PALM adds domain and functional-site annotation to ESM2 and scores missense variants zero-shot, beating AlphaMissense on a literature set
- PALM jointly predicts domain and functional-site occurrence from sequence, keeping overlapping and nested annotations through a soft-target objective; it builds on ESM2 with low-rank adaptation and a recurrent encoder.
- It was trained on roughly 15 million InterPro-annotated sequences and kept good performance on proteins with low identity to the training data, beating an ESM3 annotation baseline.
- Without any pathogenicity labels, disruption of its embeddings separated ClinVar missense variants better than ESM2 alone, and a combined score outperformed AlphaMissense on a literature-derived set absent from ClinVar.
💡 Points Based on the abstract
Variant effect predictors tend to look strongest on ClinVar because the labels and the training signals overlap, so the literature-derived set outside ClinVar is the number to watch. For a target team, the site-level annotation is the immediately usable part: it can flag whether a missense found in a patient cohort falls in a binding or catalytic site. Whether it handles variants in intrinsically disordered regions is not addressed in the abstract.
A protein language model for domain, functional site, and variant effect prediction ↗