AI for scientific discovery

Core thesis

AI is joining the scientific process as an active participant — generating hypotheses, designing experiments, and proposing new molecules. Google's AI co-scientist proposed drug repurposing candidates for AML that were experimentally validated. AlphaFold 3 predicts protein-molecule interactions. Nvidia's BioNeMo platform includes models for RNA structure, molecular synthesis, and toxicity prediction. Novo Nordisk partnered with OpenAI to integrate AI across its entire drug pipeline. The implication: the scientific method itself is being augmented by AI, and labs that don't adopt AI-driven research will fall behind within 2-3 years.

State of the art (2026)

By mid-2026 the frontier has shifted from structure prediction to the full discovery loop. AlphaFold 3 (open-sourced November 2024) and Isomorphic Labs' Drug Design Engine now read out protein-ligand interactions and novel binding pockets, while Baker Lab's RFdiffusion lineage and EvolutionaryScale's ESM3 make de novo design routine. Insilico's rentosertib became the first generative-AI-designed drug with positive Phase IIa data (Nature Medicine, June 2025). Multi-agent systems – Google's AI co-scientist and FutureHouse's Robin – propose and triage hypotheses, and DeepMind is standing up its first automated UK materials lab. The binding question is no longer whether AI generates candidates, but how fast wet-lab and clinical validation can keep pace.