The First End-to-End AI-Designed Drug Has Phase 2 Data. What Rentosertib’s Results Mean for the Pipeline.

7 min read

7 min read

7 min read

6 min read

A 118.7 mL separation in forced vital capacity between treatment and placebo. Twelve weeks. Seventy-one patients. That is the data point the AI drug discovery field has been waiting for, and it arrived in June 2025 when Insilico Medicine published Phase IIa results for Rentosertib in Nature Medicine — the first clinical proof-of-concept that an end-to-end AI-designed drug can produce a measurable efficacy signal in humans.

The result matters, and so does its context. The GENESIS-IPF trial was a Phase IIa study — designed primarily to confirm safety, not to prove efficacy at statistical power. The patient population was 71. The study ran 12 weeks in a disease measured in years. What the data establishes is a signal: TNIK inhibition appears to improve lung function in patients with idiopathic pulmonary fibrosis (IPF), and the AI-generated molecule designed to produce that effect appears to be the agent doing it. What the data does not establish is whether that signal will survive a Phase III trial sized to detect it definitively — or whether the broader AI drug discovery pipeline behind Rentosertib will follow the same path.

What GENESIS-IPF Actually Showed

The trial enrolled 71 patients with IPF across 22 sites in China, randomly assigned to placebo, 30 mg once daily (QD), 30 mg twice daily (BID), or 60 mg QD Rentosertib for 12 weeks. The primary endpoint was safety and tolerability, which was met: treatment-emergent adverse events occurred at similar rates across all groups, most were mild to moderate, and all resolved after treatment discontinuation.

The secondary efficacy signal was dose-dependent and clinically plausible. Patients on 60 mg QD showed a mean forced vital capacity (FVC) improvement of +98.4 mL compared to baseline. The placebo group declined by −20.3 mL over the same period — a separation of approximately 118.7 mL. Biomarker analysis reinforced the picture: in the high-dose group, profibrotic proteins COL1A1, MMP10, and FAP were significantly reduced, while the anti-inflammatory marker IL-10 increased. These protein changes correlated with the FVC improvements, suggesting the drug is engaging its biological target in the direction intended.

That last point is important and often missed in coverage of this trial. An efficacy signal that tracks with target-engagement biomarkers is more credible than an FVC number in isolation. It is consistent with the mechanism — it is not a statistical outlier. But consistency with the mechanism is not the same as proof of mechanism. Phase IIa is where signals are identified. Phase III is where they are confirmed or refuted.

The Target: Why TNIK Is Significant

TNIK (Traf2- and Nck-interacting kinase) is a first-in-class IPF target — no approved drug inhibits it. In IPF, TNIK drives pathological lung fibrosis by activating multiple pro-fibrotic signalling pathways simultaneously: WNT/β-catenin, TGF-β, Hippo (YAP–TAZ), JNK, and NF-κB. It promotes extracellular matrix (ECM) protein deposition, inflammatory myeloid cell recruitment, and proinflammatory cytokine production. The upstream position of TNIK in these converging pathways is what made it an attractive AI-identified target — hitting it theoretically addresses multiple fibrotic drivers at once.

The two approved IPF treatments work differently. Nintedanib (Boehringer Ingelheim’s Ofev) is a tyrosine kinase inhibitor that blocks PDGF, VEGF, and FGF receptor signalling; it slows FVC decline by approximately 50% compared to placebo but does not reverse it. Pirfenidone (Roche/Genentech’s Esbriet) has antifibrotic, anti-inflammatory, and antioxidant effects, with similar slowing of decline. Both drugs manage IPF progression; neither has shown mean FVC improvement in a clinical trial.

The clinical significance of the Rentosertib FVC signal, if it holds in Phase III, would not be incremental. A drug that improves lung function in IPF rather than merely slowing its decline would represent a different class of intervention — provided the effect is durable beyond 12 weeks in a larger and more heterogeneous patient population.

How AI Drug Discovery Produces End-to-End Designed Drugs

Insilico describes Rentosertib as the first drug in which both the target and the molecule were identified and generated by artificial intelligence. That claim is more specific — and more verifiable — than the generic “AI drug discovery” label applied to most companies in the space.

The AI drug discovery platform uses three components: PandaOmics, which analysed multi-omics datasets to identify TNIK as an IPF driver; Chemistry42, which generated inhibitor candidates optimised for potency and developability; and the overarching Pharma.AI orchestration layer. Human scientists then conducted wet-lab validation, IND-enabling studies, clinical trial design, patient enrolment, and regulatory interactions. The efficiency claim holds up on the computational phases: across 22 nominated candidates from 2021 to 2024, Insilico averaged 12–18 months from project initiation to preclinical candidate nomination, requiring only 60–200 synthesised molecules per project. Traditional discovery typically takes 2.5–4 years and screens tens of thousands of compounds.

What AI drug discovery does not compress is clinical development time. A Phase III trial for a progressive lung disease will take three to five years to complete, enrol hundreds of patients, and require an endpoint duration long enough to detect meaningful differences in disease trajectory. Insilico’s Phase III plans include a China trial of 500+ patients and a US Phase IIb trial of 200+ patients, with regulatory discussions with China’s Center for Drug Evaluation and the FDA ongoing. A first AI-designed drug FDA approval remains in the 2027–2028 horizon at the earliest.

The Pipeline Behind Rentosertib

Rentosertib is the most clinically advanced data point in the 173-program AI drug discovery field. As of early 2026, AI drug discovery has produced candidates distributed across Phase I (94 programs), Phase II (56 programs), and Phase III (15 programs), with 15–20 expected to enter pivotal trials in 2026. A first AI-designed drug FDA approval is unlikely before 2027 on current Phase III timelines.

The most capitalised clinical-stage organisations include the merged Recursion/Exscientia (eight programs across Phase I–II, following their November 2024 combination) and Relay Therapeutics (one Phase III program). BenevolentAI, acquired by Osaka Holdings in March 2025, has its lead candidate BEN-8744 — a PDE10 inhibitor for ulcerative colitis — in active trials. Insilico itself listed on the Hong Kong Stock Exchange in December 2025, reported US$56.24 million in 2025 revenue, and holds 10 programs in clinical trials across its 28 nominated preclinical candidates.

The pipeline breadth matters for one specific reason: it determines whether Rentosertib’s Phase IIa success is a proof of platform or a proof of one molecule. For the platform case to hold, the field needs multiple Phase III readouts across different disease categories. The data to answer that question will not exist until 2027 at the earliest.

The comparative clinical data does show that AI drug discovery’s molecules are not failing earlier in the clinic than traditionally discovered drugs. Phase I success rates for AI-originated candidates are reported at 80–90%, above the ~52% industry average. Phase II success rates are comparable at around 40% — though the sample remains small. The critical question is whether AI target identification produces molecules more likely to succeed in Phase II, where most drug failures occur for efficacy reasons. Rentosertib’s Phase IIa signal suggests it might, for TNIK. The field does not yet have enough Phase II completions to answer that question statistically.

The Limitations That Phase III Must Address

Geography. All 71 patients were enrolled in China. IPF prevalence, genetic background, disease stage distribution, and standard of care differ between Chinese and Western populations. The US Phase IIa trial (NCT05975983) enrolled only 8 of a planned 60 patients as of mid-2025. Western patient data for Rentosertib remains sparse.

Duration. Twelve weeks is insufficient to characterise FVC durability in a progressive fibrotic disease. Nintedanib’s pivotal trials ran 52 weeks. Whether a +98.4 mL improvement at week 12 translates to maintained benefit at week 52 — or whether there is a rebound effect — is unknown.

Power. With 71 patients across four arms, no individual cohort exceeded approximately 18 patients. The trial was not powered to demonstrate statistical significance on FVC as a primary endpoint. The efficacy data is exploratory. That is not a criticism of the trial design — it is what Phase IIa is designed to produce — but it means the result is hypothesis-generating, not hypothesis-confirming.

What to Watch

The US Phase IIa trial (NCT05975983) primary readout, expected around mid-2026, will be the first test of whether the China Phase IIa signal replicates in a Western patient population. A consistent result across both geographies would substantially strengthen the Phase III investment case. A divergent result would raise questions about patient population specificity that Phase III would need to design around.

Insilico’s Phase III trial protocol disclosure — expected as part of CDE and FDA submission filings — will reveal the chosen primary endpoint, comparator arm, and trial duration. Whether the Phase III compares Rentosertib against placebo only, or against nintedanib/pirfenidone as an active comparator, will determine both the approval pathway and the eventual commercial positioning.

For the broader AI drug discovery sector, the 2026–2027 Phase III completion window across 15+ programs will produce the first statistically meaningful comparison of AI vs. traditional discovery success rates in late-stage trials. Phase II trials fail at a rate that makes any positive signal meaningful — if AI-originated molecules hold below the industry baseline in Phase III as well, the platform value proposition strengthens materially. If they mirror industry norms, the case for AI reduces to speed and cost compression in the preclinical phase — still significant, but different in kind from the transformative narrative the sector has projected. Rentosertib’s Phase IIa results are the first credible data point from that experiment. They are not the answer. They are the reason the question is now worth asking at scale.

For context on the broader AI infrastructure driving these pipelines, see our analysis of hyperscaler capex and AI infrastructure financing.

This article was produced with AI assistance and reviewed by the editorial team.

Related posts:

AI-Designed Drugs Are Entering Clinical Trials. The Pipeline Is Larger Than Most Engineers Realize. The GLP-1 Manufacturing Race Is a Supply Chain Problem, Not a Drug Problem Three BCI Architectures, One Regulatory Finish Line: What the 2025 Trial Data Actually Shows AI’s Impact on Developer Hiring: What the Data Shows