Protein-Ligand Empirical Interaction Components for Virtual Screening.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Empirical scoring functions often inaccurately predict protein-ligand binding affinity.
• Filtering false positives from docking results is a significant challenge in structure-based virtual screening.
• Postdocking filters utilizing experimental data can aid in resolving this issue.

Purpose of the Study:

• To develop and evaluate a novel postdocking filtering method using machine learning.
• To improve the accuracy of structure-based virtual screening by reducing false positives.
• To create a classification model that distinguishes true from false positive docking predictions.

Main Methods:

• Detailed decomposition of protein-ligand interactions was performed.
• Protein-Ligand Empirical Interaction Components (PLEIC) were used as descriptors.
• Support Vector Machine (SVM) learning was employed to build a classification model (PLEIC-SVM).
• Model training utilized experimentally derived activity information.

Main Results:

• The PLEIC-SVM method demonstrated superior performance compared to standard empirical scoring functions.
• Extensive benchmarking on 36 diverse datasets from the DUD-E database validated the method's effectiveness.
• The trained PLEIC-SVM model successfully identified critical interaction patterns for specific targets.
• The method proved effective in discarding false positives during postdocking filtering.

Conclusions:

• The PLEIC-SVM method offers a significant advancement in structure-based virtual screening.
• This approach enhances the reliability of docking results by accurately filtering false positives.
• The model's ability to capture target-specific interaction patterns provides valuable insights for drug discovery.