DeepBSP-a Machine Learning Method for Accurate Prediction of Protein-Ligand Docking Structures.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Machine learning scoring functions enhance molecular docking but struggle with distinguishing native poses from decoys.
• Existing methods often lack decoy structural information in training data, limiting their effectiveness.

Purpose of the Study:

• To develop a machine learning model, DeepBSP, capable of directly predicting the root mean square deviation (rmsd) of ligand docking poses relative to their native binding poses.
• To improve the accuracy of identifying correct ligand-protein complex structures from docking simulations.

Main Methods:

• Developed DeepBSP, a machine learning model for predicting root mean square deviation (rmsd).
• Trained the model on a dataset comprising 11,925 native complexes and over 165,000 docked poses.
• Evaluated model performance on test sets and the CASF-2016 docking decoy set.

Main Results:

• DeepBSP demonstrated superior performance in distinguishing native poses from decoys compared to major scoring functions.
• The model achieved excellent docking power on both internal test sets and the CASF-2016 benchmark.
• Direct prediction of rmsd proved effective for pose evaluation.

Conclusions:

• DeepBSP significantly enhances the ability to select accurate ligand binding poses from molecular docking.
• Combining molecular docking with DeepBSP allows for more precise prediction of native-like complex structures.
• The DeepBSP model offers a valuable tool for improving the reliability of virtual screening and drug design pipelines.