WASNCI:一种基于多模态深度学习的NCIs计算方法  

WASNCI:A calculation method for NCIs based on multimodal deep learning

作　　者：赵恩杰 李文泽 柴旭清[1,2,3] 毛文涛 ZHAO En-jie;LI Wen-ze;CHAI Xu-qing;MAO Wen-tao(School of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China;Engineering Laboratory of Intelligent Business and Internet of Things Technology,Xinxiang 453007,China;High Performance Computing Center,Henan Normal University,Xinxiang 453007,China)

机构地区：[1]河南师范大学计算机与信息工程学院,河南新乡453007 [2]智慧商务与物联网技术河南省工程实验室,河南新乡453007 [3]河南师范大学高性能计算中心,河南新乡453007

出　　处：《化学研究与应用》2024年第9期2064-2072,共9页Chemical Research and Application

基　　金：国家自然科学基金项目(12274117)资助;河南省科技攻关项目(222102210333)资助;中国高校产学研创新基金-新一代信息技术创新项目资助课题计划书项目(2020ITA07040)资助;产学合作协同育人项目(202102089014,202102533043)资助。

摘　　要：非共价相互作用(NCIs)的识别和研究,特别是NCIs值大小的测量和计算,对于药物的设计、超分子体系以及功能材料的设计都有重要意义。本文提出基于二代小波和AE的自注意力多模态特征融合NCIs计算方法(WASNCI)。引用二代小波方法将分子的电子密度特征按多尺度分解为具有能量意义的不同频带信息,之后计算频带重要度对频带信息自适应加权,提高特征利用率。同时,使用自编码器对分子样本的基本化学性质提取特征。最后,构建自注意力特征融合模块,其多头自注意力机制捕捉两种特征的复杂关系,使特征既能表达分子的电子密度,又能充分利用分子的化学性质。该方法在公用数据集上进行实验验证。实验结果表明,与最新的计算方法Deep NCI相比,本文提出的WASNCI方法计算NCIs的RMSE降低到了0.109 kcal/mol,精度提高了42%。本文所提计算方法可以准确地对分子NCIs进行计算,为非共价相互作用的研究提供了技术支持。The identification and investigation of non-covalent interactions(NCIs),particularly the measurement and computation of NCIs values,hold significant implications for drug design,supramolecular systems,and functional material design.This paper proposes a Deep Adaptive Wavelet and Autoencoder based Self-Attention Multimodal Feature Fusion method for NCIs computation(WASNCI).The Deep Adaptive Wavelet Packet Net decomposes the electron density features of molecules into dfferent frequency band information with energy significance through multiscale analysis.Subsequently,the importance of each frequency band is calculated,and adaptive weighting is applied for further feature extraction.Simultaneously,an Autoencoder is employed to extract features related to the basic chemical properties of molecular samples.Finally,a Self-Attention feature fusion module is constructed.Its multi-head Self-Attention mechanism captures the complex relationships between features,enabling the expression of both electron density and chemical information effectively.The proposed method is experimentally validated on a public dataset.The experimental results demonstrate that compared to the state-of-the-art computational method DeepNCI,the proposed WASNCI method reduces the root mean square error of NCIs calculations to 0.109 kcal/mol,achieving a 42%improvement in accuracy.The computational approach presented in this paper accurately computes molecular NCIs,providing technical support for the study of non-covalent interactions.

关 键 词：非共价相互作用 深度学习 二代小波 多头自注意力 

分 类 号：O641[理学—物理化学]