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

Chunyan Li,Wei Wei,Jin Li,Junfeng Yao,Xiangxiang Zeng,Zhihan Lv

DOI: https://doi.org/10.1109/jbhi.2021.3089162

IF: 7.7

2021-01-01

IEEE Journal of Biomedical and Health Informatics

Abstract:Studying the deep learning-based molecular representation has great significance on predicting molecular property, promoted the development of drug screening and new drug discovery, and improving human well-being for avoiding illnesses. It is essential to learn the characterization of drug for various downstream tasks, such as molecular property prediction. In particular, the 3D structure features of molecules play an important role in biochemical function and activity prediction. The 3D characteristics of molecules largely determine the properties of the drug and the binding characteristics of the target. However, most current methods merely rely on 1D or 2D properties while ignoring the 3D topological structure, thereby degrading the performance of molecular inferring. In this paper, we propose 3DMol-Net to enhance the molecular representation, considering both the topology and rotation invariance (RI) of the 3D molecular structure. Specifically, we construct a molecular graph with soft relations related to the spatial arrangement of the 3D coordinates to learn 3D topology of arbitrary graph structure and employ an adaptive graph convolutional network to predict molecular properties and biochemical activities. Comparing with current graph-based methods, 3DMol-Net demonstrates superior performance in terms of both regression and classification tasks. Further verification of RI and visualization also show better robustness and representation capacity of our model.

computer science, interdisciplinary applications,mathematical & computational biology,medical informatics, information systems

What problem does this paper attempt to address?

The paper attempts to address the problem of how to effectively utilize the three-dimensional structural features of molecules in molecular representation learning. This is particularly crucial in drug screening and new drug discovery, where the 3D structure of molecules is essential for predicting molecular properties, biochemical functions, and activity. However, most existing methods rely only on one-dimensional or two-dimensional attributes, neglecting the 3D topological structure, which leads to a decline in molecular inference performance. Therefore, this paper proposes a new architecture, 3DMol-Net, which enhances molecular representation by combining Adaptive Graph Convolutional Networks (AGCN) with Rotation Invariance (RI) to better capture the 3D structural features of molecules.

Specifically, the main contributions of 3DMol-Net include:
1. **3D Graph Laplacian Matrix Based on Soft Relations**: An efficient method is proposed to capture the 3D spatial relationships between atoms in a molecule.
2. **Rotation Invariance Mapping Method**: Based on the 3D Graph Laplacian Matrix, an adaptive graph convolutional model, 3DMol-Net, is constructed to extract rotation-invariant features. This model can handle graphs of arbitrary sizes and demonstrates strong capabilities in processing non-Euclidean data.
3. **Experimental Validation**: Extensive experiments were conducted on multiple commonly used molecular datasets, and the results show that 3DMol-Net outperforms current state-of-the-art models in both regression and classification tasks, achieving better performance with fewer training parameters.

Through these innovations, 3DMol-Net can more accurately represent the 3D structure of molecules, thereby improving the accuracy of molecular property prediction and drug design.