A Bayesian statistical approach of improving knowledge-based scoring functions for protein-ligand interactions

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Knowledge-based scoring functions are crucial for predicting molecular interactions in docking and structure prediction.
• These functions suffer from sparse data, limiting their accuracy, especially for rare chemical interactions.

Purpose of the Study:

• To address the sparse data problem in knowledge-based scoring functions.
• To develop a novel method for estimating inaccuracies in scoring functions.
• To improve the accuracy of protein-ligand binding predictions.

Main Methods:

• Developed a novel approach to estimate inaccuracies in knowledge-based scoring functions.
• Integrated inaccuracy estimation to weight knowledge-based scoring functions with force-field-based potentials (FFPs).
• Applied the method to develop STScore, a protein-ligand scoring function.

Main Results:

• STScore achieved a 91% success rate in binding mode prediction for 100 complexes.
• Demonstrated a binding affinity correlation of 0.514 with experimental data from PDBbind.
• The method effectively handles sparse data and improves approximations for rare chemical groups.

Conclusions:

• The novel inaccuracy estimation method effectively overcomes sparse data limitations in scoring functions.
• The approach enhances protein-ligand interaction predictions and shows potential for broader applications.
• This method can be extended to other FFPs, knowledge-based scoring functions, protein-protein docking, and protein structure prediction.