Artificial intelligence in drug discovery: from target to clinical signal
Machine learning now permeates every stage of pharmaceutical research, from target identification to molecular design and trial readiness. This review maps where the methods carry evidentiary weight, where they remain investigational, and why a governance model that separates what a model may learn from what it may decide is the precondition for moving any computational result toward a patient-facing study.
The friction problem AI is asked to solve
Computational methods are recruited against a pipeline defined by cost, duration, and failures concentrated at the end of development.
The economics of drug development frames the entire field. A new therapeutic agent typically demands more than a decade of work and a development cost measured in billions, and the great majority of candidates that enter clinical testing never reach approval. Most failures concentrate in two areas, insufficient efficacy and unanticipated toxicity, both reflecting incomplete biological understanding at the moment a molecule is selected.
Machine learning enters precisely here, as a family of methods for extracting structure from data too large or too high-dimensional for manual analysis. A comprehensive review in Nature Reviews Drug Discovery catalogued applications in target validation, prognostic biomarker discovery, and digital pathology, and was equally explicit about the limitations, namely the limited interpretability and reproducibility of model-generated results and the persistent scarcity of systematic high-dimensional data (Vamathevan et al., 2019). This tension between demonstrated capability and disciplined validation organizes the sections that follow, and connects directly to my research program in governed clinical artificial intelligence.
Where machine learning enters drug discovery
A step-by-step view clarifies which tasks are mature and which remain investigational.
It is more accurate to speak of many narrow applications than of a single general capability. Each stage of the pipeline imposes a distinct computational task, with its own data regime and its own standard of proof. The table below maps the principal entry points.
Maturity varies sharply across these rows. Structure prediction has become a reliable instrument, whereas clinical stratification remains an area in which most claims are still retrospective and await prospective confirmation.
| Pipeline stage | Computational task | Representative outcome |
|---|---|---|
| Target identification | Mining associations in genomic, transcriptomic, and clinical data | Prioritized, plausible target hypotheses |
| Structure determination | Prediction of protein and complex structure | Atomic models for proteins lacking an experimental structure |
| Hit generation | Generative chemistry and virtual screening | Novel candidate molecules with predicted activity |
| Lead optimization | Prediction of property, selectivity, and toxicity | Synthesizable, ranked analogs |
| Clinical development | Patient stratification and trial enrichment | Biomarker-defined subpopulations |
Structure prediction and generative molecular design
The most consequential advances repositioned structure and design from bottlenecks to routine inputs.
The clearest demonstration that deep learning can deliver atomic precision came when AlphaFold predicted three-dimensional protein structures from sequence alone, at a level competitive with experimental methods, closing a problem that had remained open for half a century (Jumper et al., 2021). The subsequent model extended prediction to the joint structure of proteins with small molecules, nucleic acids, and ions, precisely the interactions that matter for binding and selectivity (Abramson et al., 2024).
Generation followed prediction. By adapting diffusion models to the geometry of the protein backbone, RFdiffusion enabled the de novo design of structures and binders, several of them experimentally validated, including a binder whose cryo-electron microscopy structure matched the computational model almost exactly (Watson et al., 2023). Together, these methods shift the limiting factor from the availability of a structure to the quality of the biological hypothesis under study.
| Method | Year | Capability | Reference |
|---|---|---|---|
| AlphaFold | 2021 | Single-chain structure prediction with atomic precision from sequence | 10.1038/s41586-021-03819-2 |
| RFdiffusion | 2023 | De novo design of protein backbone and binders | 10.1038/s41586-023-06415-8 |
| AlphaFold 3 | 2024 | Joint structure of proteins with ligands, nucleic acids, and ions | 10.1038/s41586-024-07487-w |
From the in silico hit to the clinical signal
Two programs mark the path from a computational prediction to evidence in living systems.
The decisive question is whether computational predictions survive contact with biology. An early and instructive case was the identification of halicin, in which a neural network trained to predict antibacterial activity revealed a structurally unusual antibiotic in a repurposing library, subsequently active against resistant pathogens in murine models (Stokes et al., 2020). The result drew attention because the molecule was chemically distant from known antibiotics, a region that human intuition tends to overlook.
More consequential for human evidence is rentosertib, a TNIK inhibitor, the target itself nominated by generative methods, developed for idiopathic pulmonary fibrosis. In a randomized, blinded, placebo-controlled phase 2a trial (registration NCT05938920), the highest-dose arm recorded a mean change in forced vital capacity of plus 98.4 ml (95 percent confidence interval of 10.9 to 185.9) at 12 weeks, versus minus 20.3 ml for placebo, with a tolerability profile comparable across the arms (Xu, Ren et al., 2025). The cohort is small and the readout preliminary, but it is among the first controlled human results for a molecule-target pair originating in generative chemistry.
| Program | Computational modality | Evidence stage | Source |
|---|---|---|---|
| Halicin | Discriminative screening over chemical libraries | Preclinical, efficacy in murine models | 10.1016/j.cell.2020.01.021 |
| Rentosertib | Generative chemistry, target nominated by generation | Randomized phase 2a, forced vital capacity signal | 10.1038/s41591-025-03743-2 |
Governance, failure modes, and decision authority
The reliability of a computational result is inseparable from the controls placed around its use.
The recurring lesson is that the predictive performance reported in a paper does not automatically transfer to a sound decision within an ongoing program. Several failure modes are already well characterized. Data leakage inflates retrospective accuracy. Distribution shift degrades a model when the chemical or patient space at deployment differs from the training distribution. Automation bias leads teams to over-rely on a ranked list. And irreproducibility, the constraint emphasized in the Nature Reviews Drug Discovery review, undermines independent validation (Vamathevan et al., 2019).
A governance posture that addresses these points begins by separating learning authority from decision authority. A model may classify, predict, and propose. The decision to synthesize a compound, to administer a dose to a participant, or to advance a candidate remains with accountable human review, operating behind explicit stage gates, with traceable records at each transition. This separation is the organizing principle of the externally governed learning architecture under development in this research program, and it is what makes it possible to audit a computational pipeline rather than merely trust it.
None of this slows discovery. It makes the result defensible, which is the only form of speed that matters when a candidate approaches a human study.
| Failure mode | Mechanism | Governance control |
|---|---|---|
| Data leakage | Contamination between training and evaluation data | Temporal holdout validation and provenance logging |
| Distribution shift | Deployment space differs from the training distribution | Applicability domain checking and continuous monitoring |
| Automation bias | Uncritical trust in the model's rankings | Human decision gate with recorded justification |
| Irreproducibility | Unstable, undocumented, or unshared pipelines | Versioned artifacts and reproducible execution |
Why this matters for your care
This note documents the conceptual foundation of an active research line at the pre-seed stage. The governance model described here corresponds to United States provisional patent 63/975,551 (Externally Governed Learning Systems), and the generative discovery pipeline corresponds to United States provisional patent 64/018,624 (DrugSynthAI Discovery), a multi-agent architecture for de novo molecular design targeting genetic diseases, together with the DrugSynthAI trademark registered with the USPTO in Class 042. The broader record of publications is indexed under ORCID 0009-0001-9929-3135. This is a pre-seed research and development product. It is not a commercial offering nor clinical guidance.
References
- Vamathevan et al. Applications of machine learning in drug discovery and development. Nature Reviews Drug Discovery. 2019. doi:10.1038/s41573-019-0024-5
- Jumper et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021. doi:10.1038/s41586-021-03819-2
- Abramson et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024. doi:10.1038/s41586-024-07487-w
- Watson et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023. doi:10.1038/s41586-023-06415-8
- Stokes et al. A Deep Learning Approach to Antibiotic Discovery. Cell. 2020. doi:10.1016/j.cell.2020.01.021
- Xu, Ren et al. Rentosertib, a generative AI-discovered TNIK inhibitor, for idiopathic pulmonary fibrosis: a randomized, double-blind, placebo-controlled phase 2a trial. Nature Medicine. 2025. doi:10.1038/s41591-025-03743-2
Educational and scientific content. It does not constitute diagnosis, prescription or individual clinical guidance, and does not replace a medical consultation. Management decisions must be individualized by a physician.