SynGFN: learning across chemical space with generative flow-based molecular discovery

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Artificial intelligence (AI) has accelerated the design-make-test-analyze cycle in molecular discovery.
• A key bottleneck remains the compartmentalized approach to computer-aided molecular design and synthesis.
• Optimization of the design-make-test-analyze cycle requires integrated computational and synthetic strategies.

Purpose of the Study:

• To introduce SynGFN, a novel computational framework for molecular design.
• To model molecular design as a cascade of simulated chemical reactions for building molecules from synthesizable components.
• To enhance the exploration of chemical space and identification of high-performance molecules.

Main Methods:

• SynGFN models molecular design as a sequence of simulated chemical reactions.
• It utilizes a hierarchically pretrained policy network for accelerated learning across diverse molecular distributions.
• A multifidelity acquisition framework is employed to reduce the cost of reward evaluations.

Main Results:

• SynGFN explores chemical spaces up to an order of magnitude larger than existing synthesis-aware generative models.
• The model identifies diverse, synthesizable, and high-performance molecules.
• SynGFN successfully designed inhibitors for GluN1/GluN3A, a therapeutic target for neuropsychiatric disorders.

Conclusions:

• SynGFN represents a significant advancement in synthesis-aware molecular design.
• The framework overcomes limitations of compartmentalized approaches in computer-aided molecular design.
• SynGFN demonstrates potential for accelerating the discovery of novel therapeutics.