In October 2024, the Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper from DeepMind for their breakthroughs with AlphaFold. It was the formal recognition that AI had solved a problem biochemists considered intractable for computers: protein folding. In 2026, the field has advanced beyond structural prediction — and reached the most ambitious frontier: designing therapeutic molecules from scratch.
AlphaFold 3 and IsoDDE: From Structure to Drug
AlphaFold 3, released by DeepMind and Isomorphic Labs in May 2024, radically expanded its scope compared to AlphaFold 2. Instead of merely predicting protein structure, AF3 predicts interactions between proteins, DNA, RNA, and small molecules — the fundamental components of drug design. It is 50% more accurate than the best traditional methods on the PoseBusters benchmark without needing prior structural information.
In February 2026, Isomorphic Labs released IsoDDE — a unified drug design engine that combines structure prediction, ligand binding, affinity prediction, and antibody-antigen interaction modeling in a single pipeline. IsoDDE approximately doubles AlphaFold 3's accuracy in the most difficult ligand binding cases (less than 20% sequence similarity to training data): ~50% accuracy versus AF3's ~23.3%. Some experts informally call IsoDDE the "proprietary AlphaFold 4."
The First AI-Designed Drug in Clinical Trials
The most concrete milestone of the field in 2026: Isomorphic Labs is targeting the first human clinical trials of an AI-designed drug candidate by late 2026. At the World Economic Forum in January 2026, Demis Hassabis confirmed that the timeline had been adjusted from 2025 to late 2026 — but the commitment remains.
Insilico Medicine is already ahead: its INS018_055, a drug candidate for idiopathic pulmonary fibrosis generated entirely by AI, is in Phase 2 clinical trials. It is the most advanced outcome of AI drug design in history — and the benchmark every company in the sector uses to calibrate ambitions.
GNoME: 2.2 Million New Materials
AlphaFold does not stand alone. In 2023, DeepMind released GNoME (Graph Networks for Materials Exploration), which discovered 2.2 million new crystal structures — including 52,000 new lithium-ion conductors with potential for next-generation batteries. External researchers have already synthesized 736 of these predictions in the lab.
GNoME represents a qualitative shift in what AI can do for materials science: instead of assisting existing experiments, it generated a catalog of materials that humans would not have discovered over the next decades of conventional research. AlphaFold serves over 3 million researchers in more than 190 countries — and GNoME's scale of impact could be comparable.
The Bifurcation of the Field: Discovery vs. Clinic
The most honest assessment of the AI drug discovery field in 2026 recognizes a clear bifurcation. On one hand, the discovery victories: AI-designed molecules reaching the clinic faster and cheaper than industry benchmarks. On the other, the clinical reality: AI does not shorten Phases 2 and 3, does not bypass the fundamental biology of efficacy and safety, and has not yet produced an approved drug.
AI has revolutionized the discovery stage — candidate identification, structure prediction, virtual screening of compounds. But the drug development process takes 10–15 years and billions of dollars in costs for a reason: the biology of the human body is the ultimate validator, and that validator has not been replaced by any model.
The "AI for Science" field in 2026 is at the most promising moment in its history — and also the most honest about where the promise has yet to turn into a product.
