BABAPPAlign: A Multiple Sequence Alignment Engine with a Learned Residue-Level Scoring Function

Abstract

Multiple sequence alignment (MSA) underpins comparative genomics, evolutionary analysis, and structural inference. Despite extensive methodological development, most widely used alignment algorithms rely on static amino acid substitution matrices that encode global average substitution tendencies and are inherently context-agnostic. Such models are limited in their ability to capture sequence-specific evolutionary constraints, particularly for divergent or low-signal protein families. BABAPPAlign is a multiple sequence alignment framework that replaces static substitution matrices with a learned, context-aware residue-level scoring function while preserving exact affine-gap dynamic programming. Residue compatibility is inferred using a neural scoring model, BABAPPAScore, operating on fixed contextual embeddings derived from a pretrained protein language model. Neural inference is performed entirely outside the dynamic programming recursion, ensuring that alignment optimization remains exact, deterministic, and reproducible. Evaluation on the BAliBASE benchmark using strict family-wise paired comparisons demonstrates statistically significant improvements in sum-of-pairs and total column scores over ClustalW, MUSCLE, MAFFT, and T-Coffee across 386 reference families.