The problem. AlphaFold2 is accurate but leans on a multiple-sequence alignment per target, which is slow to build and thin for orphan or designed sequences. Could a model learn enough biology from raw sequences alone to skip the MSA entirely?

The idea. Train a large protein language model — ESM-2, scaled up to ~15B parameters — on hundreds of millions of sequences with a masked-token objective: predict the hidden amino acid from context. No structural labels. Structure emerges implicitly in the internal representations, and a folding head (ESMFold) reads those representations to predict atomic coordinates directly from a single sequence. It’s less accurate than AlphaFold2 on hard targets but roughly an order of magnitude faster, and — because there’s no MSA step — it scales to sequences evolution barely samples. They used it to fold ~600 million metagenomic proteins into the ESM Metagenomic Atlas.

Why it matters. This is the paper that makes the “proteins are a language” analogy literal, and it maps directly onto the LLM machinery I already work with in Bench: tokens (residues), self-supervised pretraining, emergent structure in embeddings, and the scale-versus-accuracy trade you make when you want to run over everything rather than one target well. The representations themselves — not just the folds — are the reusable asset, exactly like text embeddings.

Verdict. Important as an idea and as infrastructure for scale, with an honest ceiling: on individual well-aligned targets AlphaFold2 still wins on accuracy, and single-sequence folding degrades on the hardest cases. The right read is “different tool for a different regime,” not “replacement.” A clean conceptual pairing with AlphaFold2 and with the inverse-design models.