The AI-Designed Drug That Reached Phase 3 Isn't an Antibiotic

The claim has been circulating for months, usually in a screenshot: an antibiotic designed by artificial intelligence, built on a molecular scaffold no chemist ever drew, has entered Phase 3 trials against a drug-resistant superbug. It is a good story. It is not a true one.

As of August 2026, no antibiotic discovered or designed by AI has reached Phase 3. None has reached Phase 1. The compounds most often named — halicin, abaucin, and the newer generative-AI molecules NG1 and DN1 out of James Collins lab at MIT and the Broad Institute — are all preclinical. Meanwhile, the one genuinely novel-class antibiotic that did enter Phase 3 this year, Roche zosurabalpin, was found the old way: a phenotypic screen of roughly 45,000 compounds.

What actually happened in 2026 is a different milestone, and a more defensible one.

The drug that got there first

On July 7, 2026, Insilico Medicine (HKEX: 3696) announced initiation of a Phase III trial for rentosertib, an oral small-molecule inhibitor of TNIK for idiopathic pulmonary fibrosis. The trial is registered as CTR20262475 and NCT07687459. It is a prospective, randomized, double-blind, placebo-controlled study enrolling 320 patients across 47 centers in China, with a primary endpoint of annual rate of decline in forced vital capacity over 52 weeks and a key secondary endpoint of time to first disease-progression event. The leading principal investigator is Professor Zuojun Xu of Peking Union Medical College Hospital, with Academician Nanshan Zhong and Professor Chang Chen of Shanghai Pulmonary Hospital as co-leads.

Rentosertib — formerly ISM001-055 — matters because of how much of it came from software. Its target, the serine/threonine kinase TNIK, was prioritized by Insilico PandaOmics engine from multi-omics fibrosis data, biological network analysis, and aging-relevant target scoring; TNIK was the top-ranked candidate in the kinase discovery scenario, a target class largely unexplored in IPF. The molecule itself was generated and optimized by Chemistry42, the company generative chemistry platform. The discovery-to-clinic path was published in Nature Biotechnology in 2024; the medicinal chemistry appeared in the Journal of Medicinal Chemistry.

The efficacy case rests on GENESIS-IPF, the randomized Phase IIa published in Nature Medicine in 2025: 71 patients across 22 sites, 12 weeks of dosing. The 60 mg once-daily arm showed a mean forced vital capacity change of +98.4 mL, against -20.3 mL for placebo. Profibrotic proteins including COL1A1, MMP10 and FAP fell in the high-dose group; the anti-inflammatory marker IL-10 rose. Adverse events were mostly mild to moderate and resolved on discontinuation. The FDA granted rentosertib orphan drug designation for IPF in February 2023.

Insilico is careful, in its own materials, about what this proves. Chief scientific officer Feng Ren framed the origin story rather than the outcome: “Rentosertib was not discovered by starting from a conventional target and simply screening more compounds. It came from a biology-first, aging-informed AI workflow.” Founder and CEO Alex Zhavoronkov was blunter about the shift in burden of proof: “For the AI drug discovery field, this is no longer only a speed story. It is a testament to the ability of AI to create truly novel therapeutics with novel target, novel molecule, not just discover me-better molecules for known targets.”

What AI compresses, and what it does not

The honest accounting is narrow. Insilico says it reaches preclinical candidate nomination in 12 to 18 months, against an industry norm of 2.5 to 4 years, and does it while synthesizing only 60 to 200 molecules per program. That is a real compression of the most wasteful part of the pipeline: the years spent guessing which target is worth pursuing and the thousands of compounds made to find one that behaves.

None of that shortens a trial. Rentosertib Phase III still runs 52 weeks, because lung function declines at the rate lungs decline, not at the rate models converge. It still needs 320 patients, because that is what statistical power costs. It still needs 47 sites and the years of monitoring, data cleaning and regulatory review that follow. Generative chemistry can produce a molecule in months; it cannot make a patient breathe on an accelerated schedule.

The pipeline data reflect exactly that asymmetry. A peer-reviewed analysis presented at ASCO in 2026 counted 117 AI-enabled therapeutic assets across 63 companies that had entered interventional human trials. Of those, 60 — about 51 percent — had completed Phase 1. Only 8, or 6.8 percent, had completed Phase 2. Phase 2 success rates for AI-derived assets sit near 40 percent, indistinguishable from the historical industry average. AI has moved the front of the funnel. The clinical middle has not budged.

Antibiotics show the gap most starkly. In Cell in August 2025, the Collins group described models trained on roughly 40,000 antibacterial small molecules that generated more than 36 million candidate compounds, yielding NG1 (active against multidrug-resistant Neisseria gonorrhoeae) and DN1 (active against MRSA and resistant gonorrhoea). Phare Bio, the nonprofit Collins co-founded, backed by ARPA-H and Google philanthropic arm, is still modifying both molecules for further testing. Zosurabalpin, by contrast, is already comparing against standard of care in roughly 400 patients with carbapenem-resistant Acinetobacter baumannii — and it got there through a conventional screen, because it started a decade earlier.

What to watch

Three things. Whether rentosertib Phase III readout, expected no earlier than 2028, converts a 12-week FVC signal in 71 patients into a 52-week benefit in 320 — the step at which most promising IPF drugs have historically failed. Whether Phare Bio files the first IND for an AI-designed antibiotic, which would start the clock the viral claim assumed had already run. And whether any of the 117 AI-enabled assets in trials clears Phase 2 at a rate meaningfully above 40 percent. Until one does, the correct summary is that AI has learned to pick better starting points, and nothing more.