How is AI designing proteins that never existed in nature?

The answer

Beyond AlphaFold's structure prediction, tools like RFdiffusion from David Baker's lab use diffusion models to design completely novel proteins. EvolutionaryScale's ESM-3, a 98-billion-parameter model trained on 2.78 billion proteins, generated a fluorescent protein so different from any known example that it represents the equivalent of 500 million years of natural evolution. Published in Science in January 2025, ESM-3 reasons jointly across protein sequence, structure, and function. Baker won the 2024 Nobel Prize in Chemistry for this pioneering work in computational protein design.