Optimizing multiple objective properties while satisfying structural constraints is a major challenge in 3D molecular discovery. This difficulty arises because optimization objectives can be non-differentiable and the structure-property relationship is often unknown. Evolutionary algorithms (EAs) are widely used for multi-objective optimization to find Pareto fronts and can naturally handle structural constraints without any explicit modelling; however, in the 3D molecular space they lack mechanisms to guarantee chemical validity and are therefore prone to producing invalid structures. Conversely, diffusion models excel at generating chemically valid 3D molecules but typically require modifying the model and retraining to incorporate structural constraints. Moreover, diffusion models are not inherently designed for direct multi-objective optimization and struggle to explore the Pareto front of the learned property distribution - a critical capability for discovering novel, high-performing molecules. To bridge this gap, we propose a novel 3D molecular multi-objective evolutionary algorithm that leverages the generative power of a pretrained diffusion model. Instead of manipulating molecules directly in the complex chemical space, our method performs crossover operations in the noise space defined by the diffusion model's forward process, thereby enabling parental features or desired fragments to be fused into offspring. The pretrained model's denoising process then restores structural validity. The approach is highly composable and, requiring no retraining, can be readily integrated with existing guidance methods to improve discovery. Experimental results demonstrate strong performance on single-objective, multi-objective, and structurally constrained optimization tasks.
Sun et al. (Fri,) studied this question.