RosENet

RosENet predicts absolute binding affinities of protein–ligand complexes by integrating voxelized molecular mechanics energies and molecular descriptors with a three-dimensional convolutional neural network to support structure-based drug discovery.

Key Features:

• Voxelized energy representation: Converts molecular mechanics energies into three-dimensional voxel grids suitable for convolutional neural network input.
• Molecular descriptors from force fields: Incorporates physico-chemical molecular descriptors derived from molecular force fields.
• 3D Convolutional Neural Network: Employs a three-dimensional CNN architecture to learn spatial relationships between energy and descriptor voxels.
• Absolute binding affinity prediction and performance: Predicts absolute binding affinities and reported a Root Mean Square Error (RMSE) of 1.26 on the PDBBind v2016core dataset.
• Validation across experimental and virtual datasets: Evaluated on nearly 500 structures, including NMR-determined structures and virtual screening experiments, demonstrating robustness across diverse datasets.
• Enables rapid screening: Voxelization and CNN processing transform energy landscapes into formats that facilitate rapid screening of potential lead compounds.
• Dataset analysis: Identifies limitations within the PDBBind dataset relevant to binding affinity prediction.

Scientific Applications:

• Structure-based binding affinity prediction: Predicts absolute binding affinities for protein–ligand complexes to inform drug discovery and lead optimization.
• Virtual screening: Supports high-throughput screening of candidate compounds using voxelized energy representations and 3D CNN inference.
• Framework extensibility: Provides a framework that can be extended to incorporate molecular dynamics simulations and other biophysical or biochemical model features.

Methodology:

Voxelization of molecular mechanics energies and derivation of molecular descriptors from molecular force fields are used as input to a three-dimensional convolutional neural network; evaluation was performed on the PDBBind v2016core dataset (RMSE 1.26) and on nearly 500 structures including NMR and virtual screening experiments.