The attraction of applying machine learning to drug target discovery is real: the evidence base is large, the patterns are complex, and the stakes are high. But the specific ways ML systems fail in this domain are worth understanding in detail, both to use them more carefully and to avoid treating their outputs with a false confidence that can lead to expensive mistakes.
Distribution Shift: The Core Problem
Most ML models trained for target discovery are trained on historical data: targets that have been extensively studied, evidence from databases built over decades, clinical outcomes for drugs against previously validated targets. This training data represents the past; what you need to predict is the future -- whether a novel target, which may have been minimally studied, is genuinely disease-relevant.
Distribution shift occurs when the test data (new targets being assessed) is systematically different from the training data in ways that affect the model's reliability. In target discovery, there are several predictable forms of distribution shift. Novel targets in understudied biological domains have sparser evidence than the training distribution, which means the model is making predictions in a region of feature space it has not learned well. Targets in rare diseases face GWAS power limitations that were not present for the common diseases overrepresented in training data. New disease areas may involve biological mechanisms that are genuinely different from the mechanisms represented in the training set.
A model that performs well on targets similar to its training set may perform much worse on the novel targets you most want to find -- which are precisely the ones that are different from existing knowledge.
Label Leakage and Survivorship Bias
Evaluating ML models for target discovery requires carefully constructed training and test sets, and this is often done inadequately in published work. The most common problem is label leakage: information about the target's eventual outcome in clinical development is indirectly accessible through evidence features that are themselves a consequence of the target's historical trajectory.
A target that proceeded to Phase II trials has typically accumulated substantial evidence in public databases, including expression studies, functional characterization, and mechanism-of-action papers that were generated as part of the clinical program development. Using this evidence to predict the target's clinical outcome creates an artificially high accuracy estimate: the model has learned to identify targets with rich evidence bases, which is partly a proxy for "targets that have been extensively developed." A model evaluated this way will perform worse on new targets where the evidence base reflects genuine uncertainty rather than the post-hoc documentation of a development program.
Negative Training Data is Systematically Missing
Training a supervised model requires both positive examples (valid targets) and negative examples (invalid targets). In drug discovery, negative examples are severely underrepresented in public data. When a compound fails in a clinical trial, the failure data is often proprietary, not published, or attributed to factors other than target invalidity. The databases used to train target discovery models are therefore dramatically biased toward targets that were selected and advanced, not toward a representative sample of all evaluated targets.
This means models trained on available data have learned to distinguish highly-studied targets from less-studied targets, not valid targets from invalid ones. These are correlated but distinct properties, and the conflation matters when you want to use the model to find valid targets among candidates that are not yet extensively studied.
Interpretability and the Black Box Problem
When an ML model ranks a target highly, the ranking is only actionable if you understand why the model assigned that ranking. A ranking driven by strong genetic evidence and consistent functional data in a relevant disease model is a very different scientific claim from a ranking driven primarily by literature co-occurrence statistics and protein interaction network centrality. The composite score is the same; the evidentiary basis is fundamentally different.
Models that produce scores without decomposing their evidence contributions make it difficult to distinguish these cases. A research team that moves a target into wet-lab work based on an opaque high score, without understanding that the score is primarily driven by a piece of evidence that is less relevant to their specific disease indication, has made a resource commitment on a misunderstood basis.
What Responsible Use Looks Like
None of this argues against using ML in target discovery; it argues for using it with appropriate epistemic caution. Specifically: treat model outputs as evidence-synthesis tools rather than predictive oracles, demand decomposed evidence contributions rather than accepting composite scores alone, prefer models with explicit uncertainty quantification that flags predictions with sparse training support, and maintain a skeptical prior for predictions in biological domains that are underrepresented in the training data.
The most productive use of ML in target discovery is as a system that integrates more evidence sources more consistently than any human analyst could, with explicit uncertainty reporting and transparent evidence attribution, feeding into domain-expert review of the top candidates. That workflow combines the scale advantages of automated evidence integration with the judgment advantages of experienced researchers. The failure mode is treating the ML output as an endpoint rather than as a structured input to expert deliberation.
The Importance of Human Override
One structural safeguard against the failure modes described above is maintaining a genuine human expert review step as a non-negotiable part of the target selection workflow, not as a rubber-stamp of the ML output. This means reviewers who are empowered to push back on high-scoring targets when their domain knowledge identifies a specific reason for skepticism, and who have enough context about the model's methodology to understand which aspects of a high ranking are most and least reliable.
The most common failure is treating the review step as a formality: experts are shown the ranked list, asked if they agree with the top candidates, and typically confirm them because they are the same candidates that would have emerged from informal literature review. That review process does not add value over the ML ranking. A review that is designed to actively challenge the ranking -- asking specifically about the evidence basis for the top candidates, identifying which evidence sources are highest quality, and probing for reasons the ranking might be wrong -- extracts the value that domain expertise can add to computational prioritization. The combination of systematic evidence integration and skeptical domain review is substantially more reliable than either approach alone.