Generative pretraining for drug molecule design with bidirectional structure-property optimization

Designing drug-like molecules that satisfy both properties requirements and structural constraints remains challenging. Current generative approaches typically reduce properties to numerical constraints and treat molecular structures as deterministic functions of properties, failing to capture complex, nonlinear structure-property relationships and limiting generative diversity and controllability. Therefore, we propose BiSP-GP, a bidirectional structure-property generative pretraining framework that unifies molecular generation and property prediction as a single sequence modeling task. BiSP-GP serializes continuous properties into semantic token sequences for joint modeling with molecular structures in a shared sequence space. Molecular generation and property prediction are cast as autoregressive sequence modeling tasks, with a cross-modal decoder supporting bidirectional mapping. The framework also incorporates scaffolds as conditional inputs to guide structure generation. Experiments demonstrate that BiSP-GP achieves strong performance gconditional molecular generation, property prediction, and downstream tasks. A case study on PAK1 further validates the model's generative ability and shows improved binding capacity in molecular docking evaluation.