The Target Identification Problem

The pharmaceutical industry’s most expensive problem is not chemistry. It is not clinical trial design. It is not even regulatory complexity. The most expensive problem is choosing the wrong target — selecting a protein or pathway as the focus of years of drug development effort, only to discover in Phase II or Phase III that modulating that target doesn’t produce meaningful benefit in patients.

Target selection failure is responsible for an estimated 40-50 percent of late-stage clinical failures, depending on which analysis you trust. (The estimates vary because companies have limited incentive to publish that their target selection was the problem — it’s easier to attribute failure to “challenging biology” or “patient heterogeneity.”) Each late-stage target failure represents years of work and hundreds of millions of dollars committed to a direction that the biology was never going to reward.

AI companies working in drug discovery have largely focused their narrative on molecule design — the step that comes after target selection. The algorithmic generation of novel chemical matter is technically impressive and visually compelling. Target identification is less photogenic. It is also, arguably, where AI could matter more.

Why Targets Fail

The most common reason a well-designed drug fails in clinical trials is not that the drug doesn’t hit its target — modern medicinal chemistry is quite good at making potent, selective binders. The most common reason is that hitting the target doesn’t produce the clinical effect that the biological hypothesis predicted.

This failure mode has several variants. The target may be genuinely important in the disease mechanism, but only in a subset of patients whose disease is driven by that mechanism, while the clinical trial enrolled a broader population where the mechanism is irrelevant for most of them. The target may play a compensatory role in healthy biology — blocking it disrupts normal function in ways that produce toxicity or that the body compensates for, eliminating the therapeutic effect. The animal model that appeared to validate the target may have been misleading, because the mouse disease model doesn’t recapitulate the human disease with sufficient fidelity. The target may be downstream of the actual driver, making it responsive to that driver but not causal in a way that intervention reverses.

Related reading

The PCSK9 story is the exception that proves the rule. PCSK9 inhibitors — Repatha (evolocumab) and Praluent (alirocumab) — succeeded because a large population genetic study identified PCSK9 loss-of-function variants in humans with dramatically lower LDL cholesterol and dramatically lower cardiovascular event rates. The genetics provided direct causal evidence in humans that PCSK9 activity causally drives LDL levels and cardiovascular risk. Inhibiting PCSK9 pharmacologically replicated the genetic loss-of-function, and the cardiovascular benefit followed. This is the ideal target selection story: human genetic validation of the causal mechanism before the drug program was ever launched.

Most targets don’t have that genetic validation. They are chosen based on mouse genetics, cell biology, or biomarker associations that have a higher rate of translational failure.

What AI Can Do for Target Selection

The most direct AI application to target selection is mining large genetic datasets to identify human genetic variants that are associated with disease risk and that implicate specific molecular mechanisms. Genome-wide association studies (GWAS) have identified thousands of loci associated with common diseases, but the translation from a statistical association to a specific drug target requires mechanistic interpretation that GWAS alone cannot provide.

AI tools can assist this interpretation by integrating GWAS results with gene expression data, protein interaction networks, tissue specificity information, and published literature to generate hypotheses about which gene at an associated locus is most likely to be functionally relevant to the disease, and through what mechanism. This type of multi-modal data integration — combining genomics, transcriptomics, proteomics, and phenotype data — is a task where ML models can handle more data and more complex interdependencies than manual analysis.

Open Targets (a public-private collaboration including Wellcome Sanger Institute, GSK, Pfizer, and several other partners) has built exactly this type of integrated platform and made it freely available. As of 2025, Open Targets contains associations for over 60,000 targets across hundreds of diseases, with ML-derived tractability assessments and evidence strength scores. This is genuine public infrastructure for target selection, and several drugs that successfully targeted GWAS-implicated genes — the BET inhibitor programs building on NME7 associations, the LRRK2 programs in Parkinson’s — used this type of genetic validation as foundational evidence.

The Causal Inference Problem

The fundamental limitation of all current AI approaches to target identification is that they excel at correlation and struggle with causation. A model trained on genetic associations, expression data, and disease phenotypes can identify factors that vary together with disease. It cannot reliably distinguish which of those factors are causally upstream of disease versus downstream consequences of it.

This distinction matters enormously in drug development. A protein whose expression is elevated in disease tissue may be elevated because it drives disease (good target) or because it is induced as a consequence of disease progression (irrelevant) or because it is part of a protective response to disease (potentially harmful target — blocking it might make things worse). Statistical associations, however sophisticated the model, cannot resolve this without experimental causal validation.

The gold standard for causal target validation in humans is Mendelian randomization — using genetic variants that naturally vary the activity of a gene to estimate the causal effect of modulating that gene’s activity on disease outcomes. Mendelian randomization effectively uses human genetics as a natural randomized experiment, and it has successfully validated or invalidated targets before expensive clinical programs were launched. The technique has been applied to PCSK9, LPA, IL-6, CRP, and dozens of other targets with consequential results.

AI can accelerate Mendelian randomization analyses by processing large-scale genetic and outcome datasets more efficiently, and by integrating MR results with other evidence streams. But the technique requires appropriate genetic instruments (variants that affect the target gene specifically and not through correlated mechanisms), adequate statistical power, and careful interpretation — requirements that machine learning can support but cannot substitute.

Related reading

The Neglected Target Problem

One place where AI genuinely expands the practical scope of target investigation is in exploring less-studied proteins. The “drugged genome” — the set of human proteins that have been the subject of serious drug development programs — is a small fraction of the roughly 20,000 proteins encoded by the human genome. Most drug programs focus on proteins with established pharmacology: kinases, GPCRs, ion channels, proteases. Proteins outside these well-studied families are underrepresented in drug discovery partly because they are less well understood, making it harder to design drug programs against them.

AI tools trained on structural data (AlphaFold structures for the entire proteome), interaction networks, and genetic associations can identify under-explored proteins that have strong genetic evidence linking them to disease and structural features suggesting druggability, even if they have no established pharmacology. This is a genuine expansion of what can be considered in target selection — not just faster analysis of the usual candidates, but identification of candidates that human scientists would not have prioritized from the published literature.

Genomics England’s rare disease program and several academic collaborations have used this approach to identify novel targets for rare diseases where the genetic cause was known but the therapeutic angle was unclear. Whether these targets survive to drug development and clinical proof-of-concept is the open question, but the identification stage — historically a bottleneck for rare disease drug development — has genuinely accelerated.

The Honest Assessment

AI has not solved target selection. It has provided better tools for hypothesis generation and prioritization, but it cannot substitute for the experimental biology required to validate that a target is causal, druggable, and safe to modulate in humans. The companies that treat AI target selection as equivalent to validated human genetics are taking more risk than their models can quantify.

The companies building genuine causal validation into their target selection — using human genetics, disease models, and mechanistic experiments to confirm the biology before committing to a full drug program — are using AI to accelerate and enrich the process rather than to shortcut it. The distinction matters not just for scientific rigor but for the probability of clinical success, which is still the metric that matters most for the patients waiting for treatments.

The target identification problem is hard because disease biology is hard. No model can make the biology less complicated. What AI can do, at its best, is help scientists ask better questions of the biology faster. That is valuable. It is also, on the evidence of 2026, not yet transformative at the level of clinical outcomes. The transformation, if it comes, will be evident in Phase III success rates three to five years from now.

The Animal Model Trap

There is an upstream failure mode in target identification that AI makes more visible without yet solving: the animal model trap. Most target validation happens in mouse models of disease, and the history of drug development is rich with examples where a target that appeared highly validated in mice proved irrelevant in humans. Alzheimer’s disease is the most famous case: dozens of amyloid-targeting drugs that worked beautifully in the APP/PS1 transgenic mouse models failed in human Phase III trials over a span of 20 years and billions of dollars. The mice had amyloid. They didn’t have Alzheimer’s disease.

AI models trained predominantly on mouse data will identify targets that are valid in mice. Whether those targets are valid in human disease is not knowable from the mouse data alone. The key input that AI target identification needs — and often lacks — is human-tissue, human-outcome data that provides ground truth about which molecular features of disease drive clinical progression in actual patients.

This is why the investments in human biological datasets — UK Biobank, GTEx (the Genotype-Tissue Expression project mapping gene expression across human tissues), the Human Cell Atlas — are so critical to making AI target identification more reliable. The more diverse and representative the human biological data that feeds into AI target identification models, the less the models will be led astray by the simplifications that animal models impose. Building those datasets takes years and tens of millions of dollars, and the returns are diffuse across many research programs. They are, in a real sense, public goods for the drug development enterprise.

Related reading

The Competitive Dynamics of Target Knowledge

Target identification in pharmaceutical research operates in a competitive environment that has unusual properties. When a company identifies a highly validated target for a common disease, the commercial pressure is to move rapidly into drug development while competitors are behind. This creates incentives to advance programs with incomplete target validation — to bet on a target that’s promising rather than waiting for validation that competitors might also access.

The result is an industry-wide tendency toward premature commitment to incompletely validated targets, particularly for common diseases where the commercial prize is large. AI target identification tools, which can generate large numbers of plausible target hypotheses quickly, may accelerate this tendency rather than correct it. More candidate targets, available faster, without proportionally more validation capacity, means more programs entering drug development based on incomplete evidence.

The correction would require either better validation tools (Mendelian randomization at scale, human organoid and organ-on-chip systems that better replicate human physiology, longer natural history datasets) or industry norms that reward patience in validation over speed of program entry. Neither is an AI problem. Both require scientific and organizational decisions that AI cannot make.

What AI target identification will ultimately prove itself on is not the number of targets it generates, but the clinical success rate of the programs that follow from those targets. That data will arrive gradually, as the current wave of AI-identified programs progresses through clinical development over the next decade. The field is running a distributed experiment with many arms, varying degrees of rigor, and no shared control condition. The results will be informative, but they will take time to read clearly.