What's next for AlphaFold: A conversation with a Google DeepMind Nobel laureate (www.technologyreview.com)

Five years after AlphaFold 2 upended structural biology, Google DeepMind’s John Jumper reflects on what the system actually changed: it solved a 50‑year “grand challenge” by using transformer neural networks and evolutionary information to predict protein folds with near-atomic accuracy, and DeepMind has since released Multimer and AlphaFold 3 and deposited roughly 200 million predicted structures into UniProt. That scale and speed have democratized access to structural hypotheses, accelerating protein design (helping groups like David Baker’s speed design cycles ~10×), enabling surprising “search‑engine” uses (screening thousands of candidate partners to find binding proteins), and powering niche studies from honeybee disease to fertilization biology. Jumper credits rapid prototyping—getting fast, wrong answers early—for enabling bold experimentation and adoption. But AlphaFold is not a panacea: its predictions are probabilistic, less reliable for protein–protein complexes, dynamics, or small‑molecule binding, and can mislead with high confidence much like LLMs. That shortfall matters for drug discovery, where sub‑angstrom errors can flip binding predictions; startups and labs (e.g., Boltz‑2, Pearl, Genesis Molecular AI) are building on AlphaFold to push errors below ~1 Å and to jointly model binding affinity. Jumper’s next aim is to fuse structure models with language models for richer scientific reasoning—foreshadowing integrated systems that propose hypotheses, check them, and close more of the loop between prediction and experiment.