AI designed a drug. Now comes the hard part.
Nov. 20, 2026 · The Science Bits
A computer program picked the biological target and drew the molecule for a drug now being tested in people with a fatal lung disease. In a small trial, patients on the highest dose did better on a breathing test. That is real, and it is the furthest this kind of drug has ever got. But the trial was built to test safety rather than whether the drug works, a third of the people on that dose stopped early, and no drug designed this way has been approved by the FDA. The larger trial that gives a real answer is not expected to finish until late 2029.
A computer helped design a pill for a disease that slowly turns people’s lungs to scar tissue.
In a small trial, patients taking the highest dose improved on a breathing test, while the comparison group got slightly worse. That is the hopeful part, and it is the part that made headlines.
Here is what to know before you celebrate. The trial’s main job was checking whether the drug was safe, not proving it worked. Six of the 18 people on that top dose stopped taking it before the study ended. And the drug, rentosertib, has only just moved into the larger trial that will actually decide whether it helps anyone.
It is the best test yet of a question buried under billions of dollars of headlines: how close is artificial intelligence to producing a medicine a real patient can take?
First, the disease
Idiopathic pulmonary fibrosis, or IPF, scars the lungs. As the scarring builds, the lungs stiffen and take in less air. People often become breathless doing ordinary things — crossing a room, getting dressed, carrying groceries. A dry cough that will not quit and deep fatigue are common. Some people decline slowly over years; others have sudden, dangerous episodes where their breathing gets much worse within days.
It mostly affects older adults, and estimates of how common it is vary depending on how cases are counted, landing somewhere in the range of a few dozen people per 100,000.
There is no cure. Two drugs are approved to slow it down: Ofev (nintedanib) and Esbriet (pirfenidone). Both are antifibrotics, meaning they slow how fast lung function declines. Neither reverses scarring that has already happened.10
Median survival after diagnosis is often given as three to five years.1011 You will still see two to four years quoted; that figure comes from data collected before the antifibrotic drugs existed. Either way, these are averages across large groups, and individual cases vary enormously.
That is the gap a new drug would be trying to fill.
What the trial found
Rentosertib comes from a company called Insilico Medicine. Its software flagged a protein called TNIK, which had not been pursued as a drug target in IPF before, and then generated the molecule aimed at it. That combination — AI picking the target and designing the molecule — is what makes it unusual.1
The results were published in the journal Nature Medicine in June 2025.1 (Peer-reviewed means other scientists checked the work before publication. It is a real filter, but it does not make a study independent: this one was funded by Insilico, with company scientists among its authors.)
Seventy-one patients at hospitals in China took either the drug or a placebo for 12 weeks, in four groups — three doses and placebo.2 A placebo is a dummy pill with no medicine in it, so researchers can tell whether the drug itself is doing anything.
On safety, which is what the trial was built to measure: medical problems during treatment were common in every group, including placebo — roughly 7 or 8 people in 10 across the board. The more telling figure is how many of those problems investigators judged to be related to the drug. On placebo, about 3 in 10. On the highest dose, closer to 8 in 10.2
On breathing, the number that made headlines: patients on the highest dose gained about 98 millilitres on average over 12 weeks, while the placebo group lost about 20.6 The measurement is forced vital capacity, or FVC — how much air you can push out of your lungs in one big breath. In a disease where that number normally falls, seeing it rise is worth paying attention to.
Now the caveats, and there are several.
It was not the question the trial was designed to answer. Lung function was a planned secondary measurement.7 The primary endpoint is the main outcome a trial is built around, and the number of patients is chosen with that outcome in mind. A 71-person safety study does not include enough people to reliably separate a real treatment effect from ordinary variation between patients. So a good secondary result is a lead worth chasing, not a conclusion.
The 98 is not the drug’s advantage over placebo. It is how much that one group changed compared with where it started. The researchers also reported a 95% confidence interval of about 11 to 186 millilitres.6 That is a measure of uncertainty around the group’s average — not the spread among individual patients, and not a prediction for any one person. Subtracting the two groups’ averages gives about 119 millilitres, but that is simple arithmetic, not an analysed comparison the trial was built to make.
And a better FVC number is not the same as feeling better. It does not automatically mean someone walks farther, coughs less, needs less oxygen, or has less scarring. Those are separate questions, and this trial did not settle them.
The headline came from 18 people, six of whom stopped early.
The safety picture is not clean
Sixteen of the 71 patients stopped treatment before the end. For twelve of them, a medical problem was the reason.34
The most common was liver injury or liver dysfunction, in seven people across the treatment groups. Four of those seven were also taking nintedanib, one of the existing IPF drugs, and the authors say more work is needed to untangle which drug was doing what.5 What the published summary does not tell us is how severe those cases were — whether they were abnormal blood tests or diagnosed injuries, whether anyone was hospitalised, whether they resolved after stopping. Those distinctions matter, and I could not establish them. If you already take several prescriptions, that is a question for a doctor rather than one to assume an answer to.
There was also one death, from heart failure, which investigators judged unrelated to the treatment.
And one more signal that got almost no coverage: three patients in the highest-dose group had an acute exacerbation of their IPF and were hospitalised, compared with one in the placebo group.1 An acute exacerbation is a sudden, dangerous worsening of the disease. Three out of 18 against one out of 17 is far too small to show the drug caused anything. But in a trial this size it is the kind of imbalance a lung specialist notices, and it belongs in any honest summary.
Did the company and the paper tell the same story?
Mostly — with one place worth understanding, because it is a useful lesson in reading this kind of news.
The paper reports that the quality-of-life and physical-ability measures — a cough questionnaire and a six-minute walking test — were broadly similar across the groups. It also reports that when researchers adjusted the cough scores statistically, the highest-dose group came out better than placebo, by a margin that just squeaked past the usual threshold for statistical significance.1
Insilico’s own website highlights that cough improvement.8 The paper contains it. So this is not a company inventing a result.
The fair criticism is narrower and more useful: it is a borderline result, on a secondary measure, in a 71-person study that looked at a number of different outcomes. The more things you measure, the better the odds that one of them crosses the line by chance. It needs confirming in a bigger trial before it means much.
Which is the general lesson. A company’s summary leads with its best number. The paper usually contains that number and the context around it. Reading both is how you find the difference.
Who was in the study
The average participant was about 67. But 64 of the 71 were men, all were Asian, and all were treated at hospitals in one country.1 That is a narrow group. It does not make the result wrong, but it limits how confidently anyone can apply it to, say, a 67-year-old American woman, or someone managing several other conditions at once.
What happens next
Insilico announced the Phase 3 trial in July 2026 and dosed the first patient in September.1213 It plans to enrol about 320 patients at 47 centres in China, each treated for a year, with the rate of lung-function decline as the main question this time. It is registered as NCT07687459, and the registry estimates completion in October 2029.14
Three hundred and twenty people is a modest Phase 3, incidentally. “Phase 3” does not automatically mean thousands of patients.
Phase 3 is where researchers try to confirm that a drug’s benefits outweigh its risks in a larger group, usually over one to four years. Trials at this stage still fail, and some produce results nobody can interpret cleanly.
And finishing is not approval. After a trial ends, researchers analyse the data, the company submits it to a regulator, and the regulator decides whether the evidence justifies the risks. Because this trial is running in China, success there would not automatically produce an FDA approval in the United States; the FDA would need to evaluate an application of its own. The lead investigator, Professor Zuojun Xu, has said approval could come three to four years after Phase 3 started, under favourable conditions13 — but that is an investigator’s estimate, not a schedule.
The drug is not approved anywhere today.
What does “AI designed a drug” even mean?
Loosely, in practice. The phrase can describe picking the target, generating the molecule, predicting how it binds, optimising something that already exists, or helping design the trial. Those are very different contributions — and an AI-generated design is still just a proposal until a laboratory makes the chemical and tests it.
Takeda’s psoriasis pill zasocitinib shows the problem. It is often called an AI-designed drug. Nimbus Therapeutics developed it and sold it to Takeda in 2022, using machine learning alongside physics-based computer modelling to evaluate thousands of compounds against a protein scientists already knew about. Nimbus’s head of research told the trade publication PharmaVoice he would not call the result an AI-designed drug.17
It matters because zasocitinib is further along than anything else in this conversation. The FDA accepted it for priority review in September, with a decision expected in the first quarter of 2027.1516 Priority review is a faster review, not an approval.
For this article I am using a strict definition: AI picked the target and designed the molecule. Rentosertib is the clearest publicly documented example that has reached final-stage testing. I cannot tell you it is the only one in the world, because nobody maintains a registry sorting drug candidates by how much AI went into them.
So is AI actually better at this?
A 2024 analysis by Jayatunga and colleagues, published in Drug Discovery Today and summarised by Boston Consulting Group, looked at the clinical pipelines of AI-focused biotech companies. In Phase 1 — the first round of human testing, which examines safety, tolerability, dosing and how the body handles a drug — their molecules passed at around 80 to 90 percent, well above historical industry rates.18
In Phase 2, where you start finding out whether a drug helps, the rate was about 40 percent. That is roughly what the industry has always managed, and only around ten molecules had finished Phase 2 when the analysis was done.
Be careful what you take from the Phase 1 number. It does not prove AI makes safer molecules. These companies choose which candidates to put into people in the first place, the sample was small, the companies are not a cross-section of the industry, and nobody randomly assigned candidates to AI and non-AI discovery. The authors themselves flag the limits.
The failures belong in the record too. Exscientia announced the first AI-designed drug to enter human testing in 2020 — DSP-1181, for obsessive-compulsive disorder, developed with Sumitomo Dainippon Pharma — and that programme was later dropped; a separate cancer candidate was wound down after it. BenevolentAI’s eczema drug BEN-2293 met its main safety goal in mid-stage testing but missed its efficacy goals.19 After Recursion and Exscientia merged, Recursion dropped three clinical programmes in May 2025, paused a fourth and wound down a preclinical one, and its shares fell sharply the day it announced the cuts.20
Meanwhile the money keeps moving. Eli Lilly agreed in March 2026 to pay Insilico $115 million up front, with milestone payments and royalties that could bring the headline figure to $2.75 billion21 — money largely owed later, and only if things go well. Isomorphic Labs, spun out of Google DeepMind, raised $2.1 billion in May and another $5 billion on November 17, three days before I wrote this.22 Its own announcement says that money is to move its first candidates toward clinical trials; it does not have one in human testing yet.23 And Anthropic, the company behind the Claude AI I used to research this story, was reported to have bought a drug-discovery startup for around $400 million in stock.24
One number you will see quoted: FDA officials said this year that the agency’s drug and biologics centres have received more than a thousand submissions containing AI components.25 That is real, and it means AI turning up anywhere in a filing — lab work, trial design, manufacturing, safety monitoring. It does not mean a thousand AI-invented drugs are queued up for approval.
How to read the next AI-medicine headline
- Which phase is it? Phase 1 is first-in-human testing of safety, dosing and how the body handles the drug. Phase 2 looks for early signs of benefit. Phase 3 is where researchers try to confirm it. A press release about early results is not a breakthrough.
- What was the trial built to measure? If the exciting number was a secondary measurement, as the lung-function result was here, it is a lead rather than a finding.
- How many people, for how long, and who were they? If everyone in the study was one sex, one country or one age group, the result may not transfer to you.
- Read the paper, not just the announcement. A company leads with its best number. The paper usually has that number and the context — how borderline it was, how many other things were measured, what the authors themselves warn about.
If you or someone you love has IPF
Rentosertib is experimental. It is not available, and nothing here is a reason to change anything you are taking.
If your options feel limited, clinical trials are worth raising with your own specialist — not this trial specifically, which is enrolling in China, but trials generally. Eligibility depends on your own situation, and there are real risks as well as potential benefits. The federal registry at clinicaltrials.gov lists what is recruiting, and your doctor can help you judge whether any of it fits.
This is not medical advice. I am a tech writer, not a doctor.
What would count as a real breakthrough? Not another molecule on a screen, and not another funding round. A larger, longer, more varied trial showing that patients actually do better, at a level of harm worth accepting. For rentosertib, that answer is due around 2029.
My take
The day is coming when AI designs better drugs, faster than people can. Today is not that day.
What is here is a real first. Software picked a target nobody had gone after for this disease, drew a molecule to block it, and that molecule is now in a 320-person trial with real patients. That is not nothing. But it is one drug, in one trial, in one country, with a safety picture that still needs explaining, and the honest answer to “does it work” is that nobody knows yet.
There are a lot of obstacles left. The hard part was never the chemistry. It is finding out whether a drug helps sick people without hurting them, and that still takes years and a lot of patients no matter what designed the pill. No amount of computing power speeds that part up.
But at the rate this is moving, I do not think it will be long. And the only way to get there is to travel that road. Somebody has to run the first trials, publish the first disappointing numbers, and find out what the software got wrong. That is what this is.
Ask me again in 2029.
How this was made: AI helped research and draft this story. Claude, made by Anthropic, did the research and the first draft. Four other AI systems — Perplexity, Grok, Gemini and ChatGPT — checked it against the sources. They caught real errors, and they also contradicted each other more than once; where they did, I went back to the original documents. One correction is worth describing, because it is the same mistake this article warns about. An earlier draft claimed Insilico’s website contradicted its own published study on the cough results. It does not — the paper reports that number too, with a lot more hedging around it. I was pleased enough with the catch that I nearly printed it. Being too eager to find a company spinning is its own kind of bias, and it nearly cost me. Disclosures: Anthropic has its own drug-discovery business, mentioned above. The trial described here was funded by Insilico Medicine, whose scientists are among the paper’s authors. I could not open the full Nature Medicine paper directly; the figures here come from the published abstract, the journal’s article page, and independent coverage that agrees with them. This story is not medical advice.
Watch the 44-second version: AI-made drugs: miracle or marketing hype? (YouTube Short)
References & credits
- Xu Z., Ren F., Wang P. et al., "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial," Nature Medicine 31:2602–2610 (June 3, 2025).
- PubMed, abstract of the phase 2a trial.
- Europe PMC, full-text record (discontinuations by group).
- Healio, "Q&A: AI-established drug breakthrough in the treatment of IPF" (July 7, 2025).
- Drug Discovery & Development, "Insilico’s AI-designed rentosertib shows promise in first phase 2a trial results" (June 5, 2025).
- American Journal of Managed Care, "AI-Derived Therapy for IPF Shows Potential in Phase 2a Trial."
- Genetic Engineering & Biotechnology News, coverage of the trial design and endpoints.
- Insilico Medicine, company case study.
- Peking Union Medical College Hospital, statement on the trial.
- Merck Manual (professional), "Idiopathic Pulmonary Fibrosis."
- American Journal of Managed Care, "The IPF Journey."
- Insilico Medicine, Phase III initiation announcement (July 7, 2026).
- News-Medical, "Rentosertib enters Phase III trial" (Sept. 10, 2026).
- Clinical Trial Vanguard, registry details and completion estimate.
- Takeda, "FDA Accepts NDA Under Priority Review for Zasocitinib" (Sept. 14, 2026).
- Dermatology Times, "FDA Grants Priority Review to Zasocitinib."
- PharmaVoice, interview with Nimbus R&D president Peter Tummino.
- Jayatunga M.K.P. et al., Drug Discovery Today (2024); summary by Boston Consulting Group.
- Endpoints News, "After years of hype, the first AI-designed drugs fall short in the clinic."
- Fierce Biotech, "Recursion reworks pipeline after Exscientia merger" (May 2025).
- Pharmaceutical Executive, "Eli Lilly and Insilico Enter $2.75 Billion Agreement" (March 30, 2026).
- MobiHealthNews, "Isomorphic Labs raises $5B for AI drug discovery" (Nov. 17, 2026).
- Reuters via Investing.com, "Isomorphic Labs delays clinical trial timeline" (Jan. 20, 2026).
- Fierce Healthcare, "Anthropic acquires Coefficient Bio in $400M deal."
- U.S. Food and Drug Administration, remarks by Mark Abdoo (May 6, 2026).