基于神经网络的分子性质预测算法研究进展

分子性质预测是材料化学领域的热点问题，基于第一性原理的计算方法虽然可以明确地描述体系中电子分布，但计算过程过于复杂，且计算复杂度随分子中原子增加呈指数级增长。近年来，随着相关研究的不断深入，涌现出多种多样的深度学习算法，将算法分为基于多层感知机(multi-layer perceptron, MLP)和图神经网络(graph neural network, GNN)两大类及六个子类，研究不同算法的特点。分析表明，MLP类算法结构简单，算法扩展性有限，与分子内部结构关联度不高；相反，GNN类算法融合消息传递机制，将分子间相互作用转换为结点、边之间的特征传递，在各向评价指标中占优。目前，基于深度学习的分子性质预测算法正从MLP类算法向GNN类算法过度。最后，提出基于深度学习的分子性质预测算法未来在数据集、各向异性特征传递、指导材料科学与生命科学中的实际应用等方面的发展方向。

Overview of Molecular Property Prediction Algorithms with Neural Network

Molecular property prediction is a hot issue in the field of material chemistry. Although the calculation method based on the first principle can clearly describe the electron distribution in the system, the calculation process is too complex, and the calculation complexity increases exponentially with the increase of atoms in the molecule. In recent years, with the deepening of related research, a variety of deep learning algorithms have emerged. In this paper, the algorithms are divided into two categories based on multi-layer perceptron (MLP) and graph neural network (GNN) and six sub categories, and the characteristics of different algorithms are studied. The analysis shows that MLP algorithm has simple structure, limited expansibility and low correlation with the internal structure of molecules; On the contrary, GNN class algorithms integrate the message passing mechanism, and transform the interaction between molecules into the feature transfer between nodes and edges, which is superior in the evaluation indexes of all directions. At present, the molecular property prediction algorithms based on deep learning is developing from MLP algorithm to GNN algorithm. Based on the above analysis, this paper puts forward the development direction of molecular property prediction algorithm based on deep learning in data set, anisotropic feature transfer, guiding the practical application in material science and Life Science in the future.