ESM2 Protein Language Models
ESM2 (Evolutionary Scale Modeling 2) is a family of transformer-based language models from Meta Research, trained on 250 million protein sequences from the UniProt database. By learning the patterns of amino acid co-occurrence across evolutionary variants, ESM2 can predict which mutations are likely to maintain or improve protein function — a zero-shot capability with no training data specific to any individual protein.
Proteins as Language
A protein sequence is a string in a 20-character alphabet (one letter per amino acid). ESM2 is trained with a masked language modeling objective — it sees sequences with randomly masked positions and learns to predict the missing amino acids. This forces the model to learn co-evolutionary constraints: which amino acids appear together, which positions are conserved across species, and which substitutions evolution tolerates.
Mutation Effect Scoring
For any amino acid substitution, ESM2 computes a log-likelihood ratio: log P(mutant) / P(wildtype). Positive values indicate the mutation is evolutionarily plausible; negative values indicate it may disrupt function. This score correlates strongly with experimental fitness effects measured by deep mutational scanning across dozens of protein families.
Scoville Splice uses ESM2 as Layer 2 of IPAP — after genomic data ingestion maps target loci, ESM2 scores every codon-optimized variant of the biosynthesis enzymes (AT3, pAMT, BCAT, CS) before synthesis. Only variants above a score threshold advance to wet lab validation.