Protein sequence domain annotation using a language model

Abstract

Protein domain annotation underlies large-scale functional inference and is commonly performed by scanning sequences against libraries of profile hidden Markov models (profile HMMs). We describe PSALM, a protein domain annotation method that combines (i) a pretrained protein language model (ESM-2) with (ii) a per-residue domain-state classifier and (iii) a structured probabilistic decoder that produces a single, non-overlapping set of domain calls with explicit boundaries and scores. On a benchmark of 89M protein sequences with 107M annotated domains, PSALM attains a domain-detection sensitivity-specificity tradeoff comparable to HMMER. We characterize sequence and residue-level coverage on UniProtKB, observing higher coverage for HMMER at stringent expected false positive counts (E-values) and higher coverage for PSALM at relaxed E-values. We release code for data processing, training, and inference, along with the model weights and datasets used for training, validation, and benchmarking.