LigVoxel: inpainting binding pockets using 3D-convolutional neural networks.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Structure-based drug discovery relies on protein structural data to design small molecules.
• Manually designing compounds requires chemists to intuitively place atoms within protein pockets.

Purpose of the Study:

• To propose a data-driven, structure-based approach for imaging ligands as spatial fields in protein pockets.
• To develop a deep learning framework that mimics chemists' atom placement intuition.

Main Methods:

• An end-to-end deep learning framework was trained on experimental protein-ligand complexes.
• The model generates spatial images representing ligand chemical properties (occupancy, aromaticity, donor-acceptor matching).

Main Results:

• Predicted spatial fields showed significant overlap with unseen ligands in target pockets.
• The method successfully recovered original ligand crystal poses within 2 Å RMSD in 70 out of 85 cases.
• Ligand poses were predicted by maximizing overlap between predicted fields and known ligands.

Conclusions:

• The developed deep learning models can generate spatial representations of ligand chemical properties.
• This approach shows promise for guiding structure-based drug discovery efforts.
• The LigVoxel tool is integrated into the PlayMolecule.org web application suite.