1. “AI-discovered” is not a single scientific event
Drug discovery is a chain of decisions rather than one invention. Researchers must identify a disease mechanism, select a tractable target, find or design a molecule, optimise potency and selectivity, assess absorption and toxicity, manufacture the candidate, and then test it in people. Machine learning can support nearly every step, but success at one step does not guarantee success at the next. A model may predict that a compound binds a protein while missing metabolism, off-target effects, tissue exposure, or the fact that the selected target is not causally important in patients. Reviews of the field therefore distinguish computational performance from therapeutic performance (Bender & Cortés-Ciriano, 2021a, 2021b; Vamathevan et al., 2019).
The phrase “AI-designed drug” can describe very different contributions. In one programme, AI may rank known targets. In another, a generative model may propose a novel chemical structure. Elsewhere, machine learning may optimise trial recruitment or predict which patients are most likely to respond. These are all meaningful uses, but they carry different evidentiary burdens. A target-selection model should be judged by prospective biological validation; a molecular generator by synthesis, potency, selectivity, and developability; and a clinical model by performance on representative patients. Treating all of them as one category makes both praise and criticism scientifically imprecise.
The most useful question is therefore not whether AI “works” in drug discovery. It is where the technology improves a decision compared with a strong non-AI baseline. This framing matters for students because spectacular demonstrations can distract from ordinary sources of failure: poorly defined labels, narrow datasets, information leakage, and optimisation against endpoints that are convenient rather than clinically meaningful. Explainability does not automatically solve these problems, but interpretable evidence can help researchers identify whether a model has learned chemistry and biology or merely repeated correlations in its training data (Jiménez-Luna et al., 2020).
2. Where the evidence is already strong
AI is particularly effective when a problem has abundant, well-curated data and a measurable objective. Molecular property prediction, virtual screening, image analysis, literature mining, and protein-structure prediction fit this pattern better than open-ended questions about complex disease causality. AlphaFold 3, for example, extended structure prediction to biomolecular complexes containing proteins, nucleic acids, ligands, ions, and modified residues, creating richer starting hypotheses for medicinal chemistry and structural biology (Abramson et al., 2024). A predicted complex is not proof of binding, but it can guide mutagenesis, assay design, and prioritisation.
Generative chemistry provides another concrete example. In 2019, researchers reported a deep-learning workflow that identified potent inhibitors of discoidin domain receptor 1 within a compressed design-and-test cycle (Zhavoronkov et al., 2019). The result did not establish a medicine; it demonstrated that generative methods could propose synthesizable compounds with relevant activity. This distinction is important. Early discovery is dominated by search, and algorithms can reduce the number of compounds that must be made. The value is real even when the final molecule still requires extensive human-led optimisation.
AI is also becoming useful in less visible parts of development. Models can help identify trial sites, match participants to eligibility criteria, monitor safety signals, and standardise regulatory documents. These applications may lack the drama of de novo molecule generation, yet they address delays that directly affect development cost and duration. The U.S. Food and Drug Administration has accordingly emphasised a risk-based approach to AI and machine learning in drug and biological product development, with attention to context of use, data governance, model performance, and lifecycle management (U.S. Food and Drug Administration, 2023).
3. Human trials are the necessary reality check
The field entered a more demanding phase when AI-originated drug programmes reached clinical testing. The TNIK inhibitor rentosertib, developed for idiopathic pulmonary fibrosis, became a prominent case because AI contributed to both target identification and molecular design. Preclinical and early clinical work reported a coherent antifibrotic rationale, and a randomised phase 2a study provided evidence about tolerability and exploratory efficacy in patients (Ren et al., 2025; Xu et al., 2025). This is more informative than a retrospective benchmark because the programme had to survive synthesis, toxicology, manufacturing, regulatory review, and human dosing.
The correct interpretation is neither triumph nor dismissal. A phase 2a signal does not equal regulatory approval, and a single programme cannot establish that AI broadly increases success rates. Drug development is affected by indication, trial design, dose selection, patient heterogeneity, and chance. What the programme does show is that an AI-supported pipeline can produce a clinically testable candidate on a compressed timetable. The next question is whether multiple programmes demonstrate reproducible advantages in probability of success, not merely speed to nomination.
This is why prospective evidence matters more than headline counts of generated molecules. A model can create millions of structures, but the laboratory can synthesise and test only a small fraction. The bottleneck shifts from generation to selection. Strong platforms therefore integrate computation with repeated experimental feedback, sometimes described as a design-make-test-analyse loop. The algorithm proposes; chemistry and biology measure; the new data update the next proposal. AI becomes most valuable when embedded in this cycle rather than treated as an oracle.
4. The unresolved problems are biological and institutional
Training data in drug discovery are not neutral. Published datasets overrepresent successful experiments, well-studied protein families, and chemical series that companies are willing to disclose. Negative results are often missing. Assay conditions vary, labels can be noisy, and compounds that appear inactive may simply have been tested in unsuitable systems. A model trained on these records can reproduce historical attention rather than reveal new biology. Data diversity is therefore not an administrative detail; it determines which diseases, populations, and molecular mechanisms are visible to the system.
Generalisation is equally difficult. Random train-test splits can place closely related compounds in both sets and produce optimistic results. More realistic evaluations separate chemical scaffolds, time periods, laboratories, or therapeutic programmes. Prospective testing is stronger still because predictions are frozen before experiments are conducted. Bender and Cortés-Ciriano (2021a, 2021b) argue that unrealistic benchmarks and poorly matched baselines are major sources of inflated expectations. A useful model must outperform the decision process it is intended to replace, not merely a weak comparator.
Institutional incentives can also distort evaluation. Start-ups benefit from describing platforms as revolutionary, while pharmaceutical companies may publicise speed but withhold failures and cost data. Academic groups are rewarded for novel architectures more than for careful replication. As a result, the evidence base contains impressive case studies but limited portfolio-level comparison. Transparent reporting should include how targets were chosen, how many compounds were generated and synthesised, which candidates failed, what experiments changed the model, and whether timelines or attrition improved against historical controls.
5. A more scientific way to judge the AI-drug revolution
The first criterion should be decision quality. Did the model identify a target with stronger human genetic or mechanistic support? Did it reduce unnecessary synthesis while maintaining chemical diversity? Did it detect toxicity earlier? Did it improve patient selection without excluding underrepresented groups? These questions connect algorithmic outputs to the actual reasons drug programmes fail. They also make it possible to recognise partial value. An AI tool may be worthwhile even if it does not create the final drug, provided it reliably removes bad options sooner.
The second criterion should be reproducibility. Models should be evaluated on temporally separated data and, where possible, in prospective studies. Reporting must describe datasets, uncertainty, failure modes, and human interventions. The third criterion is clinical translation. Ultimately, medicines are judged by benefits and harms in patients. A faster route to a weak candidate is not progress; a faster route to a better-supported candidate is. Clinical outcomes will therefore remain the final audit of every claim about AI-driven discovery.
The emerging evidence supports a balanced conclusion. AI is already changing how scientists search chemical and biological space, and some programmes have progressed far enough to challenge the idea that the field is only hype. At the same time, no algorithm can remove the uncertainty of living systems. The most credible future is not autonomous drug discovery but a tighter scientific loop in which computational models prioritise experiments, experimental data correct the models, and human judgement defines the questions worth asking. That is less cinematic than a machine inventing a cure, but it is a more plausible route to better medicines.