Generative Biology's AI-Designed Drugs Enter Phase III Trials, Rewriting Rules of Discovery

For a decade, artificial intelligence in drug discovery has run on promises of faster timelines and novel molecules that no human chemist would have drawn. Now, that computational theory is facing the hardest test in medicine. Generative biology has officially moved from the laboratory bench to late-stage clinical testing, with multiple AI-designed drugs—including a novel treatment for a fatal lung disease—entering Phase III trials. It is the final hurdle before regulatory approval, and it will prove whether AI can actually produce medicines that work in the human body, rather than just impressive benchmark scores on a server.[3][4]

The clearest test case is rentosertib, an experimental pill for idiopathic pulmonary fibrosis (IPF) developed by Hong Kong-based Insilico Medicine. In conventional drug discovery, researchers start with a known biological target and screen massive libraries of existing compounds hoping for a hit. Insilico inverted the process. The company used its AI biology engine, PandaOmics, to analyze massive datasets of aging and disease pathways, identifying a protein called TNIK as a novel driver of lung scarring. Then, a generative chemistry platform called Chemistry42 designed a brand-new molecular structure from scratch specifically to inhibit that protein.[1][2][7]

Rentosertib is now the first medicine to reach Phase III where both the biological target and the chemical structure were entirely generated by artificial intelligence. On July 7, 2026, Insilico initiated a 320-patient, 52-week randomized, double-blind trial across sites in China. The stakes for IPF patients are immense; the disease causes progressive, irreversible lung scarring, and current therapies only slow the decline rather than halting it.[1][2][4]

The data pushing rentosertib into Phase III is promising but preliminary. In a 12-week Phase IIa study published in Nature Medicine, patients taking a 60-milligram daily dose saw their forced vital capacity (FVC)—a key measure of lung function—improve by a mean of 98.4 milliliters, while the placebo group declined by 20.3 milliliters. However, as clinical investigators note, a 12-week signal in 71 patients is vastly different from proving long-term disease modification over a full year in a large population.[2][5]

Insilico is not alone at the Phase III threshold. Generate:Biomedicines, a company treating biology as a programmable language, has advanced an AI-engineered antibody called GB-0895 into global Phase III trials for severe asthma. Their AI system simultaneously optimized the antibody for binding affinity, half-life extension, and manufacturability before a single molecule was synthesized—a multi-parameter balancing act that traditionally requires years of iterative lab work.[5][6]

Across the industry, the pipeline is maturing rapidly. Industry trackers count more than 150 AI-discovered or AI-designed molecules currently in human trials. Roughly 15 of those programs have reached Phase III. The speed of the preclinical phase is undeniable: AI platforms are consistently delivering developmental candidates in 12 to 18 months, compared to the traditional 2.5 to 4 years, while synthesizing only dozens of molecules instead of thousands.[4][6]

Yet, the ultimate question remains unanswered: does this accelerated preclinical speed translate to a higher success rate in humans? As of mid-2026, no fully AI-designed drug has received FDA approval. The field has already absorbed high-profile failures, including the quiet abandonment of the very first AI-designed drug to enter trials back in 2020. Biology is notoriously complex, and a molecule that binds perfectly in a cloud simulation can still fail due to unforeseen toxicity or lack of efficacy in a complex human organism.[5][6]

If these Phase III trials succeed, it will fundamentally rewrite the economics and timelines of the pharmaceutical industry. It would prove that biology can be engineered programmatically, shifting drug discovery from a process of artisanal trial-and-error to one of predictable generation. But regulators and scientists are holding the line: the fact that an AI designed a drug does not lower the evidentiary bar. The molecules must now survive the brutal, binary reality of late-stage clinical trials.[3][7]