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

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Designing molecules with desired properties and structures is complex.
• Current methods struggle with nonlinear structure-property relationships, limiting diversity.

Purpose of the Study:

• To develop a unified framework for molecular generation and property prediction.
• To improve generative diversity and controllability in drug design.

Main Methods:

• Proposed BiSP-GP, a bidirectional structure-property generative pretraining framework.
• Serialized properties into semantic token sequences for joint modeling with molecular structures.
• Utilized a cross-modal decoder for bidirectional mapping and scaffold-guided generation.

Main Results:

• BiSP-GP demonstrated strong performance in conditional molecular generation and property prediction.
• Achieved success in downstream tasks and improved binding capacity in molecular docking evaluations.
• A case study on PAK1 validated the model's generative capabilities.

Conclusions:

• BiSP-GP effectively unifies molecular generation and property prediction as a single sequence modeling task.
• The framework enhances the design of drug-like molecules by addressing limitations of previous generative approaches.
• BiSP-GP shows promise for accelerating drug discovery and development.