MSA Transformer

MSA Transformer encodes protein multiple sequence alignments to produce structure-aware sequence embeddings, tied row attention maps, and contact probability matrices. The API accepts aligned protein MSAs (up to 1024 positions × 256 sequences per item, batch size 4) and returns mean or per-position embeddings from selected layers, plus optional row attentions and contact maps. This enables analysis of coevolutionary constraints and extraction of features for downstream structure, fitness, and function modeling.