HybridMolGen: a unified framework for goal-directed molecular generation via multi-objective reinforcement learning

Area of Science:

• Computational chemistry and cheminformatics.
• Artificial intelligence in drug discovery.
• Deep learning for molecular generation.

Background:

• De novo molecular design faces challenges in optimizing conflicting objectives like drug-likeness, synthetic accessibility, and novelty while ensuring chemical validity.
• Existing methods struggle to balance multiple criteria effectively in molecular generation.

Purpose of the Study:

• To introduce HybridMolGen, a unified deep learning framework for simultaneous multi-objective optimization in de novo molecular design.
• To enhance the efficiency and effectiveness of drug discovery pipelines through advanced molecular generation.

Main Methods:

• Synergistic combination of diffusion probabilistic models, SE(3)-equivariant graph neural networks, and property-conditioned transformers.
• Multi-objective reinforcement learning paradigm for discovering optimal property trade-offs without manual weight tuning.
• Enforcement of geometric and topological constraints for structural validity and molecular diversity.

Main Results:

• State-of-the-art performance on MOSES, GuacaMol, and ZINC-250k datasets, achieving 99.7% validity and 94.3% novelty.
• HybridMolGen discovers 1.57x more molecules satisfying all target criteria and generates 2.23x more Pareto-efficient solutions compared to traditional methods.
• Comprehensive ablation studies show the three-way integration outperforms two-component combinations by 6.5%.

Conclusions:

• HybridMolGen demonstrates significant architectural synergy, outperforming component aggregation and traditional scalarization methods.
• The framework offers a powerful tool for accelerating drug discovery by enabling efficient, multi-objective molecular design.
• The unified approach successfully addresses key challenges in generating chemically valid and functionally optimized molecules.