CryptoBench: cryptic protein-ligand binding sites dataset and benchmark

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Protein-ligand binding site prediction is vital for research and medicine.
• Existing methods often use ligand-bound (holo) protein structures, which is problematic for cryptic binding sites (CBSs).
• This reliance on holo states leads to unrealistic performance expectations for CBS detection.

Purpose of the Study:

• To introduce CryptoBench, a comprehensive benchmark dataset for training and evaluating novel CBS prediction methods.
• To establish a performance baseline for existing CBS prediction methodologies using CryptoBench.
• To compare the efficacy of sequence-based versus structure-based methods for CBS detection.

Main Methods:

• CryptoBench was constructed using apo-holo protein pairs with significant structural changes in binding sites.
• The dataset includes 1107 structures with predefined cross-validation splits.
• Sequence-based methods utilized protein language model embeddings, while structure-based methods included PocketMiner and P2Rank.

Main Results:

• The developed sequence-based approach outperformed PocketMiner and P2Rank in predicting CBS residues.
• Key metrics such as AUC, AUCPR, MCC, and F1 scores demonstrated the superiority of the sequence-based method.
• CryptoBench serves as the most extensive dataset for CBS prediction to date.

Conclusions:

• The sequence-based method provides a strong baseline for future CBS prediction research.
• CryptoBench is a foundational resource for advancing the field of cryptic binding site detection.
• The dataset and code are publicly available to facilitate further research and development.