【作者】 李文杰；

【导师】 王晗；

【作者基本信息】 东北师范大学 ， 计算机应用技术， 2025， 硕士

【摘要】 跨膜蛋白结合口袋是小分子药物与跨膜蛋白之间发生交互的实际结构区域,精准预测结合口袋对药物研发和蛋白质功能解析具有重要的科学意义。跨膜蛋白在细胞膜中发挥着重要的生物学功能,然而跨膜蛋白具有不同于水溶蛋白的生理特性,难以使用传统实验手段测定空间结构及其结合作用,相应的结合口袋样本量不足,由此基于跨膜蛋白结合口袋的分子作用研究是当前热点科学和应用问题。课题组前期研究人员曾提出蛋白质交互结构域是蛋白质与其他物质发生特异性结合的特殊局部结构区域,并已在前期工作证明此类区域的确存在空间结构共性。此结论作为本文克服跨膜蛋白小样本问题,为有效提升跨膜蛋白结合口袋预测性能提供理论基础。针对目标科学问题,本文设计实现迁移学习方法,将蛋白质交互结构域特征学习迁移用于跨膜蛋白结合口袋目标问题,主要进行如下研究:首先,设计并实现高精度的跨膜蛋白交互结构域预测模型。通过局部空间区域球采样算法,构建了高质量交互结构域样本集,设计了基于图卷积神经网络的模型框架,成功实现对跨膜蛋白交互结构域的精确预测,为实现迁移学习获取足够的先验知识。其次,将所获得的交互结构域特征通过迁移学习的方法,设计并实现有效的跨膜蛋白结合口袋预测模型。通过计算TM-score衡量样本与真实结合口袋的结构相似度,从而提高预测精度与可解释性。通过系统消融和数据消融的实验结果表明迁移学习策略显著提高了结合口袋的预测性能;通过与多个主流方法在独立测试集上进行对比,显示出本研究所提出的跨膜蛋白结合口袋预测模型在泛化能力上的明显优势。通过差异性分析进一步验证了交互结构域模块的引入为结合口袋预测提供了通用的先验知识。综上所述,本研究通过构建跨膜蛋白交互结构域和结合口袋数据集,结合图卷积神经网络和迁移学习技术,成功实现在有限数据集上相较于目前已有方法,表现出更优的预测性能。本研究不仅为药物靶点的识别提供了新的方法,同时也为未来在蛋白质功能研究、药物开发以及疾病机制解析等领域提供了技术支持和理论基础。更多还原

【Abstract】 Transmembrane(TM)protein binding pockets represent the actual structural regions where small-molecule drugs interact with TM proteins.Accurate prediction of these binding pockets holds significant scientific value for drug discovery and protein function analysis.TM proteins play crucial biological roles within the cell membrane;however,their unique physiological properties,distinct from those of soluble proteins,make it challenging to determine their spatial structures and binding interactions using conventional experimental techniques.As a result,the availability of binding pocket samples is limited,making molecular interaction studies based on TM protein binding pockets a key scientific and practical challenge.The previous researchers in our lab previously proposed that protein interaction domains are specific local structural regions where proteins engage in selective binding with other molecules.In earlier studies,we demonstrated that these regions exhibit common spatial structural characteristics.This conclusion provides a theoretical basis for addressing the small-sample challenge of transmembrane proteins and enhancing the predictive performance of their binding pocket identification.To tackle this scientific problem,this study develops and implements a transfer learning approach,wherein feature representations of protein interaction domains are transferred to the prediction of TM protein binding pockets.The main contributions of this work are as follows.First,we designed and implemented a high-accuracy prediction model for TM protein interaction domains.Using a local spatial spherical sampling algorithm,we constructed a high-quality dataset of interaction domains.A graph convolutional neural network(GCN)-based model framework was designed to achieve precise predictions of TM protein interaction domains,enabling the acquisition of sufficient prior knowledge for transfer learning.Second,we transferred the extracted interaction domain features to the prediction of TM protein binding pockets via a transfer learning strategy.The TM-score was employed to assess structural similarity between predicted binding pockets and experimentally validated pockets,thereby enhancing prediction accuracy and interpretability.Systematic ablation and data ablation experiments demonstrated that the transfer learning strategy significantly improved binding pocket prediction performance.Furthermore,comparisons with multiple state-of-the-art methods on independent test datasets confirmed that our proposed TM protein binding pocket prediction model exhibits superior generalization ability.Differential analysis further validated that the incorporation of the interaction domain module provides generalized prior knowledge for binding pocket prediction.In summary,this study constructed a dataset of transmembrane protein interaction domains and binding pockets,and successfully integrated graph convolutional networks with transfer learning techniques to achieve superior predictive performance compared to existing methods on a limited dataset.Our work not only provides a novel approach for drug target identification but also establishes a technical and theoretical foundation for future research in protein function analysis,drug development,and disease mechanism exploration.更多还原