AI in Drug Discovery: The First Clinical Evidence and an Honest Reckoning for 2025-2026

Artificial intelligence (AI) in drug discovery has long been trapped between two extremes: a promised revolution that would solve disease within months, and the counter-claim that it is little more than a marketing bubble. The years 2025 and 2026 have, at last, added hard data to this debate. For the first time, a drug whose target and molecule were both discovered end-to-end by generative AI produced an efficacy signal in a randomized, placebo-controlled clinical trial. In the same period, independent studies clarified the genuine limitations of the field's most celebrated tools. This article aims to offer a reckoning stripped of marketing language and traceable to its sources: what, precisely, has AI achieved in drug discovery, and what has it not yet achieved?

The first clinical evidence: the rentosertib Phase 2a trial

The field's turning point is the rentosertib (ISM001-055) study published in Nature Medicine on 3 June 2025. What distinguishes this molecule, developed by Insilico Medicine, is that both its drug target (the TNIK kinase) and the molecule itself were discovered entirely through generative AI platforms (PandaOmics, Chemistry42). Conducted in patients with idiopathic pulmonary fibrosis (IPF), this randomized, double-blind, placebo-controlled Phase 2a trial constitutes the first randomized evidence that an AI-discovered drug can show efficacy in humans.

The study followed 71 patients across 21 centers, all in China, over 12 weeks. Patients were randomized into four arms: placebo (n=17), 30 mg once daily (n=18), 30 mg twice daily (n=18), and 60 mg once daily (n=18). The primary endpoint was safety, and treatment-emergent adverse event rates were similar across arms (70.6% on placebo; 83.3% at the highest dose). The most common reason for discontinuation was liver toxicity, and most patients who discontinued were taking the standard antifibrotic nintedanib concomitantly.

Efficacy was measured as a secondary endpoint via forced vital capacity (FVC). In the 60 mg once-daily arm, FVC increased by +98.4 mL (95% CI 10.9–185.9), whereas the placebo arm declined by −20.3 mL. The percentage change (+2.82%) met the accepted threshold for clinical meaningfulness in IPF (2–6%). The most striking signal came from patients not on standard antifibrotic therapy: in this subgroup, FVC rose by +187.8 mL (95% CI 68.6–306.9) on the 60 mg dose. No significant change was seen in those taking it alongside standard of care, pointing to a possible drug-drug interaction signal.

Rentosertib's journey continues. Insilico has announced plans to begin a Phase 3 trial of the oral formulation in the second half of 2026. A Phase 2a trial has begun enrolling in the United States, and an inhaled formulation has received investigational approval (IND) from China's regulatory authority.

Strong in Phase 1, average in Phase 2

Looking beyond a single trial to the field as a whole, the most reliable data come from an analysis by Jayatunga and colleagues published in Drug Discovery Today in 2024. This study was the first to systematically examine the clinical success rates of AI-discovered molecules, and it revealed a striking picture.

These figures capture the central truth of the field. AI is genuinely superior at making a molecule "drug-like," safe, and pharmacokinetically sound, which is why AI-discovered drugs clear Phase 1 at notably high rates. But Phase 2 success remains on par with traditional drugs. In other words, AI is good at predicting whether a molecule will be safe; it does not yet make a difference in predicting whether it will work in humans. Moreover, this Phase 2 figure still rests on a small sample. A 2025 follow-up commentary in Nature Biotechnology confirms the same picture.

As for sheer growth: the number of AI-derived programs in clinical development reportedly rose from around three in 2016 to roughly 67 in 2023, and to more than 170 by 2026. Because this figure comes from industry tracking reports rather than a peer-reviewed source, it should be treated with caution. Even so, the curve is a tangible indicator of the interest and investment flowing into the field.

AlphaFold 3: a powerful tool, but not a "solved" problem

The most visible AI success in structural biology is AlphaFold 3, introduced in Nature in 2024 by teams from DeepMind and Isomorphic Labs. Unlike its predecessor, which predicted protein structure alone, AlphaFold 3 can predict proteins together with DNA, RNA, ligands, ions, and post-translational modifications in a single model. For protein-ligand interactions, it proved roughly 50% more accurate than traditional docking methods on the PoseBusters benchmark. Its code and model weights were released for academic use in November 2024.

Yet independent evidence that balances this picture also accumulated in 2025, and reading the two side by side is essential to avoid overstatement. The most important critique concerns memorization. A preprint published in 2025 showed that these methods exhibited an accuracy gap of more than 10% between common ligands seen over 100 times in the Protein Data Bank and rarer ligands. Performance drops sharply on novel complexes that do not resemble the training data; even when the binding pocket is modeled correctly, the ligand's pose can be wrong.

The details complete this image: AF3's success rate on a realistic, drug-like ligand set sits well below what was reported on PoseBusters (around 40%); on orphan proteins it surpassed the previous version in only 57.5% of cases; and it remained weak on large ligand-induced conformational changes (>5 Å) and on dynamic molecules such as RNA and DNA. Because the test set uses a time-based split but no ligand-similarity filter, it is difficult to measure true performance on genuinely novel targets. In short, AlphaFold 3 is an extraordinary tool, but presenting protein-ligand prediction as a "solved problem" would be misleading.

The regulatory framework takes shape

The years 2025-2026 also brought the first concrete steps on the regulatory side. On 6 January 2025, the U.S. Food and Drug Administration (FDA) published draft guidance on the use of AI to support regulatory decision-making for drug and biological products. The document proposes a seven-step "context-of-use" credibility framework and calls for transparency in model architecture and training data. One important detail: the guidance excludes early discovery from its scope, targeting only AI that directly informs a regulatory decision. The final guidance is expected in the second quarter of 2026.

Another first came on 8 December 2025: the FDA formally qualified its first AI tool for use in clinical trials. Known as AIM-NASH, this tool is used in the histological assessment of MASH (metabolic dysfunction-associated steatohepatitis). A critical distinction applies here: this is a measurement and assessment tool, not a drug-design tool. It should therefore not be confused with a headline announcing "the first AI drug approved."

Consolidation, failures, and an honest picture

A realistic portrait of the field must include not only its successes but also its failures and corporate movements. Two pioneers, Recursion and Exscientia, merged in November 2024 in a stock swap worth roughly $688 million; the combined company operates under the Recursion name. What was once held up as "the first AI-designed clinical molecule," DSP-1181 (2020), had long since been a discontinued program.

Several candidates were shelved in 2025. Recursion's REC-994, a candidate for cerebral cavernous malformation, was halted in May 2025 when long-term data failed to confirm an early trend; other programs such as REC-2282 and REC-3964 were also discontinued. Meanwhile, although Google's AI company Isomorphic Labs has signed deals worth roughly $3 billion with Novartis and Eli Lilly, as of June 2026 it has not a single clinical trial underway; it is "preparing" for its first human studies. It would therefore be inaccurate to say this company has "produced a drug."

What it has and hasn't achieved

On the proven side: AI is accelerating target discovery and candidate nomination. In the rentosertib case, the path from target to IND candidate was completed in roughly 18 months; the gains in time and cost at the preclinical stage appear real. AI-discovered molecules clear Phase 1 at high rates, meaning they are strong in safety and pharmacokinetic optimization. And, for the first time, an AI-discovered target-molecule pair produced an efficacy signal in Phase 2a. AlphaFold 3, for its part, is measurably superior to traditional docking in static protein-ligand interactions.

On the unproven side, the picture is more cautious: it has not yet been shown that AI improves a drug's clinical efficacy; Phase 2 success sits at the industry average. Not a single AI-discovered drug has completed Phase 3 or received approval (as of June 2026). The systemic limitations are also serious: models perform well within the distribution of their training data but weaken in novel chemical-biological space (domain shift and overfitting); independent, multi-institutional, prospective external validation remains weak; and the risk of over-trusting AI nominations and reducing experimental verification (automation bias) is a genuine danger.

Conclusion

As of 2025-2026, AI in drug discovery has moved from the "promise" stage to the "first evidence" stage, but it remains far from "maturity." The rentosertib Phase 2a results delivered the first randomized efficacy signal the field had awaited for years, and that is no trivial milestone. At the same time, the memorization limits of AlphaFold 3 and the industry's average Phase 2 success rate provide a sturdy counterweight to the hype. The honest assessment is this: AI has proven that it adds real value in accelerating the discovery of a drug and making it safe; but it has not yet confirmed its promise of predicting that a drug will work in humans through large, replicated, independent studies. The right posture in watching this field is neither dismissal nor celebration, but the evaluation of each new claim against its source, its level of evidence, and its sample size. The next two to three years, as Phase 3 results and independent validations arrive, will bring this picture into sharper focus.