Predicting binding sites from unbound versus bound protein structures.

Area of Science:

• Structural Biology
• Computational Chemistry
• Bioinformatics

Background:

• Accurate prediction of protein binding sites is crucial for drug discovery and understanding protein function.
• The choice of protein structure (ligand-bound vs. ligand-free) may influence the performance of binding site prediction algorithms.
• A comprehensive evaluation requires a diverse dataset covering various protein conformations.

Purpose of the Study:

• To evaluate the performance of seven diverse binding-site prediction algorithms.
• To investigate the impact of using ligand-bound versus ligand-free protein crystal structures on prediction accuracy.
• To assess the consistency of prediction performance across different conformations of the same protein.

Main Methods:

• Application of seven algorithms (Surfnet, Ghecom, LIGSITEcsc, Fpocket, Depth, AutoSite, Kalasanty) to a curated dataset.
• Dataset comprises 2528 crystal structures from 304 unique protein sequences, including multiple ligand-bound and ligand-free forms per protein.
• Performance was assessed using F scores and Matthew's correlation coefficients.

Main Results:

• Most algorithms showed no significant performance difference between ligand-bound (holo) and ligand-free (apo) structures.
• Fpocket demonstrated a minor performance improvement on holo structures.
• Prediction success varied significantly across different crystal structures of the same protein, irrespective of quality metrics like resolution.

Conclusions:

• The choice between holo and apo structures has minimal impact on the performance of most binding site prediction tools.
• Algorithm performance is inconsistent across different conformations of the same protein, highlighting challenges in predicting binding sites reliably.
• Protein structure quality metrics do not reliably predict the success or failure of binding site prediction algorithms.