Google DeepMind's AlphaFold isn't the AI science template

Google DeepMind's AlphaFold cracked one of biology's toughest nuts—the protein folding problem—by training on the Protein Data Bank, a 53-year, $21B effort compiling 170,000 experimentally validated structures. While this template worked for proteins (where crystallography is unusually reliable), most sciences lack such clean, replicable data. Lab conditions drift, contaminants skew results, and datasets rarely meet the consistency needed for modern neural networks.

This limitation explains why AI agents—systems that actively reason under uncertainty using multiple imperfect tools—are gaining traction. Unlike AlphaFold's data-hungry approach, agents mirror how scientists already work: combining docking calculations, molecular dynamics, and assays, then iteratively refining hypotheses. Startups like Recursion ($1.2B funding) and Generate Biomedicines ($370M) are betting on this hybrid approach, where AI doesn't replace but augments human expertise. The shift from 'big data' to 'smart reasoning' could democratize AI-driven science, but it requires new standards for measurement and collaboration—areas where government funding and coordination will be critical.