The promise of AI accelerating pharmaceutical development has been discussed for years, but 2025-2026 has produced the clearest evidence yet that the technology is moving beyond protein structure prediction into actual drug candidates entering clinical testing.
Isomorphic Labs — the drug discovery company spun out of DeepMind alongside AlphaFold 3 — announced in Q1 2026 that two of its AI-designed small molecule candidates had advanced to Phase I clinical trials in partnership with Eli Lilly and Novartis. These are not AI-assisted drug candidates where a human designed the molecule and AI helped optimize it. These are molecules where the binding target was identified, and the molecular structure was generated and ranked entirely by AI models, with human scientists validating the outputs rather than originating them.
AlphaFold 3 itself represents a significant expansion from AlphaFold 2's protein structure focus. The third generation models interactions between proteins and DNA, RNA, and small molecule ligands — precisely the interactions that determine whether a potential drug molecule will bind to its target and how strongly. AstraZeneca and Pfizer have both confirmed research partnerships with Isomorphic Labs to use these models in their discovery pipelines.
Insilico Medicine is further along the clinical pipeline. Its INS018_055 — a drug for idiopathic pulmonary fibrosis, a serious lung disease, designed almost entirely by the company's AI platform — is in Phase II trials. Insilico reports that from target identification to clinical candidate took 18 months, compared to an industry average of four to six years for this phase. The company's approach combines generative chemistry models for molecular design with AI-predicted ADMET properties (absorption, distribution, metabolism, excretion, toxicity) to filter candidates before any physical synthesis.
Recursion Pharmaceuticals, a Utah-based company, is taking a different approach: running millions of biological experiments at scale, using robotic lab automation and machine learning to map disease biology and identify novel targets. Its partnership with Nvidia to build a dedicated AI infrastructure platform for biology — combining BioNeMo foundation models with Recursion's experimental data — reflects the scale of investment now flowing into the space.
What remains genuinely uncertain: clinical success rates for AI-designed drugs versus traditionally discovered ones. AlphaFold can predict whether a molecule will bind to a protein, but biological systems are vastly more complex — off-target effects, immune responses, metabolic pathways — and AI models do not yet reliably predict clinical efficacy. The industry will need several more years of Phase II and Phase III data to know whether AI drug discovery actually changes the attrition rate, which historically runs around 90% from candidate to approval.
The optimistic case is that AI compresses the front end of the pipeline — finding candidates faster and cheaper — while leaving clinical biology as challenging as ever. Even that more modest outcome would be enormously valuable given that a traditional drug discovery program costs upwards of $2 billion and takes over a decade.