An AI-designed drug just reached Phase 3
An AI-designed drug just reached Phase 3
It treats a lung disease, not diabetes. It is still the clearest look yet at what artificial intelligence has changed about how new drugs get made — and what it hasn’t.
On September 9, 2026, the first patient was dosed in a Phase 3 trial of a drug called rentosertib. The trial, GENESIS-IPF-3, is planned to enroll 320 people with idiopathic pulmonary fibrosis across 47 centers in China, dosed once daily for 52 weeks, with the annual rate of lung-function decline as its primary endpoint. What makes it unusual is where the drug came from. AI systems were used both to identify the target it hits and to generate and optimize the molecule itself.
Insilico Medicine used one AI platform to rank drug targets and land on TNIK, a protein that had not been taken into clinical testing for lung fibrosis, then used a generative chemistry platform to design a compound against it. Humans still selected, synthesized, tested and optimized the molecules that followed. When the company published its Phase 2a results in Nature Medicine in June 2025, it noted that no AI-discovered drug had yet made it through a Phase 3 trial. This program is now in one.
There is no diabetes anywhere in this story. It is worth following anyway. “AI is going to speed up drug discovery” has been said to people with chronic conditions for years, and this is among the first programs far enough along to check the claim against something real.
What the Phase 2a trial showed
The Phase 2a trial, published in Nature Medicine in June 2025, randomized 71 adults with idiopathic pulmonary fibrosis to one of three rentosertib doses or placebo for 12 weeks. Insilico Medicine, which developed the drug, sponsored the trial, and company employees were among the researchers who designed the study and analyzed and interpreted the data. That is worth knowing before reading any of the numbers.
The primary endpoint was safety, not benefit. Lung function was a secondary measure: the 60 mg once-daily group gained a mean of 98.4 mL of forced vital capacity over 12 weeks while the placebo group lost 20.3 mL. The uncertainty was wide: the 95% confidence interval around the high-dose group’s mean change runs from 10.9 to 185.9 mL. That is a within-group estimate, not a confidence interval for the treatment effect against placebo, and this small safety-focused trial was never designed to establish efficacy. Sixteen of the 71 patients stopped treatment before week 12. Liver injury or dysfunction accounted for seven of those sixteen, and four of those seven were also taking nintedanib, an existing antifibrotic. The authors list the limits themselves: small arms, all participants in China, 12 weeks. One early trial does not settle whether AI changed anything. The pattern across the field is more informative.
The argument underneath all of this
The real question is not whether the software works. It plainly produces molecules, and those molecules keep clearing Phase 1. The question is which part of drug development was the bottleneck to begin with.
Two readings of the same evidence, both held by serious people:
THEORY 1 The slow part was chemistry
Finding and optimizing a molecule used to take years of synthesis and screening. Insilico reports reaching a preclinical candidate in 18 months. If chemistry was the real constraint, compressing it pulls in every timeline downstream, and the Phase 1 rates across AI-native companies are the first sign of it.
THEORY 2 The slow part was biology
Picking the right target has always been the hard part, and no platform has yet published evidence that it does that better than people do. The Phase 2 rates reported so far are in line with industry norms, and the 2026 Nature Reviews Drug Discovery Perspective finds no established clinical impact yet. On this reading, AI has sped up a step that was never the reason drugs fail.
Nobody can settle that yet, and the more careful reviewers say so directly. The number of AI-discovered drugs that have reached mid-stage trials is small enough that a couple of readouts would move the percentages noticeably. The 2026 Nature Reviews Drug Discovery Perspective calls this an absence of evidence rather than evidence of absence — not proof that AI fails to help, just no demonstration yet that it does, where it counts. Rentosertib’s Phase 3 runs 52 weeks before anyone sees a result.
Software has gotten much faster at proposing molecules. Nothing yet shows it is better at knowing which biology matters — and that is the step that usually decides whether a drug helps anyone.
An editorial read of the evidenceWhat this means if you live with type 1
Practically, nothing changes today. Rentosertib is being tested in a lung disease, is not approved anywhere, and has no application to diabetes. The useful part is calibration. When a headline says AI designed a drug, what has been demonstrated so far is a faster front end — target to candidate to first-in-human — not a faster or surer path through the trials that determine whether it helps.
That distinction lands harder in type 1 than in most conditions. Speeding up molecule discovery addresses one part of the problem, and the Diabetologia review lists the others: long and expensive trials, difficulty predicting who will progress, responses that vary widely between people, imperfect disease models, and unsettled endpoints. The same review notes that teplizumab, the first therapy approved for delaying type 1 diabetes, took about 30 years of testing to reach approval, including a Phase 2 trial that ran close to a decade. Cutting discovery from years to months does not shorten a trial that has to run long enough to see whether beta cells hold. It is reasonable to find this encouraging and still expect the same waiting. And if a product, test, or supplement is being sold to you on the strength of being AI-designed, that phrase describes how it was made, not what it has been shown to do. Talk with your doctor before making any changes to your care.
AI-discovered drugs in the clinic. Kp Jayatunga M, Ayers M, Bruens L, Jayanth D, Meier C. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons Drug Discovery Today, 2024. Authors are consultants at Boston Consulting Group. Retrieved via PubMed.
Discovery and Phase 1. Ren F, et al. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models Nature Biotechnology, published online March 2024. Insilico Medicine authors.
Aging-clock analysis. Zhavoronkov A, Galkin F, et al. Integration of proteomic aging clocks in a phase 2a clinical trial supports simultaneous geroprotective assessment Nature Biotechnology, September 2026. Insilico Medicine authors; exploratory analysis of 42 consenting trial participants.
Type 1 diabetes. Retrieved via PubMed. Shapiro MR, Tallon EM, Brown ME, Posgai AL, Clements MA, Brusko TM. Leveraging artificial intelligence and machine learning to accelerate discovery of disease-modifying therapies in type 1 diabetes Diabetologia, 2025;68(3):477-494 (published online December 2024).
Phase 3 first dosing. Insilico Medicine, Insilico Medicine doses first patient in GENESIS-IPF-3, company announcement, September 9, 2026 (NCT07687459 / CTR20262475). Trial size, center count and duration are as stated by the company.
The wider field, 2026. Bender A, Thomas MC, Scannell JW, et al. Artificial intelligence in drug discovery — what it is, where we stand and the path forward Nature Reviews Drug Discovery, August 2026. Peer-reviewed Perspective; academic and industry authors.
How to read this evidence. The Phase 2a trial is randomized and placebo-controlled, the strongest design here, but it was powered for safety rather than benefit, ran 12 weeks, and enrolled 71 people at sites in one country. The pipeline analysis is a survey of company pipelines rather than a controlled comparison, and its authors flag the small sample behind the Phase 2 figure; the 2026 Nature Reviews Drug Discovery Perspective is a peer-reviewed critique of that same evidence base. The aging-clock analysis is exploratory; its authors state plainly that they cannot separate aging effects from the drug’s anti-fibrotic activity within a fibrosis cohort. Rentosertib is an investigational drug. It is not FDA approved, not approved for any use in the United States, and has no indication in diabetes.
This post is for educational purposes only and is not medical advice. Nothing here is a recommendation to start, stop, or change any medication. Talk with your doctor before making any changes to your care.
An AI-designed drug reached Phase 3
The first patient was dosed in a Phase 3 trial of rentosertib on September 9, 2026. AI systems were used to pick its target and to design the molecule. It is being tested in a lung disease, not diabetes, and it is not approved anywhere.
What the research shows
- A randomized trial of 71 patients over 12 weeks, published in Nature Medicine in 2025, was designed to test safety. Side-effect rates were similar across the drug and placebo groups.
- Lung function was a secondary measure in that trial: the highest-dose group gained an average 98.4 mL while placebo lost 20.3 mL. The interval around that gain (10.9 to 185.9 mL) is a within-group estimate, not a comparison against placebo.
- A 2024 review in Drug Discovery Today found AI-discovered molecules clear Phase 1 at 80-90%, above industry norms, but succeed in Phase 2 at about 40%, the same as everyone else.
- Insilico Medicine, which developed the drug, sponsored the trial, and company employees were among those who designed and analyzed it.
What it does not show
- Anything about diabetes. Rentosertib is investigational, is not FDA approved, and has no diabetes indication.
- That AI picks better targets. Phase 2 is where drugs usually fail, and a 2026 Nature Reviews Drug Discovery Perspective finds no established clinical impact from AI yet.
- Long-term safety. In the trial, 16 of the 71 patients stopped early; liver injury or dysfunction accounted for seven of them.
If you live with type 1
- What AI has sped up is the front end: choosing a target and designing a molecule. The company reports 18 months from starting target discovery to naming a candidate.
- Faster discovery fixes one bottleneck. A 2025 Diabetologia review lists the others: long trials, hard-to-predict progression, varied responses, imperfect models, unsettled endpoints. Teplizumab took about 30 years to reach approval.
- No AI-designed molecule is in type 1 trials. That same review reported none of the repurposing candidates it surfaced were in a registered type 1 diabetes trial at the time of writing.
- “AI-designed” describes how something was made, not what it has been shown to do.
Share this
Listen to the Juicebox Podcast
Conversations about diabetes, five days a week, since 2015.
Read the full disclaimer
