What made rentosertib an "AI drug"

Back in July I wrote about why AI-designed drugs ace phase 1 and then stall. The argument was that AI had gotten good at designing molecules, but picking the right biological target was still a human bet, and that bet is what phase 2 punishes. Rentosertib is the most interesting test of that argument so far, because Insilico says AI picked the target too. So it's worth reading exactly what the software did.

Step one: the target

The discovery paper, published in Nature Biotechnology (Ren et al., online March 2024), describes running Insilico's PandaOmics platform over gene-expression datasets from fibrotic tissue. The platform combines several scoring methods (network analysis, causal inference, expression changes, text mining of the literature and more) and outputs a ranked list of genes.

Here's the detail that matters. Before the ranking came out, the team chose which filters to switch on: protein class (kinases only), novelty, and small-molecule druggability. Under that "protein and receptor kinase" setting, a kinase called TNIK came out number 1 among the top five candidates. The paper says TNIK "has not been studied as a therapeutic target in IPF and was thus selected by the AI algorithm."

So it's fair to say the AI nominated TNIK. It's also fair to say people framed the question: they told it to look for druggable, novel kinases, which is a very sensible human prior about what a drug company can actually go after. The model searched a space that scientists defined. That's not a knock. It's what using a tool well looks like, and it's the part headlines tend to drop.

Step two: the molecule

For the chemistry, the team used Insilico's Chemistry42 platform, which the paper says "first applied 30 generative models in parallel to generate compound structures." But the inputs were again human choices. The team picked the ATP-binding pocket of TNIK's kinase domain as the place to bind, a well-established strategy for kinases. They set the generator to produce structures that form a hydrogen bond with a specific hinge residue (Cys108) and fill a pocket near the gatekeeper residue.

Then came ordinary medicinal chemistry. The paper describes compounds "selected based on synthetic accessibility, novelty and medicinal chemistry properties" being synthesized and assayed. The first potent hits had problems: they cleared too fast in liver microsome tests, inhibited CYP enzymes and dissolved poorly. A lead-optimization round focused on those properties produced the compound now called rentosertib. The paper puts the timeline from target discovery to preclinical candidate at roughly 18 months.

That's genuinely fast, and generating good starting chemistry against a fresh target is a real contribution. But the honest description is AI-assisted drug discovery with humans steering at every gate, not a model that handed over a finished drug.

Step three: the phase 2a trial

The clinical paper (Xu et al., Nature Medicine, June 2025) reports a double-blind, placebo-controlled trial of 71 patients with IPF, all in China, randomized to placebo or one of three rentosertib doses for 12 weeks. Read it with one fact up front: the primary endpoint was safety, specifically the share of patients with at least one adverse event that emerged during treatment.

On safety, those rates were similar across arms (70.6% on placebo; 72.2%, 83.3% and 83.3% on the three doses). Adverse events judged treatment-related were more common on the drug: 29.4% on placebo versus 77.8% at the top dose. Seven patients stopped because of liver injury or dysfunction. Three patients on the top dose had an acute exacerbation of IPF and were hospitalized, versus one on placebo who wasn't. The authors note that 12 weeks is too short to judge exacerbations.

On lung function, a secondary endpoint, forced vital capacity (FVC) fell by an average 20.3 mL on placebo and rose by 98.4 mL at 60 mg once daily (95% CI 10.9 to 185.9). The lower doses didn't separate: −27.0 mL at 30 mg once daily and +19.7 mL at 30 mg twice daily. This is the number in every press release, and in IPF, where existing drugs only slow decline, a gain would be a big deal.

The fine print on +98.4 mL

Eighteen patients per arm. The paper states plainly that "a sample size calculation based on statistical power considerations was not performed." The arms were sized to see safety signals, not to prove efficacy.

Lots of missing data at the top dose. Six of the 18 patients on 60 mg stopped treatment early, and 8 of the 18 had no week-12 FVC measurement. Those values were filled in by multiple imputation, which assumes the missing patients resemble the ones who stayed. When the dropouts include people who stopped for side effects, that assumption carries real weight.

A wide interval. A confidence interval running from 10.9 to 185.9 mL says the true effect could be anywhere from barely there to large.

The subgroup puzzle. Patients on 60 mg who weren't also taking a standard antifibrotic gained 187.8 mL, while those taking nintedanib or pirfenidone alongside it showed no significant change. That could be a drug interaction, as the authors suggest, or noise from slicing an 18-patient arm in two.

Other measures were flat. Changes in diffusion capacity, FEV1 and six-minute walk distance were similar across groups. A cough questionnaire improved at the top dose with p = 0.0495, right at the edge.

None of this is hidden. The authors list "the small cohort size of each arm," the homogeneous population and the short follow-up as limitations, and call for longer, larger trials. Eleven of the 27 authors are Insilico employees and Insilico sponsored the trial, which the paper discloses. The paper's own conclusion is measured: rentosertib is "generally safe and potentially effective."

What phase 3 will actually test

According to Insilico's 7 July 2026 announcement, the phase 3 trial plans to enroll 320 participants at 47 centers in China, dosing once daily for 52 weeks, with the annual rate of FVC decline as its primary endpoint. That design asks the efficacy question properly: a pre-specified lung-function endpoint, a year of treatment, enough patients. The company's release also states that the drug "remains investigational and has not been approved by any regulatory authority."

Here's why I'll be watching it. If rentosertib works in phase 3, it's the first real evidence that an AI-assisted pipeline can pick a target that holds up in patients, the exact step I argued was still a human bet. If it fails, it fails at the same stage most drugs do, and the 18-month discovery timeline becomes a story about speed, not about better biology.

How to read "AI-discovered drug"

Ask what the AI chose, and inside what limits. A ranked list under human-set filters is still useful. It just isn't the same as an autonomous discovery.

Ask what the trial's primary endpoint was. A safety-primary phase 2a with an encouraging secondary result is a reason to run phase 3, not a result to bank.

Count patients and dropouts in the arm that "worked." Here it was 18 patients, with 8 missing the measurement that made headlines.

Wait for the trial designed to answer the question. For rentosertib, that's the 52-week phase 3.