Look Before You Leap: Improving Structure-Based Drug Optimization with Attribution-Guided Genetic Operators

Graph-based molecular property predictors are increasingly central to drug discovery, yet the atom-level information they encode remains largely unexplored for guiding molecular optimization. We propose attribution-guided site selection, a modular modification to the Graph-Based Genetic Algorithm (GB-GA) that biases crossover and mutation toward atoms estimated to have the greatest potential to improve predicted fitness, using attribution scores derived from a directed message-passing neural network. On the DOCKSTRING molecular design benchmarks, the three attribution-guided variants tested systematically outperform unmodified GB-GA across targets. The modification is computationally lightweight, requires no retraining, and is orthogonal to outer-loop optimization strategies.