Protein Sequence Domain Annotation using Language Models

AO_SCPLOWBSTRACTC_SCPLOWProtein function inference relies on annotating protein domains via sequence similarity, often modeled through profile Hidden Markov Models (profile HMMs), which capture evolutionary diversity within related domains. However, profile HMMs make strong simplifying independence assumptions when modeling residues in a sequence. Here, we introduce PSALM (Protein Sequence Annotation using Language Models), a hierarchical approach that relaxes these assumptions and uses representations of protein sequences learned by protein language models to enable high-sensitivity, high-specificity residue-level protein sequence annotation. We also develop the Multi-Domain Protein Homology Benchmark (MDPH-Bench), a benchmark for protein sequence domain annotation, where training and test sequences have been rigorously split to share no similarity between any of their domains at a given threshold of sequence identity. Prior benchmarks, which split one domain family at a time, do not support methods for annotating multi-domain proteins, where training and test sequences need to have multiple domains from different families. We validate PSALMs performance on MDPH-Bench and highlight PSALM as a promising alternative to HMMER, a state-of-the-art profile HMM-based method, for protein sequence annotation.