A Transformer for Reaction-Aware Compound Explorations with GFlowNet in QSAR-Guided Molecular Design.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Molecular modeling and design

Background:

• Balancing biological activity and synthetic feasibility is crucial in drug discovery.
• Existing deep learning models for molecular generation often neglect synthetic routes.
• Exploring vast chemical spaces requires efficient and practical methods.

Purpose of the Study:

• To introduce TRACE-GFN, a novel deep learning model for molecular optimization.
• To incorporate synthetic reaction pathways into quantitative structure-activity relationship (QSAR)-guided molecular design.
• To generate diverse and synthetically feasible drug candidates with high biological activity.

Main Methods:

• TRACE-GFN integrates a transformer model for learning chemical reactions with a generative flow network (GFlowNet).
• The model performs QSAR-guided molecular design, explicitly considering chemical reaction pathways.
• Benchmark experiments were conducted on dopamine receptor D2 (DRD2), AKT1, and CXCR4 targets.

Main Results:

• TRACE-GFN successfully identified compounds with high QSAR values across tested targets.
• The model demonstrated strong diversity in generated molecular candidates.
• TRACE-GFN outperformed existing molecular generation models in benchmark experiments.

Conclusions:

• TRACE-GFN enables efficient exploration of promising compounds by accounting for real-world chemical reactions.
• The model addresses the critical need for synthetic feasibility in deep learning-based drug discovery.
• The approach facilitates the discovery of novel therapeutics with improved drug-likeness.