【作者】 吴昊；

【导师】 姜海涛；

【作者基本信息】 山东大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 酶促反应在生物体内起着关键作用,加速代谢过程和生物化学反应。它们不仅调节细胞内的化学平衡,还参与身体各种生理功能,如消化、呼吸和免疫等。因此,对酶催化效率的理解和预测对于医学、生物工程和药物设计等领域具有极其重要的意义。酶的周转常数(Kcat)是衡量酶的催化效率的重要指标,传统的通过实验技术获取周转常数的方式耗时耗力。随着人工智能技术的发展,高通量计算酶周转常数的方法成为可能。利用计算方法预测周转常数的流程一般为对小分子和蛋白质进行特征提取,通过机器学习或者深度学习模型完成回归预测。然而,现有的预测方法对于表征底物的人工设计的分子指纹特征比较宽泛,而且酶和底物的交互缺少针对性。考虑到酶促反应的生物催化过程,如何使用专门针对底物设计的特征表征方法,挖掘酶催化活性的结构信息,设计契合预测酶周转常数的模型是亟待解决的问题。针对以上分析,本文提出一种新的基于图注意力网络和Transformer模型的酶的周转常数预测方法,称之为GAT-Trans-Kcat。首先,通过分子图的形式来表示参与反应的底物,利用图注意力网络更好地表征底物的官能团信息,通过信息传递图神经网络捕获契合表征酶催化过程的底物信息。其次,使用大规模蛋白质语言预训练模型ESM获取表征蛋白质的进化和结构特征。考虑到酶的催化过程是酶和底物的结合和反应,本文通过引入交叉注意力机制,将酶的活性部位信息传递给底物,实现两者的交互。最后,将蛋白酶和底物的向量表征进行拼接,通过前馈神经网络预测酶的周转常数。酶-底物基准数据集的实验表明,GAT-Trans-Kcat能够有效预测酶的周转常数,并且在均方根误差、R2决定系数、平均绝对误差和皮尔逊相关系数四个指标的表现上更为优秀。在底物特征提取任务和蛋白质特征提取任务上的对比结果也验证了本模型的合理性。为了更好地理解GAT-Trans-Kcat,本文还设计了长度鲁棒性实验和线性回归实验对模型进行可解释性分析。最后,本文输出了交叉注意力层权重矩阵的热力图来证明模型能够更好地捕捉反应物和催化酶的结合位点。总而言之,本文不仅验证了 GAT-Trans-Kcat的有效性,还展现了其在模拟酶促反应过程中的可解释性,为蛋白酶周转常数预测提供一定借鉴。更多还原

【Abstract】 Enzymatic reaction play a key role in living organisms,accelerating metabolic processes and biochemical reactions.They not only regulate chemical balance within cells,but also participate in various physiological functions of the body,such as digestion,respiration,and immunity.Therefore,the understanding and prediction of enzyme catalytic efficiency is extremely important for fields such as medicine,bioengineering,and drug design.The enzyme turnover constant(Kcat)is an important indicator to measure the catalytic efficiency of enzymes.The traditional way of obtaining turnover constants through experimental techniques is time-consuming and labor-intensive.With the development of artificial intelligence technology,highthroughput calculation of enzyme turnover constants has become possible.The process of predicting turnover constants using computational methods is generally to extract features from small molecules and proteins and perform regression predictions through machine learning or deep learning models.However,existing prediction methods are relatively broad for the artificially designed molecular fingerprint features that characterize substrates,and the interactions between enzymes and substrates lack specificity.Considering the biocatalytic process of enzymatic reactions,how to use feature characterization methods specifically designed for substrates,mine the structural information of enzyme catalytic activity,and design a model that is suitable for predicting enzyme turnover constants is an urgent problem to be solved.In view of the above analysis,this paper proposes a new enzyme turnover constant prediction method based on graph attention network and Transformer model,called GAT-Trans-Kcat.Firstly,the substrate involved in the reaction is represented in the form of molecular diagrams,the functional group information of the substrate is better characterized by the graph attention network,and the substrate information suitable for the enzymatic catalytic process is captured by the information transfer graph neural network.Secondly,the large-scale protein language pre-trained model ESM was used to obtain the evolutionary and structural features of the proteins.Considering that the catalytic process of enzyme is the combination and reaction of enzyme and substrate,this paper introduces the cross-attention mechanism to transfer the information of the active site of enzyme to the substrate,so as to realize the interaction between the two.Finally,the vector representations of protease and substrate were spliced to predict the turnover constant of the enzyme by feedforward neural network.Experiments on enzyme-substrate reference data show that GAT-Trans-Kcat can effectively predict the turnover constant of enzymes,and has better performance on four indexes:root mean square error,R2 determination coefficient,mean absolute error and Pearson correlation coefficient.The comparison between the substrate feature extraction task and the protein feature extraction task also verified the rationality of the model.In order to better understand GAT-Trans-Kcat,length robustness experiment and linear regression experiment are also designed to analyze the interpretability of the model.Finally,the thermal map of the weight matrix of the cross attention layer is output to prove that the model can better capture the binding site of the reactant and the catalytic enzyme.In conclusion,this paper not only verified the effectiveness of GATTrans-Kcat,but also demonstrated its interpretability in the process of simulating enzymatic reaction,which provided some reference for the prediction of protease turnover constant.更多还原