Genomic and protein language models: Evo 2, Enformer, ESM

ESM-2, from Lin and colleagues at Meta, is a protein language model whose embeddings predict atomic-level structure and are reused as gene tokens by cell models such as UCE. Enformer, from Avsec and colleagues at DeepMind, predicts expression and chromatin signals from 200 kilobases of DNA by combining convolutions with transformer attention. Evo 2 is a 40-billion-parameter genomic model on the StripedHyena 2 architecture with a one-million-base context; its zero-shot variant-effect claims are the kind of result that must be checked on clinical variant sets, and at least one independent check found them wanting.