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

Abstract Motivation Multiple sequence alignment (MSA) remains a core problem in bioinformatics, yet most widely used alignment methods still rely on static amino acid substitution matrices that cannot adapt to sequence-specific context. Results BABAPPAlign is a progressive MSA engine that replaces static substitution scoring with a trained residue-level scorer operating on fixed protein-language-model embeddings, while retaining exact affine-gap dynamic programming. It also provides an integrated codon-aware alignment mode. Using BAliBASE as the primary inferential benchmark, with supporting external validation on deterministic subsets of PREFAB and HOMSTRAD, the learned backend outperformed matched in-engine EBA-style cosine and BLOSUM62 controls, and also exceeded MAFFT. Availability and Implementation Implemented in Python. Source code: https://github.com/sinhakrishnendu/BABAPPAlign. Archived release: https://doi.org/10.5281/zenodo.17934124. Pretrained weights: https://doi.org/10.5281/zenodo.18053200. Supplementary Information Supplementary data are available online.