When Simulations Meet Machine Learning: Redefining Molecular Docking for Protein-Glycosaminoglycan Systems

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Glycosaminoglycans (GAGs) are crucial extracellular matrix components with complex interactions.
• The flexibility and specificity of GAGs present significant challenges for computational modeling and drug discovery.
• Accurate prediction of protein-GAG binding poses is essential for understanding biological processes and designing therapeutics.

Purpose of the Study:

• To evaluate the efficacy of repulsive scaling replica exchange molecular dynamics (RS-REMD) and molecular mechanics generalized Born surface area (MM-GBSA) in predicting protein-GAG binding.
• To develop and assess machine learning (ML) models for improving the accuracy of binding pose prediction for flexible ligands like GAGs.
• To explore the integration of simulation data with ML for enhanced molecular docking strategies.

Main Methods:

• Implemented RS-REMD simulations with the CHARMM36m force field for seven protein-GAG complexes.
• Applied MM-GBSA to analyze binding energy components.
• Trained five ML models (FCNN, linear regression, LightGBM, Random Forest, SVR) using simulation-derived features to predict binding accuracy (RMSatd).

Main Results:

• MM-GBSA showed weak to moderate correlation with binding accuracy (RMSatd).
• Trained ML models significantly improved the selection of native-like binding poses compared to MM-GBSA alone.
• The Random Forest model demonstrated the highest accuracy in predicting native GAG binding poses.

Conclusions:

• Integrating RS-REMD simulations with ML offers a powerful approach to refine molecular docking for flexible ligands.
• Machine learning models, especially Random Forest, can effectively enhance the prediction of protein-GAG interactions.
• This combined strategy holds promise for advancing drug discovery and understanding extracellular matrix biology.