AI Drug Discovery Moves Into the Clinic: What One Phase 2a Trial Shows, and Which Numbers Still Need Checking

"AI pharma" is most often flattened into one sentence by mixing company announcements, conference posters and the fact that a molecule was "found by AI" into the same claim. The checkable approach is narrower: look only at what a journal paper or a trial registry supports, then read it inside a specific indication, phase and endpoint.

A randomised trial you can actually check#

Insilico Medicine's rentosertib (formerly ISM001-055) is a case that can be traced line by line. The paper describes it as a small-molecule TNIK inhibitor designed with generative AI, targeting TNIK — itself a target discovered with generative AI — in idiopathic pulmonary fibrosis (IPF). The results were published in Nature Medicine (paper, online 2025-06-03, August 2025 issue).

The design was a multicentre, double-blind, randomised, placebo-controlled phase 2a trial with 12 weeks of treatment and four arms: 30 mg once daily (18 patients), 30 mg twice daily (18), 60 mg once daily (18) and placebo (17). The registration number is NCT05938920, where the sponsor is listed as InSilico Medicine Hong Kong Limited, the phase as Phase 2 and the status as completed.

What the paper says, and what it does not#

The primary endpoint was safety: the share of patients with at least one treatment-emergent adverse event was similar across arms — 72.2% (13/18), 83.3% (15/18), 83.3% (15/18), against 70.6% (12/17) on placebo. The paper's wording is that treatment-related serious adverse event rates were low and comparable across groups, with the most common events leading to discontinuation related to liver toxicity or diarrhea.

Efficacy was a secondary endpoint, and it is where retellings drift most: mean forced vital capacity (FVC) change was +98.4 ml (95% CI 10.9–185.9) in the 60 mg once-daily group versus −20.3 ml (95% CI −116.1–75.6) on placebo. The paper's own conclusion is that the results suggest rentosertib is safe and well tolerated and warrants further investigation in larger-scale, longer-duration trials — not that efficacy is proven, and not a phase 3 result.

Where to look before citing a pipeline#

The paper opens by stating that despite substantial progress in generative chemistry, few novel AI-discovered or AI-designed drugs have reached human clinical trials. So three things are worth holding onto when retelling it: the indication (IPF only, in this paper), the phase (2a), and the endpoint type (safety primary, FVC secondary). Figures such as "18 months versus 5–7 years," "48% objective response" or "timelines cut 60% by 2030" still lack a first-hand source and should not be written as industry fact. To cite a specific asset, go back to the trial registry and the journal paper; this site does not vouch for "AI-discovered" claims in company announcements.

(Revised October 12, 2026: the original listed Insilico, Recursion, Exscientia and XtalPi as all having public pipelines; first-hand sources could not be checked for each, so the article now covers only the case with a paper and a registry entry, and the unsourced trend claim that discovery loops are shortening was removed.)