融合序列和结构特征的图神经网络的分子性质预测方法  

Molecular property prediction method of graph neural network based on fusion of sequence and structural features

作　　者：孟庆洁 杨东旭 逯洋[1] MENG Qingjie;YANG Dongxu;LU Yang(College of Mathematics and Computer,Jilin Normal University,Siping 136000,China)

机构地区：[1]吉林师范大学数学与计算机学院,吉林四平136000

出　　处：《湖南文理学院学报（自然科学版）》2024年第4期12-18,56,共8页Journal of Hunan University of Arts and Science(Science and Technology)

基　　金：吉林省自然科学基金项目(YDZJ202301ZYTS157);吉林省创新创业人才项目(2023QN31);吉林省发展和改革创新项目(2021C038-7)。

摘　　要：高效准确地预测分子性质能够帮助实验人员快速评估和预测新材料性能,从而节省大量实验时间和成本。然而传统实验方法存在计算成本高且精度低等问题,近年来图神经网络的引入极大提高了预测精度,但现有的模型也存在着参数量大和键信息提取不足等问题。由此提出了一种新的图神经网络模型,该模型结合改进的三头注意力机制和双向长短期记忆网络,分别将结构和序列信息相融合来更全面捕获分子的特征信息,提取的信息更为全面。模型在QM9数据集上进行了验证,实验结果表明,该模型的MAE比基准模型降低了1.95%,且参数量对比基准模型降低了30倍,大大降低了预测的时间成本,从而证明了所提出的图神经网络模型的有效性。Efficient and accurate prediction of molecular properties can help experimenters quickly evaluate and predict the performance of new materials,thereby saving a lot of experimental time and cost.However,problems like high computational costs and low accuracy exist in traditional experimental methods.In recent years,the introduction of graph neural networks has greatly improved prediction accuracy.However,there are also problems in existing models,such as large parameter quantities and insufficient extraction of key information.So,a new graph neural network model that combines an improved three head attention mechanism and a bidirectional long short-term memory network was proposed to comprehensively capture molecular feature information by fusing structural and sequence information,resulting in a more comprehensive extraction of information.The model was validated on the QM9 dataset,and the experimental results showed that the MAE of the model was reduced by 1.95%compared to the benchmark model,and the number of parameters compared to the benchmark model has been reduced by 30 times,greatly reducing the time cost of prediction,thus proving the effectiveness of the proposed graph neural network model.

关 键 词：分子性质预测 图神经网络 图注意力机制 双向LSTM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]