquired schizophrenia asset emraclidine missed both of its Phase II readouts in 2024, a reminder that even at multi-billion-dollar conviction, central nervous system (CNS) Phase II/III failure rates still run around 85%.

Two data framings need to stay distinct. The industry-wide LOA numbers (BIO 2011–2020; Citeline 2014–2023) sample all disclosed programs, including small biotech single-asset programs that fail early and exit quietly. The major-pharma-only numbers tell a different story. A 2025 Drug Discovery Today analysis of 18 leading pharma companies over 2006–2022 found an average LOA of about 14.3%, median 13.8%, range 8–23%. Those estimates are higher because they reflect internal selection. Large organizations kill their weakest programs before they ever appear in public disclosures. When you see LOA numbers in a blog post or investor deck, check which population the author is using. The two framings differ by roughly a factor of two for the same underlying biology.

The crowded-vs-abandoned map of therapeutic areas

The BIO/Informa/QLS analysis of 2011–2020 bucketed clinical-development transitions into 21 major disease areas covering 623 indications. Oncology, autoimmunity, infectious diseases, neurology, cardiology, hematology, metabolic, endocrine, psychiatry, respiratory, ophthalmology, allergy, gastroenterology (non-IBD), and urology make up the 14 “major” categories, with everything else (dermatology, renal, obstetrics, rheumatology for non-autoimmune indications, ENT/dental, orthopedics) swept into “Other”.

The headline findings from that decade of data still hold directionally after the Citeline refresh, per Item 8 of the 2026 Addendum:

• Oncology is the most crowded. Over 30% of all phase transitions recorded in 2011–2020 were oncology programs. Of the 12,728 total transitions in the dataset, oncology accounted for roughly 33%.
• Oncology is also among the least productive. Phase I LOA for oncology was 5.3%, the third-lowest of the 14 major areas and about 57% of the non-oncology average of 9.3%. Oncology had the lowest success rate at every clinical transition except one.
• The exception is regulatory. The NDA/BLA → approval transition ran at 92.0% for oncology vs 90.2% for non-oncology, a slight oncology advantage. Here, NDA means New Drug Application and BLA means Biologics License Application. One plausible interpretation is that oncology filings that reach the regulator are often supported by large effect sizes in biomarker-defined populations, and the FDA’s oncology center has accumulated substantial experience reviewing this kind of evidence. That reading fits the approval data in 2023 and 2024, where oncology dominated CDER’s novel-NME approvals in both years at 29.1% and 26.0% of total novel approvals respectively.
• Hematology is the highest-yielding major area. Phase I LOA sits close to 25%, roughly five times the oncology rate. Much of that performance comes from hemophilia research, which has unusually tractable underlying biology. The coagulation cascade is one of the best-characterized physiological networks in medicine, and 32 of 34 hemophilia A phase transitions in the dataset were successful. Anemia (multiple sub-indications) also runs well above average.
• Urology is the most abandoned. Less than 1% of all phase transitions.

Concept Translation: The coagulation cascade is a sequence of enzymatic activations that ends in fibrin clot formation. It has been mapped in detail since the 1960s, with most components purified, structures solved, and mutations cataloged. Hemophilia A is a defined deficiency of factor VIII, a single protein in the cascade; hemophilia B is a deficiency of factor IX. The combination of a clear genetic cause, a single missing protein per disease, and a measurable bleeding-time biomarker makes this one of the most ML-friendly subdomains in medicine. The “well-characterized” status pays off as base-rate Phase II success.

Why oncology is crowded and hard is the question most practitioners intuit but rarely state clearly. Commercially, first-in-class and best-in-class oncology wins can command some of the largest addressable markets in medicine, which keeps capital flowing in. Scientifically, cancer is hundreds of diseases collected under one label, many with multi-driver genetics, rapidly evolving resistance, and tissue-of-origin heterogeneity that shows up in Phase II when tumor biology diverges from the preclinical model. The 2024 corpus makes the same point directly. Scientific complexity and competitive dynamics continue to depress oncology success rates.

Hematology shows the opposite pattern. Programs often target a well-characterized protein or cell lineage in a circulating system, without the spatial heterogeneity of solid tumors or the neuroanatomical specificity of CNS disease. Patient populations are also often genetically well defined. A larger share of hematology biology is tractable enough to drug than in fields such as neurology.

Concept Translation: “Spatial heterogeneity of solid tumors” means that within a single tumor, different regions can carry different mutations, different gene expression profiles, and different sensitivity to drugs. A biopsy is one sample; the tumor is many samples that happen to share an organ. Hematological malignancies live in blood and bone marrow, which mix; the sampling problem is comparatively easier. CNS targets add their own tax, because the blood-brain barrier rejects many drugs at the door.

The big commercial players and therapeutic-area concentration. The top 10 pharmaceutal companies by revenue around 2021 all listed oncology as a primary therapeutic area. As of today, the position of oncology-related diseases as the most popular therapeutic areas because they generate the most revenue is only partly true. Cardiometabolic, powered by GLP-1s, has overtaken oncology in revenue growth. Oncology still holds the largest therapeutic-area concentration in R&D investment and CDER novel approvals.

Rare vs chronic, and why rare wins on a per-program basis

The BIO 2011–2020 analysis also split non-oncology programs into rare diseases (affecting fewer than 200,000 people in the US, or 1 in 2,000 people in the EU) and chronic, high-prevalence diseases (identified from the CMS, or Centers for Medicare & Medicaid Services, Chronic Conditions Data Warehouse, with >1 million US-affected patients). Both groups exclude oncology, which is itself 43% rare.

The finding was striking. Rare-disease programs had roughly three times the Phase I LOA of chronic, high-prevalence programs (17.0% vs 5.9%). Rare-disease programs outperformed the overall non-oncology cohort at every transition. The single largest gap was Phase II, at 44.6% for rare diseases vs 28.9% for the overall cohort. Phase I was also markedly better (67.4% vs 52.0%), with Phase III and NDA/BLA broadly comparable.

Three forces usually explain the difference. Rare-disease trials often enroll genetically stratified populations with cleaner disease biology. Effect sizes tend to be larger because untreated baseline disease is more severe. Regulators offer accelerated pathways such as orphan designation (the US rare-disease status that brings tax credits, fee waivers, and post-approval exclusivity), breakthrough therapy designation, and accelerated approval. The 2024 FDA annual report noted that more than 50% of CDER’s novel approvals carried orphan designation, quantitative evidence that the rare-disease pipeline is now a structural majority of new-product output rather than a niche. The claim is narrower than “rare disease is easier.” Rare-disease programs more often begin where the target-disease link is already visible in humans. Target discovery in a rare monogenic disease may start with a causative gene. Target discovery in heart failure with preserved ejection fraction often starts with clinical heterogeneity that has to be resolved before a target becomes legible.

Concept Translation: Orphan designation, breakthrough therapy, and accelerated approval are the FDA’s incentive system for diseases with small populations, severe unmet need, or strong early-phase evidence. Sponsors can stack them. Each one removes a different friction point: orphan removes some regulatory and economic disincentives, breakthrough partners the FDA review team with the sponsor and allows rolling submission, and accelerated approval lets the agency act on a surrogate endpoint with a confirmatory post-market study. The combination matters for the LOA picture because each removed friction point translates into a higher conditional success rate.

For a working drug hunter, rare-disease structure and biomarker enrichment often reinforce each other. A program with both a small, genetically defined population and a prospective enrichment biomarker sits inside the strongest historical success pattern in the data.

Biomarker stratification and likelihood of approval

Across all disease areas in the BIO 2011–2020 data, programs that used a patient-preselection biomarker achieved Phase I LOA of 15.9%, versus 7.6% for programs without one. A patient-preselection biomarker is a measurable feature used before enrollment to select patients more likely to respond, such as a mutation, receptor level, or lab value. This remains one of the clearest findings in the industry success-rate literature, and the 2026 Addendum notes that it still holds through the Citeline refresh.

The phase-by-phase breakdown shows where the difference appears.

• Phase I: no meaningful difference. Both biomarker-enriched and unselected programs clear Phase I at approximately the overall 52.0% rate, because Phase I is about safety and PK, not efficacy.
• Phase II: the gap is large. 46.3% for biomarker-supported programs vs 28.3% for unselected ones. This is the first direct test of whether “this target matters in this disease in these patients” is true.
• Phase III: the advantage persists, though it is smaller. 68.2% vs 57.1%. Programs that survived Phase II with a biomarker-defined cohort have already shown that their effect sizes are real in the relevant population.
• Filing → approval: a small residual difference. By then, Phase III design and data quality determine most of the outcome.

The causal chain begins with target selection. A biomarker is a program-level commitment to a specific subpopulation of patients in whom the target is hypothesized to matter. Finding a target whose biology supports that commitment is the target-discovery problem; the clinical-development problem comes later. When the target–disease evidence frameworks, including driver-vs-passenger distinctions, oncogene addiction, human genetic support, and multi-omics convergence, produce a target with a natural stratifier built in, the LOA doubles. Without that stratifier, the program falls back toward the industry base rate. Human genetic support shows a parallel pattern. Both deliver roughly 2× LOA improvement on average, and the multiplier scales toward 3× for high-confidence causal evidence (Mendelian traits, high-impact coding variants). When the same target carries both a strong genetic rationale and a stratifying biomarker, the gains compound.

Concept Translation: Biomarker-stratified trial design is roughly the clinical analogue of training a model on a curated subpopulation where the predictive signal is known to be present, rather than on the full distribution. The cost is generalizability (the result strictly applies to biomarker-positive patients) and the benefit is statistical power (the effect size is larger in the conditioned cohort). LOA doubling is the empirical reward for accepting that trade-off. ML practitioners will recognize the same logic from any setting where conditional models on labeled subgroups outperform a single model trained on a heterogeneous population.

This is the strongest quantitative support for the argument made in the front-of-funnel post. The biomarker-LOA finding is the number that argument rests on.

Biomarker stratification improves the chance of an apparent Phase II win. It does not guarantee that the biomarker is measuring the biology you think it is measuring. The PARP-inhibitor reversal is the clearest recent case study. Niraparib, olaparib, and rucaparib achieved early FDA approvals in advanced ovarian cancer based on strong progression-free-survival (PFS) signals in BRCA-mutant and HRD-positive populations, that is, patient groups selected for DNA-repair-related genomic features, and between late 2022 and late 2023 the FDA narrowed or withdrew six of those indications because mature randomized controlled trial data showed an overall survival detriment despite the earlier PFS gains. PFS is friendly to biomarker-enriched trial design; overall survival is the outcome that matters most to patients. The advantage is real in expectation, but the endpoint still has to track the disease.

Concept Translation: PARP inhibitors block poly(ADP-ribose) polymerase, an enzyme that helps cells repair single-strand DNA breaks. Cells with BRCA1/BRCA2 mutations or other homologous-recombination-deficiency (HRD) defects already cannot repair certain double-strand breaks well, so blocking PARP forces them into “synthetic lethality.” Losing two repair routes at once is fatal. PFS (progression-free survival) is time until the tumor grows or the patient dies; OS (overall survival) is time until death from any cause. PFS reads out faster and is easier to power, which makes it attractive for biomarker-enriched Phase II/III. The PARP reversal is what happens when the biomarker correctly selects responders on the proxy outcome (PFS) but the proxy fails to track the outcome that actually matters (OS).

Indication prioritization: the strategic companion to target selection

A good target is only half the program. The other half is indication prioritization, the choice of which disease to run the first trial in when a target is relevant to several. The course notes frame it plainly. “If the first indication is chosen incorrectly and clinical trials fail, subsequent trials with the next indication become difficult to launch”. A failed Phase II costs more than the trial itself. It can also stigmatize the asset when the team tries to relaunch it in a different indication.

The course gives four criteria for indication prioritization:

• Pathway relevance. Does the target play a key role in the pathogenesis of the candidate indication? Analyze omics and experimental data. Animal models may or may not recapitulate the human disease biology; treat them as useful but biased training labels.
• Competitive landscape. What does the existing-therapy and competitor-pipeline density look like? Diseases with no effective treatments may be high priority even if the science is harder; diseases with five approved comparators raise the efficacy bar.
• Unmet medical need and societal impact. Disability-adjusted life years (DALYs) are the standard unit for quantifying which diseases are most important to address. They roll prevalence, mortality, and morbidity into one number.
• Commercial effectiveness. Patient-population size, willingness-to-pay, time to market. Rare-disease indications with small populations can support high prices and accelerated paths; chronic-disease indications need volume and price to clear the hurdle.

Concept Translation: A DALY (disability-adjusted life year) is one year of healthy life lost. It combines years of life lost to premature death with years lived with disability, weighted by severity. The WHO and the Global Burden of Disease project use it as a single scalar for comparing burden across very different conditions. One death from leukemia at age 40 is many DALYs; a hundred years lived with mild migraines is fewer DALYs than the population count would suggest. Indication prioritization uses DALYs as a public-health weight on top of the commercial and scientific weights.

Those four criteria define a multi-criteria ranking problem over a set of candidate indications, with noisy partial signal in each criterion. Rank the N indications a target might plausibly treat by a learned weighting of DALY burden, pathway-relevance score, competitive density, and commercial headroom. The training signal is weak and slow. You get a label every five to seven years when a Phase II reads out. In practice, ML indication-prioritization tools in 2026 are decision supports rather than decision makers. They surface hypotheses the human strategy team should weigh, at a breadth a human team could not otherwise cover.

Indication expansion is the other lifecycle move. As of 2020, roughly 52% of the global drug pipeline was indication-expansion entries (new indications for already-approved drugs) versus 48% new molecular entities. In 2014 the split was roughly 50-50; by 2018–2019 expansion had risen to 65–67% of total entries. That trend fits directly into the LOA story. Indication expansion starts with a compound that has already cleared safety and PK, so the expansion-indication program inherits everything except the efficacy question. The 2014 BIO data has off-patent therapies running a Phase I LOA of 14.7% versus novel agents at 6.8%, roughly double, for the same reason.

Both prioritization and expansion are places where target-discovery knowledge reaches directly into commercial strategy. If you know which subnetworks of the human disease graph a given target modulates, you know which indications it could plausibly address, and you know where the strongest target-disease evidence exists to guide the first-indication choice. That is the rationale for the knowledge graphs for target discovery¹.

Worked example: Keytruda as indication prioritization done right

Pembrolizumab (Keytruda, Merck) is the most cited indication-expansion case in the course notes, and its trajectory follows the four criteria above in sequence.

First indication, 2014: advanced melanoma. At the time, advanced melanoma met two conditions that made it a sensible first indication. Profound unmet need existed (median survival of roughly a year or less for advanced metastatic disease in the pre-checkpoint era), and the competitive landscape had limited options even after ipilimumab’s 2011 approval. Melanoma is also immunogenically “hot,” meaning more visible to the immune system than many other tumors; tumor mutational burden is on average high. That matters for a drug whose mechanism (blocking PD-1 to unleash anti-tumor T cells) depends on there being a T-cell response to unleash. PD-1 and PD-L1 form an immune checkpoint, a receptor-ligand pair that dampens T-cell activity. Pathway relevance was strong, unmet need was severe, and the competitive landscape was manageable. Those were the right priorities.

Concept Translation: PD-1 sits on T cells. PD-L1 sits on many normal cells and on a fraction of tumor cells. When PD-L1 binds PD-1, the T cell pulls back its attack, which is how the body normally stops T cells from attacking healthy tissue. Tumors that learn to display PD-L1 hijack this brake to evade immune destruction. Pembrolizumab is a monoclonal antibody that physically blocks PD-1, releasing the brake and letting T cells attack. The mechanism only works on tumors the immune system can already “see,” which is why tumor mutational burden (the count of non-self-looking mutations per tumor genome) and PD-L1 expression both predict response. From an ML angle: the drug is a feature mask on a learned pathway, but only useful when the pathway has signal to mask in the first place.

Second indication, NSCLC. Non-small cell lung cancer became, per the course, the single largest contributor to Keytruda revenue and the dominant indication share in U.S. sales. NSCLC shows how expansion changes the economics. The patient population is an order of magnitude larger than advanced melanoma, the PD-L1 expression biomarker gives a way to enrich for likely responders before treatment, and the checkpoint-responsive fraction is substantial. The same mechanism entered a much larger market with a built-in biomarker.

Expansion footprint. Keytruda has been approved for more than 40 indications and tested in over 1,500 clinical trials as of early 2026. The 2026 Addendum notes that Keytruda is the most combined drug in clinical oncology history and places the looming 2028 composition-of-matter loss of exclusivity (the point at which the core patent protection runs out) at the center of Merck’s strategy. Three pieces define that strategy. Aggressive perioperative-indication expansion (movement into treatment settings before and after surgery), the subcutaneous Qlex formulation (given under the skin rather than by intravenous infusion, approved September 2025 across 38 indications), and biosimilar defense via formulation patents extending to 2039.

What Keytruda illustrates for the LOA story. Keytruda succeeded because its target had deep pathway relevance across many tumor types, its first indication offered both severe unmet need and a biomarker-defined responder population, and later expansions followed tumor types whose biology suggested they would respond to immune-checkpoint blockade. That is indication prioritization done well. It explains why Keytruda’s revenue trajectory looks so different from the median oncology asset even though oncology as a field carries a 5.3% industry-wide Phase I LOA. Program-level LOA depends on target quality and indication choice. Crowded therapeutic areas can still produce standout programs; they are unforgiving to average ones.

The non-oncology mirror is botulinum toxin (Botox, Dysport), a smaller revenue story but a long-arc example of the same logic. It was initially approved for narrow neuromuscular indications, then systematically expanded into migraine, overactive bladder, and cosmetic indications over two decades. One mechanism, blocking acetylcholine release at neuromuscular junctions, supported a widening set of indications. Pharmacology and indication prioritization stayed aligned.

Concept Translation: Acetylcholine is the chemical messenger that motor neurons release at the synapse where they meet a muscle fiber (the neuromuscular junction). The release tells the muscle to contract. Botulinum toxin cleaves the proteins that allow that release, and the muscle stops getting the contract signal. In facial cosmetics, this smooths wrinkles by paralyzing small muscles. In migraine, it interferes with sensory-nerve signaling. In overactive bladder, it relaxes the detrusor muscle. Same mechanism, different muscle or nerve, different indication.

What likelihood of approval means for an ML practitioner

For practitioners entering drug discovery from an ML background, LOA and therapeutic-area economics matter in three operational ways.

First: it defines the loss function at the program level. Approval is a rare, slow, expensive label: 6.7% base rate, 5–10 year lag, and per-asset cost estimates whose means span roughly $0.95B to $2.23B depending on whether high-cost outliers are excluded, with medians closer to $0.7B–$1.0B (RAND 2025 median: $708M; JAMA 2020 median: $985M; Deloitte 2024 mean for the top-20 biopharma cohort: $2.23B; Tufts CSDD’s widely cited $2.8B mean is heavily weighted by ~$1.16B in foregone-investor returns over the 10–15 year development window plus ~$312M in post-approval R&D). You cannot optimize directly against that outcome the way you optimize a docking score against a binding-affinity benchmark. You work instead with proxies that correlate with approval: human genetic support, target-disease evidence quality, druggability scores, and biomarker-enrichability. The biomarker-stratification result is the clearest single proxy in the literature. Any target-prioritization pipeline that scores candidate targets on whether a target has a plausible patient-stratification biomarker is, in effect, modeling the 15.9%-vs-7.6% split. The quantifying drug target novelty² is a second proxy axis. The target–disease evidence frameworks systematize the first.

Second: it sets the prior on therapeutic-area choice. If your platform is modality-agnostic or pathway-focused, therapeutic-area choice becomes a modeling decision. Oncology is data-rich (TCGA, The Cancer Genome Atlas; ICGC, the International Cancer Genome Consortium; GEO oncology collections, the Gene Expression Omnibus; and the Cancer Dependency Map, a large set of gene-essentiality screens) and low base rate. Rare genetic disease is data-poor (small patient cohorts, narrow genomic catalogues for any single disease) and high base rate, with accelerated regulatory paths. The ML-productivity question is where the data sits in a therapeutic area whose LOA can justify the modeling work. Many 2024–2026 platform biotechs that started in oncology have explicitly added rare-disease indications for exactly this base-rate reason.

Third: it quantifies the cost of being wrong. Funnel cost estimates put numbers behind upstream decision quality. Target validation absorbs roughly $353M over 2.5 years; lead optimization runs over half a billion; Phase II/III adds roughly half a billion more. Phase 2 completion, successful or unsuccessful, absorbs around 20% of the per-drug discovery pipeline spend. The directional point about Phase II cost concentration is consistent with the contemporary RAND/JAMA/Deloitte cost-distribution work even where the exact magnitudes differ. A target-assessment tool that trims Phase II failure by a few percentage points can pay for itself in industry-average units. The biomarker-stratification result shows the scale of the upside when upstream decisions improve.

Chapter 1 of Build AI Drug Discovery Pipelines frames the scale problem: 10⁶³ drug-like molecules, ~10⁵ potential human protein targets, a $1–3B and 10–15 year cost to move from ideation to market, and Eroom’s Law as the secular productivity decline that this book’s methods are intended to push against. The LOA economics in this post are the companion to the chemical search space. Even if chemistry were solved, the pipeline’s output rate would still be constrained by target and indication choice. Part of the productivity gap that Eroom’s Law names sits at the front of the funnel.

LOA numbers to watch through 2027

Three developments are worth watching over the next twelve to twenty-four months.

The Citeline numbers will refresh. The 2014–2023 cohort is the current industry reference; a 2024 or 2015–2024 update will arrive. Watch whether the biomarker-stratification doubling holds up. The front-of-funnel ML argument leans heavily on that result. Also watch whether Phase II transition rates continue to slide or whether the tentative post-2018 recovery reported in the 2025 Nature Communications analysis (cited but contested in the Citeline coverage) proves to be signal rather than artifact.

The AI-discovered-drug cohort will produce its first registrational readouts. Registrational readouts are late-stage trial results intended to support approval. Rentosertib (Insilico Medicine’s INS018_055 / ISM001-055), the first widely recognized drug whose biological target and molecular structure were both nominated by generative-AI systems, is the clearest signal program. Insilico’s platform identified TNIK (TRAF2 and NCK-interacting protein kinase) as a master regulator of fibrosis and inflammation, then designed a selective small-molecule inhibitor; the discovery-to-clinical-trial timeline ran under 30 months versus a 4.5–6 year traditional industry average. Phase IIa results published in Nature Medicine in June 2025 showed a +98.4 mL change in forced vital capacity (FVC), a standard lung-function measure, in the 60 mg arm versus a −20.3 mL placebo decline at week 12 (n=71 patients, 12 weeks, not powered for pivotal efficacy). More than 173 AI-discovered drug programs are tracked in clinical development as of early 2026, with 15–20 expected to reach pivotal Phase III this year (industry tracker reporting; the cohort is not yet mature enough to compute its own LOA). The field will be judged on cohort LOA. Ask again in 2028.

The therapeutic-area revenue-vs-success-rate gap may widen. GLP-1 economics are rewriting cardiometabolic R&D priorities, and cardiometabolic remains a large, chronic, high-prevalence area where, per the BIO rare-vs-chronic analysis, Phase I LOA runs around 5.9%. The next wave of GLP-1-adjacent targets (GLP-1/GIP dual agonists, amylin analogs, triple agonists, oral formulations) will be crowded and, on historical base rates, hard. Oncology, meanwhile, is structurally unchanged. Whether cardiometabolic’s post-GLP-1 base rate improves because GLP-1-like programs offer clearer mechanisms and better stratification, or stays near the chronic-disease average, is an open empirical question for 2027–2030 cohort readouts.

The LOA picture will not get simpler. Every additional year of data makes the cohorts denser, more modality-heterogeneous (small molecules, biologics, ADCs, cell therapies, gene therapies, oligonucleotides, each with its own base rates), and more biomarker-stratified on average. Approval counts are at historical highs while Phase I success is at historical lows, and interpreting both trends depends on selection into Phase I. Both possibilities matter (sharper programs entering, noisier programs entering), and in 2026 the most defensible answer is that both effects are present, with biomarker-stratified programs accounting for much of the productive output.

Further reading

• BIO, Informa Pharma Intelligence, and QLS Advisors. (2021). Clinical Development Success Rates and Contributing Factors 2011–2020. [URL: https://www.bio.org/clinical-development-success-rates-and-contributing-factors-2011-2020]. The foundational industry-wide analysis covering 12,728 phase transitions across 21 major disease areas. Source for disease-area rankings, the biomarker-stratified LOA doubling (15.9% vs 7.6%), and the rare-vs-chronic disease comparison (17.0% vs 5.9% Phase I LOA). The analytical framework the industry still uses a decade after the data cutoff, and the specific report this post’s crowded-vs-abandoned map section is built on.

• Citeline / Norstella. (2024). Why are clinical development success rates falling? Industry analysis. [URL: https://www.norstella.com/why-clinical-development-success-rates-falling/; accessible coverage at https://pharmaphorum.com/rd/clinical-development-rates-are-falling-its-not-all-bad-news]. The current industry reference for headline LOA figures, covering 2014–2023. The primary source for the 6.7% Phase I LOA figure, the phase-by-phase transition rates (47%/28%/55%/92%), and the Phase I success-rate decline from >75% (2006–2008) to <40% (most recent cohort). Citeline is the successor entity to Informa Pharma Intelligence, now part of Norstella; cite Citeline/Norstella for 2026-current blog posts and BIO for the 2011–2020 historical baseline.

• FDA Center for Drug Evaluation and Research. (Annual). Novel Drug Approvals [year]. [URL: https://www.fda.gov/drugs/novel-drug-approvals-fda]. Annual summaries link from this page. The canonical per-year approval counts, orphan and expedited-program breakdowns, and PDUFA-goal-date statistics that the post’s 2023/2024/2025 approval-count figures (55, 50, 46 respectively) reference, along with the 29.1% (2023) and 26.0% (2024) oncology share of CDER novel approvals discussed in the crowded-vs-abandoned section. Pair with the Nature Reviews Drug Discovery annual FDA approvals analysis; search “Nature Reviews Drug Discovery FDA approvals annual analysis [year]” for readable per-approval commentary.

• Lai, R., et al. (2025). Productivity of the top 18 pharmaceutical companies measured by likelihood of approval of their compounds 2006–2022. Drug Discovery Today. [URL: https://www.sciencedirect.com/science/article/pii/S1359644625000042]. Empirical analysis that produced the major-pharma-only LOA figure (mean 14.3%, median 13.8%, range 8–23%) the post cites for contrast against the industry-wide Citeline numbers. Essential reading for the “industry-wide vs major-pharma-only” distinction, the factor-of-two framing the post uses to warn against mixing cohorts in LOA comparisons.

Concept Translation: Tissue expression follows a long-tailed distribution. A few thousand genes are tightly localized; most are spread broadly across the body. Drug targets are disproportionately drawn from the localized end, for a direct reason: a drug only acts where its target lives. If the target lives everywhere, the drug acts everywhere, and so do its side effects.

Why tissue specificity matters in 2026

Two shifts have moved tissue specificity closer to the center of target assessment in 2026.

First, the modality landscape has widened. Antisense oligonucleotides, siRNAs, PROTACs, ADCs, and CAR-T cells all create different on-target, off-tissue failure modes and now sit inside routine target assessment conversations (see druggability in the AlphaFold 3 era). On-target, off-tissue means the drug is acting on its intended target in a tissue where that biology causes harm. A PROTAC that degrades its target in tumor tissue will also degrade it in healthy tissue where the target is expressed. An ADC whose surface antigen is not strictly tumor-restricted creates a toxicity risk. The January 2024 FDA class-wide boxed warning on CD19- and BCMA-directed autologous CAR-T products for T-cell malignancies reflected how deeply the modality can penetrate immune-cell populations, rather than a chemistry failure in the CAR construct. Tissue-specificity evidence now has to travel across more modalities, and each modality adds its own interpretation layer on top of the same expression data.

Concept Translation: “Off-target” and “off-tissue” are commonly conflated. Off-target means the drug is hitting proteins it shouldn’t, which is a selectivity problem solved by better chemistry. On-target, off-tissue means the drug is hitting the right protein, in the wrong place. The drug is doing exactly what it was designed to do, on the protein it was designed to do it to, but in a tissue where that action causes harm. Tissue-specificity analysis is the screen for the second kind. No amount of selectivity tuning fixes the second problem; it has to be addressed at target selection.

Second, safety-evidence standards are shifting. The FDA Modernization Act 2.0 (December 2022) authorized alternatives to traditional animal testing in IND submissions, but did not require them, and the follow-on FDAMA 3.0 passed the Senate by unanimous consent in December 2025. In April 2025, the FDA announced a roadmap to phase out routine animal testing starting with monoclonal antibodies, with faster review timelines as an incentive for sponsors that adopt validated organ-on-chip, organoid, and AI-based safety evidence. Animal testing remains the default in practice, and NAMs (new approach methodologies) are usually used alongside it rather than in place of it. Computational tissue-safety evidence, which depends heavily on the databases and methods in this post, is likely to carry more weight in future target prioritization.

In the target-discovery series, tissue specificity sits inside the safety criterion. The pillar post (the front-of-funnel decision) listed five target-assessment criteria, and the evidence-of-linkage post (target-disease evidence frameworks) covered one of them. This post covers the part of safety that can be estimated early: where a target is expressed, and where on-target, off-tissue toxicity is therefore most likely to appear.

What “tissue-specific” means when you look at the numbers

Only a modest fraction of human protein-coding genes are tissue-specific. The Human Protein Atlas classifies roughly 15% of protein-coding genes as tissue-enriched, while around 40% are not tissue-specific at all. About 5% are detected in just a single tissue; another ~17% are restricted to a small group of tissues that may count as a coherent compartment, such as the several structures of the brain. Compared with housekeeping genes (genes expressed consistently across tissues), tissue-specific genes are more tightly tied to organ biology and are overrepresented among drug targets.

Marketed non-cancer drugs tend to be more enriched for tissue-specific targets than anti-cancer drugs. We’ll work through an example to see why.

GABRB2, the brain-specific β2 subunit of the GABA-A receptor, is the target of clomethiazole, a sedative and anticonvulsant. Because GABRB2 is largely confined to brain, clomethiazole’s on-target effects are concentrated where they are wanted, while on-target activity elsewhere is limited. That is the narrow-expression logic in its clearest form: when the protein largely resides in the desired tissue, the program starts with a built-in safety advantage.

The counterexample is BCL-XL. It is an anti-apoptotic protein over-expressed in many cancers, and inhibitors of the pathway, most notably navitoclax, were designed to remove that survival defense from tumor cells. The pharmacology worked, then met a hard clinical limit. Healthy platelets, the blood cell fragments required for clotting, also depend on BCL-XL for survival, so navitoclax produced severe dose-limiting thrombocytopenia (a dangerously low platelet count) at therapeutic doses. This is on-target, off-tissue toxicity. The drug is acting on the intended protein in tumor cells and in platelets, and that broad expression compresses the therapeutic index, the gap between an effective dose and a toxic dose. Tissue-specificity analysis is meant to identify this problem before the program reaches humans.

The logic also shifts across therapeutic areas. In oncology, the primary goal is to eliminate or suppress cancer cells, so side effects that would be unacceptable for a chronic non-oncology therapy, including vomiting, alopecia, and hematological toxicity, are often tolerated because the alternative is untreated cancer. Tumor-agnostic therapies, which are selected by a shared molecular lesion rather than organ of origin, introduce a second shift. When a molecular lesion defines the target, patient selection moves from organ of origin to the cells carrying the driver alteration.

The two canonical tissue-expression databases: GTEx and HPA

Most target-selection workflows that evaluate tissue specificity rely heavily on two public resources.

GTEx (Genotype-Tissue Expression). A reference atlas of bulk RNA-seq across normal human tissues from post-mortem donors, covering more than 50 tissues across a wide sample population. GTEx is the standard source for baseline RNA expression of a candidate target across healthy human tissues, and its eQTL calls (tests of whether genetic variants shift gene expression) link genotype to tissue-specific expression. The current major release at the time of writing is GTEx V10 (released November 2024), which expanded the prior version with roughly 12% more RNA-seq samples, a small RNA-seq dataset spanning approximately 16,760 tissue samples, and high-coverage (~195×) deep somatic whole-genome sequencing on a substantial donor subset. The somatic-WGS layer adds a second question to the usual expression lookup. In addition to asking where a target is expressed in normal tissue, teams can ask where somatic variants in or near that gene appear in healthy donors, which helps frame whether off-tissue expression is genetically modulated in ways that matter for safety.

Human Protein Atlas (HPA). A parallel resource with stronger protein-level coverage. HPA integrates RNA-seq data with immunohistochemistry (antibody staining of tissue sections) across a large number of cell types and tissues, and increasingly includes cell-type-specific expression data generated by single-cell RNA sequencing. It is the resource to consult when the question is whether a candidate target is actually translated in a given tissue, rather than merely transcribed there.

The two resources are complementary. Most contemporary target-assessment templates expect both to be consulted before a target advances to the feasibility stage. A common workflow maps the tissue-expression footprint of a candidate target with GTEx, then checks HPA protein and cell-type views for organs where on-target modulation would create a safety risk. That sequence is now a standard pre-IND analysis (work done before an Investigational New Drug filing).

Why both mRNA and protein matter

The central dogma’s DNA → RNA → protein chain suggests that mRNA abundance should be a reasonable proxy for protein abundance. In practice, the correlation is often modest, and sometimes systematically so. Three reasons matter most.

• Protein stability and half-life vary. A protein with a short half-life can have a much lower steady-state abundance than its transcript would predict.
• Post-translational modifications regulate activity. A protein may be abundant yet inactive because it is dephosphorylated, sequestered, or proteolytically cleaved, making it less druggable than its mRNA profile suggests.
• The tumor microenvironment is non-stationary. Cancer tissues tend to show lower mRNA-protein correlation than normal tissues, because the proteome is changing faster than the transcriptome at any given biopsy snapshot, and post-translational modifications in the tumor microenvironment further decouple the two.

Concept Translation: A reasonable mental model: DNA is the static codebase; mRNA is a dispatched job request; protein is the running worker. Counting RNA reads is like counting jobs spawned per hour, not workers currently active. Protein lifetime, modifications, sequestration, and degradation all live between dispatch and execution. The two metrics are correlated, sometimes loosely, but they answer different questions. Drug targets are proteins, so a tissue-safety claim built only on RNA is solving a related-but-not-identical problem.

A tissue-specificity claim built only on GTEx RNA is fragile for a meaningful minority of targets. The obvious next check is protein-level evidence. HPA immunohistochemistry is the most common starting point, though targeted mass spectrometry and emerging single-cell mass-spectrometry methods are becoming more relevant.

The species-conservation problem

A related complication gets less attention than it should. Tissue-specific genes are less evolutionarily conserved than broadly expressed genes. Tissue-specific genes identified in humans have substantially fewer one-to-one mouse orthologs (genes in different species descended from the same ancestral gene) than non-tissue-specific genes, regardless of how tight the tissue specificity is. Tissue-specificity correlations between species degrade quickly below chimpanzee, with fly-to-human correlations near the bottom of the ordering.

Concept Translation: The species-conservation problem is a domain-shift problem. Mouse models are “in distribution” for biology that is shared with mouse, namely broadly expressed, deeply conserved genes, and “out of distribution” for human-specific or primate-specific tissue biology. The narrow, tissue-enriched targets that the framework most wants to validate in mouse are also the ones where mouse is least informative. The mismatch is structural, not a question of model quality.

The result is a familiar blind spot in preclinical development. Mouse safety data may not faithfully reflect the human safety profile for exactly the tissue-enriched targets where the framework most wants confidence. This mismatch is one reason alternative safety-evidence methods (iPSC-derived models, where adult cells are reprogrammed into stem-like cells and redifferentiated in culture, organ-on-chip systems, organoids, and human-cell-based in silico approaches) have been gaining regulatory traction in parallel with the FDA Modernization Acts.

Single-cell and why bulk isn’t enough

Bulk RNA-seq, the data type that underlies the classical GTEx and HPA views of tissue specificity, averages expression across millions of cells in a tissue sample. If the target of interest is expressed by a rare cell population within that tissue, or at different levels in different cell types, bulk data can miss it or average it away. In diseases where cellular heterogeneity drives phenotype, such as most cancers, most autoimmune diseases, and many neurodegenerative conditions, that averaging becomes a real limitation.

Concept Translation: Bulk versus single-cell is the difference between a dataset summary and a per-row dataset. Bulk RNA-seq returns mean expression across the cells in a tissue sample, the way a query that returns AVG(feature) across a million rows returns one number. Single-cell RNA-seq returns one row per cell. If the disease is driven by 1% of the cells, the average is a useless statistic.

Single-cell RNA-sequencing (scRNA-seq) addresses that problem, with substantially more analytical overhead. Sample preparation requires tissue dissociation followed by isolation of individual cells. The resulting data then goes through denoising, clustering, dimensionality reduction, and visualization pipelines that bulk data does not require, and the per-sample cost is roughly 10× that of bulk RNA-seq. The result is a per-cell-type expression profile that lets you ask the question bulk data cannot answer directly: is this target really only in hepatocytes (the main functional cells of the liver), or is it also in the rarer hepatic stellate cell population?

Three technologies, three trade-offs

Practitioners typically choose among three classes of single-cell assay, each with distinct trade-offs.

Plate-based (SMART-seq-like). Cells are isolated by sorting, deposited into wells, lysed, and their RNA amplified with unique identifiers before cDNA synthesis and sequencing. The method sequences thousands of cells with good per-cell gene coverage. Best when you need deep transcript characterization of a small population.

Droplet-based (10x Genomics Chromium being the most common implementation). Microfluidics places individual cells in droplets carrying unique barcodes; first-strand cDNA synthesis happens in-droplet, followed by sequencing. This scales to hundreds of thousands of cells, at the cost of lower gene coverage per cell. The default in most contemporary scRNA-seq-based target discovery work.

Spatial transcriptomics. Links sequencing data to spatial position in a histological section by placing tissue on a substrate coated with barcoded probes whose position is known. RNA is captured onto the barcoded probes, and the barcode sequence tells you where each transcript came from. This matters when tissue organization carries biological information, such as tumor margins, crypt structures in the gut, and laminar layers in the brain, that dissociated scRNA-seq destroys.

Each of these technologies carries its own analysis burden. Single-cell data is sparse and zero-inflated rather than normally distributed, meaning many measured values are zeros. That produces dropout events, where a gene that is present in a cell is missed in sequencing and appears absent. Droplet-based methods can capture two or more cells in one droplet, producing multiplets, mixed-cell profiles that inflate expression estimates or fabricate apparent cell populations. A scRNA-seq analysis that does not handle these artifacts explicitly will produce tissue-specificity claims that do not hold up on replication.

Concept Translation: Two terms here collide with ML vocabulary in ways that confuse cross-disciplinary readers. Dropout in scRNA-seq means a gene is genuinely present in the cell but its transcripts were missed during sampling, a sparse-data artifact with nothing to do with the regularization technique. Multiplets are closer to hash collisions: two distinct cells assigned to the same barcode, indistinguishable downstream. Both produce systematic biases that look like real signal if not corrected, which is why standard scRNA-seq pipelines include explicit dropout-imputation and doublet-detection steps.

Multi-modal single-cell beyond CITE-seq

Single-cell methods can be combined to cover more than just transcriptomics.

• CITE-seq uses antibodies conjugated to oligonucleotide barcodes to measure cell-surface proteins alongside the transcriptome. This matters most when the candidate target is a cell-surface antigen, for example when assessing the tumor selectivity of a candidate CAR-T or ADC antigen, where protein presence on the cell membrane matters more than transcript presence in the cytosol.
• Single-cell mass spectrometry and mass cytometry are moving from development into production use, with the goal of getting protein-level resolution at single-cell scale.
• Combined assays pair scRNA-seq with ATAC-seq, a readout of open and potentially regulatory DNA, or methylation, and extend the single-cell approach into multi-omic tissue-specificity analysis.

What single-cell data unlocks for target selection

Three applications are especially relevant for tissue-specificity evaluation.

Identifying disease-driving subpopulations. In heterogeneous diseases, including most cancers and autoimmune disorders, a small population of cells can drive the phenotype. Single-cell data can resolve these populations and identify their expression signatures, defining both the cell type a drug should target and the tissue distribution of that cell type.

Gene regulatory network inference. Methods like SCENIC (Aibar et al., Nat Methods 2017) reconstruct transcription-factor regulatory networks from single-cell data, mapping which transcription factors appear to control which downstream genes. The follow-on SCENIC+ (2023) extends the framework to multiomic enhancer-driven networks. Networks reconstructed from patient single-cell data can highlight altered pathways in disease. That is useful for candidate target discovery and for understanding which cell populations a known target acts within.

Cell-cell communication inference. Ligand-receptor analysis on scRNA-seq data can predict how cell populations signal to each other. In cancer, it can show that cancer-associated fibroblasts are regulating EMT-program expression in adjacent tumor cells, where EMT refers to epithelial-to-mesenchymal transition (a cell-state program associated with invasion). That, in turn, raises the question of whether the fibroblast ligand or the tumor-cell receptor is the better target.

Worked example on the NTRK fusion-driven cancer story

A useful counterexample to the “narrow expression equals safer target” framing is larotrectinib and the NTRK fusion story.

NTRK (neurotrophic tyrosine receptor kinase) is the canonical tumor-agnostic target, a target selected by molecular lesion rather than organ of origin. In most cancers where it has been implicated, the driver lesion is a fusion, typically with a partner that leaves the kinase in an always-on state, producing an abnormally active TRK fusion protein. NTRK fusions appear across multiple cancer types; in 2018 the FDA approved larotrectinib (Vitrakvi, Bayer/Loxo Oncology) for any cancer driven by an NTRK gene fusion, making it one of the first tumor-agnostic targeted therapies.

Larotrectinib is now one of several NTRK-fusion options. Entrectinib (Rozlytrek, Roche) was approved in 2019 with broader spectrum activity including ROS1 fusions; repotrectinib (Augtyro, BMS) was approved in 2023 with improved potency and activity against resistance mutations. Tumor-agnostic NTRK-fusion treatment is now a three-drug class.

Concept Translation: Tumor-agnostic targeting is closer to predicting on a feature pattern than predicting on a category label. The classical model says: if the organ of origin is X, treat with drug Y. NTRK-style tumor-agnostic targeting says: if the tumor expresses the fusion-protein feature pattern Z, treat with drug Y, regardless of where the tumor sits. The diagnostic problem becomes pattern recognition over molecular features, and the tissue label drops out of the prediction.

The lesson for tissue-specificity reasoning is direct. Patient identification for NTRK-targeted therapy is molecular, based on NTRK fusion by IHC (immunohistochemistry) or NGS (next-generation sequencing), rather than tissue-specific. TRK kinase activity is tolerated at the doses used because the fusion-expressed form is the version doing most of the clinically relevant signaling in the tumor, while the normal distribution of wild-type TRK is not the main thing the drug is hitting. In this setting, the key question is whether the driver form of the target is restricted to disease tissue by the genetic lesion the patient carries.

Computationally, tumor-agnostic programs depend on cross-cancer analysis: assembling genomic and expression data across many tumor types to identify which lesions recur across tissues and which do not. The Cancer Genome Atlas (TCGA) and the pan-cancer compendia it seeded (datasets that pool tumors across many cancer types) are the backbone data for that work. Tissue specificity still matters here, but in a different form from the narrow-expression framing.

Cross-modal tissue safety in 2026

Up to this point, the discussion has mostly assumed a small-molecule drug, the modality where tissue-expression analysis has the longest methodological history. The modality landscape has broadened, and each new modality adds its own tissue-safety question.

• PROTACs and molecular glues. Catalytic degradation means even transient target expression in a non-target tissue can produce a lasting effect. Degradation half-life, the identity of the recruited E3 ligase (the ubiquitin-tagging enzyme brought in by the degrader), and the rate of resynthesis of the target protein all enter the off-tissue-exposure calculation in ways that reversible binding does not.
• Antisense oligonucleotides and siRNA. Distribution is heavily tissue-biased by chemistry. Unconjugated ASOs partition to liver and kidney; GalNAc-conjugated siRNA, where GalNAc is a sugar ligand used to direct uptake into hepatocytes, partitions almost exclusively to hepatocytes; lipid-nanoparticle-delivered siRNA is dominated by liver uptake. Tissue-distribution safety here is partly defined by where the chemistry delivers the molecule, not merely by target expression.
• CAR-T. The 2024 boxed warning is a direct consequence of CAR-T products hitting antigens that are shared with normal tissues, combined with the general genotoxic risk of lentiviral integration (the insertion of vector DNA into the engineered T-cell genome). CAR-T tissue safety therefore requires thinking about the modified T-cell’s own trafficking and persistence (where those engineered cells travel and how long they remain active) alongside the antigen’s tissue distribution on tumor cells.
• ADCs. The antibody’s target antigen must be sufficiently tumor-selective because payload leakage at non-target sites is a known failure mode. There are 15 FDA-approved ADCs as of mid-2025, and both the linker chemistry (how the antibody is attached to its toxic cargo) and the payload class shape the tissue-safety calculation.

Concept Translation: ASOs and siRNAs are an interesting case for ML readers because the delivery vehicle enforces a routing constraint that the target does not control. A GalNAc tag is functionally a hardcoded address: regardless of where in the genome the matching transcript lives, the molecule will be delivered to liver. The drug acts on its target only in the tissue the chemistry can reach. This decouples target tissue distribution from drug tissue distribution in a way that small-molecule pharmacology rarely does.

All of these modalities draw on tissue-expression data from GTEx, HPA, and scRNA-seq. What changes is the interpretation layer. A single generic tissue-safety score is therefore less useful than a modality-conditional one. Graph-based representations fit this structure well, because tissue, cell type, target, modality, and interaction type can all be represented as different node classes connected by learned relations.

What’s next in tissue-specificity analysis

Multi-modal reference atlases and their integration. The Human Cell Atlas, Tabula Sapiens, and several disease-specific scRNA-seq consortia are converging toward reference panels that include tissue-level, cell-type-level, and in some cases spatial resolution. For a candidate target, the 2026 question is how to integrate these reference atlases with patient-specific single-cell data into a tissue-safety assessment that predicts rather than merely describes. Cell-type annotation transfer, using a reference atlas to assign identities to cells in a new dataset, is an active ML problem and the benchmarks are improving, but there is no single accepted pipeline for computing a target’s tissue-safety score.

Concept Translation: Cell-type annotation transfer is transfer learning on atlas-scale labeled data. The reference atlas provides expert-curated cell-type labels for hundreds of thousands of cells; a new dataset comes in unlabeled. The goal is to project the new cells into the reference space and inherit the labels, accounting for batch effects, technology differences, and cell-state shifts induced by disease. The same general toolkit as cross-domain classification, applied to a domain where the label space (cell types) is itself contested and evolving.

Modality-specific tissue-safety modeling. The next obvious ML problem is predicting modality-specific on-target, off-tissue toxicity. The field still does not have a standard benchmark. It is a good fit for anyone with access to regulatory safety datasets and the willingness to build modality-conditional representations on top of the tissue-expression atlases.

Regulatory relevance of computational tissue-safety evidence. FDAMA 3.0 and the FDA’s April 2025 roadmap to phase out routine animal testing for monoclonal antibodies both push computational safety evidence up the regulatory priority list. In practice, animal testing remains the default and NAMs are used alongside rather than in place of animal studies. Tissue-expression databases, now enlarged with GTEx V10’s deep somatic-WGS layer, are among the most mature inputs to a computational safety dossier.

Tissue-distribution safety for target pairs. The synthetic-lethality case raises a question the single-target framing does not. If the therapy is a combination of two agents that hit different targets, the relevant safety question is the tissue distribution of both targets simultaneously. Tissue specificity of a target pair is less developed than tissue specificity of a single target and is a natural next step for methods that already reason over single targets: synthetic lethality.¹

At the narrowest level, tissue specificity retires candidates before they ever get expensive. More broadly, “where does this protein live?” can no longer be answered with a single database lookup. It now requires a stack that runs from bulk RNA and protein atlases through single-cell and spatial measurements to modality-specific safety reasoning. Much of the ML opportunity in front-of-funnel target discovery sits in stitching those layers together.

Further reading

• GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science (2020). doi:10.1126/science.aaz1776. The canonical primary citation for GTEx as a tissue-expression resource; the GTEx Portal (gtexportal.org) is the companion hands-on reference for practitioners working directly with the data. Cite the 2020 Consortium paper for methodological grounding, and link the portal for the live tissue-expression viewer that readers new to the resource will actually use. 2026 update: the current major release is GTEx V10 (November 2024), which expanded the prior version with ~12% more RNA-seq samples, small RNA-seq across roughly 16,760 tissue samples, and deep (~195×) somatic whole-genome sequencing on a substantial donor subset. For any release-version-specific claim, verify against the portal at publication time.

• Uhlén, Fagerberg, Hallström, Lindskog, Oksvold, Mardinoglu et al. (2015). Tissue-based map of the human proteome. Science 347, 1260419. doi:10.1126/science.1260419. The foundational Human Protein Atlas publication that established the tissue-classification system this post’s opening statistics rely on. The proteinatlas.org portal is the companion hands-on reference.

• Tabula Sapiens Consortium. The Tabula Sapiens: a multiple-organ, single-cell transcriptomic atlas of humans. Science (2022). doi:10.1126/science.abl4896. The natural 2026 single-cell reference atlas for target-discovery work. The Human Cell Atlas (Regev et al., eLife 2017; search “Human Cell Atlas Regev eLife 2017”) is the parallel international consortium; both are actively expanding and coverage grows with each release. Cite one or both depending on how much real estate you want to give the single-cell reference-atlas story; Tabula Sapiens is the more concise recent citation.

Why drug target novelty matters in 2026

Target novelty matters because it sits alongside biomarker strategy as one of the few front-of-funnel choices that can materially change a program’s odds. Target selection itself is the subject of front-of-funnel decisions. Biomarker stratification, covered in likelihood of approval by therapeutic area, roughly doubles approval odds. The 2011–2020 BIO / Informa Pharma Intelligence / QLS Advisors Clinical Development Success Rates data put Phase I likelihood of approval for off-patent therapies at 14.7% versus 6.8% for novel therapies. Off-patent here means programs built on drugs whose core patents have expired. Most of that gap appears in Phase III, where non-originator products - follow-on or reused therapies rather than brand-new ones - transitioned at 70.3% and novel products at 52.9%. Vaccines ran 9.7%, biologics 9.1%, new molecular entities (entirely new drugs) 5.7%, and biosimilars (the follow-on versions of biologic drugs) 32.3%.

Concept Translation: The novel-vs-off-patent split is the drug-discovery version of an exploration-vs-exploitation trade-off. Off-patent / repurposing programs exploit known biology and known compounds at lower per-program cost and higher per-program success rate, but with crowded competition and limited upside. Novel-target programs explore - paying more for validation, accepting higher attrition, and reaching for the larger payoff that comes from being first to a target. Most portfolios mix both, and the mix is one of the more visible commercial decisions a research organization makes.

A second reason is structural. Over seventy-five percent of protein research still focuses on the ten percent of proteins known before the human genome was mapped. The skew reflects ease of work more than biological importance. Antibodies exist. Assays exist. Knockout mice exist. Crystal structures exist. Grant reviewers recognize the name. For the other ninety percent of proteins - the ignorome, the large pool of understudied proteins - little of that infrastructure is waiting for you. Recent ML gains in structure prediction, protein language models, and literature-scale NLP matter most here because they substitute computed evidence for missing infrastructure.

A third reason is NTRK. In 2018 the FDA approved larotrectinib (Vitrakvi, Bayer/Loxo Oncology) for any cancer driven by an NTRK gene fusion, a rearrangement that creates an abnormal TRK-family driver protein. It was the first tumor-agnostic targeted therapy - approval was tied to the molecular lesion rather than the tissue where the tumor started. Entrectinib followed in 2019 with broader spectrum activity; repotrectinib followed in 2023 with improved potency and activity against resistance mutations. The sequence is the commercial problem in miniature: a novel target, a first-in-class drug, and several approved drugs in the same class within five years. After the first-in-class drug lands, the target is no longer novel in the commercially useful sense. The window to own it is short.

Two kinds of drug target novelty

A useful starting point is Agarwal’s working definition, because it still shows up in platform documentation and industry decks. A novel target is a gene or protein with no approved-drug association. A proven target has at least one. In Agarwal’s analysis of roughly a thousand drug targets, 247 were proven and 712 novel; more than half of the novel targets had no competitor programs, whereas nearly ninety percent of proven targets had at least one. Competition on novel targets increased as programs advanced through clinical trials, likely because early positive signal drew followers. At portfolio level, novelty correlates with solitude, and solitude carries commercial value.

The definition is useful, but it collapses very different situations. A protein with hundreds of papers, a crystal structure, knockout-mouse phenotypes, three failed clinical programs, and no approval is “novel” by Agarwal’s definition. So is a protein with no papers, no structure, no knockout, no tool compound, and no drug-discovery history. Both are technically first-in-class opportunities. They do not require the same investment.

A more practical framework is the Target Development Level (TDL) classification from the NIH Illuminating the Druggable Genome (IDG) initiative, launched in 2014. The IDG Knowledge Management Center defined four categories.

• Tclin. Targets linked to at least one approved drug by mechanism of action - the drug’s therapeutic effect is understood to run through that target. (Agarwal’s “proven.”)
• Tchem. Proteins known to bind small molecules with high potency - compounds can hit them strongly enough to be useful probes or starting points - but without approved-drug links. Small-molecule tool compounds exist; approval does not. Still “novel” in Agarwal’s sense, but far from the ignorome.
• Tbio. Proteins with a confirmed Mendelian disease phenotype (where inherited variants in the gene cause a recognizable human disorder) or other biological evidence, but no demonstrated tractable chemistry - no convincing small-molecule starting point. “Novel” and biologically interesting; tools may be limited to genetic manipulation.
• Tdark. Proteins meeting none of the above. Minimal biological characterization, no tool compounds, often no crystal structure. The ignorome.

Concept Translation: TDL is closer to a label-quality gradient than a binary class. Tclin proteins have the equivalent of fully labeled, audited training data: drugs, structures, assays, papers, mouse knockouts. Tdark proteins have almost nothing labeled at all. ML methods trained on protein-level features inherit this skew - performance is highest where the data is densest, lowest in the long tail. The novelty-vs-confidence trade-off below is partly the question of whether your team can create the missing labels for a Tbio or Tdark program before the budget runs out.

Moving from Tdark to Tclin usually means a decade and several hundred million dollars of infrastructure investment per target. In practice, the “novelty vs confidence” decision is usually a Tchem-versus-Tbio decision. Few organizations fund Tdark programs without some other reason to believe, and Tclin is already proven.

One IDG finding matters here. Tdark proteins receive less NIH funding than other categories, and the pattern reinforces itself: less funding yields fewer papers, which lowers priority in the next grant cycle. That is the computational-tractability story. If the ignorome is under-studied because it is under-resourced, then any method that substitutes cheaply for experimental characterization - predicted structures, NLP-aggregated literature, network-inferred functional context - shifts the cost curve more for Tbio and Tdark than for Tclin.

The trade-off nobody gets to dodge

For any candidate target, novelty and confidence usually move in opposite directions. High-confidence targets are well studied, well characterized, and often already drugged. Validation costs are lower because many of the needed experiments have already been done somewhere. Competition is also higher: more organizations have programs, patents are crowded, and the commercial ceiling is divided. High-novelty targets reverse that profile. They demand more in-house validation and carry more biological risk, but they offer more commercial upside if the biology holds. In the Tchem-versus-Tbio decision, the key question is direct - how much of the target-validation burden does your team still have to carry itself?

Three patterns matter more in practice than they do on slides.

• The validation cost of a Tbio target is frequently underestimated. Teams look at the novel-target peak-sales story and forget that the tool-compound, pharmacokinetic/pharmacodynamic (PK/PD), and off-target profile work that comes free with a Tchem/Tclin target has to be rebuilt from scratch.

• The competitive cost of a Tclin target is frequently underestimated. Teams look at the proven-target LOA numbers and forget that the LOA is averaged across a crowd. The odds that your program, third-to-market, captures enough peak sales to justify R&D are not the class odds. That is why the BCR-ABL kinase-inhibitor space progressed from imatinib in 2001 through dasatinib, nilotinib, bosutinib, ponatinib, and asciminib, each positioned as an improvement on its predecessors and each competing for a smaller per-drug revenue share. Imatinib is a textbook first-in-class success story. It is also now generic, off-patent since 2015–2016, and no longer first-line in many cases.

• “Novel” and “understudied” are not the same axis. A protein can be proven in one disease (Tclin) and effectively novel in another. A kinase with approved inhibitors for oncology may have strong preclinical evidence in neurodegeneration and no programs targeting it there. The novelty framework needs to be applied at the (target × disease) pair level, not the target level alone. This is where the text-mining methods in the next section matter.

Quantifying target novelty

Novelty is a continuum, so the practical question is how to measure it. The standard approach uses NLP over text-based evidence - publications, grants, clinical-trial registrations - to produce single-number scores that track a target’s attention level in the scientific community.

The basic workflow is direct. Build a named-entity-recognition (NER) system that can identify gene, protein, disease, and drug names in biomedical text. Apply it to a corpus, typically PubMed, NIH RePORTER grant records, ClinicalTrials.gov trial registrations, and patent filings. For a candidate (target × disease) pair, count:

• Publication volume. How many papers mention this target in the context of this disease? Trend over time - is the area growing, flat, declining?
• Funding signals. How many grants are active on this target? Is NIH funding for this target growing relative to the field? Large, growing, well-funded areas are the opposite of ignorome.
• Clinical-trial involvement. How many trials reference this target directly or as a mechanism of action? In which phases? In which indications?
• Drug-interaction evidence. How many investigational or approved compounds are annotated as hitting this target? Any of them in the disease of interest?

Concept Translation: Novelty scoring is information retrieval over the scientific corpus, with the target as the query and “attention paid to it” as the relevance signal. Most practical implementations are a layered NER + relation-extraction pipeline followed by aggregation. From an LLM perspective, this is exactly the territory modern biomedical language models - BioBERT, PubMedBERT, BioGPT, and successors - are pretrained for. The scores that come out are corpus-attention summaries, not biological judgments. They tell you how much the field is paying attention to a target, which is informative but not equivalent to “is this target real?”

Two database patterns come up repeatedly for this kind of work. The TIN-X platform from the IDG consortium quantifies (target × disorder) pairs along two axes: a novelty index (a publication-scarcity score) and an importance index (which measures how strongly the literature links that target to a specific disorder). The scatter of targets in novelty-importance space gives you the portfolio view. Well-studied important targets are crowded, well-studied unimportant targets are dead ends, under-studied important targets are the opportunity, and under-studied unimportant targets are probably in the ignorome for good reason.

Concept Translation: TIN-X reduces every (target × disease) pair to a 2D point. The plot is a portfolio embedding - you can see at a glance which corner the candidate sits in, and where the unworked opportunities cluster. The same kind of low-dimensional projection that ML practitioners use to inspect representation spaces (UMAP, t-SNE on learned embeddings) maps onto a real product question here: where on the plot do programs that pay off live, and which corner does this candidate occupy?

The NER-plus-RE components underlying this work - named-entity recognition plus relation extraction - are useful but still imperfect. Biomedical named-entity recognition reaches about 90% F-score in benchmarks for well-structured entity types (gene, disease, chemical), where F-score is the standard balance of precision and recall. Relation extraction for claims such as “drug X binds target Y with Kd Z” sits closer to 50% F-score. The asymmetry matters for novelty scoring. Counting publications and extracting entities is tractable; extracting structured claims about target-disease mechanisms is where errors still accumulate.

LLM-based methods have moved the RE frontier since 2023, with transformer-based relation-extraction systems - typically fine-tuned BERT-family or domain-pretrained biomedical LLMs - now deployed in several commercial and open-source target-discovery pipelines.

Quantified novelty is decision support. A target that scores “low publication volume, high disease-association evidence in genetics” is genuinely interesting (the structured frameworks for evidence frameworks for target-disease linkage are the subject of the companion post). A human reviewer still has to read the papers to confirm that the NLP has not confused gene-symbol ambiguities, missed a key negative result, or counted reviews as primary papers. The scores compress the literature. They do not conclude.

Drug repurposing as the opposite axis

Novel-target discovery asks “can we find new biology to drug?” Drug repurposing asks “can we find new diseases for chemistry we’ve already validated?” The problems are asymmetric. Repurposing starts with an approved or well-characterized compound and searches across diseases; novel-target discovery starts with a disease and searches across targets. The intersection is direct: by the Agarwal definition, the target of a repurposed drug is already proven because the protein already has an approved-drug association. The novelty sits in the indication.

Repurposing has been sold as cheaper and faster for a decade. The claim is directionally right, but it leaves out structural costs. Roughly one-third of recent drug approvals have been repurposing-adjacent, and the Phase I LOA advantage over novel is meaningful, with the largest separation at the Phase III transition (70.3% for non-originator products vs 52.9% for novel). Those advantages do not erase the following costs.

• Bringing a repurposed compound to market still runs to hundreds of millions or low billions of dollars. Preclinical-research savings help, but they do not carry the later-phase load. Repurposed compounds still go through the same regulatory path as novel ones.
• Out-licensing an existing compound for a new indication carries “in-kind” costs - remanufacturing active product and placebo, compiling new study reports, and establishing pharmacovigilance (ongoing safety monitoring) for a new indication. Academic repurposing advocates often underestimate those costs.
• IP and regulatory exclusivity are weaker for new indications than for new molecules. The US 505(b)(2) pathway - a regulatory route that lets sponsors rely partly on prior evidence for an already approved drug - typically secures three to five years of market exclusivity, versus five years of new-chemical-entity exclusivity on the 505(b)(1) pathway used for novel molecules. For indication-only repurposing claims (where the same molecule is already approved for a different use and is now claimed for a new disease), exclusivity is usually at the lower end of that range. Skinny labelling - the practice of a generic manufacturer launching without the new indication on the label and thereby avoiding patents that cover only that indication - further weakens the sponsor’s position. The legal and commercial landscape around skinny labelling continues to evolve.
• Patents for repurposing-specific indications are harder to obtain and often expire in inconvenient windows. The common pattern is that the most valuable repurposing discoveries happen after the original patent expires, which is exactly when commercial protection is weakest.
• Pharmaceutical organizations often decline repurposing opportunities that don’t align with core therapeutic areas, and almost always decline opportunities in compounds they’ve previously discontinued. These are organizational constraints, not scientific ones, and they stay constant regardless of how good the computational repurposing signal is.

Concept Translation: 505(b)(2) is recognizable as fine-tuning rather than training from scratch. The full 505(b)(1) NDA path is the equivalent of training a new model end-to-end on a fresh problem - every safety dataset, every dose-finding study, every preclinical model has to be rebuilt. 505(b)(2) lets the sponsor reuse the prior approval as a frozen backbone and add a smaller “scientific bridge” study, typically a pharmacokinetic comparison, to fit the new claim. The cost reduction is roughly the gap between training a 70B model from scratch and fine-tuning one for a new task: large, but not as large as the headline numbers suggest, because the downstream regulatory infrastructure still has to be built.

Seen in concrete terms, “cheaper and faster” buys a specific regulatory shortcut. The 505(b)(2) pathway, established by the Hatch-Waxman Amendments of 1984, allows the FDA to rely on existing safety and efficacy data - including published literature and prior FDA findings on a Reference Listed Drug (the previously approved product that anchors the comparison) - so long as a scientific bridge such as a pharmacokinetic study connects the new use to the established record. Development through approval via 505(b)(2) typically runs three to eight years and roughly $8 million to $200 million. A full 505(b)(1) NDA - a new drug application for a fully new product - typically runs ten to fifteen years and $1 billion to $2.2 billion. That is a 70–90% reduction in out-of-pocket capital expenditure, with three to five years of market exclusivity at the end. 505(b)(2) approvals routinely exceed forty per year and peaked at sixty-eight in 2020, often outnumbering new molecular entities approved via 505(b)(1). One caveat remains: FDA review time for a 505(b)(2) submission can exceed that of an NME because evaluating the scientific bridge between old and new data creates its own regulatory burden. The development calendar often shrinks. The regulatory calendar does not always follow.

These frictions explain why the biggest repurposing success stories tend to be atypical. Massive unmet need creates market-making urgency (AZT in early HIV, remdesivir in COVID-19); a side effect eclipses the original indication (Viagra); or a mechanistic insight opens a distinct indication for an off-patent drug (thalidomide in erythema nodosum leprosum, then multiple myeloma). The ordinary case - a statistically significant signal in a computational repurposing screen without unusual market dynamics - often fails the business-case analysis even when the science is sound.

Methods for computational drug repurposing

If the goal is to predict new disease-drug pairings systematically, current pipelines usually fall into three methodological families.

Structure- and similarity-based methods. Molecular docking against alternative targets, pharmacophore matching (which compares the 3D binding features a molecule presents rather than its exact structure), and fingerprint-based similarity to drugs with known activity in a target disease. The classical ML framing here is nearest-neighbor retrieval in chemical or target space - an approved drug whose structure or induced transcriptomic profile resembles that of a known therapeutic for disease X is a repurposing candidate for X.

Genetic-association and omics-based methods. Compounds whose target list overlaps with disease-associated loci (from GWAS, rare-variant analysis, or differential-expression studies) are candidates for repurposing into that disease. Connectivity-map methods compare the gene-expression changes caused by a drug with the changes seen in disease and look for reversals - the canonical ML instantiation of this and the single most cited ML method in drug repurposing.

Concept Translation: The Connectivity Map / L1000 framing is signature matching in transcriptomic space. Every drug perturbation produces a characteristic gene-expression vector - a “signature” - and every disease state produces another. A drug whose signature is anti-correlated with the disease signature is a candidate for treating that disease, on the heuristic that reversing the disease’s transcriptional state may reverse the disease itself. Mechanically, this is nearest-neighbor in a high-dimensional vector space with a sign flip, exactly the kind of retrieval problem that contrastive learning is built for.

Network-based methods. Biomedical knowledge graphs encoding drug-target, target-pathway, target-disease, and pathway-disease relationships allow repurposing to be formulated as a link-prediction or proximity problem. The model either predicts useful missing links in the graph or asks which drugs sit closest to the disease module in network space. Drugs “close” to a disease in that sense - whether close is measured by random walks, network diffusion, or learned embeddings - are candidates.

All three families lean on unsupervised and semi-supervised ML because labeled examples are scarce. You have lots of unlabeled structural data, lots of unlabeled transcriptomic profiles, and a few thousand confirmed drug-disease pairs. The standard move is to learn representations in which similar things cluster, then use sparse labels to rank. This is the frame Build AI Drug Discovery Pipelines Chapter 7 takes up in detail, with drug repurposing as the running example.

Databases that wire drugs and targets together

Three databases recur in the drug-target querying that any novelty or repurposing workflow needs.

DrugBank. Searchable by drug or target, returning interactions, chemical properties, mechanism of action, related diseases, clinical-trial status, and brand names. Basic search is free; advanced query and bulk download require a license, with free academic licenses available. DrugBank’s strength is the curation quality on approved drugs and late-stage candidates.

Pharos (the public portal for IDG data, funded by the NIH Common Fund). Searchable by target, drug, or disease, with target pages that include the TDL classification, tissue expression across healthy human tissues, protein structure when available, associated drugs, and pathway memberships. Pharos is the canonical source for the Tclin/Tchem/Tbio/Tdark categorization and is freely available. For any novelty-scoring workflow that wants to anchor “novel” to something other than ad-hoc thresholds, the TDL assignment from Pharos is the standard reference.

DGIdb (Drug-Gene Interaction Database). Organizes the druggable genome into two classes: genes with known drug interactions, and genes that are potentially druggable. The second class is particularly useful for prioritizing gene lists from omics studies by which ones have a plausible path to small-molecule tractability. Publicly available, free to download.

For a practitioner-built novelty scoring pipeline, the useful pattern is to join across all three: DGIdb for the druggable-genome filter, Pharos for the TDL assignment, and DrugBank for the approved/investigational drug inventory. Layer PubMed, NIH RePORTER, and ClinicalTrials.gov on top for the text-signal component. These are mature components, and most real pipelines depend on them.

A worked example in network-based drug repurposing for SARS-CoV-2

An early-pandemic SARS-CoV-2 study gives a concrete view of computational repurposing at scale. Researchers combined three data sources - the human protein-protein interaction network (the “interactome”), experimentally characterized SARS-CoV-2 viral target interactions, and drug-target interaction data - and used network-diffusion and network-proximity algorithms (which score how near a drug’s targets sit to disease-relevant proteins in the network) to rank 6,340 approved and investigational drugs by expected efficacy against SARS-CoV-2.

The ranking was then tested experimentally in human cells. The top-ranked drugs showed a 62% success rate in reducing viral infection. Of the six drugs that were effective, four could be directly repurposed for COVID-19 treatment. The more striking result was that seventy-six of the seventy-seven drugs that reduced viral infection did not bind to proteins targeted by SARS-CoV-2 (primary paper: Morselli Gysi et al., PNAS 118(19):e2025581118, 2021). The efficacy came from indirect, network-based mechanisms - perturbations to host pathways that happen to be close in network space to viral-target pathways, rather than direct antiviral activity in the classical sense.

Concept Translation: Network proximity is graph-based propagation rather than 1:1 binding. A drug “near” a disease in network space is one whose protein targets share many shortest paths with the disease’s known protein targets. The propagation is computed as a random walk, a Laplacian smoothing, or a learned graph-embedding distance - different algorithms, same intuition. The mechanism of action can be entirely indirect: the drug perturbs a host protein that is in the right neighborhood of the viral-target proteins, and the perturbation ripples through the network. A classical docking screen would never find these candidates because they don’t bind the obvious targets.

Two points matter.

First, the hit rate is high relative to a naïve virtual screen (a broad in silico ranking of compounds with little biological context). The network-proximity frame encodes a prior - drugs already approved for other indications have cleared a basic tractability bar. Restricting the search space to those compounds and ranking by network relevance to the target biology concentrates candidates on the right side of that bar. That is why repurposing can work computationally even when it fails commercially.

Second, the network result distinguishes repurposing as target identification from repurposing as drug matching. A classical docking-based screen looks for drugs that bind viral proteins. A network-diffusion screen finds drugs that perturb host pathways whose topology matters for viral infection - a different mechanism, and one that a target-centric analysis would easily miss. For an ML practitioner entering drug discovery, the lesson is direct: the representation choice - chemical similarity, network proximity, or transcriptomic signature - encodes a hypothesis about mechanism. Repurposing screens are strongest when they triangulate across representations.

The NTRK case when novelty works

Imatinib succeeded as the first BCR-ABL inhibitor because CML (chronic myeloid leukemia) was a rapidly fatal disease with few good alternatives; the rational kinase-inhibitor paradigm it validated then took two decades to spread across other kinases. The more recent parallel is NTRK.

The target-discovery rationale for NTRK was unusually clean. In a subset of cancers across multiple tissues, a gene fusion produces a TRK-family protein that is constitutively active and functions as a driver. Patients identifiable by a single molecular criterion - NTRK fusion detectable by IHC (immunohistochemistry) or NGS (next-generation sequencing) - share a common therapeutic vulnerability across many tumor histologies. Larotrectinib’s 2018 approval marked the first tumor-agnostic targeted therapy. Entrectinib followed in 2019 with a broader spectrum (ROS1 fusions as well as NTRK); repotrectinib followed in 2023 with improved potency and resistance-mutation coverage (activity against variants that emerge after treatment with earlier drugs). Within five years of first-in-class approval, the class had three approved drugs.

The NTRK arc is a case study in novelty half-life. Larotrectinib’s team owned a genuinely novel target-disease linkage for a narrow window. Once the tumor-agnostic framing proved out regulatorily, the competitive landscape tightened quickly. Entrectinib had been in clinical development before larotrectinib’s approval, and repotrectinib was engineered specifically to address resistance mutations that emerge on the earlier agents. By 2026, NTRK-fusion oncology is no longer a first-in-class opportunity for new entrants; the novelty lives in resistance mechanisms, new fusion partners, and non-oncology TRK indications that remain lightly explored.

Cross-cancer analysis is the computational work that keeps tumor-agnostic programs coherent. It assembles expression and genomic data across TCGA (The Cancer Genome Atlas) and similar pan-cancer compendia to identify which molecular lesions recur across tissues and which do not. NTRK would not be the NTRK story without a decade of pan-cancer sequencing work laying that groundwork. The generalizable point for target-novelty reasoning is that the lesion-centric view, organized around the molecular abnormality itself, often outruns the tissue-centric view in finding novel-but-tractable therapeutic categories. Pan-cancer analysis is the computational engine for finding the next NTRK.

What’s next in drug target novelty

The Tdark problem remains open. Every infrastructure advantage that makes Tclin targets easy to work on - crystal structures, tool compounds, knockout models, antibodies - was built over decades at enormous cost. Protein-language-model embeddings, AlphaFold-class structure prediction, and NLP over sparse literature are the field’s cheaper substitutes. The operational question for the next few years is whether those substitutes are good enough to de-risk Tdark programs until they compete with Tchem programs on expected value.

Repurposing is stronger as a computational problem than as a commercial one. Every methodological advance in computational repurposing produces more candidates than the industry funds. The bottleneck is structural - patents, exclusivity, skinny labelling, and organizational fit. A less explored research direction is program design that works with that commercial reality. Rare-disease orphan-drug programs, neglected-tropical-disease partnerships, and academic-industry hybrid models are common proposals; none has yet produced the volume of repurposing approvals that the computational hit rates would suggest.

LLM-based literature intelligence is changing novelty scoring faster than the surrounding tooling. NER and relation extraction on biomedical text are substantially more accurate now than they were when the IDG TDL system was first operationalized. The open question in 2026 is whether that translates into better target nominations or only larger target-evidence databases. The risk is familiar - higher-recall text extraction can produce confident-looking summaries of uncertain or contradictory evidence. Explicitly flagging what the ML does not know will matter more than pushing raw scores higher.

Novelty scoring shares the validation problem of the rest of target-discovery ML. The ground truth for “was this a good novel-target nomination?” arrives as Phase III readout a decade later. Retrospective benchmarks are biased toward targets that entered pipelines, and failed targets get less documentation than successful ones. Prospective benchmarks are limited by timescale. Until target-discovery ML has something like the PDBbind / CASP benchmark - a widely accepted held-out benchmark set or community challenge for the task that actually matters - the discipline will remain calibrated as much by industry credibility as by methodology papers. That is a property of the field, not of any specific method.

The common commitment underneath these methods is compression. Every credible target-novelty score is a summary of what the literature collectively knows, and every credible repurposing candidate is a summary of what chemistry has already paid to validate. The ML’s job is to make those summaries trustworthy enough for a program committee to act on them. That is a narrower claim than the usual “AI-discovered novel target” narrative. It is also closer to what target-discovery ML actually does well.

Further reading

• Agarwal, P. (circa 2013–2016). Novel vs proven drug targets: portfolio-level analysis of ~1,000 targets. doi:10.1038/nrd4089 The foundational portfolio-level analysis that defines “novel” as a gene/protein with no approved-drug association and quantifies the novelty-vs-competition relationship across ~959 analyzed targets. Source for the 247 proven / 712 novel split, the >50% competitor-free rate among novel targets, and the >90% competitor-present rate among proven targets.

• NIH Illuminating the Druggable Genome (IDG) initiative. Pharos: the IDG Knowledge Management Center portal. [URL: https://pharos.nih.gov/]. The canonical public portal for Target Development Level (TDL) classifications (Tclin, Tchem, Tbio, Tdark), target-level aggregation of biological and pharmacological data, and the IDG’s understudy-bias framing. For the canonical academic references, see Nguyen et al., Nucleic Acids Research database issue (multiple years - search “Pharos IDG Nucleic Acids Research database issue” for the most recent), and the 2014 IDG launch paper (search “Illuminating the Druggable Genome 2014 launch Nature Reviews Drug Discovery OR Nature Chemical Biology”).

• Morselli Gysi, D., do Valle, Í., Zitnik, M., Ameli, A., Gan, X., Varol, O., Ghiassian, S. D., Patten, J. J., Davey, R. A., Loscalzo, J., Barabási, A.-L. (2021). Network medicine framework for identifying drug-repurposing opportunities for COVID-19. Proceedings of the National Academy of Sciences 118(19):e2025581118. DOI: 10.1073/pnas.2025581118. The primary source for the SARS-CoV-2 network-medicine worked example: ranked 6,340 approved and investigational drugs by network-diffusion / network-proximity scores, experimentally validated 918 in VeroE6 cells, and produced the 62% top-rank hit rate and the 76-of-77 indirect-mechanism finding. The methods section also documents the three-pipeline framework (AI, network diffusion, network proximity) the draft summarizes.

• Pushpakom, S., Iorio, F., Eyers, P. A., Escott, K. J., Hopper, S., Wells, A., Doig, A., Guilliams, T., Latimer, J., McNamee, C., Norris, A., Sanseau, P., Cavalla, D., Pirmohamed, M. (2019). Drug repurposing: progress, challenges and recommendations. Nature Reviews Drug Discovery 18:41–58. [URL: https://doi.org/10.1038/nrd.2018.168]. The most-cited general-audience review of drug-repurposing methodology. Source for the three-family framing of computational repurposing methods (structure/similarity, genetic/omics, network) that the draft’s Methods section extends, and for the commercial-versus-computational tension the post’s “Drug repurposing as the opposite axis” section takes up. Important pairing for the “one-third of recent approvals are repurposing-adjacent” claim - the specific fraction depends heavily on counting methodology; Pushpakom gives the operational definition and supporting counts for that framing.

Concept Translation: “Undruggable” is an artifact of the tool - it means the current inhibitor toolkit (small molecules that fit deep pockets) can’t act on this protein. It does not mean the protein is intrinsically resistant to intervention. The ML analogy: a feature you cannot perturb with your current intervention API. Once the API expands (PROTACs, ASOs, ADCs), the same protein moves from undruggable to druggable without any biological change.

Why druggability assessment matters in 2026

Two shifts have moved the frontier of druggability at roughly the same time, and together they change the analysis.

The first is structural. AlphaFold 2 turned single-chain protein structure prediction from a research problem into routine infrastructure in 2020. AlphaFold 3 in May 2024 extended that capability to biomolecular complexes - proteins with small-molecule ligands, with DNA and RNA, with ions, and with post-translational modifications, all in a single diffusion-based model (Abramson et al., Nature 630:493–500, 2024). The AlphaFold 2 work earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry, with David Baker taking the other half for computational protein design. AF3 now sits inside a broader field that includes Boltz-1 and Boltz-2 (MIT Jameel Clinic), Chai-1 and Chai-2 (Chai Discovery), Protenix (ByteDance), HelixFold-3 (Baidu), and OpenFold3. Predicted structures are now a default input to virtual screening - the computational ranking of large compound libraries against a target - rather than a special case.

The second shift is modality. Antisense oligonucleotides (ASOs), small interfering RNAs (siRNAs), targeted protein degraders (PROTACs and molecular glues), antibody-drug conjugates (ADCs), and cyclic peptides all have approved products in the market or late-stage assets in development. Each of these expands the “druggable” envelope along a different axis. An RNA-binding ASO can modulate a protein that has no structured binding pocket at all. A PROTAC can take down a scaffolding protein - one that organizes other proteins rather than catalyzing a reaction - even when there is no catalytic activity to inhibit. An ADC can deliver a cytotoxic payload selectively to tumor cells whose surface antigen (a protein displayed on the cell surface) is otherwise non-actionable.

A 2015-era druggability analysis might ask does this target have a small-molecule binding pocket, and is its structure known? A 2026 analysis has to ask that question across at least five modality classes. It also has to answer the harder follow-up: if several modalities are tractable, which one fits this target’s biology, safety profile, and commercial context best?

What “druggable” actually means

The vocabulary needs care because the field uses these terms loosely.

Druggability refers to the ability of a protein, peptide, or nucleic acid to be modulated by a drug. The 2002 analysis that anchored the field estimated that about 3,000 of the ~20,000 human protein-coding genes - roughly 15% - are druggable in the small-molecule sense, and that only around 27% of those druggable genes actually have an approved drug. Small-molecule drugs target roughly 2–5% of the human genome in total. That figure marks the reach of current chemistry.

Druggability decomposes into two more specific questions.

Target quality asks whether modulating the target will actually change the disease, and whether doing so is safe. The relevant variables are the target’s role in essential cellular processes, its network centrality (how many pathways and interactions depend on it), its tissue-expression pattern (a target essential for heart-muscle function is a bad target, almost regardless of its affinity profile), its approval-adjacent precedent (whether closely related targets or modalities have already cleared regulatory review), and its cellular localization (an intracellular target requires permeability that a cell-surface target does not). Target quality is what target–disease evidence gives you the machinery to evaluate.

Ligandability asks whether a drug molecule can bind the target in a way that modulates it. Ligandability depends on three things: the availability of the target’s structure (experimental or predicted), the presence of closely related proteins with known ligands (if a close cousin of your target has been drugged, the odds are better), and the accessibility of a binding site (a small-molecule pocket for small molecules, a regulatory region for nucleic-acid-targeted drugs, or a surface-exposed patch for antibodies).

Concept Translation: Target quality and ligandability are independent axes. A target can be highly ligandable - it has a deep, well-defined pocket, a structure, and a binder - while being a terrible drug target if hitting it doesn’t change disease or causes serious toxicity. Many of the 80–90% of proteins called “undruggable” in older analyses are really not yet ligandable by small molecules; their target quality may be excellent. New modalities are mostly ligandability-extenders, not target-quality-extenders.

Classical small-molecule druggability tools such as DoGSiteScorer (Volkamer et al., 2012), fpocket (Le Guilloux et al., 2009), and SiteMap (Halgren, 2009) turn that third question into a quantitative pocket score. They compute geometric and chemical descriptors of candidate cavities. These remain widely cited workhorses. Newer ML-based approaches (PockDrug, PocketMiner) have emerged and are worth tracking, but the original three are still common reference points in 2026 ligandability analyses. Structure prediction enters the picture at both stages of ligandability: it provides a structure for targets that have never been crystallized, and it increasingly provides one with a candidate ligand already docked - computationally placed into a proposed binding pose.

The classical druggability landscape by protein family

Before structure prediction started moving the boundary, the druggable genome was shaped largely by which protein families had accessible binding sites and tractable chemistry. A handful of families dominate approved drugs.

G protein-coupled receptors (GPCRs) are the single most targeted class. They sit in the cell membrane and relay outside signals inward. Their extracellular ligand-binding domains can be reached by small molecules, peptide-like molecules, antibodies, and allosteric modulators - drugs that bind away from the main ligand site and tune receptor activity.

Ion channels provide a good contrast. The sequenced human genome identified more than 400 putative ion channels, though only a fraction have been cloned and functionally characterized. Their widespread tissue distribution and the physiological consequences of opening or closing them make them both compelling and treacherous targets.

Kinases are the third-most-targeted group. Kinases transfer phosphate groups from ATP to substrate proteins, and the ATP-binding pocket is the canonical small-molecule site. Five inhibitor strategies recur across the literature: ATP mimics (the default), covalent inhibitors (which form a bond with a cysteine in or near the pocket and are irreversible), bivalent inhibitors (binding the catalytic site and a second stability or localization site), allosteric inhibitors that bind distal regulatory sites (type III and IV), and degraders or molecular glues that hijack the ubiquitin–proteasome system to remove the kinase rather than inhibit it.

Concept Translation: Kinases are useful to keep in mind because they’re the canonical “well-modeled” target family - thousands of crystal structures, deep mutagenesis data, and a consistent pocket geometry across the family. They’re roughly the ImageNet of structure-based drug design: the family where everything was tuned first and where most ML benchmarks were established. The frontier targets are unlike kinases - flat surfaces, no obvious pocket, sparse training data - and that’s why the modality expansion matters.

Kinase degraders blur the classical taxonomy. A “kinase inhibitor” and a “kinase degrader” act on the same target through different mechanisms, so the druggability criteria differ as well. That overlap sets up the rest of this post.

Structure prediction and the 2020/2024 inflections

For most of drug discovery’s history, structure-based drug design was available only for targets that had been successfully crystallized. The c-Met/crizotinib program is a useful illustration of that earlier workflow. c-Met is a receptor tyrosine kinase with elevated levels and abnormal activity in several cancers. Pfizer’s program started from a series of pyrrole-substituted 2-indolinones (oxindoles), ATP-competitive tyrosine kinase inhibitors first explored in the mid-1990s, that showed activity against c-Met but poor drug-like properties. Iterative co-crystallography with PHA-665752 (an early lead) revealed the c-Met ATP-pocket environment in detail; that structure motivated a switch to a novel 2-amino-5-aryl-3-benzyloxypyridine series, which yielded the 3,5-disubstituted 2-aminopyridine clinical candidate crizotinib (PF-02341066, Cui et al., J. Med. Chem. 54:6342–6363, 2011). Each iteration of the design cycle was gated on a new crystal. Crizotinib, a c-Met/ALK dual inhibitor, was approved by the FDA in 2011.

That workflow still works, and for targets that crystallize well it remains the standard. Crystallography, however, is slow and expensive. For many targets - especially membrane proteins, intrinsically disordered regions, and transient complexes - it either fails or produces low-resolution artefacts that are hard to use for design.

Two inflection points changed the practical calculus.

2020: AlphaFold 2. At CASP14, the community benchmark competition for protein-structure prediction, DeepMind’s AlphaFold 2 produced predictions indistinguishable from experimental structures for a majority of single-chain proteins. By 2022 the AlphaFold Database had released ~800,000 predicted structures. By 2025 it was around 200 million predicted structures, covering the bulk of UniProt, the standard protein-sequence database. Single-chain structure prediction stopped being a bottleneck.

2024: AlphaFold 3. AF3 (Abramson et al., Nature 630:493–500, 2024) moved from single-chain prediction to complex prediction using a diffusion-based architecture built around a Pairformer block - which updates pairwise residue relationships - plus a diffusion module. It jointly predicts proteins with small-molecule ligands, DNA, RNA, ions, and post-translational modifications in a single forward pass. On the PoseBusters ligand-pose benchmark, a test of whether models place ligands in the right binding pose, DeepMind reports AF3 as roughly 50% more accurate than the best traditional physics-based docking tools. That pushes protein–ligand pose-prediction success rates to about 76% versus roughly 50% for the docking programs AutoDock Vina and Gold on the same benchmark, and makes AF3 the first AI system to outperform physics-based tools for biomolecular structure prediction. Model weights were released for non-commercial academic use in November 2024, with broader (still non-commercial) availability in February 2025. Commercial use remains gated behind Isomorphic Labs partnerships.

Concept Translation: The AF2-to-AF3 jump is conceptually familiar to anyone who has watched generative-model architectures evolve. AF2 was a transformer-style model that emitted a single deterministic structure. AF3 swaps the structure-emission head for a diffusion module - the same general framework as image diffusion models - that samples from a distribution over plausible structures. The Pairformer is a transformer-style block that operates on pairwise residue relations rather than tokens. The architectural pattern (encoder over pairwise relations + diffusion decoder) is now the template for every AF3-class model in the field.

The competitive field matters because AF3’s non-commercial licensing makes commercially usable alternatives important for industrial drug discovery:

• Boltz-1 (MIT, Nov 2024): MIT license, fully open source, matches AF3 accuracy on standard benchmarks. Boltz-1x added physics-based steering to reduce steric clashes; Boltz-2 (2025) added binding-affinity prediction, claiming roughly 2× the precision of standard ML and docking baselines on hit discovery.
• Chai-1 and Chai-2 (Chai Discovery): Apache 2.0 license, inference-only weights, can run with or without a multiple-sequence alignment. Chai-2 (2025) introduced zero-shot de novo antibody design with reported experimental hit rates around 16–20% on novel targets.
• Protenix (ByteDance, 2025), HelixFold-3 (Baidu), and OpenFold3 (in preview as of late 2025) round out the field.

By 2026, teams treat predicted structures as routine inputs to virtual screening and structure-based design. AF3-class models have substantially reduced the need for experimental crystallographic structures in early-discovery programs, particularly for targets where co-crystal structures are slow or expensive to generate. Virtual screening campaigns at 10K-compound scale now run in days on cloud GPUs.

A worked example: AlphaFold-enabled design for an undrugged target

A 2022 hepatocellular carcinoma (HCC) program provides the first published demonstration of using AlphaFold-predicted structures for hit identification in practice (Ren, Ding, Zheng et al., Chemical Science 14:1443–1452, 2023). The target, CDK20, also known as cell cycle-related kinase (CCRK), had no experimental structure and was effectively undrugged. Traditional structure-based design on CDK20 would have required first solving its crystal structure, a multi-year project in itself.

Instead, the team used the AlphaFold-predicted structure (AF-Q8IZL9-F1-model_v1) of CDK20 directly. The prediction was high-confidence across most of the protein except the C-terminus, which they removed because it occluded the solvent-exposed region of the ATP pocket. Pocket analysis identified a shallow ATP binding pocket of approximately 150 ų in the DFG-in conformation - one common active-state arrangement of kinase active-site residues - with MET84 as the hinge residue and PHE81 occupying the gatekeeper position (the hinge is where many kinase inhibitors make key hydrogen bonds; the gatekeeper residue helps control access to the back pocket). They fed the AlphaFold structure into Insilico Medicine’s generative chemistry platform (Chemistry42), which generated 8,918 candidate molecules. After docking, clustering, and pose inspection, seven were synthesized for the first round of testing. The first hit (ISM042-2-001) bound CDK20 with a Kd of 9.2 ± 0.5 µM (lower Kd means tighter binding) and was identified 30 days from target selection. A second optimization round produced ISM042-2-048, with Kd ≈ 567 nM and a CDK20 kinase-inhibition IC50 of 33.4 ± 22.6 nM - the concentration required to cut enzyme activity by half. ISM042-2-048 also showed selective antiproliferation against the CDK20-overexpressing HCC line Huh7 (cellular IC50 ≈ 209 nM) versus the HEK293 counter-screen (~1707 nM), a comparison cell line used to check selectivity.

AlphaFold converted a target that would previously have been deferred until crystallography succeeded into a target that could be designed against immediately. In programs that use this workflow, the structural-modeling portion of the pipeline has moved from years to weeks. Hit-to-lead optimization, preclinical pharmacology, and everything downstream still take the time they always did.

Compared with the crizotinib workflow described above, the difference is sharp. Crizotinib needed a crystal of each intermediate compound; CDK20’s lead compound was designed against an AI-predicted structure before any experimental structure existed. Structure-based drug design used to be gated on crystallography. In 2026 it is gated on whether the structure-prediction model is accurate enough at the specific pocket you care about. That is a softer gate, and the tooling keeps improving.

Beyond small molecules: modality choice is now a first-class question

Categories of therapeutic molecule that did not have approved products a decade ago now have market-stage drugs and late-stage candidates, each reaching targets that small molecules cannot and expanding the druggable envelope.

Targeted protein degradation (PROTACs and molecular glues)

Targeted protein degradation (TPD) shifts pharmacology from occupancy-driven - where you block a protein’s function by occupying its binding site - to event-driven, where you induce the cell’s endogenous waste-disposal machinery to destroy the protein outright. The relevant machinery is the ubiquitin–proteasome system (UPS), the main pathway cells use to tag and remove unwanted intracellular proteins. A cascade of three enzymes - E1 (activating), E2 (conjugating), and E3 (ligase) - tags target proteins with polyubiquitin chains and marks them for destruction by the 26S proteasome. TPD drugs hijack this cascade.

Concept Translation: The occupancy-vs-event shift is closer to a control-flow change than a chemistry change. An occupancy-driven inhibitor is a function call that holds an object in a blocked state for as long as the call is on the stack - release the call, the object returns to normal. An event-driven degrader is a single instruction that triggers garbage collection on the object. The first scales with drug concentration; the second scales with the rate at which the cell remakes the protein. That difference is why PROTACs can be effective at much lower concentrations than equivalent inhibitors and why their pharmacology is often dominated by protein resynthesis rates rather than drug residence time.

A PROTAC (PROteolysis TArgeting Chimera) is a heterobifunctional molecule - it has one binding end for the target and another for an E3 ligase, joined by a linker. By forming a ternary complex (a three-part assembly of target, PROTAC, and ligase), it brings the target into proximity with the ubiquitination machinery. Because of this tripartite structure, PROTACs are large molecules, typically 700–1,000 Da. A PROTAC also acts catalytically. Once the target is ubiquitinated and released into the proteasome, the PROTAC dissociates and recruits another target copy, so one degrader molecule can remove many target molecules. It can therefore hit targets that have no functional small-molecule binding site at all, as long as some binder exists for the target.

Molecular glues are the close cousin: small molecules typically <500 Da, with no linker, that bind directly to an E3 ligase and induce a conformational change in the ligase surface, creating a novel interaction interface that recruits “neosubstrate” proteins (proteins the ligase would not normally bind) for degradation. The original IMiDs (thalidomide, lenalidomide, pomalidomide) are retrospectively understood as molecular glues. In 2026 the line between PROTACs and molecular glues is blurring; newer-generation degraders combine features of both.

The opportunity space is large. The human genome encodes over 600 E3 ligases, of which fewer than 2% have been successfully exploited for TPD to date. Most current PROTACs recruit just a few E3 ligases, especially CRBN or VHL; DCAF15 and MDM2 are emerging; the rest of the ligase repertoire is largely untouched. Tumor-specific E3 ligases are an active frontier for tissue-selective degradation. The motivation is direct: a CRBN-recruiting PROTAC can degrade its target in healthy tissue as well as diseased tissue.

The state of TPD in 2026 is a useful test case for “modality choice as a first-class question”:

• Bavdegalutamide (ARV-110), Arvinas’s androgen-receptor PROTAC, was the first PROTAC to enter human clinical trials (Phase 1, 2019). In patients with AR T878X/H875Y mutations in ARDENT, it produced 46% PSA50 response rates, where PSA50 is a prostate-cancer endpoint based on prostate-specific antigen decline. Development was discontinued in 2024 in favor of ARV-766, a next-generation AR degrader with broader mutant coverage. The first clinical entrant did not become the lasting lead program.
• Vepdegestrant (ARV-471), Arvinas/Pfizer’s estrogen-receptor PROTAC, reached Phase 3 (VERITAC-2) in estrogen-receptor-positive, HER2-negative breast cancer after prior CDK4/6 inhibitor failure. Topline results in March 2025 met the primary endpoint in the ESR1-mutant subpopulation, the subgroup whose tumors carried ESR1 mutations (median PFS 5.0 vs 2.1 months, HR 0.57, p<0.001). PFS is progression-free survival, and HR is the hazard ratio comparing event rates over time. The study did not reach statistical significance in the intent-to-treat population, meaning all enrolled patients. NDA, the New Drug Application, filed June 2025. PDUFA action date: June 5, 2026. If approved, vepdegestrant will be the first FDA-approved PROTAC.
• In September 2025, Arvinas and Pfizer publicly sought a third-party partner to commercialize vepdegestrant. That move suggests cooling internal commitment despite the positive result in the ESR1-mutant subgroup. The modality has clinical signal, but the commercial case remains unsettled.

The broader TPD space has expanded well beyond ER and AR. CC-94676 (BMS, AR PROTAC), the CELMoDs mezigdomide and iberdomide - cereblon E3-ligase modulators now in Phase 3 in multiple myeloma - and a growing number of Phase 1 molecular-glue degraders for BCL6, BTK, and other targets are all active programs.

For a target-assessment document, this changes the screen. For a scaffolding protein or transcription factor without an inhibitable catalytic site, the relevant question becomes “does it have a surface patch we can bind at all?” If yes, a PROTAC may be tractable.

RNA-targeting drugs (ASOs and siRNAs)

Antisense oligonucleotides (ASOs) are 15–30 nucleotide sequences that bind target RNA via Watson–Crick base pairing - that is, ordinary sequence complementarity. Two mechanisms dominate:

• RNase H1 cleavage. The ASO–RNA duplex is a substrate for endogenous RNase H1, an enzyme that cleaves the RNA strand. Most approved ASOs work this way.
• Steric blockage / splice switching. The ASO binds pre-mRNA (the RNA transcript before splicing) and prevents ribosomal assembly or alters splicing. Exon-skipping ASOs (Duchenne muscular dystrophy) and exon-inclusion ASOs (spinal muscular atrophy) are the canonical examples.

siRNAs act through the RNA-induced silencing complex (RISC), a cellular complex that uses the guide RNA to find and silence matching transcripts. Approved examples include patisiran (hereditary transthyretin amyloidosis), givosiran (acute hepatic porphyria), lumasiran (primary hyperoxaluria type 1), and inclisiran/Leqvio - Novartis’s siRNA targeting PCSK9, originally approved by the FDA in December 2021 for adults with ASCVD (atherosclerotic cardiovascular disease) or heterozygous familial hypercholesterolemia, expanded in 2023 to broader primary hyperlipidemia, and updated again in July 2025 to permit first-line monotherapy use in hypercholesterolemia.

Concept Translation: ASOs and siRNAs target messenger RNA rather than the protein itself. The matching is sequence-based - Watson–Crick base pairing is exact-match string lookup over a 4-letter alphabet (A, U, G, C). For target assessment this is a different design space entirely: the target is a string, the drug is a string, and the binding rule is a known function of the two strings. Compare this to small-molecule binding, where the binding rule is a learned function of two 3D shapes. Sequence-matching is much easier to design but introduces a new failure mode - off-target hits anywhere in the transcriptome that match closely enough.

For target assessment, ASOs and siRNAs require an accessible sequence rather than a structured protein binding pocket. A target with no deep pocket, no covalent warhead opportunity, and no PROTAC handle can still be tractable if its RNA is accessible. That rule has caveats: small-molecule splice modulators like risdiplam for SMA exploit RNA secondary structure - the local folded shape of the RNA strand - so structure does come back into play at the RNA level for some programs.

Antibody-drug conjugates (ADCs)

An ADC is a monoclonal antibody covalently linked to a cytotoxic payload via a chemical linker. The antibody provides selectivity; the payload provides potency. An ideal ADC stays stable in circulation, internalizes on binding to the target (the antibody–target complex is pulled into the cell), and releases the payload intracellularly in the vicinity of the target.

Concept Translation: An ADC is a delivery system. The antibody is a routing label that says “deliver to cells expressing this surface marker”; the payload is the actual cytotoxic drug, often too toxic to administer freely. The linker controls when the payload separates - ideally only after the package has been internalized. The pattern is roughly the same as targeted delivery in any system: route on a label, release on arrival. The engineering challenges are also analogous (label specificity, in-transit stability, controlled release).

2026 status: 15 ADCs are FDA-approved, with 19 ADCs approved globally across FDA, EMA, NMPA, and PMDA. Pipeline scale is over 400 ADCs in development, 200+ in clinical trials, and at least 24 in Phase 3. Recent FDA approvals include Datroway (datopotamab deruxtecan, January 2025) and Emrelis (telisotuzumab vedotin-tllv, May 2025, targeting c-Met-overexpressing NSCLC, or non-small cell lung cancer). Topoisomerase I inhibitor payloads (DXd, SN-38) now dominate late-stage development due to strong bystander effects, where the released drug can also kill neighboring cells.

For target assessment, ADCs open up targets where selective expression matters more than pocket tractability. A tumor-surface antigen that cannot be usefully modulated by a naked antibody (an antibody without an attached payload) can still be a good ADC target if it internalizes on binding.

Therapeutic peptides and cyclic peptides

Therapeutic peptides occupy the middle ground between small molecules and biologics. They typically fall in the 500–5,000 Dalton range, often offer high target specificity, are less likely to trigger an immune response than antibodies, and have better membrane permeability than most proteins. Linear peptides suffer from poor stability and short half-life. Cyclic peptides address both by constraining the peptide into a ring, which improves rigidity, stability, and permeability. Cyclic peptides sit within the broader “beyond-Rule-of-Five” (bRo5) chemical space - shorthand for compounds that break the usual small-molecule size and polarity heuristics. Macrocycles (large rigid rings of 12 or more atoms), PROTACs, peptides, and metallodrugs all populate this space. By 2024, the field broadly accepted that strict adherence to Lipinski’s Rule of Five was not a prerequisite for intracellular engagement, and that bRo5 chemistry was central to engaging the flat protein–protein interaction surfaces that dominate the “undruggable” set. Cyclic peptides have shown activity in oncology, antiviral, antibacterial, and antimalarial contexts.

CAR-T, gene therapy, cell therapy

CAR-T and TCR-T therapies genetically modify a patient’s T cells to express a chimeric antigen receptor or T-cell receptor, then infuse the modified cells back. Kymriah and Yescarta were the first two FDA approvals in 2017; Tecartus, Breyanzi, and Abecma followed for mantle cell lymphoma, large B-cell lymphoma, and multiple myeloma respectively. By early 2026, the FDA reported having approved close to 50 cell and gene therapy products over the prior decade across the Center for Biologics Evaluation and Research (CBER) Office of Therapeutic Products, the cumulative result of approvals such as Casgevy and Lyfgenia (sickle cell disease), Beqvez (hemophilia B), Elevidys (Duchenne muscular dystrophy), Vyjuvek (dystrophic epidermolysis bullosa), Aucatzyl (B-ALL, or B-cell acute lymphoblastic leukemia), and an ongoing run of CAR-T approvals.

For target assessment, CAR-T targets are a different class of object from small-molecule targets. A CAR-T target is a surface antigen expressed, ideally exclusively, on disease cells - the “druggability” question is about on-tumor versus off-tumor expression (whether the antigen appears on cancer cells, healthy cells, or both), not pocket geometry. That leads to the other half of the 2026 druggability question.

The druggable genome is expanding, and the ignorome remains large

There are more than 10,000 known human diseases. The original 2002 estimate put 3,000–10,000 disease-related genes in the genome, with roughly 10% of those being disease-modifying on knockout (disabling the gene changed a disease-relevant phenotype in model systems). That yields the 600–1,500 small-molecule-druggable target estimate cited at the top of this post.

The Rule of Five (Lipinski et al., 1997) shaped small-molecule design for a generation: no more than 5 hydrogen-bond donors, 10 acceptors, molecular weight under 500, and LogP under 5, with LogP being a measure of lipophilicity. By 2026, approved drugs increasingly exceed Rule-of-Five boundaries. The main alternative is the Rule of Three for fragment-based discovery, a strategy that starts from very small weak-binding molecules (Congreve et al., Drug Discov. Today 8:876–877, 2003: MW < 300, ≤3 H-bond donors, ≤3 H-bond acceptors, cLogP ≤ 3, where cLogP is the computed version of LogP). The empirical boundaries of “drug-like” are shifting as chemistry expands to PROTACs, macrocycles, ADCs, and covalent warheads. These categories routinely break Lipinski’s original rules and still get approved.

Concept Translation: Rule of Five was a hand-crafted feature filter - a set of property thresholds that empirically separated absorbable drugs from non-absorbable ones, fitted on roughly 2,200 oral drugs in the late 1990s. Like any hand-crafted filter, it was tuned to its training distribution. As the chemistry expanded into larger and more polar molecules, the filter started rejecting things that worked. The 2026 stance is to use Lipinski thresholds as a soft prior for the small-molecule region of chemical space and to ignore them entirely outside it.

Meanwhile, over 75% of protein research still focuses on the 10% of proteins known before the human genome was mapped. The NIH’s Illuminating the Druggable Genome (IDG) initiative, launched in 2014, created a four-level Target Development Level classification to make this visible:

• Tclin: targets linked to at least one approved drug by mechanism of action.
• Tchem: proteins known to bind small molecules with high potency, but without approved-drug links.
• Tbio: proteins with a confirmed Mendelian disease phenotype or meeting certain experimental criteria.
• Tdark: proteins meeting none of the above.

IDG found that Tdark proteins receive less research funding than other categories, which perpetuates the knowledge gap. The ignorome - the Tdark set - is both the largest opportunity space for novel target discovery and the hardest place to work, because almost every downstream tool (antibodies, assays, knockout models) has to be built from scratch. The computational tractability gains from AF3-class structure prediction matter most here: a target with no crystal structure and minimal biological characterization now at least comes with a predicted structure to work from.

Concept Translation: The TDL classification is a label hierarchy on the human proteome - closer to a knowledge-graph annotation than a clean training-data split. Tclin proteins have the most data on them (drugs, structures, papers, assays) and Tdark have the least. ML methods trained on protein-level features inherit the same skew as the labels: they perform best on Tclin and worst on Tdark, the long tail. The “ignorome” framing emphasizes that the tail is where novel drug targets actually live.

A 2026 assessment of a novel target should include, at minimum, an AF3-class structure prediction with pocket analysis, a TDL classification, a scan across modality options (small molecule, PROTAC, ASO, ADC, peptide), and an explicit note on whether Rule-of-Five chemistry is even the right design envelope for this target. The question novelty scoring and the ignorome¹ takes up how novelty itself is now quantified across the TDL spectrum.

What’s next in druggability assessment

By 2026, the central druggability question is “which of five or six modalities fits this target best, and what is the safety profile of hitting it that way?” Three follow-ups stand out.

Safety. A PROTAC that catalytically degrades its target in tumor tissue may also degrade the same target in healthy tissue where it is essential. An ASO that knocks down a transcript in the liver may behave differently in the kidney. Modality-specific on-target, off-tissue toxicity - where the intended target is hit in the wrong tissue - is the next frontier, and it’s directly continuous with the tissue-specificity question the next post in this series picks up: tissue specificity.²

Two targets, not one. The entire framing of druggability assumes you’re choosing a single target and asking what modality hits it. For some diseases, most notably cancers with defined synthetic-lethal vulnerabilities (where a cell tolerates either perturbation alone but not both together), the right answer is that you need to hit two targets together, and each of those targets individually is a poor drug candidate. That shifts the druggability question to a two-body problem, with its own computational and experimental methods: synthetic lethality.³

Benchmarks for modality-choice models. We don’t yet have a standard benchmark for “given this target, which modality will work best?” The PoseBusters benchmark answers a narrower question (given this target and this ligand, is the predicted pose right?). Benchmarks for PROTAC ternary-complex geometry, ASO off-target profiles, and ADC internalization kinetics exist in pieces but do not yet form a coherent evaluation suite. That gap is a natural target for the next round of infrastructure investment. For the pillar-post view of where this fits in the overall workflow: the pillar post on drug target discovery.

AlphaFold 3 was an inflection point. The remaining work now sits in modality selection, safety, and benchmark design.

Further reading

This Further Reading section leans on primary papers rather than review-heavy framing because the topic dates quickly. AF3 is two years old at the post’s publication, PROTAC/TPD clinical status moves every few months, and the ADC field adds approvals on a rolling basis. Primary papers and canonical databases age more gracefully.

• Abramson, Adler, Dunger, Evans, et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500. DOI: 10.1038/s41586-024-07487-w. The primary source for every AF3-specific claim in the post, including the PoseBusters benchmark, the Pairformer-plus-diffusion architecture, and the joint protein–ligand–DNA–RNA prediction capability. Cited rather than a more recent review because the benchmark numbers in reviews inherit from this paper anyway. The competing AF3-class models (Boltz-2, Chai-2, Protenix, OpenFold3) have released benchmark head-to-heads since 2024; readers building a current-day pipeline should consult the relevant tool’s release notes for the most up-to-date head-to-head numbers.

• Hopkins, A. L., and Groom, C. R. (2002). The druggable genome. Nature Reviews Drug Discovery 1, 727–730. The foundational paper that anchors the “10–15% of the proteome is druggable” framing used throughout the post’s definitional section. Oprea, Bologa, Brunak et al. (2018), Unexplored therapeutic opportunities in the human genome, Nature Reviews Drug Discovery 17, 317–332, is the closest-to-current methodological update and provides the Tclin/Tchem/Tbio/Tdark classification that the ignorome section discusses - consider citing it alongside Hopkins & Groom if the post needs a more current primary reference.

• Ren, F., Ding, X., Zheng, M., et al. (2023). AlphaFold accelerates artificial intelligence powered drug discovery: efficient discovery of a novel CDK20 small molecule inhibitor. Chemical Science 14, 1443–1452. DOI: 10.1039/d2sc05709c. The primary source for the worked CDK20/HCC example used in the post - the first published demonstration of AlphaFold-predicted structures driving hit identification end-to-end. Verifies the 30-day target-to-hit timeline, the 8,918-molecule generative pool, the MET84/PHE81 hinge–gatekeeper geometry, and the ISM042-2-048 33 nM IC50 figure.

• Cui, J. J., Tran-Dubé, M., Shen, H., et al. (2011). Structure based drug design of crizotinib (PF-02341066), a potent and selective dual inhibitor of mesenchymal–epithelial transition factor (c-MET) kinase and anaplastic lymphoma kinase (ALK). Journal of Medicinal Chemistry 54, 6342–6363. DOI: 10.1021/jm2007613. The medicinal-chemistry case history behind the crizotinib comparison - documents the path from pyrrole-substituted oxindoles through PHA-665752 to the 2-aminopyridine clinical candidate, including the iterative crystallography that the AF3-era CDK20 workflow makes optional.

• NIH Illuminating the Druggable Genome Program. Pharos: the IDG Knowledge Management Center’s public resource for target development-level classification. pharos.nih.gov. The canonical open-access source for the TDL classification (Tclin/Tchem/Tbio/Tdark) the post uses in its ignorome section. Cited as the database home page rather than a specific IDG paper because the TDL categories are continuously updated against new drug approvals and omics evidence - a release-notes or current-database link ages better than a one-time publication link.

Concept Translation: Think of a drug target as a feature in a model that you’ve decided to intervene on. Picking a feature that strongly correlates with disease tells you nothing about whether modulating it will change the outcome - that’s the difference between a predictive feature and a causal one. Most of this post is about the evidence techniques the field uses to tell these apart for biological “features” (genes and proteins).

Why target–disease evidence matters in 2026

Industry-wide Phase I likelihood of approval has fallen from about 10.4% to 6.7%. The drivers include a shift toward first-in-class and riskier targets, greater use of biomarkers as surrogate efficacy endpoints (early readouts used as stand-ins for clinical benefit), and tougher regulatory scrutiny. As easy targets are mostly drugged, what remains rests on thinner evidence bases. Biomarker-stratified programs run at roughly double the likelihood of approval of unstratified ones, and targets backed by human genetic evidence run at roughly double the success rate of targets without that support.

Two industry-side facts motivate greater evidence rigor. First, the pipeline is crowded around a small number of targets. About 25% of the global R&D pipeline (roughly 13,600 drug-target pairs) concentrates on just 38 unique biological targets, and in oncology specifically the number of assets per target has grown from 1.8 to 9 over the last two decades (Plenge, 2026). When nine companies are pursuing the same protein, evidence becomes the differentiator - which program has the clearest human-genetic case linking target to disease, the most defensible biomarker, and the cleanest safety picture after accounting for off-target effects (unwanted activity at proteins other than the intended target). Second, human genetic evidence now accompanies approximately two-thirds of recently FDA-approved therapeutics (Ochoa et al., 2022, Nat Rev Drug Discov; Trajanoska et al., 2023, Nature). What was a competitive advantage a decade ago is now table stakes for serious target programs.

What target–disease association evidence actually means

Target-disease evidence is a structured argument built from heterogeneous data sources. A credible argument has four parts:

1. A mechanistic hypothesis. A story about how modulating this protein would alter the disease phenotype. We seek a plausible causal chain, going beyond naive correlation with the disease state.
2. Human evidence that the hypothesis is correct. Genetic when possible, since human genetic variation imitates the design of a randomized controlled trial without requiring a drug intervention. Multi-omics, patient-tissue, and literature evidence all count, with known caveats.
3. Experimental evidence that the modulation produces the therapeutic effect. Knockout, knockdown, or pharmacological perturbation - that is, genetic removal, genetic reduction, or chemical modulation of the target - in disease-relevant models, ideally more than one, that reverses the disease phenotype rather than only altering a biomarker.
4. A defensible position on the competing hypothesis. Why this mechanism rather than the three adjacent ones the literature also supports?

Concept Translation: Knockouts and knockdowns are the biology field’s version of ablation studies. Knockout = the gene is removed entirely (full ablation). Knockdown = expression is reduced but not eliminated (partial ablation, more like dropout). Pharmacological perturbation = a chemical inhibits the protein, which is closer to “intervening at inference time” than “retraining without the feature.” All three answer the same question: what does the system do when this component is missing or weakened?

Teams that skip one of these four risk failure. The most common failure is #4 - the team has strong data on its favored mechanism and has never seriously engaged with the alternatives.

Two frameworks for organizing this argument are the AstraZeneca 5R framework and the GOT-IT recommendations.

The AstraZeneca 5R framework: “right target” as the first R

The 5R framework defines five criteria a drug program has to satisfy: right target, right tissue, right safety profile, right patient, right commercial potential. The “right target” criterion is defined narrowly as a strong link between target and disease, predictive biomarkers that identify likely responders, and demonstrated differentiated efficacy relative to alternatives. Lack of efficacy has been the most important cause of project failure in clinical trials. The 5R framework forces teams to separate evidence of target-disease linkage from evidence of tissue availability, safety, and patient stratification (the choice of which patient subgroup to treat). A single piece of data often speaks to only one of these categories, and conflating them makes the evidence base look stronger than it is.

The framework originated at AstraZeneca in the mid-2010s with Cook et al. (2014) Nature Reviews Drug Discovery, followed by Morgan et al. (2018) in Drug Discovery Today, which documented an approximately five-fold improvement in the company’s small-molecule project survival from candidate-selection to Phase III decision after the framework was institutionalized.

GOT-IT: four assessment blocks

The GOT-IT (Good Target Identification and Validation) recommendations, published by Emmerich et al. (2021) in Nature Reviews Drug Discovery, organize target assessment into four blocks. The first is the focus of this post:

• Disease linkage: evaluating the relationship between the target and the disease of interest and determining whether the target is involved in the underlying disease biology.
• Target-related safety: covered in a later post on tissue-specificity.
• Strategic issues: commercial potential, portfolio fit, competition.
• Technical feasibility: covered in a later post on druggability.

5R and GOT-IT overlap, but they aren’t redundant. 5R is a pipeline-level framework, like a checklist for whether a program should advance. GOT-IT is a target-level framework, a checklist for whether the target-assessment work has been done thoroughly. Many organizations use both, sometimes customizing one internally and treating the other as the external benchmark, or using them in tandem with additional frameworks.

Four important distinctions within target–disease linkage

Driver vs passenger mutations

In cancer genomics, driver mutations provide selective advantages for cancer initiation and progression. Passenger mutations accumulate alongside them without contributing to disease fitness. A single gene can host both classes. The APC gene, a highly mutated tumor suppressor in colorectal cancer, carries both driver and passenger mutations across its coding region. Distinguishing them is a computational problem: which variants observed in a tumor caused the malignant phenotype, and which reflect background mutation noise?

Concept Translation: Drivers and passengers are the cancer-genomics version of signal vs. correlated noise. A tumor genome contains hundreds to thousands of mutations, only a handful of which actually move the phenotype. The other mutations co-occur with the drivers because they accumulated in the same lineage. Picking drivers out of the noise is the same kind of feature-selection problem as identifying which input variables actually drive a model’s predictions when the inputs are highly correlated.

A list of “genes mutated in disease X” is raw material. A target-discovery workflow has to rank those genes by probability of being a driver, using statistical tests on mutation frequency against background, functional impact scoring, recurrence across patients, pathway enrichment, and so on. Tools like MutSigCV, OncodriveCLUST, and later deep-learning classifiers exist for this. The output of a cancer-genomics pipeline is a ranked driver list. Moving from that list to a target list requires additional evidence.

A related wrinkle is that some passengers are clinically informative. Weak drivers that fall below detection thresholds, and accumulated “passenger” burden that shifts cell phenotype cumulatively, can matter clinically. Some passengers also provide evolutionary evidence of tumor lineage and can be used for timing inferences.

Causative, supportive, and symptomatic treatment

A second distinction concerns the kind of therapeutic effect a target is meant to produce:

• Causative treatment: therapy directed against the cause of disease. Antiviral agents against SARS-CoV-2. Ritonavir inhibiting HIV protease. A PARP inhibitor in a BRCA-mutated tumor. Targets for causative treatment live upstream in the disease mechanism.
• Symptomatic treatment: therapy that eases symptoms without addressing the underlying cause. Pain relievers. Cough suppressants. Targets for symptomatic treatment live downstream of the disease cause, typically in host response pathways.
• Supportive treatment: care that addresses broader patient needs (physical, existential, palliative) rather than targeting disease biology directly.

For target discovery, this distinction clarifies what “linkage to the disease” has to establish. A target for causative treatment has to be causally upstream of the disease process. A target for symptomatic treatment has to be causally upstream of a symptom, which is a weaker claim. A program’s target-evidence burden depends on which kind of drug it’s trying to develop.

Alzheimer’s provides a standard example of a disease where causative target discovery has resisted easy success. Acetylcholinesterase inhibitors (donepezil, rivastigmine, galantamine) and memantine provide symptomatic therapies that target downstream neurotransmitter biology rather than disease etiology. The amyloid-targeting antibodies aducanumab (accelerated approval 2021, voluntarily withdrawn in 2024), lecanemab (Leqembi, FDA traditional approval July 2023, EMA approval late 2024), and donanemab (Kisunla, FDA approval July 2024) aim at a putative causal mechanism. Phase III data for both lecanemab and donanemab show statistically significant slowing of cognitive decline relative to placebo over 18 months. Debate remains over whether those effect sizes are clinically meaningful. The safety burden of amyloid-related imaging abnormalities (ARIA) has also been a major clinical concern, and the EMA has had a mixed reception of the two drugs. The target-disease linkage evidence for amyloid is therefore stronger than it was a decade ago - there is now Phase III evidence that targeting amyloid produces a measurable clinical effect - but this does not clarify whether amyloid is the causative driver of cognitive decline, a downstream marker, or one of several parallel drivers.¹ The earlier generation of amyloid-targeting programs provides a cautionary tale, though continuing studies test whether the hypothesis was right but underpowered, or only partly right with diminishing returns.

Oncogene addiction

Oncogene addiction describes cancer cells’ dependence on individual oncogenes - genes whose altered activity drives tumor growth - to sustain the malignant phenotype. A tumor that depends on an activated oncogene is selectively vulnerable to inhibition of that oncogene. Tumor cells need the oncogenic signal to survive; healthy cells do not. This has been a successful organizing concept in targeted oncology. EGFR mutations in a subset of non-small cell lung cancer, BCR-ABL fusion in chronic myeloid leukemia, and BRAF V600E mutations in melanoma all produce tumors with oncogene-addicted phenotypes, and targeted inhibitors in each context have produced durable clinical responses.

For target-discovery methodology, oncogene addiction is useful because it predicts a specific clinical-trial readout: patients whose tumors harbor the addiction should respond, and others should not. That prediction is testable in Phase II with biomarker-stratified cohorts, and it helps explain why biomarker-stratified oncology programs show higher LOA than unstratified ones (we’ll quantify this in a later post on LOAs by therapeutic area). An evidence package for an oncogene-addiction target should therefore include the candidate biomarker - the measurable feature that marks the likely responder subgroup.

Target-disease correlation vs causal target-disease linkage

The most consequential distinction is also the most abstract. A gene whose expression differs between disease tissue and healthy tissue is associated with the disease. A gene whose loss-of-function variants - variants that reduce or abolish a gene’s activity - segregate with disease risk in human populations at a genetic locus that survives multiple-testing correction, with a plausible protein-level mechanism and Mendelian-randomization-style evidence that inherited variation in the gene affects disease rather than the reverse, is a causal candidate. Curating such evidence is hard. Transcriptomic association is relatively cheap; causal genetic evidence is expensive and sparse.

Concept Translation: Mendelian randomization is the genetics field’s instrumental-variable method. Because alleles are randomly assorted at conception, an inherited variant acts like a randomized treatment assignment - it isn’t influenced by lifestyle, environment, or downstream disease state. If carriers of a loss-of-function variant in gene X are healthier on average, that’s strong evidence that reducing X’s activity reduces disease risk, not just that X is correlated with it. This is closer to an A/B test than to an observational dataset, which is why genetic evidence carries more causal weight than transcriptomic evidence.

Referring back to the 2× headline from earlier, the Nelson et al. (2015) result and the King et al. (2019) revised analysis both confirmed the population-average ~2× clinical-success multiplier for genetically supported drug programs. The same general 2× factor also enriches for labeled side effects, giving it predictive value for toxicology programs (Minikel & Nelson, 2024; Carss et al., 2023). The multiplier is not constant across evidence types. King et al. found that when the causal gene is unambiguous - Mendelian traits, single-gene disorders with clear inheritance patterns, or GWAS associations driven by coding variants where the variant-to-gene mapping is direct - the approval probability multiplier rises above 2× and into the 3× range. The 2026 update from Minikel and Nelson refined this, showing that the multiplier scales with confidence in the causal gene assignment and is largely independent of genetic effect size, minor allele frequency, or year of discovery. In practice, a low-effect-size GWAS hit with confident causal-gene mapping is worth more than a high-effect-size hit at a locus where the variant-to-gene call is ambiguous. Trajanoska et al. (2023, Nature) place this in longer historical context: they identified 40 germline genetic observations that translated directly into approved therapies for 36 rare and 4 common conditions, with a median 25-year interval between target discovery and drug approval. The genetic-anchor strategy works, but it compounds slowly.

When assembling a target’s evidence package, the quality of the genetic anchor matters more than the quantity of associated data. A coding variant with established protein consequences is worth more than a half-dozen non-coding GWAS signals at adjacent loci with ambiguous fine-mapping.

The omics stack as target–disease evidence

Target-disease linkage evidence comes from an omics stack, and each layer has a characteristic failure mode. Knowing when each layer can mislead you is the core methodological skill.

Genomics and GWAS

Genomics is usually the strongest layer for target-disease linkage. Human genetic variation provides a natural experiment: people who carry a loss-of-function variant in gene X are, on average, similar to people who would be pharmacologically treated with an inhibitor of gene X. If carriers of the variant have lower disease risk, that is a strong argument that inhibiting gene X would lower disease risk too. This is the Mendelian-randomization logic that gives genetic evidence special weight.

Genome-wide association studies (GWAS) are the dominant method for finding these signals. GWAS identifies statistical associations between genetic variants and traits or diseases across large population cohorts. The method has identified thousands of risk loci across hundreds of diseases, now aggregated in resources like the NHGRI-EBI GWAS Catalog.

Concept Translation: GWAS is in some ways a giant feature-importance scan over the genome - for each of millions of variants, test whether allele frequency differs between cases and controls. The genome-wide significance threshold of p < 5×10⁻⁸ is a Bonferroni-style correction for testing roughly a million effective independent variants. The catch is that statistical association picks the neighborhood the causal variant lives in, not the variant itself. Most GWAS hits sit in non-coding DNA - regulatory regions outside protein-coding sequence - which is why follow-up fine-mapping is needed to narrow the signal. Fine-mapping is a feature-attribution problem for a region of the genome: which variant in the correlated cluster is actually doing the work?

The practical limitations matter. GWAS has to apply heavy multiple-testing correction (the classic p < 5×10⁻⁸ genome-wide significance threshold), which means it misses real signals at smaller effect sizes, especially in underpowered cohorts. Most GWAS hits are in non-coding regions, so the protein-level mechanism is often implicit. Identifying the causal gene from a genome-wide-significant locus often requires additional work - fine-mapping (narrowing the locus to the most plausible causal variants), eQTL analysis (testing whether a variant shifts gene expression), and experimental follow-up.

The populations represented in GWAS are also overwhelmingly European-ancestry, which creates a real equity problem when target discovery built on GWAS evidence is applied clinically to more diverse populations. The June 2021 GWAS Catalog inventory put European-ancestry participants at approximately 86.3% of all GWAS subjects, with East Asian at ~5.9% and African at ~1.1%. By 2023 the GWAS Catalog European share had only modestly declined, to around 86.5%, while the GWAS Diversity Monitor, which tracks the running participant counts more broadly, recorded European representation of ~94.5% of total participants by September 2024 (Mills & Rahal, Nat Genet, 2019; Corpas et al., Cell Genomics, 2025). The two metrics differ because the GWAS Diversity Monitor weights large biobank cohorts heavily, but the conclusion is the same: target discovery built on this evidence base will systematically under-serve non-European populations unless deliberately corrected with fine-mapping in diverse cohorts and downstream validation across ancestries. This is a target-evidence problem as much as a deployment problem. A target whose causal gene is well mapped only in European cohorts may not support a defensible target-disease linkage claim for patients of other ancestries until cross-ancestry replication is done.

Concept Translation: This is the same dataset-bias problem that ML practitioners deal with constantly. A model trained on a non-representative dataset can be highly accurate on the in-distribution group while failing on out-of-distribution groups. The fix is the same in principle: collect more representative data, evaluate per-subgroup, and don’t ship a model whose validation doesn’t cover the deployment population.

Transcriptomics

Transcriptomics measures RNA levels, sitting between genotype and protein as a sensitive readout of cellular state. Bulk and single-cell RNA sequencing, along with older microarray data, make up the largest public omics archives (GEO, ArrayExpress) and provide differential-expression evidence: which genes change in disease tissue versus healthy tissue?

The main failure mode of transcriptomics for target discovery is simple. Many transcriptomic changes are bystanders. A gene that changes expression in disease is not necessarily causing the disease; it may be responding to the disease, or changing in parallel with it because of upstream regulation. A transcriptomic hit generates a hypothesis; it does not establish causal linkage. Evidence becomes much stronger when transcriptomics and genetics agree, and the gene is both differentially expressed and under genetic selection in disease.

Concept Translation: Bystander vs. causal in transcriptomics is the same trap as confounded correlation in observational ML. A feature that strongly correlates with the label can still be downstream of the label rather than upstream - predictive of it without driving it. RNA-level changes pick up both kinds and can’t separate them on their own. This is why GWAS-supported transcriptomic hits are weighted more heavily than transcriptomic hits in isolation.

Proteomics and metabolomics

Proteomics measures protein abundance directly. Since most drugs target proteins, proteomic evidence sits closer to the mechanism-of-action question than transcriptomic evidence. The limitations are technical: mass-spec proteomics is expensive, less sensitive than RNA-seq, and struggles with membrane proteins, which include many of the cell-surface receptor and transporter classes that matter most in drug discovery.

Metabolomics measures small-molecule metabolites. It often points to disease-associated metabolic changes that can be traced back to protein targets - which enzymes are producing or consuming these altered metabolites. Metabolomics is particularly useful for metabolic diseases and for target discovery in contexts where disease biology is about pathway flux more than signaling.

Epigenomics

Epigenomics (DNA methylation, histone modifications, chromatin accessibility) provides a reversible regulatory layer above the genome. Epigenetic enzymes are themselves drug targets (HDAC inhibitors, DNMT inhibitors, both with approved drugs), and epigenetic data can identify non-genetic mechanisms driving disease. The characteristic limitation is interpretability: it is often unclear what functional impact a given epigenetic change has on gene expression in the specific disease context.

Concept Translation: A useful mental model: if the genome is the source code, epigenomics is the runtime configuration. Methylation marks and chromatin states change which parts of the code get executed in which cell types, without altering the code itself. Same source, different behavior - and the configuration is reversible, which is why epigenetic enzymes are drug targets.

Multi-omics integration and its limits

Combining these layers sounds straightforward. In practice, it’s a difficult technical problem. Batch effects across studies, confounding by disease severity or patient demographics, differential dropout in single-cell data, and different noise models across modalities all complicate naive integration. The best-case design, when available, analyzes multi-omics data from the same set of patients rather than integrating across cohorts, so sample-level confounders are controlled. That is rarely possible at GWAS-cohort scales, where multi-omics profiling of every participant is prohibitively expensive. Target-discovery workflows that build evidence from public multi-omics resources need to state these limitations explicitly. Workflows that report “our multi-omics integration identified target X” without addressing them are not making a credible argument.

A useful target-disease evidence package usually includes (a) genetic evidence where it exists, as the anchor, (b) transcriptomic and proteomic evidence in disease tissue that is consistent with the genetic signal, (c) tissue-specificity data (see the tissue-specificity post²), and (d) network or pathway context that situates the target in plausible biology. No single layer is enough - the layers have to agree.

Networks and knowledge graphs: evidence beyond single-gene associations

A gene-at-a-time view of the omics stack misses mechanistic context. A gene with moderate genetic support and moderate transcriptomic support may look like a weak target in isolation, yet sit at a critical node in a disease-implicated pathway. In that setting, network position is itself evidence. Target-discovery methodology increasingly builds evidence at the network level as well as the gene level.

The method family is biological graphs and knowledge graphs. Protein-protein interaction networks, gene regulatory networks, gene co-expression networks, metabolic networks, and signaling networks each represent a different axis of biological relationship, and each carries evidence relevant to target linkage. Knowledge graphs integrate these with literature-derived associations, drug-target relationships, disease-gene associations, and related data, producing heterogeneous graphs with multiple types of nodes and relationships that can be queried for evidence patterns around a candidate target.

Concept Translation: A biomedical knowledge graph is a heterogeneous graph in the standard ML sense - multiple node types (genes, proteins, diseases, drugs, pathways) and multiple edge types (binds, expressed-in, treats, associated-with). The same graph-ML toolkit applies: node classification (is this gene a likely target?), link prediction (does this gene-disease edge exist but is undiscovered?), graph neural network embeddings, and metapath-based random walks. The biology adds structure but doesn’t change the modeling problem.

Two network features are especially useful for target-linkage arguments. Hubs are highly connected nodes. They are enriched for essential genes and are often promising targets in diseases driven by loss of function, but they are also more likely to be toxic to modulate pharmacologically. Bridging nodes are high-betweenness nodes - nodes that sit on many shortest paths between modules. They are often more attractive targets because their inhibition tends to be non-lethal while still disrupting specific disease-relevant pathways.

Concept Translation: Hubs are nodes with high degree centrality (think high-PageRank pages); knock them out and a lot of the graph collapses, which is why they’re both attractive and dangerous. Bridging nodes are nodes with high betweenness centrality - they sit on many shortest paths, like a bottleneck router connecting subnets. Inhibiting a bridge can disconnect a specific module without crashing the whole network, which is roughly what a clean drug effect looks like.

A knowledge graph over disease-gene-drug-pathway data also lets teams aggregate and score evidence across heterogeneous sources. The Open Targets Platform is the canonical public example: an open-source platform that integrates evidence from genetics (GWAS, ClinVar, ClinGen), somatic mutation data (Cancer Gene Census, COSMIC), known drug-target relationships, expression data, text mining, and animal models, producing a quantitative association score between each gene and each disease. As of the most recent platform release, the Open Targets evidence layer comprises over 27.8 million timestamped evidence assertions across the genes-and-diseases space.

The platform’s scoring methodology matters in practice. The overall target–disease association score is computed as a harmonic sum of source-weighted evidence scores - a scheme in which each additional piece of similar evidence adds less than the previous one, rather than a raw sum or arithmetic mean. The harmonic-sum construction keeps heavily studied targets from accumulating inflated scores through redundant literature citations. Adding the 50th literature reference does not move the score the way the 1st reference does. Within data types that contain multiple sources, the scaling factor uses an inverse-squares construction so that source contributions saturate gracefully rather than letting any single high-throughput source dominate.

Concept Translation: The harmonic-sum aggregator is a regularizer against redundant evidence. Each additional similar source contributes less than the previous one, much like a saturating activation function. Without it, well-studied targets - those with thousands of papers - would dominate the rankings purely by literature volume, and the score would be a popularity ranking instead of an evidence ranking. This is the same problem that plagues citation-count metrics in academia and is solved with the same general technique: diminishing returns on duplicate signal.

Source weights vary by evidence type. ClinGen clinical-validity curations at the “definitive” tier receive an absolute weight of 1.0, the strongest possible single-source evidence. Somatic mutation evidence from the Cancer Gene Census uses a tiered modifying scheme that adjusts the base score by mutational frequency and disease-specificity context: a mutation observed in only 1 sample in the dataset receives a −0.25 modifier (penalizing weak recurrence), while a mutation that occurs more frequently in a particular disease relative to others receives a +0.25 modifier (rewarding disease-specificity). The platform’s “Associations on the Fly” (AOTF) feature, introduced in recent releases, lets users dynamically reweight the evidence contributions to formulate custom therapeutic hypotheses without forking the underlying scoring schema.

The consortium itself is worth noting for field-guide purposes. Open Targets is a pre-competitive public-private partnership - direct competitors contribute to a shared evidence resource before they diverge into proprietary drug programs. As of 2026 the active partner roster includes EMBL-EBI, the Wellcome Sanger Institute, Genentech (Roche Group), GSK, MSD (Merck & Co.), Pfizer, Sanofi, and Bristol Myers Squibb (which joined in late 2022). Biogen and Takeda were earlier members who exited in 2020. The platform aggregates signals that no single company would assemble alone.

Formalizing evidence aggregation as a graph problem also opens the door to graph machine learning. Target-disease association can be formulated as a link-prediction problem on a heterogeneous biomedical knowledge graph: given a graph where some gene-disease edges are known, predict which missing edges are real. Graph neural networks, network-embedding methods, and random-walk-based algorithms have all been applied to this problem. The knowledge graphs and the rentosertib case study picks up this thread in detail.³

A worked example: senolytic target identification in idiopathic pulmonary fibrosis

To make the frameworks concrete, consider a target-identification workflow for a specific disease-hypothesis pair: cellular senescence as a target class in idiopathic pulmonary fibrosis (IPF).

The mechanistic hypothesis. Cellular senescence is a permanent state of cell-cycle arrest with a secretory phenotype (SASP) that promotes inflammation and disrupts tissue homeostasis. Senescent cell accumulation is associated with many age-related diseases, and in IPF specifically, senescent cell programs have been documented in disease tissue. The working hypothesis: compounds that selectively kill senescent cells (senolytics) should slow IPF progression by removing the SASP-producing source of inflammation and fibrotic signaling. This is a causative hypothesis in the sense of the earlier framework - senescent cells are positioned as upstream drivers rather than symptoms.

Concept Translation: Senescent cells are sometimes called “zombie cells” - they’ve stopped dividing but won’t die, and they leak inflammatory signals (the SASP) into surrounding tissue. The senolytic strategy is to selectively kill them. From a target-discovery angle, the question is which proteins keep senescent cells alive that are not equally important for keeping healthy cells alive. Those proteins are the senolytic targets.

Evidence layer 1: tissue pathology. Senescent cells are identifiable in IPF lung tissue by senescence-associated markers; their burden correlates with disease severity. This is differential-expression-style evidence - the disease tissue differs from healthy tissue in a specific, characterizable way. It is necessary but insufficient on its own. Senescence could still be a bystander response to lung injury rather than a driver of ongoing fibrosis.

Evidence layer 2: mechanism. Senescent cells are apoptosis-resistant by design; that is part of what makes them persistent. The resistance depends on anti-apoptotic pathways collectively called SCAPs (senescent cell anti-apoptotic pathways). That yields a specific target-class hypothesis: proteins involved in SCAPs are candidate senolytic targets, because inhibiting them should remove the apoptosis resistance and allow senescent cells to die. This narrows the search from “any gene differentially expressed in IPF” to “genes in a specific functional class whose inhibition would have a mechanistically predicted effect.”

Evidence layer 3: computational target prioritization. A knowledge-graph-based target discovery platform can operationalize the combined evidence. The workflow:

1. Start from a disease-specific dataset (IPF patient tissue vs healthy controls) and score candidate targets by disease-association strength.
2. Apply a gene-set filter restricting candidates to pre-specified functional classes (apoptosis-related pathways, SCAPs).
3. Apply a small-molecule druggability filter. This asks whether a target can plausibly be modulated by a conventional small-molecule drug. For senolytic drug development, the delivered modality is small molecule, so candidates that are only tractable by antibody or biologic are de-prioritized.
4. Apply an expression filter that focuses on protein classes suitable as drug targets (e.g., kinases, enzymes).
5. Apply an optional novelty filter that preferentially surfaces targets with fewer published references, for teams pursuing first-in-class programs against new mechanisms rather than fast-follower strategies against established mechanisms.

The output, checked against known biology. A representative output of this workflow is PTK2 (also known as FAK, focal adhesion kinase) as a high-ranked senolytic candidate. PTK2 is a tyrosine kinase, druggable by small molecules, and expressed in relevant contexts. The biological cross-check comes from independent experimental literature. Dasatinib, a multi-kinase inhibitor originally approved for chronic myeloid leukemia, has been identified as a senolytic in human dermal fibroblasts, and its senolytic activity has been attributed in part to inhibition of PTK2/FAK among other kinases. This independent wet-lab evidence predates the computational prioritization and converges on the same target class.

What this example illustrates. The evidence package for PTK2/FAK as a senolytic target is an aggregation of four sources: disease-tissue pathology (senescent cell burden in IPF), a mechanistic hypothesis (SCAP dependence), computational target prioritization weighted by disease association, pathway membership, and druggability, and convergent experimental evidence from a repurposed clinical-stage compound. No individual layer would be sufficient; together they make a credible argument.

The claim is narrower than it might sound. PTK2 remains a candidate rather than a validated senolytic target in IPF patients; validation would require dosing studies, biomarker responses, and clinical evidence. On the evidence presented here, PTK2 clears the 5R “right target” and GOT-IT “disease linkage” bars for advancing into target validation. Moving from candidate to validated target requires druggability assessment⁴, and the downstream posts in this series.

Open questions in target–disease evidence work

The benchmarking problem. Target-discovery ML is unusually hard to benchmark. The ground-truth signal - did this target hold up in Phase III? - takes a decade and nine figures per data point. Surrogate benchmarks exist (does the method rediscover known targets for known diseases when given only pre-discovery literature? does it recover genetic signals already in Open Targets? does it predict held-out drug-target relationships?) but all share the same weakness: the evaluation corpus overlaps with the corpus the ML methods were trained on. Truly unbiased benchmarking for target-discovery ML, where the test set is future-clinical rather than past-published, remains a live methodological problem.

The ignorome problem. Research attention across the ~20,000 human protein-coding genes is highly skewed: a small fraction of genes accounts for most of the literature, the annotations, the available assays, and the training data for ML methods. A target-discovery method that only surfaces candidates adjacent to already-well-studied proteins is solving a weaker version of the problem than the one the industry actually needs. The novelty vs repurposing post picks up this thread.⁵

The causality problem. Most of the evidence stack is associational. Genetic evidence - especially Mendelian-randomization-style evidence - carries more causal weight because it approaches causality in a way other omics layers do not. Methods that can extract more causal structure from non-genetic data (perturbation screens, single-cell perturbation readouts like Perturb-seq, large-scale CRISPR screens) are one of the most active areas of methodological development, and they feed directly into target-disease evidence work. The synthetic lethality post gets into CRISPR-screen analysis in depth.⁶

Evidence frameworks for target-disease linkage separate disciplined target discovery from assertion. Teams that internalize them are less likely to skip steps, and skipped steps are rarely recoverable later. The framework languages (5R, GOT-IT) and the evidence-stack discipline - genomics as anchor, corroborated by transcriptomics and network context, challenged by alternative hypotheses - are what distinguish a defensible target list from one that merely sounds plausible.

Further reading

Foundational framework papers

• Cook, D., Brown, D., Alexander, R., March, R., Morgan, P., Satterthwaite, G., & Pangalos, M. N. (2014). Lessons learned from the fate of AstraZeneca’s drug pipeline: a five-dimensional framework. Nature Reviews Drug Discovery, 13(6), 419–431. The canonical 5R paper, published from inside AstraZeneca’s R&D leadership at the moment the framework was institutionalized.

• Morgan, P., Brown, D. G., Lennard, S., Anantharaman, M. J., Gilbert, J., Cook, D., et al. (2018). Impact of a five-dimensional framework on R&D productivity at AstraZeneca. Drug Discovery Today, 23(8), 1395–1399. The quantitative follow-up documenting an approximately five-fold improvement in small-molecule project survival from candidate-selection to Phase III decision after the 5R was institutionalized - the empirical basis for the framework’s enduring methodological status.

• Emmerich, C. H., Gamboa, L. M., Hofmann, M. C. J., Bonin-Andresen, M., Arbach, O., Schendel, P., et al. (2021). Improving target assessment in biomedical research: the GOT-IT recommendations. Nature Reviews Drug Discovery, 20(1), 64–81. The definitive methodological reference for the four-block target-assessment framework (disease linkage, target-related safety, strategic issues, technical feasibility) that this post’s structure mirrors.

Genetic evidence and the success multiplier

• Nelson, M. R., Tipney, H., Painter, J. L., Shen, J., Nicoletti, P., Shen, Y., et al. (2015). The support of human genetic evidence for approved drug indications. Nature Genetics, 47(8), 856–860. The original analysis establishing that drug programs with human genetic support are roughly twice as likely to succeed in clinical development. This is the source for the 2× headline that anchors most subsequent literature.

• King, E. A., Davis, J. W., & Degner, J. F. (2019). Are drug targets with genetic support twice as likely to be approved? Revised estimates of the impact of genetic support for drug mechanisms on the probability of drug approval. PLOS Genetics, 15(12), e1008489. The revised estimate using updated pipeline data, confirming the ~2× general result but showing that high-confidence causal evidence (Mendelian traits, coding-variant GWAS hits) yields a larger >2× multiplier - the source for the 2× / 3× distinction in the post’s hook.

• Trajanoska, K., Bhérer, C., Taliun, D., Zhou, S., Richards, J. B., & Mooser, V. (2023). From target discovery to clinical drug development with human genetics. Nature, 620(7975), 737–745. Identified 40 germline genetic observations that translated directly into approved therapies for 36 rare and 4 common conditions, with a median 25-year interval between target discovery and drug approval - the long-horizon companion to the multiplier finding.

• Minikel, E. V., Painter, J. L., Dong, C. C., & Nelson, M. R. (2024). Refining the impact of genetic evidence on clinical success. Nature, 629, 624–629. The 2024 update showing that the genetic-support multiplier scales with confidence in the causal-gene assignment but is largely independent of effect size, allele frequency, or year of discovery.

• Ochoa, D., Karim, M., Ghoussaini, M., Hulcoop, D. G., McDonagh, E. M., & Dunham, I. (2022). Human genetics evidence supports two-thirds of the 2021 FDA-approved drugs. Nature Reviews Drug Discovery, 21(8), 551. The source for the “two-thirds of recent FDA approvals are underpinned by human genetic evidence” statistic in the introduction.

• Carss, K. J., Deaton, A. M., Del Rio-Espinola, A., Diogo, D., Fielden, M., Kulkarni, D. A., et al. (2023). Using human genetics to improve safety assessment of therapeutics. Nature Reviews Drug Discovery, 22(2), 145–162. Companion reference for the side-effect-prediction half of the genetic-evidence story.

GWAS diversity

• Mills, M. C., & Rahal, C. (2019). A scientometric review of genome-wide association studies. Communications Biology, 2, 9. Foundational quantitative analysis of European-ancestry overrepresentation in the GWAS literature; ongoing version available at the GWAS Diversity Monitor (gwasdiversitymonitor.com).

• Corpas, M., Pius, M., Poburennaya, M., Guio, H., Dwek, M., Nagaraj, S., et al. (2025). Bridging genomics’ greatest challenge: the diversity gap. Cell Genomics, 5(1), 100724. The most recent quantitative update on GWAS diversity, with the September 2024 GWAS Diversity Monitor numbers (~94.5% European participants) and the 2023 GWAS Catalog comparison (~86.5%) used in this post.

Open Targets

• Open Targets Platform Documentation: target–disease associations and target–disease evidence. platform-docs.opentargets.org. The canonical reference for the harmonic-sum scoring construction, the source-weight schema (including the ClinGen 1.0 absolute weight and the Cancer Gene Census ±0.25 modifiers), and the Associations on the Fly (AOTF) feature. The release-notes page is the right link because evidence-category weighting changes across releases - cite release notes rather than a static platform URL so readers land on the version-of-record for whatever’s current at your publish date.

• Open Targets news and partner roster: opentargets.org/partners and opentargets.org/news. For the partner-consortium roster cited in the body (EMBL-EBI, Wellcome Sanger, Genentech, GSK, MSD, Pfizer, Sanofi, BMS) and the historical exits (Biogen and Takeda, both 2020).

Anti-amyloid antibodies in Alzheimer’s (for the causative-vs-symptomatic example)

• van Dyck, C. H., Swanson, C. J., Aisen, P., Bateman, R. J., Chen, C., Gee, M., et al. (2023). Lecanemab in early Alzheimer’s disease. New England Journal of Medicine, 388(1), 9–21. The pivotal Phase III result for lecanemab; the source for the “statistically significant slowing of cognitive decline” claim with the precise CDR-SB effect size.

• Sims, J. R., Zimmer, J. A., Evans, C. D., Lu, M., Ardayfio, P., Sparks, J., et al. (2023). Donanemab in early symptomatic Alzheimer disease: the TRAILBLAZER-ALZ 2 randomized clinical trial. JAMA, 330(6), 512–527. The pivotal Phase III result for donanemab. The post deliberately frames both drugs as test cases of the amyloid hypothesis rather than as resolved validation of it.

Why drug target discovery matters in 2026

When it comes to understanding chemistry and biology, the models are getting better. Generative models can design and optimize novel small molecules with greater success. AlphaFold 3 and the open-weights field around it, meaning models released with their parameters, have made routine structure prediction a solved problem for many classes of proteins. Docking, virtual screening, and ADMET prediction (absorption, distribution, metabolism, excretion, and toxicity) continue to mature as well.

Yet clinical success rates have worsened. Industry-wide likelihood of approval (LOA) for a Phase I compound fell from about 10.4% for the 2014 single-year cohort to around 6.7% (average LOA for 10-year period from 2014-2023). Phase II remains the dominant attrition step, with only about 28% of drugs that enter Phase II making it to Phase III (phase-by-phase transitions: Phase I = 47%, Phase II = 28%, Phase III = 55%, Filing-to-approval = 92%).

To assist instruction, front-of-funnel ML now has a concrete clinical example. In June 2025, Nature Medicine published Phase IIa results for rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF), whose target (identified from multi-omics analysis of IPF patient tissue, meaning joint analysis of several molecular data types) and molecule (generative chemistry) were both AI-derived end-to-end. The study was small and short: 71 patients over 12 weeks (it was hypothesis-generating rather than a study designed to support approval) with reported improvements in standard lung-function measures relative to a placebo. Still, it marked the first time a fully AI-discovered drug, target included, produced a direction-of-effect efficacy signal in humans. Industry trackers now count more than 173 AI-discovered drug programs in clinical testing as of early 2026.

This post lays out the series: what target discovery is, why it matters, what makes it hard, and where ML helps. Each subsequent post will take one axis of target discovery and examine it in detail:

• How to tell a drug target matters using evidence frameworks
• Druggability, ligandability, and modality choice in the AlphaFold 3 era
• Tissue specificity: the safety half of target selection
• The most crowded and abandoned therapeutic areas
• Novelty vs repurposing: when to invent a target and when to reuse one
• Synthetic lethality and combination targets: ML methods for finding drug pairs that work together
• Knowledge graphs and case studies in AI-driven target discovery
• Virtual cells for target discovery, perturbation models, and benchmarks

What a drug target is

A drug target is a biomolecule, usually a protein and sometimes a nucleic acid, whose activity we want to modulate to produce a therapeutic effect. “Modulate” covers a range: inhibit an overactive enzyme, block a receptor from binding its ligand, degrade a disease-driving protein, replace a missing transcript. Binding the target has to produce the therapeutic effect. A molecule that binds to the target without changing disease outcomes is hitting a decoy.

Nearly half of oral drugs on the market target enzymes, with kinases as the single most productive sub-family. About a third target cell-surface receptors, especially G-protein-coupled receptors. Ion channels, transporters, and nuclear hormone receptors round out the major protein classes. Nucleic acids form a smaller but fast-growing non-protein class. If you’ve read chapter 1 of Build AI Drug Discovery Pipelines, we derive a ~10⁵ protein “biological search space” covering on the order of 100,000 potential human protein targets once splice variants and post-translational modifications, that is, alternative versions of a protein and chemical changes added after it is made, are counted. A drug-like molecule has to find, bind, and modulate the right one.

Target discovery, the front-of-funnel activity this series focuses on, is the work of deciding which of those proteins to go after for a given disease. It splits into two phases that casual usage often folds together:

• Target identification: generating candidate target-disease links. “Here are ten proteins that seem to be involved in pancreatic fibrosis; let’s prioritize the top three.”
• Target validation: building the evidence that one of those candidates is actually causal. Does knocking it out in a disease model reverse the phenotype? Does a human genetic signal support it? Is there pharmacological precedent?

Both phases happen before a single compound is screened. Get either wrong and the downstream pipeline is solving the wrong problem.

Funnel Economics: The Cost of Being Wrong at Target Selection

The drug-development funnel runs roughly like this:

• Disease hypothesis and early target identification: years to decades, often academic.
• Target validation: up to ~2.5 years, averaging around $353M when done end-to-end including follow-on studies.
• Lead discovery and optimization: up to ~2 years, over half a billion dollars.
• Preclinical development: about a year, ~$340M.
• Clinical development: Phase I about 1.5 years, Phases II/III about 2.5 years combined, roughly half a billion dollars in total.
• Regulatory approval: ~1.5 years, ~$3M, the cheapest transition once you’re there.

Aggregate “cost to develop a new drug” figures vary widely because analysts treat time costs, failure allocation, and capitalization differently. The Tufts Center for the Study of Drug Development’s widely cited $2.8 billion mean per-approved-drug cost includes roughly $1.16 billion of foregone-investor returns over the development window plus $312 million in post-approval R&D. More recent cost estimates from RAND (2025), JAMA (Wouters et al., 2020), and Deloitte (2024) report different distributions, with medians closer to $0.7 billion to $1.0 billion and means weighted by high-cost outliers in the $0.95 billion to $2.23 billion range. Though reported headline figures depend heavily on which costs are included, the point is consistent. This is an expensive process, and being wrong costs a lot of cash!

That makes the target-selection decision unusually consequential. It spends comparatively less money, yet determines whether everything underneath it is wasted. Successful or unsuccessful completion of Phase 2 for an individual drug costs pharmaceutical companies around 20% of the sum spent on the drug discovery pipeline recalculated for each individual drug. Phase III is a larger absolute commitment, but Phase II is where we first get an answer to “is this target right?” and where many disappointments occur.

In a 2024 report, GlobalData attributes roughly half of clinical trial failures across phases to lack of efficacy and another ~30% to safety or toxicity. If efficacy collapses because the chosen target never had enough evidence behind it, then better target selection compounds downstream. Prevent one Phase II failure and the savings can exceed what you spent on target-assessment work by an order of magnitude. Catch the same mistake at target selection instead of in Phase II and the savings are larger still.

The 2014–2023 industry data also points to a second problem. Phase I likelihood of approval fell from 10.4% for the 2014 cohort to 6.7% for the most recent ten-year cohort, partly driven by an industry shift toward first-in-class, riskier targets where the evidence base is thinner. Many easier targets have already been drugged. What’s left is harder, and harder problems put more weight on front-of-funnel methods. Biomarker-stratified programs, which enroll patient subgroups defined by a measurable biomarker, run at roughly double the LOA of unstratified ones. This result follows from stronger target-selection evidence, as the biomarker identifies the patient subgroup in whom the target matters. For this reason, target and biomarker development often move together.

Why drug target discovery is hard

There is no single solution or approach as every target hunting strategy is context dependent. Useful target-assessment work has to satisfy at least five criteria simultaneously.

Target–disease linkage

Does the target play a causal role in the disease rather than merely correlate with it? Genetic evidence carries the most weight in most organizations. That includes GWAS hits, population-scale studies that link genomic loci to disease, Mendelian randomization, which uses inherited variants as a quasi-natural experiment, and rare-variant burden tests, which ask whether damaging variants in a gene accumulate more often in cases than controls. Multi-omics integration, which combines several molecular data layers such as RNA, protein, and chromatin measurements, acts as supporting evidence.

Target-related safety

Can humans tolerate modulation of this protein for years? Evolutionary conservation predicts essentiality and toxicity, wwhereas tissue specificity predicts off-target effects. Knockout phenotypes also provide a useful preview into safety implications if we were to neutralize the target.

Commercial strategy

Is there a path to reimbursement? Is this a first-in-class program, meaning a new mechanism, a best-in-class program, meaning a superior drug against an established mechanism, or a repurposing play, meaning an existing drug or target used in a new disease? These questions inform which target to prioritize even when several are scientifically equivalent.

Technical feasibility

Is the target druggable, meaning some therapeutic class can plausibly bind or modulate it? Can the organization develop the right modality, such as a small molecule, antibody, or oligonucleotide? A protein-protein interaction with no obvious pocket may be validated yet undruggable.

Data quality

Is the evidence base solid, or built on a few underpowered studies that never replicated? Manual curation, field-standard evaluation, and in some cases meta-analyses feed into this.

Taken together, a “good target” satisfies a multi-objective optimization problem across all five properties, involving a variety of data modalities. In practice, these criteria produce weighted votes rather than verdicts. Few targets get a clean yes across all five. More often, we see mixed signals and have to decide whether one target candidate’s profile beats the other eight candidate targets on our list. Target-discovery ML tries to improve this decision-making process.

Where ML helps in drug target discovery

The chemistry-side ML playbook of generating candidates, predicting properties, and optimizing molecules does not transfer cleanly to target discovery. There is not an explicity accuracy metric to optimize against, training data is fragmented across a dozen modalities, and ground truth, i.e., “was this the right target?”, takes a decade and often a Phase III trial to establish. Despite these limitations, ML does three things well here.

Scale

Datasets for target discovery are enormous and heterogeneous:

• GenBank contains over 300 billion base pairs of sequencing data.
• UniProt has more than 180 million protein sequences.
• The ENCODE multi-omics database exceeds one petabyte.
• PubMed has more than 35 million citations
• The USPTO database has hundreds of thousands of patents
• ClinicalTrials.gov has more than 438,000 clinical studies

No human team can read or integrate all of this. ML can convert it into searchable, queryable representations. Knowledge graphs built from biomedical text, which turn papers into linked entities and relationships, compressed numerical representations of multi-omics datasets, and network-proximity algorithms that rank genes by their topological distance from disease-associated nodes are all methods that scale to the size of the actual evidence base.

Cryptic patterns

A well-known example comes from SARS-CoV-2 repurposing, where we consider existing drugs that might work in a new or unaddressed disease. One study used network-diffusion and network-proximity algorithms on a combined dataset of the human interactome, the network of protein-protein interactions in cells, viral targets, and drug interaction data to rank 6,340 drugs for expected efficacy. The top-ranked drugs, tested experimentally, showed a 62% success rate in reducing viral infection, and 76 of the 77 drugs that worked did not bind to proteins directly targeted by the virus, suggesting network-based mechanisms rather than direct target engagement. That kind of signal, which requires multiple hops of complex reasoning, is hard to recover by manual review. In this case, graph algorithms operating across the interactome were able to surface it.

Automation

Literature review for a new target can take days or months. Named-entity recognition, which identifies mentions of genes, diseases, drugs, authors, and institutions in text, can be automated with usable quality (state-of-the-art NER reaches roughly 90% F-score). Relation extraction, which asks whether a paper asserts a specific link between those entities, is harder, around 50%. A knowledge graph that updates as new papers are published treats the evidence base as a living object, instead of freezing it at the last manual review.

These are front-of-funnel tasks that ML handles well. Final prioritization is a different matter*. In practice, target selection remains heavily human-curated. ML produces ranked candidate lists, aggregated evidence dashboards, and network-scored hypotheses, but the decision to allocate a program’s budget to target #3 rather than #7 still depends on organizational expertise, portfolio balance, competitive intelligence, and the intellectual-property landscape. Those factors are usually absent from training data. For now, target-discovery ML compresses the evidence into a form a human committee can reason about.

*If you’re coming from a background in recommender systems, this is similar to a cascading systems approach where we care about maximizing recall at the start (i.e., we don’t want to preemptively filter out a candidate and cut a potential multibillion dollar revenue stream) and, by the end, are most trying to maximize precision (i.e., avoid a costly false positive that damages our corporate brand, wallet, and sanity).

Worked Example: Rentosertib

Insilico Medicine’s rentosertib (also known as ISM001-055 / INS018_055) for idiopathic pulmonary fibrosis (IPF) serves as an instructional walkthrough (though I do not go so far as to categorize it as a gold-standard benchmark) of one end-to-end AI-driven target-and-molecule program.

Target identification began with multi-omics analysis of lung tissue from IPF patients versus healthy controls, combined with text mining across IPF literature and knowledge-graph reasoning over protein-protein interactions and pathway data. TNIK, a TRAF2- and NCK-interacting kinase, emerged as a novel regulator of fibrotic pathways. I.e., TNIK was not an established IPF target when the program started. The timeline is also notable. Project initiation to preclinical candidate nomination, the point where a program chooses the molecule it will advance into formal preclinical testing, took roughly 12–18 months, and the asset reached human clinical trials in under 30 months total, against an industry average of 4.5–6 years for early-stage discovery. Target validation ran in parallel with early chemistry; preclinical models confirmed that TNIK inhibition reduced fibrotic phenotypes, and generative-chemistry tools designed a series of inhibitors. The knowledge-graph workflow behind this will be the subject of a future article in this series.

Phase I established safety and pharmacokinetics, meaning how the body absorbs, distributes, and clears the drug. The Phase IIa trial (GENESIS-IPF, NCT05938920) enrolled 71 IPF patients across 21 sites in China over 12 weeks. The secondary endpoint, change in forced vital capacity (a standard measure of lung function) from baseline at week 12, showed +98.4 mL for the 60 mg daily dose versus −20.3 mL for placebo. Nature Medicine published the results in 2025, where they were positioned as the first clinical proof-of-concept for an end-to-end AI-discovered drug.

Keeping limitations in view, seventy-one patients over 12 weeks is a hypothesis-generating dataset, not a study sized or designed to support approval. No pivotal trial is running as of April 2026; Insilico is in regulatory discussions about a Phase IIb pivotal study, and a separate US Phase IIa (NCT05975983) is enrolling, with eight of the planned 60 patients having completed the 12-week treatment as of mid-2025. Keeping within the subject matter of this article series, TNIK had to be identified as a target before any chemistry could start. That identification came from a workflow combining multi-omics data, literature mining, and knowledge-graph reasoning. The Phase IIa result matters because it is consistent with the target call having been right. Whether the methodology generalizes is a question the next decade of readouts will answer.

A useful counterweight is Recursion’s REC-994, a lead pre-merger AI-discovered candidate for cerebral cavernous malformation, which was discontinued in May 2025 after long-term Phase II data failed to confirm earlier efficacy trends. High-profile AI-guided programs can fail in the clinic for the same reason other programs do. Through curating better evidence, we hope to lower the failure rate.

Who does drug target discovery

For readers whose mental model of drug discovery centers on Big Pharma, it’s motivating to keep in mind that, in the United States, nearly 60% of newly approved drugs were discovered in universities or biotechnology companies, not by Big Pharma’s own R&D. Small, often academic-adjacent biotechs take the early-stage innovation risk. Big Pharma licenses, acquires, or partners to bring the late-stage program through approval and marketing.

The reasons are structural. Big Pharma faces the “better than the Beatles” problem (the bar for new drug approval keeps rising because better drugs already exist), the “low-hanging fruit” problem (the easy targets are mostly drugged), the “cautious regulator” problem (FDA standards ratchet up after each safety scare and rarely relax), and a tendency to industrialize the wrong activities; scaling basic research and brute-force screening has not improved clinical success rates in aggregate. For more information, we discuss these problems and the related Eroom’s Law in detail within chapter 1 of “Build AI Drug Discovery Pipelines.”

For a practitioner in 2026, this means target-discovery tooling is disproportionately built inside smaller, often AI-native companies, licensed into Big Pharma programs, or run academically against public data. Big Pharma’s target-discovery groups increasingly act as evaluators and integrators rather than primary generators.

The rest of the drug target discovery series

The rest of this series goes deep on what this pillar surveys:

• Evidence frameworks — Driver vs passenger mutations, oncogene addiction (tumors becoming unusually dependent on one gene), multi-omics integration, GWAS, and practical target-assessment frameworks such as AstraZeneca’s 5R and GOT-IT.
• Druggability in the AlphaFold 3 era — Classical druggability, the expanding druggable-genome concept, therapeutic modalities such as antisense oligonucleotides (ASOs), PROTAC degraders, and antibody-drug conjugates (ADCs), and what AlphaFold 3 changed about structure-based drug design.
• Tissue specificity — Single-cell RNA-seq and reference atlases such as GTEx and the Human Protein Atlas, and how tissue-level gene-expression data feeds into safety prediction.
• Likelihood of approval by therapeutic area — The economics of repurposing, and why some disease areas are systematically more tractable.
• Novelty vs repurposing — When to invent a new target and when to reuse one. The Illuminating the Druggable Genome program, which focuses on understudied proteins, and research bias toward well-studied proteins.
• Synthetic lethality — CRISPR-based screening, synthetic lethality, where dual perturbation of two genes kills a cell even though either single perturbation does not, and ML methods for drug-synergy prediction, including tools such as MAGeCK, CRISPRi, and Perturb-seq.
• Knowledge graphs and the rentosertib case study — Biomedical named-entity recognition, relation extraction, knowledge-graph embedding models, and walkthroughs of publicly documented AI-discovered-drug programs.
• Virtual cells — Single-cell atlases, perturbation-response models, Perturb-seq benchmarks, and how simulated interventions can prioritize target-validation experiments.

How drug target discovery connects to Build AI Drug Discovery Pipelines

Build AI Drug Discovery Pipelines concentrates on methods that begin after a target has been chosen. Molecular property prediction, virtual screening, generative chemistry, protein structure prediction, drug repurposing, and multimodal pipelines are chapter-length topics because they are well-defined ML problems with benchmarks and data.

If you came here from the book and have already built a property predictor, screened a virtual library, or trained a generative model for lead compounds, this series asks a prior question: how did anyone decide that was the right target in the first place? As you read through this series, you might notice that the methods change with the problem. For example, the work may involve graph learning over biomedical knowledge graphs, natural-language processing over biomedical literature, or exploiting multi-omics integration with multimodal models. However, at their core, these methods are more like variants of the methods we discuss in the book, rather than new or alien concepts.

What’s next in drug target discovery: open questions

Near-term readouts to watch. The 15 or so AI-discovered programs expected to enter pivotal Phase III trials in 2026 will provide the first broad test of whether front-of-funnel AI improves late-stage success rates. Rentosertib is the clearest test of AI-led target identification, because the target-selection step itself was AI-derived. Schrödinger/Takeda’s zasocitinib (TYK2 inhibitor) is a corresponding test of physics-based AI design.

Methodological open questions. First, how do you benchmark target-discovery ML without waiting a decade for each label? Surrogate endpoints, or proxy benchmarks, such as asking whether a method can rediscover known targets for known diseases using only pre-discovery literature, help but do not close the loop and may suffer from data leakage, data snooping, and related problems. Second, how far does the current generation of target-discovery ML extend beyond well-studied indications? The “ignorome,” the large set of human proteins that remain thinly studied, biases the training data toward what has already been studied. A target-discovery method that mostly finds targets adjacent to known ones misses the hardest part of the problem. The ignorome is the subject of a future article on novelty vs repurposing.

Regulatory shift. The FDA’s January 2025 draft guidance on AI in drug development established a risk-based credibility framework for when and how AI contributions must be qualified for regulatory submissions. Final guidance is expected within 2026, which will influence how target-discovery AI can be cited in an IND package (the dossier submitted to request permission for human testing).

For the first time, the ML methods used in target discovery are starting to produce clinical readouts. However, the jury’s still out and the next decade will show how much they change outcomes – anyone telling you differently has a financial interest competing with transparency ;)

Further reading

• Citeline / Norstella (2024). Why are clinical development success rates falling? Industry report. Norstella/Citeline article. The primary source for the 6.7% Phase I LOA figure and the phase-by-phase transition rates cited throughout this post.

• Sun, D. et al. (2025). A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine. doi:10.1038/s41591-025-03743-2. The rentosertib Phase IIa paper that the “worked example” section walks through; the primary source for readers who want to see the trial design, endpoints, and safety data in full.

• FDA Center for Drug Evaluation and Research (2025). Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products. Draft guidance, January 2025. FDA draft guidance. The regulatory framework being written for how AI contributions - including target-discovery AI - must be qualified in IND and NDA submissions, that is, submissions to begin human testing or request approval.

• Wouters, O. J., McKee, M., Luyten, J. (2020). Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009–2018. JAMA, 323(9): 844–853. The JAMA evaluation behind the $985M median / $1.3B mean figures cited in the funnel-economics section.

• RAND Corporation (2025). Typical Cost of Developing a New Drug Is Skewed by Few High-Cost Outliers. RAND press release. The 2025 analysis that put the median direct R&D cost at $150M (rising to $708M when capitalized for failures), and showed how a small number of outliers skew the industry-wide mean.