A protein language model for domain, functional site, and variant effect prediction

PALM, a protein annotation language model trained on ~15 million InterPro-annotated UniRef50 sequences, predicts domain and functional-site occurrence with F1 scores of 0.941 and 0.877.

• Why it matters: Growing protein sequence databases outpace curated annotations, limiting functional interpretation of new proteins and coding variants.
• What they did: PALM integrates ESM2 with LoRA, learned layer aggregation, and a BiLSTM encoder, using a soft-target objective to preserve overlapping and nested annotations. It is evaluated on 2.3 million held-out proteins and trained across 17,951 domain and 1,383 site classes.
• The result: PALM embedding disruption distinguishes 151,823 ClinVar missense variants more accurately than ESM2 embedding disruption (AUROC 0.835 vs 0.710), and combining PALM disruption with ESM2 likelihood-ratio reaches AUROC 0.877.

AI summary of the abstract · may contain errors · not medical advice · read the paper.