AI-driven drug discovery using transformer-based molecular representation learning.

Area of Science:

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

Background:

• Vast chemical space limits drug discovery due to screening throughput and data limitations.
• Predictive models struggle with unexplored chemotypes, hindering drug development.
• Need for efficient methods to predict molecular potency and activity.

Purpose of the Study:

• To introduce a transformer-based molecular modeling framework for target-specific potency prediction.
• To enable accurate pIC50 regression and binary activity classification for drug discovery.
• To accelerate virtual screening and lead prioritization for diverse biological targets.

Main Methods:

• Utilized curated BindingDB bioactivity data for Alzheimer's, diabetes, and cancer targets.
• Developed a Byte Latent Transformer (BLT) trained directly on SMILES strings for activity prediction.
• Integrated a chemically described molecular search engine with guided SMILES mutations for optimization.

Main Results:

• Achieved highly accurate pIC50 regression (R² 0.95-0.98) and binary activity classification.
• Demonstrated framework's ability to capture chemical features and predict potency.
• Successfully generated optimized candidate molecules with preserved structural diversity.

Conclusions:

• The transformer-based framework significantly enhances target-specific potency prediction accuracy.
• The approach overcomes limitations in exploring chemical space and data extrapolation.
• Enables robust virtual screening and efficient lead prioritization in drug discovery.