For a patient with idiopathic pulmonary fibrosis, breathing gets harder every month. Their lung tissue gradually turns to thick, inflexible scar tissue, cutting off oxygen transfer until their lungs stop functioning entirely. FDA-approved treatments can slow the scarring down, but none can stop or reverse it.
On June 3, 2025, Nature Medicine published peer-reviewed results for a drug candidate called rentosertib. Over 12 weeks, patients receiving a 60 mg daily dose showed a mean increase in forced lung capacity of nearly 100 milliliters. Patients on placebo lost 20 milliliters over the same period.
The clinical data alone was notable. What made pharmaceutical executives stop and read the paper was how the drug was made: every stage of its early development—from identifying the biological target to generating the candidate molecule—was done by generative artificial intelligence algorithms in 18 months.
How did rentosertib succeed where competing AI-designed molecules from BenevolentAI, Exscientia, and Recursion collapsed? What do its Phase IIa results actually prove about generative chemistry? And why is the real story of healthcare AI unfolding far away from drug discovery labs?
The Headline Read — and Why It Overpromises
The surface reading of this milestone is seductive: AI has solved drug discovery. A computer identified a target, generated a molecule in a fraction of traditional time, and passed clinical testing. The old multi-billion-dollar, ten-year pharmaceutical development cycle is obsolete.
That headline version is tempting. It is also dangerously premature.
Passing a Phase IIa trial does not mean a drug works, nor does it mean it will ever reach a pharmacy shelf. It means the molecule did not poison its small test cohort and showed enough preliminary promise to justify a much larger, much riskier trial.
Inside the 18-Month Molecule
To understand what Insilico Medicine accomplished with rentosertib — originally designated ISM001-055 — you have to look at the traditional timeline of drug development.
Normally, finding a biological target and designing a molecule to bind to it requires four to five years of trial-and-error chemistry. Insilico compressed that window into 18 months using two specialized AI platforms.
First, its target discovery software, PandaOmics, analyzed biological datasets to identify an enzyme involved in tissue scarring — what scientists call Traf2- and NCK-interacting kinase, or TNIK. When TNIK signals out of control, it drives the cellular cascades that turn healthy lung tissue into scar tissue.
Second, Insilico fed that target structure into Chemistry42, a generative chemistry platform. Instead of searching existing chemical libraries, Chemistry42 generated entirely new molecular structures from scratch, optimizing them to fit into the TNIK enzyme like a key in a lock.
Insilico then ran a 12-week Phase IIa trial across 22 sites in China, enrolling 71 patients divided into placebo and three oral dosage groups. The results published in Nature Medicine met the trial’s primary safety endpoints. More importantly, the secondary endpoints showed a clear dose-dependent response: patients on the 60 mg once-daily dose gained +98.4 mL in forced vital capacity, while placebo patients lost −20.3 mL.
The Valley of Death: Why Most AI Drug Candidates Failed
Insilico’s trial result stands out primarily because almost every other high-profile AI-designed drug candidate has stumbled at this exact hurdle.
Computational chemistry platforms excel at generating molecules that bind tightly to target proteins in a test tube. But clinical trials do not happen in test tubes — they happen inside human bodies, where complex metabolic pathways, organ toxicities, and off-target interactions take over.
The history of early AI-designed candidates illustrates this clinical reality:
- BenevolentAI: Its lead candidate BEN-2293, designed for atopic dermatitis, met safety endpoints in Phase IIa in April 2023 but showed no significant improvement in itch or inflammation, failing its primary efficacy target entirely.
- Exscientia: The firm discontinued EXS-21546 in October 2023 after clinical modeling revealed it could not achieve a viable therapeutic window without causing adverse effects in patients with solid tumors.
- Recursion Pharmaceuticals: Its candidate REC-994 for cerebral cavernous malformation passed Phase II safety in September 2024. But long-term extension data showed the initial positive trends faded, forcing Recursion to wind down the trial in May 2025.
Statistically, between 50% and 60% of all small-molecule drug candidates that complete Phase II fail in Phase III due to unpredicted toxicity or lack of efficacy in larger patient populations. Overall, only 10% to 12% of molecules entering clinical development ever receive regulatory approval.
Passing Phase IIa proves that generative chemistry can design a molecule that survives early safety testing. Proving therapeutic efficacy requires surviving Phase III.
The Healthcare AI Split: Clinical Ambition vs. Administrative Scale
The hurdles facing rentosertib highlight a broader divide across healthcare AI: a stark split between high-risk clinical models and the unglamorous administrative tools achieving actual commercial scale.
On the high-risk side sit frontier models like Google’s Med-PaLM 2, which achieved 86.5% accuracy on USMLE-style medical exam questions. Yet despite expert-level benchmark scores, Google’s documentation explicitly notes that Med-PaLM 2 is not FDA-approved for clinical decision-making. The regulatory liability of a single hallucinated treatment recommendation keeps clinical AI restricted to research assistance.
On the low-risk side sit administrative tools like Microsoft’s DAX Copilot, an ambient listening tool that records doctor-patient visits and automatically drafts clinical notes into electronic health records like Epic.
Because DAX Copilot carries zero clinical decision liability — a human physician reviews every note before signing — it bypassed regulatory bottlenecks entirely. Microsoft announced in October 2024 that DAX Copilot was deployed across more than 400 healthcare organizations. While drug discovery models struggle through multi-year clinical trials, administrative AI has quietly invaded hundreds of hospitals.
Between these two extremes sits DeepMind’s AlphaFold 2. An independent study by the Innovation Growth Lab found that researchers using AlphaFold 2 submitted 45% to 49% more novel, experimentally verified protein structures to the Protein Data Bank than traditional labs. AlphaFold 2 did not replace physical biology; it gave human structural biologists better hypotheses to test in the real world.
What to Watch Next
Insilico Medicine initiated a Phase III clinical trial in July 2026, enrolling 320 patients across China to test rentosertib over a longer treatment duration.
That trial is the real test. If rentosertib maintains its lung function improvements in a 320-patient cohort without emerging toxicities, it will become the first fully AI-designed candidate to demonstrate definitive efficacy in a late-stage trial.
Until then, the pharmaceutical industry remains anchored in a basic truth: machine learning algorithms can compress early discovery timelines from four years down to 18 months, but they cannot compress the time it takes for a human body to respond to a drug.
Frequently Asked Questions
Is rentosertib the first AI-designed drug?
Rentosertib (ISM001-055) is the first drug candidate where both the biological target and the novel molecule were identified and designed by generative AI to successfully complete a Phase IIa trial showing positive safety and preliminary efficacy.
What is Idiopathic Pulmonary Fibrosis (IPF)?
Idiopathic pulmonary fibrosis is a chronic, progressive lung disease where tissue becomes thick and scarred over time, reducing lung capacity and oxygen transfer. It currently has no cure and a poor long-term prognosis.
Does passing Phase IIa mean rentosertib will be approved by the FDA?
No. Phase IIa trials test safety, dosing, and preliminary efficacy in small patient groups (71 patients in this trial). Approximately 50% to 60% of candidates that pass Phase II still fail in larger Phase III trials.
Is Google’s Med-PaLM 2 approved for treating patients?
No. Med-PaLM 2 is not an FDA-cleared medical device. It cannot provide clinical diagnoses or treatment decisions directly to patients.
Generative chemistry can design a molecule in 18 months, but it cannot algorithmically predict how human biology reacts over three years. Rentosertib’s Phase III trial will test whether human biology agrees with the computational model.
Sources
- Insilico Medicine GENESIS-IPF Trial: Published in Nature Medicine (June 3, 2025). ClinicalTrials.gov Identifier: NCT05938920
- BCC Research Report: AI Impact on Emerging Drugs Market (July 14, 2026).
- Innovation Growth Lab (IGL) Study: AI in Science: Evidence of Impact from AlphaFold 2 (Nesta).
- Microsoft DAX Copilot Deployment: Microsoft FY25 Q1 Earnings Call transcript (October 30, 2024).
- Google Research: Towards Expert-Level Medical Question Answering with Large Language Models (Singhal et al., May 2023).
About the Author
Ether Exter is an AI enthusiast with 5 years of experience testing and experimenting with AI models, breaking down what actually works. Follow on X: @EtherExperiment.
