Scalable Multiple Sequence Alignment via Genetic Algorithms and localized Deep Reinforcement Learning Agents

Multiple Sequence Alignment (MSA) is a fundamental NP-hard problem in Bioinformatics, central to numerous sequence analysis tasks. Despite extensive research, existing approaches still struggle to achieve optimal alignment accuracy. Recently, Deep Reinforcement Learning (DRL) approaches have shown promise in addressing these limitations. However, their scalability remains limited by the high computational cost and training time required for large-scale MSA tasks. The proposed MSA approach integrates bio-inspired optimization with adaptive learning. A Genetic Algorithm (GA) serves as a high-level “orchestrator”, reformulating alignment as an evolution-driven optimization process. At each evolutionary step, multiple localized reinforcement learning agents generate high fidelity sub-alignments that are then merged into a globally consistent solution. This hybridization of stochastic evolutionary search and policy-driven learning prevents the need to retrain DRL models on extensive datasets, while simultaneously enhancing alignment precision and computational scalability. Preliminary experimental results confirm the effectiveness of the proposed approach, which achieves higher pairwise sum scores across multiple benchmark datasets, highlighting its robustness and competitive advantage in sequence alignment.