基于混合深度学习的蛋白质-多肽结合位点识别研究  

Investigation of Protein-Peptide Binding Sites Recognition Based on Hybrid Deep Learning

作　　者：朱琪 郑美丽 张步忠 ZHU Qi;ZHENG Meili;ZHANG Buzhong(School of Computer and Information,Anqing Normal University,Anqing 246133,China)

机构地区：[1]安庆师范大学计算机与信息学院,安徽安庆246133

出　　处：《安庆师范大学学报（自然科学版）》2024年第3期101-106,共6页Journal of Anqing Normal University(Natural Science Edition)

基　　金：安徽省高校优秀人才支持计划项目(gxyq2020029);安庆师范大学2021年度研究生学术创新项目(2021yjsXSCX014)。

摘　　要：蛋白质和多肽间的相互作用机制非常重要,且使用计算方法预测蛋白质-多肽相互作用位点能快速定位并有效降低研发成本。然而,现有模型预测两者结合位点的准确率不高,并且难以处理数据不平衡问题。基于此,本文提出了一个基于混合深度学习策略的预测方法——ResPep。在数据集处理层面上,用K-means聚类算法对数据集中多数类样本下采样以平衡数据样本;在算法处理层面上,用残差网络、一维卷积神经网络和多头注意力网络来建立轻量级的混合深度学习模型,同时模型学习融入代价敏感性。结果表明,与多个现有方法相比,ResPep在公共测试数据集TS125上拥有更好的泛化性能。The interacting mechanism between proteins and peptides is very important.Computational methods can swiftly pinpoint protein-peptide interaction sites,thereby significantly reducing research and development expenditures.However,current models cannot predict the binding sites with high precision and face challenges in addressing data imbalance issues.To deal with these problems,a novel prediction method named ResPep based on hybrid deep learning methods is proposed in this paper.In terms of dataset processing,K-means clustering algorithm is used to down-sample majority-class samples in the dataset to balance data samples.At the algorithm level,a lightweight hybrid deep learning model is proposed,comprising residual networks,one-dimensional convolutional neural networks,and multi-head attention mechanisms.Furthermore,cost sensitivity learning is integrated in ResPep model.The experimental results show that ResPep achieves better generalization performance compared with most existing methods on the public test dataset TS125.

关 键 词：蛋白质-多肽 相互作用位点识别 深度学习 不平衡数据 蛋白质序列 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]