Summary
“Fully AI-generated” (de novo) therapeutics are now clearly in the clinic, but the most rigorous public evidence still sits in early-to-mid clinical development rather than in completed late-stage trials. The strongest peer-reviewed clinical proof-of-concept as of May 2026 is Insilico Medicine’s rentosertib (formerly ISM001-055 / INS018_055), which reported Phase 2a randomized trial results in idiopathic pulmonary fibrosis (IPF) in Nature Medicine (published online June 3, 2025).[1]
In parallel, multiple programs designed with generative AI platforms have entered Phase 1 (for example Isomorphic Labs’ first AI-designed drug in the UK, and Absci’s ABS-101) or moved into pivotal testing (for example Generate:Biomedicines’ GB-0895 with planned global Phase 3 trials, and MindRank’s MDR-001 with first patient dosed in Phase 3 in China).[2–5]
State of the art
In practice, “AI-designed” can refer to different technical contributions along the path from target selection to lead optimization and molecule engineering. For example, Insilico describes an end-to-end workflow in which it first identified a target for IPF using its PandaOmics engine and then designed and optimized a small molecule using its Chemistry42 generative chemistry engine.[6]
A separate, increasingly important capability is high-accuracy modeling of protein shapes and interactions. Public reporting around Isomorphic Labs’ first AI-designed drug explicitly frames the effort as being driven by AlphaFold 3, described as a tool that can “accurately predict the shape of proteins and how they interact with other molecules.”[2]
On the biologics side, de novo protein generation methods are demonstrating atomic-level control over binding. A 2026 report describes combining a fine-tuned RFdiffusion2 network with yeast display screening to generate antibody formats (VHHs, scFvs, and full antibodies) that bind user-specified epitopes “with atomic-level precision,” with cryo-EM confirmation of binding poses and affinity maturation yielding single-digit nanomolar binders while maintaining epitope selectivity.[7]
Clinical evidence
The table below summarizes a small set of de novo, generative-AI-designed programs with explicit, sourced clinical milestones in the provided dataset.
Furthest along in peer-reviewed clinical readouts
The most detailed and peer-reviewed clinical signal in the dataset is Insilico’s rentosertib in IPF. The published report describes “the first phase 2a multicenter, double-blind, randomized, placebo-controlled trial” testing rentosertib (formerly ISM001-055), characterized as a “first-in-class AI-generated small-molecule inhibitor of TNIK,” with TNIK described as a “first-in-class target in idiopathic pulmonary fibrosis (IPF) discovered using generative AI.”[1]
Efficacy was reported as a change in forced vital capacity (FVC), with the highest-dose arm showing a mean change of mL (95% CI 10.9 to 185.9) in the 60 mg once-daily group compared with mL (95% CI −116.1 to 75.6) in the placebo group.[1] Complementary communications around the same Phase 2a program describe meeting the primary endpoint of safety and tolerability across dose levels and reporting dose-dependent FVC improvement on a secondary endpoint.[12]
Separately, Insilico also reported regulatory progress for a direct-to-lung formulation: an inhalation solution of rentosertib received IND clearance supporting a Phase 1 study evaluating safety, tolerability, and pharmacokinetics.[13]
The wave behind it
Beyond rentosertib, multiple organizations are now publicly describing first-in-human testing for AI-designed candidates.
Isomorphic Labs (Alphabet) announced the start of human trials for its first AI-designed drug, stating that a Phase 1 trial is underway in the UK and will focus on safety and tolerability in healthy volunteers, with the project described as driven by AlphaFold 3.[2]
Generate:Biomedicines has disclosed both early clinical data and late-stage intent. For GB-0669 (a SARS-CoV-2 neutralizing antibody), first-in-human data described Phase 1 tolerability with no dose-limiting toxicities or serious adverse events and reported neutralization metrics (for example, neutralizing index 15 at 1200 mg).[11] For GB-0895 (anti-TSLP), the company announced plans to initiate two global Phase 3 trials (SOLAIRIA-1 and SOLAIRIA-2), describing the dosing regimen (300 mg subcutaneous every six months) and a 52-week design with a primary endpoint tied to annualized asthma exacerbation rate.[4, 10]
Absci has advanced multiple AI-engineered antibodies into clinical testing. It reported first healthy volunteers dosed in a Phase 1 study of ABS-101, an anti-TL1A antibody engineered with its generative AI platform.[3] It also reported first healthy volunteers dosed in a Phase 1/2a HEADLINE study for ABS-201 (anti-PRLR) and stated that interim data are expected in the second half of 2026.[14]
Reality checks
First, “in the clinic” does not mean “clinically proven.” The clearest efficacy signal in the dataset (rentosertib’s Phase 2a FVC change) is still a Phase 2a outcome and is explicitly framed as warranting further investigation in larger and longer trials.[15]
Second, even when programs reach pivotal trials, this typically reflects a transition into longer, confirmatory testing rather than completion. For example, the Phase 3 asthma program for GB-0895 is described as planned initiation and specifies a 52-week placebo-controlled design, implying that results would only come after substantial follow-up and analysis.[4, 10] Similarly, MindRank’s MDR-001 disclosure is framed as first patient dosed in Phase 3, which is a start-of-trial milestone rather than an endpoint readout.[5]
Third, the term “AI-designed” remains heterogeneous across modalities and companies. Some programs emphasize end-to-end generative workflows for both target and molecule (as described for rentosertib’s target discovery and AI-generated inhibitor), while others emphasize structure prediction (AlphaFold-driven efforts) or de novo protein/antibody generation (RFdiffusion-based approaches).[1, 2, 7]
What to watch
The near-term scientific question is whether the Phase 2a signal observed for rentosertib—especially the dose-associated FVC change—reproduces in larger and longer studies in IPF, a setting where pulmonary function decline is clinically meaningful and difficult to reverse.[1]
The near-term translational question is whether AI-engineered antibodies like GB-0895 can translate long-acting pharmacokinetics and biomarker suppression from Phase 1 into clinically meaningful reductions in exacerbations over a year-long Phase 3 design.[4, 10] Finally, as more candidates enter Phase 1 (including those described by Isomorphic Labs and Absci), the field will be judged less by claims about design speed and more by the clinical properties that early studies are explicitly set up to measure: safety, tolerability, pharmacokinetics, and—later—validated efficacy endpoints.[2, 3]
