AI-designed drugs reach Phase III: a 2026 reality check on AI in drug discovery

"AI will discover new medicines" has been a headline for a decade. In 2026 it finally became testable, because AI-designed molecules are now in late-stage human trials. Here's what has actually happened, without the hype.

This article is for general education, not medical advice. None of the drug candidates mentioned here is approved.

Why drug discovery is a good problem for AI

A new medicine has to hit the right biological target, bind to it strongly, avoid hitting other things, survive in the body long enough, and not be toxic. The space of possible molecules is astronomically large, and most candidates fail. Isomorphic Labs' president Max Jaderberg puts the failure rate at roughly 90% for medicines entering human trials.

AI helps at three stages:

1. Target discovery: finding which protein or pathway to hit, by mining biological data.
2. Molecule generation: generative chemistry models propose new molecules with the desired properties.
3. Prediction and prioritisation: models estimate binding, selectivity and toxicity, so chemists synthesise and test fewer, better candidates.

Case 1: rentosertib, from AI target to Phase III

Insilico Medicine's rentosertib is the clearest end-to-end test so far. According to Insilico's September 2026 announcement:

• The target, a protein called TNIK, was identified by AI. The trial's lead investigator notes that TNIK had never previously been linked to fibrosis.
• The molecule itself was designed with generative AI.
• A 12-week Phase IIa study, published in Nature Medicine in 2025, showed what Insilico describes as a promising dose-dependent trend, including improvements in lung function.
• GENESIS-IPF-3, a randomised, double-blind, placebo-controlled Phase III trial testing once-daily dosing over 52 weeks, dosed its first patient at Peking Union Medical College Hospital.

The honest framing comes from the lead investigator himself: from starting Phase III to final regulatory approval will take three to four years under favourable conditions. Phase III is where many drugs fail. If rentosertib succeeds, it will be a landmark. If it fails, that tells us AI hasn't solved biology, not that AI is useless.

Case 2: an AI-designed cancer drug's first human data

At the ESMO Congress on 25 October 2026, results were presented for ISM6331, an oral drug candidate designed with Insilico's Chemistry42 generative chemistry software that targets the TEAD family of proteins involved in tumour growth. Drug Discovery Trends reports:

• In previously treated mesothelioma patients, 4 of 22 evaluable patients had partial responses (18.2%), and 68.2% had disease control at the first scan.
• No dose-limiting toxicities were reported between 15 and 130 mg; three patients at 130 mg had grade 3 treatment-related side effects.
• For comparison, Vivace Therapeutics' TEAD inhibitor VT3989 produced responses in 26% of 47 mesothelioma patients at optimised doses in an earlier Phase 1/2 trial.

That last point matters. "AI-designed" does not automatically mean "better." The molecule has to beat the alternatives on the same evidence standards. Insilico plans a randomised Phase 2a study in 2027.

Case 3: Isomorphic Labs and the "AI designs every drug" claim

Isomorphic Labs builds on the protein-structure work that came out of Google DeepMind. At Bloomberg Tech on 2 November 2026, Jaderberg predicted that AI will eventually design every drug, while acknowledging that biology depends on real-world experiments and lab time, R&D World reports. Key facts from that report:

• Its internal programmes focus on oncology, immunology and inflammation, and are preclinical.
• It partners with Eli Lilly, Novartis and Johnson & Johnson, and says it does not sell model access or API tokens.
• Its CEO, Demis Hassabis, said in January that first clinical trials were expected by the end of 2026.

What AI does not change

• Biology is slow. Cells, animals and people work at their own pace. A 52-week trial takes 52 weeks.
• Data quality limits models. Public biological data is noisy and biased toward well-studied targets.
• Validation is expensive. Every promising prediction still needs synthesis and testing.
• Regulators judge evidence, not methods. A drug is approved on safety and efficacy data, however it was designed.

What this means for you: skills to learn

India has one of the world's largest pharmaceutical industries, and AI-enabled R&D teams need engineers as well as scientists:

• Python data engineering: cleaning and joining messy assay, literature and clinical data.
• Machine learning fundamentals: evaluation, overfitting and honest validation splits matter even more when data is scarce.
• Generative AI and retrieval: literature and patent search assistants, with citations and guardrails.
• Cloud and MLOps: reproducible pipelines, experiment tracking and secure data handling.
• Domain literacy: enough biology and chemistry to talk to scientists.

Our AI Architect & Cloud Master Track covers the cloud, retrieval and agent foundations these teams use. The scientific depth comes from working alongside chemists and biologists, and reading the primary papers.