A fully differentiable ligand pose optimization framework guided by deep learning and a traditional scoring function.

Area of Science:

• Computational chemistry
• Molecular modeling
• Drug discovery

Background:

• Machine learning and deep learning scoring functions (SFs) show promise for predicting protein-ligand binding affinities.
• Differentiating similar ligand conformations, including native poses, remains a challenge in molecular docking.

Purpose of the Study:

• To develop a fully differentiable, end-to-end framework for optimizing ligand binding poses.
• To enhance molecular docking accuracy by improving ligand conformation prediction.

Main Methods:

• Proposed a hybrid scoring function, DeepRMSD+Vina, combining a multi-layer perceptron (DeepRMSD) with AutoDock Vina.
• The framework is fully differentiable, enabling optimization towards the energy-lowest conformation.
• Utilized the CASF-2016 dataset for evaluation.

Main Results:

• Achieved a 94.4% success rate on the CASF-2016 dataset, outperforming existing SFs.
• Demonstrated effectiveness in redocking and cross-docking tasks.
• Identified key physical interactions, like hydrogen-bonding, in protein-ligand binding.

Conclusions:

• The DeepRMSD+Vina framework offers a powerful new paradigm for ligand conformation optimization using deep learning.
• This approach has high potential for advancing drug design and discovery.
• The model and framework are publicly available for research.