3D Molecular Optimization via a Differentiable Graph Neural Network
Keywords: 3D Molecular Optimization, Equivariant Graph Neural Network, Differentiable Graph Neural Network
TL;DR: Utilize a differentiable graph neural network to optimize 3D molecule.
Abstract: 3D Molecular optimization, also known as the structural design of functional molecules, is a crucial task in the field of chemical science and engineering, particularly in applications like drug discovery. Although deep generative models and combinatorial optimization methods have achieved some success, they still face challenges when it comes to directly modeling discrete chemical structures. Often, these approaches heavily rely on exhaustive enumeration, which can be computationally intensive. The main challenge stems from the discrete and non-differentiable nature of molecule structures. To overcome this challenge, we propose a novel approach called differentiable graph neural network (DGNN). DGNN leverages a learned 3D graph neural network to convert discrete chemical structures into locally differentiable representations. This enables gradient-based optimization on a chemical graph structure by back-propagating derivatives from target properties through a graph neural network (GNN). We conduct experiments on the QM9 dataset and verify the effectiveness of the model.
Submission Number: 14
Loading