3DMol-Net: Learn 3D Molecular Representation Using Adaptive Graph Convolutional Network Based on Rotation Invariance.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Deep learning models are crucial for predicting molecular properties, aiding drug screening and discovery.
• Molecular 3D structure features significantly impact biochemical function and activity prediction.
• Current methods often overlook 3D topological structures, limiting molecular inference performance.

Purpose of the Study:

• To enhance molecular representation by developing a model that considers both 3D topology and rotation invariance.
• To improve the accuracy of molecular property and biochemical activity predictions.

Main Methods:

• Proposed 3DMol-Net, a novel deep learning architecture.
• Constructed a molecular graph incorporating soft relations based on 3D spatial coordinates to capture 3D topology.
• Employed an adaptive graph convolutional network for property and activity prediction.

Main Results:

• 3DMol-Net demonstrated superior performance over existing graph-based methods in both regression and classification tasks.
• The model showed enhanced robustness and representation capacity, verified through rotation invariance tests and visualization.

Conclusions:

• 3DMol-Net effectively enhances molecular representation by integrating 3D topological and rotation-invariant features.
• The model offers a promising approach for advancing drug discovery and development through improved molecular property prediction.