Pocket contrastive learning and conformation fusion for robust structure-conditioned 3D molecular generation under receptor conformational variation.

Abstract

Motivation: Pocket-conditioned molecular generation incorporates the three-dimensional (3D) geometry and chemical environment of protein binding sites into molecular design. However, receptor conformational variation can introduce distribution shifts between training and inference, particularly under cross-docking conditions. Existing approaches often do not fully exploit the correspondence between related receptor conformations during training.

Results: We propose a structure-conditioned 3D molecular generation framework that combines pocket-level contrastive learning with reference-conformation feature fusion. Built on a stepwise constructive generative model, these training-stage components encourage stable pocket representations across related conformations while preserving single-pocket inference without requiring a reference conformation. Evaluation on CrossDocked2020 and three external datasets, ASB-E, DUD-E, and APObind, demonstrates improvements in generation success, structural validity, and docking-related performance, particularly under cross-dock evaluation. These results support improved robustness of pocket-conditioned molecular generation under receptor conformational variation.

Availability and implementation: The source code and reproduction instructions are available at https://github.com/LHHHHuan/MTMGen. Pretrained models and an archival code snapshot are available at https://doi.org/10.5281/zenodo.21944857.