Benchmarking Molecular Mutation Operators for Evolutionary Drug Design

This study investigates and compares different molecular mutation strategies to optimize their application as genetic algorithm operators in drug design. We evaluated five distinct mutation methods-Graph-Based Genetic Algorithm, Graph-Based Generative Model, SmilesClickChem, SELFIES Token, and SMILES Token Mutation-by assessing their computational efficiency, validity, and impact on molecular complexity and structural conservation. Our results reveal that the Graph-Based Genetic Algorithm achieves the highest molecular validity (96.5%) while maintaining computational efficiency, making it ideal for rapid iterative drug discovery. SmilesClickChem and Graph-Based Generative Model tend to increase molecular complexity, whereas SF-T simplifies molecular structures, suggesting different applications in lead optimization. Additionally, we analyzed mutation-induced changes in pIC50 potency and found that SELFIES Token caused the most substantial shifts in bioactivity, particularly in SRC-targeted molecules. These findings underscore the importance of selecting the appropriate mutation strategy to balance validity, structural diversity, and computational cost in AI-driven drug design. Our insights help refine evolutionary algorithms for molecular generation and optimize candidate selection in early-stage drug discovery.