How AI was used to discover rentosertib

Rentosertib is the drug most often cited when people say an AI-discovered medicine has reached late-stage trials. Insilico Medicine’s software did two things: it picked the protein to aim at, and it generated the chemical structure that aims at it. Both steps are genuinely unusual; neither tells you whether the drug works.

Step one: picking the target

Insilico’s target-finding software is called PandaOmics. In 2019 it was pointed at idiopathic pulmonary fibrosis (IPF).

It was given 15 public gene-expression datasets from the GEO repository covering IPF lung tissue, plus single-cell RNA sequencing data. Alongside that it ran natural-language processing across patents, publications, grant awards and trial records, which is how it estimated how crowded or novel a given target was.

The company then applied filters. It asked specifically for protein and receptor kinases, a class of enzyme medicinal chemists know how to drug. It filtered for novelty, for druggability, and for whether a crystal structure of the protein already existed, since structure-based design needs one.

Around 20 candidate targets came out. TNIK, a serine/threonine kinase, ranked first.

One more criterion decided it. Insilico looked for targets also implicated in ageing, and says TNIK touches six of the recognised hallmarks of ageing. That is why the company later ran the proteomic ageing analyses that generated headlines in September 2026; the ageing angle was in the selection criteria from the start, not discovered afterwards.

Sources differ on when the programme began: one account dates the antifibrotic target work to around September 2019, another to mid-2020.

Step two: designing the molecule

The chemistry software is called Chemistry42. It was asked to generate molecules that would fit the ATP-binding site of TNIK.

The design brief was a two-point pharmacophore: a hydrogen bond to the Cys108 backbone in the kinase hinge region, and a hydrophobic contact in the back cavity near the Met105 gatekeeper, lined by Leu73, Leu103, Ala52 and Val104. Insilico says it ran 30 generative models in parallel against that brief, each using a different algorithm.

Early compounds bound well but had poor ADME properties, meaning the body absorbed, distributed or cleared them badly. Ordinary medicinal chemistry then took over. Human chemists ran lead optimisation cycles until they arrived at INS018_055, later named rentosertib, covered by patent WO2022179528A1. Compounds were synthesised under contract by WuXi AppTec.

How many molecules were actually made

This is where the record becomes slippery. Three different figures circulate, all from Insilico:

• “fewer than 80 molecules synthesised”, the figure in most press releases;
• “78 molecules”;
• “the 55th of 79”, which is the only version that explains the “-055” in ISM001-055.

The peer-reviewed papers give no figure at all. Insilico’s own benchmark for a conventional programme is 60 to 200 molecules, so even at face value the saving sits at the modest end of the range the company itself quotes.

The timeline and cost claims

These figures are Insilico’s own claims, not independently audited numbers.

The verifiable dates around those claims are: preclinical candidate nominated in December 2020 and announced on 24 February 2021; a Phase 0 microdose study in Adelaide in November 2021; first Phase 1 dosing in Christchurch, New Zealand, in February 2022.

The US$430 million comparison is doing a lot of work. It covers a full preclinical programme including its failures, and it is not obvious that it is the right comparator for getting one molecule to candidate nomination. Insilico has not published a cost breakdown.

What is genuinely new, and what is not

What is new. In most AI drug discovery stories the software designed a molecule against a target human biologists had already chosen. With rentosertib the software chose the target too. TNIK had not previously been developed as a fibrosis target, and no antifibrotic medicine targets that kinase family. Derek Lowe, the chemist who writes Science’s In the Pipeline and who is broadly sceptical of the field, has noted that in almost every AI candidate “the targets were already known”. Rentosertib is the strongest counter-example available.

What is not new. Rentosertib is not the first AI-derived drug to reach Phase 3. Generate:Biomedicines began a Phase 3 trial of GB-0895, an AI-engineered anti-TSLP antibody for asthma, on 3 December 2025. That antibody was designed against a target humans had already validated, which is why Insilico frames its own claim as the first medicine with an AI-identified target and an AI-designed molecule to reach Phase 3. That is narrower, and more defensible, than “the first AI drug in Phase 3”.

“AI-discovered” is also not a regulatory term. No medicines agency recognises the category, and every count of AI-discovered drugs in trials comes from a private tracker using its own definition.

What the sceptics say

Derek Lowe wrote in December 2021 that ISM001-055 was “a believable target with believable chemical matter”, and “a step up from much of the other ‘AI found a drug!’ announcements, including some previous hype from Insilico Medicine itself”. On what follows he is blunt: Insilico “is going to be in here taking their shots the same way as the rest of us”. He also argues that AI helps at the stages that cost least, and that the field publishes its successes and not its failures.

Andreas Bender (professor at Khalifa University, previously at Cambridge) makes the technical criticism. Generative models are trained on chemical space humans have already explored, so they tend to return variations on what is known. He says “much of the in vivo translation is completely missing”: models predict binding well and predict what a compound does to a whole organism badly. Organ toxicity, including drug-induced liver injury, is among the worst-predicted outcomes, and liver findings are rentosertib’s main safety liability to date.

Charlotte Deane (professor at the University of Oxford), speaking to BBC News in January 2025, made the definitional point: there is “no definition yet of what exactly counts as an ‘AI discovered’ drug”.

MIT Technology Review reported on 21 August 2026 that the patents covering the molecule, US 11,795,160 and 11,739,078, name five human beings as inventors. No AI system is listed. Whatever the software contributed, the legal record describes a human invention.

The sector statistics. A Boston Consulting Group analysis published in Drug Discovery Today in 2024 found that AI-derived molecules had an 80% to 90% success rate in Phase 1, against a historic norm of about 66%, but roughly 40% in Phase 2, the same as the historic norm. Phase 1 mostly tests safety in healthy volunteers; Phase 2 is the first real test of whether a drug does anything.

Insilico’s own Nature Medicine paper concedes the point, stating that AI-discovered drugs “have experienced similar levels of phase 2 trial failure” and that “none has so far progressed through phase 3”.

Independent supportive voices

Marinka Zitnik (Harvard Medical School), who was not involved in the work, wrote an accompanying News and Views article in Nature Medicine in July 2025 describing the trial as marking “a turning point”. Her assessment concerns the discovery process rather than the clinical result; the trial she commented on enrolled 71 people, ran for 12 weeks and had safety, not efficacy, as its primary endpoint.

Timothy Cernak (University of Michigan), quoted in Chemical & Engineering News in March 2024, gave the most quoted mixed verdict: “I think Insilico’s been involved in hyping that, but I think they built something really robust here.” His point was that the company had taken a programme from target to clinic, “soup to nuts”, which no other AI drug discovery company had done.

Other AI-discovered medicines and what happened to them

Recursion acquired Exscientia for US$688 million in a deal that closed on 20 November 2024. Isomorphic Labs raised a US$2.1 billion Series B in May 2026 and has no clinical candidate. Across the sector, roughly US$60 billion has been invested since 2019 and about 175 programmes have entered the clinic. None has been approved.

Where rentosertib sits, September 2026

The AI contribution to rentosertib is better documented than most, and the target was genuinely new. The molecule then went through conventional medicinal chemistry, conventional toxicology and conventional trials, and it now sits at exactly the stage where the sector’s record is weakest. One 71-person Phase 2a trial, with safety as its primary endpoint, has been published. A 320-person Phase 3 trial began dosing on 10 September 2026 and is expected to finish in October 2029. Until it reports, what AI did or did not contribute cannot be judged by results.