DeepRMSD+Vina

DeepRMSD+Vina optimizes ligand binding poses by combining a deep learning-based RMSD predictor with the AutoDock Vina scoring function to enable differentiable pose optimization for improved docking accuracy.

Key Features:

• Hybrid scoring function (DeepRMSD+Vina): Integrates RMSD of the docking pose relative to the native pose with the AutoDock Vina score into a single hybrid scoring function.
• DeepRMSD multi-layer perceptron: Uses a multi-layer perceptron to predict RMSD values for docking poses.
• End-to-end differentiability: Provides a fully differentiable system that permits direct optimization of ligand conformations toward energy-lowest or native poses.
• Improved affinity and pose discrimination: Leverages deep learning-based scoring functions to better predict protein-ligand binding affinities and distinguish highly similar ligand conformations.
• Benchmark performance: Demonstrated a 94.4% success rate on the CASF-2016 docking power dataset.
• Docking validations: Validated in practical molecular docking scenarios including redocking and cross-docking tasks.
• Structural interaction identification: Capable of detecting critical physical interactions in protein-ligand binding such as hydrogen bonding.

Scientific Applications:

• Ligand pose optimization: Optimization of ligand binding conformations toward global energy minima or native poses in docking studies.
• Binding affinity prediction: Enhanced prediction and ranking of protein-ligand binding affinities using deep learning-augmented scoring.
• Docking benchmarking: Evaluation of scoring function docking power using datasets such as CASF-2016.
• Redocking and cross-docking workflows: Application in redocking and cross-docking tasks to validate docking protocols and poses.
• Structural interaction analysis: Identification and analysis of physical interactions (e.g., hydrogen bonds) relevant to binding specificity.
• Computational drug discovery: Integration into molecular docking pipelines for drug design and virtual screening studies.

Methodology:

Combine a DeepRMSD multi-layer perceptron with the AutoDock Vina score into a fully differentiable hybrid scoring function using RMSD relative to native poses and Vina scores to optimize ligand binding poses; evaluated on the CASF-2016 docking power dataset and applied to redocking and cross-docking tasks.