How Insilico’s AI designed a drug that reached Phase 2a.
The rentosertib numbers.
Every figure below carries its named source and a grade: A = peer-reviewed/regulatory, B = top-tier press or primary company document, C = company self-reported (not independently audited).
AI found the target, designed the molecule, and the trial was published.
Insilico Medicine fed multiomics data from patient tissue and the published scientific literature into its target-discovery engine, PandaOmics, which ranked a novel target — TNIK — first for idiopathic pulmonary fibrosis (IPF). Its generative-chemistry engine, Chemistry42, then designed a small molecule to hit it. The resulting drug, rentosertib, reached first-in-human trials in under 30 months.
In a Phase 2a trial published in Nature Medicine on June 3, 2025, rentosertib met its primary safety endpoint, and an exploratory analysis showed the top dose (60mg daily) improving lung function by +98.4 mL versus a 20.3 mL decline on placebo over 12 weeks.
‘AI-designed’ is real — and not the whole story.
AI genuinely compressed the slowest, most expensive part of drug discovery: finding and validating a target. That is a real result, peer-reviewed in one of medicine’s top journals. But the headline ‘AI-designed drug’ hides what matters to an investor or a pharma board:
- This is a Phase 2a result in 71 patients — early and small.
- The primary endpoint was safety; the lung-function gain was an exploratory signal, not a confirmatory efficacy result.
- The trial ran at 21 sites in a single country and must replicate.
The diligence question is not ‘did AI design it’ but ‘what has actually been de-risked, and what bar remains.’ That distinction is the core of AI due diligence on an AI-driven asset.
How much should this change your priors?
Phase 2a is early, and many drugs with positive 2a signals fail in larger, confirmatory trials. The efficacy readout was exploratory, the sample small, and the sites concentrated in one country. None of that diminishes the genuine point — AI nominated a novel target and designed a viable molecule faster than the historical norm — but the clinical and regulatory bar for approval is unchanged. ‘AI-discovered’ speeds the start of the race, not the finish.
Insilico & rentosertib: common questions.
Did AI really design this drug?
Yes. Insilico’s generative-AI platform nominated the target (TNIK) using PandaOmics and designed the molecule using Chemistry42. The work is documented in a peer-reviewed Nature Medicine paper (June 2025). AI drove the discovery and design; humans ran the chemistry, preclinical work, and trials.
Did rentosertib work in the Phase 2a trial?
The trial met its primary endpoint, which was safety and tolerability. On efficacy, an exploratory analysis showed the top dose improving lung function (FVC) by +98.4 mL versus a 20.3 mL decline on placebo over 12 weeks. That is a promising signal, but exploratory — not a confirmatory proof of efficacy.
How fast was it compared with traditional drug discovery?
Insilico reports the program reached a preclinical candidate about 18 months after starting and first-in-human Phase 1 in under 30 months — materially faster than the multi-year norm for novel targets. The speed advantage is in discovery and design; clinical trials still run on their own timeline.
What does this mean for AI in drug discovery?
It is among the strongest peer-reviewed evidence to date that generative AI can find a novel target and design a molecule that survives into mid-stage trials. It does not shortcut clinical risk. For investors, the lesson is to separate a genuine discovery-speed advantage from the unchanged bar of Phase 3 and approval.