Fine-Tuning a Transformer Model for METTL3 Lead Optimization.

Area of Science:

• Medicinal Chemistry
• Machine Learning
• Drug Discovery

Background:

• Transformers, initially for natural language processing, are being adapted for drug discovery.
• A transformer model pre-trained on the ChEMBL database provides a foundation for medicinal chemistry applications.
• Methyltransferase-like protein 3 (METTL3) is a target for therapeutic intervention.

Purpose of the Study:

• To fine-tune a pre-trained transformer model for enhancing the affinity and metabolic stability of METTL3 inhibitors.
• To validate the model's performance retrospectively using known METTL3 inhibitors.
• To prospectively optimize both binding affinity and metabolic stability simultaneously.

Main Methods:

• Fine-tuning a pre-trained transformer model on a dataset of ~500 METTL3 inhibitors with known binding affinities.
• Retrospective analysis to validate model performance.
• Prospective application involving multi-objective optimization for binding affinity and metabolic stability.

Main Results:

• The fine-tuned model successfully improved METTL3 inhibitor properties.
• Retrospective analysis confirmed the model's predictive capabilities.
• Prospective optimization yielded two inhibitors with low-nanomolar potency and improved stability compared to the lead compound.

Conclusions:

• Transformer models can be effectively customized for medicinal chemistry tasks.
• Multi-objective optimization using transformers shows promise for accelerating drug discovery.
• The developed approach identified promising METTL3 inhibitors with enhanced potency and stability.