Structural optimization of drug molecules with incrementally trained language models.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Automating the optimization of drug molecules for enhanced on-target potency using machine learning remains a significant challenge in chemistry.
• Chemical language models (CLMs) show promise due to their ability to learn from sequential data, enabling the design of novel molecules with specific properties.

Purpose of the Study:

• To develop and validate a machine learning strategy for automated structural optimization of drug molecules.
• To leverage chemical language models (CLMs) for de novo drug design with improved on-target potency.

Main Methods:

• A novel training strategy was established, mimicking the learning trajectory of a drug discovery program.
• Incremental fine-tuning of CLMs was performed using increasingly potent template molecules from structure-activity relationship (SAR) series.
• The technique was prospectively applied to ligand development for data-driven molecular design.

Main Results:

• Incremental CLM fine-tuning successfully biased model design towards highly active analogues.
• The data-driven approach enabled the design of molecules with potency exceeding known representatives of specific bioactive chemotypes, without external scoring.
• CLMs demonstrated an ability to capture complex SAR patterns and long-range dependencies.

Conclusions:

• CLMs can effectively exploit structure-activity relationship (SAR) knowledge to design analogues with improved on-target activity de novo.
• The findings corroborate the applicability of CLMs for the structural optimization of drug molecules in medicinal chemistry.
• This approach offers a powerful tool for accelerating drug discovery and development.