CMGN: a conditional molecular generation net to design target-specific molecules with desired properties

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Medicinal chemistry

Background:

• Rational drug design is complex, requiring molecules with specific properties for target engagement.
• Generative neural networks offer potential for inverse drug design but face challenges in generating biologically active molecules with predefined properties.
• Existing methods struggle to simultaneously optimize for biological activity and desired drug-like properties.

Purpose of the Study:

• To introduce a novel conditional molecular generation net (CMGN) for efficient and accurate drug candidate generation.
• To leverage large-scale pretraining and fine-tuning for enhanced molecular understanding and property prediction.
• To demonstrate CMGN's capability in navigating chemical space for targeted drug discovery.

Main Methods:

• Developed CMGN, a bidirectional and autoregressive transformer-based model.
• Employed large-scale pretraining for molecular representation learning.
• Fine-tuned the model on specific target datasets and incorporated fragment-based property learning.
• Utilized fragment-growth control mechanisms to navigate chemical space.

Main Results:

• CMGN successfully generated molecules with desired properties for specific targets.
• The model demonstrated proficiency in fragment-to-lead processes.
• Case studies confirmed CMGN's utility in multi-objective lead optimization.
• Learned structure-property relationships through fragment and property recovery tasks.

Conclusions:

• CMGN effectively addresses the challenge of generating molecules with both biological activity and desired properties.
• The model shows significant potential to accelerate the early stages of the drug discovery pipeline.
• CMGN offers a powerful tool for navigating complex chemical spaces in pursuit of novel therapeutics.