Improving MSA Estimation through Adaptive Weight Vectors in MOEA/D

Abstract

Accurate phylogenetic inference from biological sequences depends critically on the quality of multiple sequence alignments, yet optimal alignment for many sequences is computationally intractable and sensitive to scoring choices. In this work we introduce MOEA/D-ADF, a novel variant of MOEA/D that adaptively adjusts subproblem weight vectors based on fitness variance to improve the exploration-exploitation trade-off. We combine MOEA/D-ADF with PMAO (PASTA with many application-aware optimization criteria) to form PMAO++, where PMAO-generated solutions are used to seed MOEA/D-ADF, which then evolves a population using 30 weight vectors to produce a diverse ensemble of alignment-tree pairs. PMAO++ outperforms the original PMAO on a majority of benchmark cases, achieving better false-negative (FN) rates on 12 of 17 BAliBASE-derived datasets and producing superior best-case trees, including several instances with zero FN rate. Beyond improving single best alignments, the rich set of alignment-tree pairs produced by PMAO++ is especially valuable for downstream summary methods (for example, consensus and summary-tree approaches), allowing more robust phylogenetic inference by integrating signal across multiple plausible alignments and trees. Certain dataset features, such as large terminal N/C extensions found in the RV40 group, remain challenging, but overall PMAO++ demonstrates clear advantages for sequence-based phylogenetic analysis. Future work will explore parameter tuning, larger benchmark suites, and tighter integration with summary-tree pipelines to further enhance applicability for biological sequence studies.