Predicting the accuracy of ligand overlay methods with Random Forest models.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Molecular overlay methods are crucial for predicting ligand binding modes.
• Previous studies utilized decision trees for template selection.
• Optimizing overlay accuracy remains a key challenge in computational drug design.

Purpose of the Study:

• To extend Random Forest (RF) modeling for selecting optimal overlay templates and algorithms.
• To enhance the accuracy of binding mode prediction in molecular docking.
• To assess the predictive power of RF models using diverse structural data.

Main Methods:

• Utilized an extensive dataset of 815 ligand-bound X-ray structures.
• Generated approximately 70,000 molecular overlays using ROCS, FlexS, Phase, and FieldCompare.
• Trained RF models incorporating ligand/protein similarity and Lipinski descriptors.
• Employed protein sequence similarity to enable template selection for related proteins.

Main Results:

• RF model scores accurately predicted overlay accuracy.
• Automated template and method selection using RF achieved 57% accuracy for unseen ligands.
• Protein sequence similarity allowed successful predictions for proteins without X-ray structures.

Conclusions:

• Random Forest modeling significantly enhances binding mode prediction accuracy.
• The developed RF approach offers a robust method for template and algorithm selection.
• This strategy improves the reliability of molecular overlays, aiding in structure-based drug design.