New Method Optimizes Protein MSA Subsampling for Language Models

Protein language models rely on Multiple Sequence Alignments (MSAs) for evolutionary context, but the large size of MSAs often necessitates subsampling. Current subsampling strategies, such as random selection or diversity-driven sampling, are heuristic and offer limited control over the specific evolutionary signals retained. This research redefines MSA subsampling as an explicit optimization problem. The new approach, named AP-REASONER, utilizes an Affinity-Propagation-based factor-graph framework. It incorporates evolution-aware unary factors and exemplar-consistency factors, along with control knobs, to perform factor-graph reasoning through message passing. This process infers an MSA subset within a fixed token budget, allowing for precise control over key evolutionary measures like query identity and diversity. Experiments on tasks such as long-range contact prediction and conformational ensemble prediction demonstrate that AP-REASONER surpasses baseline subsamplers. It shows improved performance on structure-sensitive downstream tasks and enables the controllable recovery of alternative protein conformations. These findings underscore the benefits of modeling MSA subsampling as a controllable optimization problem, with factor-graph reasoning offering an effective, non-heuristic solution.