Ask AI, or jump to a tool, chat or page.
ESM3 is a frontier generative masked language model for protein biology that jointly reasons over three tracks of discrete tokens — sequence, structure, and function. The model accepts an `ESMProtein` prompt with any combination of partial tracks (sequence, coordinates, secondary structure, SASA, function annotations) and iteratively unmasks positions on a chosen track via `generate(track=..., num_steps=...)`. This single unified interface supports forward folding (sequence -> structure), inverse folding (structure -> sequence), function prediction (sequence/structure -> InterPro function annotations), and multi-chain complex folding. ESM3 also exposes `encode`/`decode` for token conversion and a `logits` endpoint that returns per-residue logits and hidden-state embeddings for any input protein.