An AI-Discovered Drug Reached Phase 2: What Rentosertib Proved — and Didn't

Artificial intelligence has moved beyond proposing molecules on a screen. Rentosertib, also known as ISM001-055, reached a randomised phase 2a trial in people with idiopathic pulmonary fibrosis (IPF). The peer-reviewed result is important: a drug program whose target and molecule were developed with an AI-led platform produced a measurable human signal.

It is not a cure, an approval or proof that algorithms have automated medicine. The same study shows why every AI-generated hypothesis must still pass through conventional clinical science.

What the trial did

IPF is a progressive disease in which scar tissue reduces the lungs' ability to work. The phase 2a study enrolled 71 participants in China and randomly assigned them to placebo or one of three rentosertib dose groups for 12 weeks. It was double-blind, meaning participants and investigators did not know which treatment each person received.

The main aim was safety and tolerability. Researchers also measured forced vital capacity (FVC), the amount of air a person can exhale after a full breath and a key measure in IPF trials.

At the highest tested dose, mean FVC increased by 98.4 millilitres from baseline, compared with a 20.3-millilitre decrease in the placebo group. That separation is encouraging. It is also an early signal from a small group over a short period—not a demonstration that the drug prevents disability, hospitalisation or death.

The safety result deserves equal attention

Sixteen of the 71 participants discontinued the study. Treatment-related adverse events became more frequent as the dose increased, and the paper discusses diarrhoea and liver-related findings among the safety concerns.

This does not automatically make the drug unsafe. Early trials are designed to expose the balance between potential activity and tolerability. It does mean a headline celebrating the FVC number while hiding discontinuations gives an incomplete picture.

Larger and longer trials must determine whether the signal holds, which dose offers the best balance and whether uncommon or delayed adverse effects appear. The study was conducted in one country, so future research also needs a broader population.

What AI contributed

Insilico Medicine's platform linked disease biology to TNIK, a target implicated in fibrosis, and helped design the small molecule that became rentosertib. Generative models can search a chemical space much larger than a human team can enumerate, prioritise candidates and iterate designs before synthesis.

That can make the front end of discovery faster and more systematic. But the label “AI-discovered” covers only part of the journey. Chemists still make the compound. Laboratories test its behaviour. Clinicians design trials. Participants accept risk. Regulators inspect data, manufacturing and the evidence for benefit.

AI can help choose what to test; it cannot declare its own output safe and effective.

Why the FDA is building AI rules

The U.S. Food and Drug Administration says it received more than 500 drug and biologic submissions containing AI components between 2016 and 2023. The technology may support discovery, trial design, manufacturing and safety monitoring—not just molecule generation.

Its current principles emphasise a defined context of use, risk-based validation, data governance, human oversight and monitoring across the model's lifecycle. A model that ranks laboratory candidates does not need the same evidence as one used to select patients or support a regulatory endpoint.

Reproducibility matters too. Developers must document training data, relevant limitations and how a model behaves when the population or process changes. “The AI predicted it” is not an audit trail.

The verdict

Rentosertib crossed a meaningful line: an AI-led discovery program produced a drug with an encouraging, peer-reviewed phase 2a signal in people. The evidence remains preliminary because the study was small, short and not designed to prove long-term clinical benefit.

The next milestone is not a more impressive demo. It is a larger controlled trial that replicates efficacy and clarifies safety. That is the recurring theme of medical innovation: AI can accelerate the search, but patients benefit only when disciplined experiments confirm the result. See our guide to autonomous surgery for the same distinction between a research breakthrough and a clinically available system.