Deciphering DEL pocket patterns through contrastive learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• DNA-encoded libraries (DELs) enable high-throughput screening of vast chemical spaces for drug discovery.
• Limited success in advancing DEL-derived compounds to clinical trials highlights the need for improved target selection.
• Current methods struggle to assess protein binding pocket flexibility and functional similarity, hindering DEL screening optimization.

Purpose of the Study:

• To develop a novel model, ErePOC, for comprehensive protein binding pocket representation.
• To improve the identification of suitable protein targets for DNA-encoded library screening.
• To enhance the success rates of DEL-based drug discovery initiatives.

Main Methods:

• ErePOC model utilizes contrastive learning with ESM-2 embeddings to capture structural and functional pocket features.
• Integration of low-dimensional physicochemical properties with high-dimensional ErePOC embeddings for comprehensive target analysis.
• Downstream classification tasks were employed to evaluate the model's precision in pocket representation.

Main Results:

• ErePOC achieved 98% precision in downstream classification tasks, demonstrating superior pocket representation capabilities.
• The model successfully identified human proteins suitable for DEL screening, uncovering enrichment across 18 functional categories.
• Analysis revealed shared characteristics among DEL targets, providing insights into DEL target space.

Conclusions:

• ErePOC provides a robust framework for enhancing DEL-based drug discovery through improved target selection.
• The model's ability to analyze pocket flexibility and functional similarity addresses limitations in current DEL screening approaches.
• This work facilitates more effective identification of drug targets and analysis of pocket similarity in DEL campaigns.