Descriptions:
Two Minute Papers host Dr. Károly Zsolnai-Fehér sits down with John Jumper, the Nobel Prize-winning chemist and DeepMind researcher who led the development of AlphaFold, for an unusually deep technical conversation about one of AI’s most consequential real-world applications. Jumper explains the core scientific problem from first principles: proteins are nanoscale machines whose 3D folded structure determines their biological function, but experimentally determining that structure has historically required a year of lab work costing roughly $100,000 per protein using synchrotrons the size of small villages.
AlphaFold, a deep learning system trained on DNA sequences and known experimental structures, can now predict those 3D structures in five to ten minutes with accuracy approaching experimental results. Since deployment, it has predicted structures for approximately 200 million proteins — effectively every protein from any organism with a fully sequenced genome — and around 3 million scientists use its public database for drug development and disease research.
Jumper shares candid behind-the-scenes moments from development, including his own surprise when AlphaFold predicted proteins with seemingly impossible geometries — giant internal voids or spiral shapes that looked completely wrong — only to discover the model had correctly inferred that these proteins naturally form trimers or heterodimers in biological contexts it was never explicitly trained on. The episode stands as a rare primary-source account of how AlphaFold actually works, directly from one of its creators.
📺 Source: Two Minute Papers · Published December 02, 2025
🏷️ Format: Interview
