【作者】 孙宁；

【导师】 杨润涛；

【作者基本信息】 山东大学 ， 控制科学与工程， 2024， 硕士

【摘要】 通过在蛋白质合成过程中添加或移除化学基团,蛋白质翻译后修饰(Post-translational modifications,PTMs)对蛋白质结构、活性调控、亚细胞定位、稳定性及相互作用网络产生影响。PTMs类别多样。其中,丙二酰化基于丙二酰辅酶作用于生物代谢调控、炎症反应、Ⅱ型糖尿病及植物抗逆,泛素化基于多种特定的酶介导反应深刻影响细胞增殖分化、周期凋亡、肿瘤发展和神经退行性疾病。因此,丙二酰化和泛素化位点准确预测对深入理解细胞内蛋白质功能调控机制、疾病机制至关重要,同时推动新药物研究开发以及临床治疗策略优化。近年,基于机器学习和深度学习的框架逐渐成为有效预测工具。然而,现有方法受限于特征提取工程及样本的规模和类别。在特征提取时,一方面,使用单一特征建模,特征提取不全面,导致模型预测性能差且不具备鲁棒性。另一方面,忽视氨基酸基团间相互作用所反映的生物学机制,导致模型无法准确捕捉到氨基酸间的语义信息。同时,基准数据集规模和样本类别失衡严重降低模型对真实少数类样本的预测能力。因此,预测翻译后修饰位点仍存在较大挑战性。针对上述局限性,本文以蛋白质翻译后修饰中的丙二酰化和泛素化为例,通过改善特征工程和数据增强模块,搭建预测模型,主要研究工作概述如下:(1)针对现有方法只应用单一物种、预测精度较低的问题,从提取氨基酸序列语义特征和解决样本严重失衡两个角度出发,提出基于生成对抗思想和卷积神经网络的跨物种丙二酰化位点预测模型(Pred-Mal)。基于氨基酸序列与人类语言的相似性,通过词嵌入方法提取氨基酸基团语义特征,并增设自注意力模块提高有效氨基酸基团的关注度,进而捕获氨基酸序列全局信息。针对样本类别不均衡问题,通过构建生成对抗网络扩充少数类样本,以提高预测真实阳性样本的精度。为提高特征提取能力同时减轻计算压力,使用卷积神经网络执行分类任务。在单物种和跨物种数据集上的实验结果显示,该模型通过改进的词嵌入和数据增强模块,模型AUC分别提升到0.9821、0.9837和0.9656,说明Pred-Mal是提升位点预测准确率、解决位点问题中样本类别失衡的有效工具,验证了使用生成对抗网络进行氨基酸序列增强的可能性。(2)针对泛素化基准数据集样本不均衡和Pred-Mal计算需求过大的问题,提出基于自编码和对抗思想的氨基酸序列增强策略(Dis-VAE)。首先,针对生成器采取卷积神经网络无法充分贴合氨基酸序列生成的问题,使用自编码器自动映射的真实阳性位点的特征空间,在此空间随机采样合成伪序列,以充分利用氨基酸序列隐含的特征信息。同时,结合生成对抗思想,对合成数据进行矫正,有效提升生成数据质量。为使生成的序列更接近真实分布,提升训练稳定性,使用Sliced Wasserstein Distance替换Jensen-Shannon散度。在拟南芥泛素化独立测试集上,该方法AUC可以达到0.8745。与现有拟南芥泛素化位点预测工具相比,Dis-VAE各项指标都得到显著提升,验证了通过Dis-VAE合成少数类样本的有效性。(3)针对目前集成模型中特征维度爆炸和计算资源浪费的问题,结合Dis-VAE的数据增强模块,设计拟南芥泛素化位点并行集成框架(Parallel-Ubi)。首先,为解决特征工程与分类器的适配问题,选取多维度氨基酸序列特征,包括序列特征、物理化学性质、语义信息和One-Hot编码。针对各项特征工程选取最有效的分类器,构成“特征工程-分类器”子网的并行集成分类模型。其次,通过主成分分析(Principal Component Analysis,PCA)对高维特征进行降维,提取有效特征信息,避免引入噪声干扰模型性能。之后,讨论子网集成、特征融合和分类算法集成的优劣性,分析集成方法在翻译后修饰位点通用模型中的可行性。实验结果表明,该方法在AUC上高达0.9688,验证了Parallel-Ubi预测的有效性。同时,该模型并行运行的结构大大降低了计算成本。更多还原

【Abstract】 Protein post-translational modifications(PTMs)impact protein structure,activity regulation,subcellular localization,stability,and interaction networks by adding or removing chemical groups during protein synthesis.For example,malonylation plays a pivotal role in regulating biological metabolism,modulating inflammatory responses,influencing Type Ⅱdiabetes,and enhancing plant resilience against stress through the action of malonyl coenzyme.Similarly,ubiquitination,driven by a series of specific enzyme-mediated reactions,critically affects cellular processes such as proliferation,differentiation,apoptosis,tumor progression,and the pathology of neurodegenerative diseases.Therefore,accurate prediction of malonylation and ubiquitination sites is crucial for in-depth understanding of intracellular protein function regulation mechanisms,disease mechanisms,as well as research to promote the development of new drugs and optimization of clinical therapeutic strategies.Recently,algorithms based on machine learning and deep learning have emerged as powerful tools for prediction.However,these algorithms are constrained by limitations in feature engineering and the diversity and size of the datasets.In feature extraction,on the one hand,the use of single feature modeling with incomplete feature extraction leads to poor and non-robust model prediction performance.On the other hand,the semantic information reflected by the interactions between amino acid groups is ignored,resulting in a model that fails to accurately capture the interactions between amino acids and ignores the underlying biological mechanisms.Compounded by the challenges of benchmark dataset scales and the imbalance in sample categories,these issues substantially decrease model efficacy.This,predicting post-translational modification sites remains challenging.To address these concerns,this thesis takes malonylation and ubiquitination in post-translational modification of proteins as an example,and builds prediction models by improving the feature engineering and data enhancement modules.The main research work is summarized as follows.(1)To solve the problems of applying only a single species and low accuracy on existing tools,this chapter proposes a cross-species frameworks(Pred-Mal)based on generative adversarial network and convolutional neural networks for predicting malonylation sites from the perspective of semantic characterization of amino acid sequences and severe imbalance of site samples.Leveraging the similarity between amino acid sequences and human language,semantic features of amino acid groups are extracted via word embedding techniques,and the self-attention module is introduced to enhance the focus on relevant amino acid groups,thereby Pred-Mal can capture the global information of amino acid sequences.To address the imbalance of sample class,the generative adversarial network is constructed to augment minority class samples,thereby enhancing the precision of predicting true positive samples.To improve feature extraction capabilities while reducing computational burden,a convolutional neural network is utilized to perform the classification task.Experimental results on both single-species and cross-species datasets have demonstrated that,through improved word embedding and data augmentation modules,the model AUC has been elevated to 0.9821,0.9837,and 0.9656,respectively.This indicates that Pred-Mal is an effective tool for enhancing the predicted accuracy and addressing the issue of sample class imbalance in site problems.(2)To address the imbalance in ubiquitination benchmark datasets and excessive computational requirements for Pred-Mal,this chapter introduces a novel approach(Dis-VAE)grounded in autoencoder and adversarial principles.Firstly,to solve the problem that the generator taking convolutional neural network cannot adequately fit the amino acid sequence generation,this chapter uses the potential space automatically mapped by the self-encoder to synthesize the pseudo-sequence by randomly sampling in the feature space of the true positive samples,to make full use of the feature information implied by amino acid sequences.Subsequently,by integrating generative adversarial concepts,timely corrections to the synthesized data are made,effectively enhancing the quality of the generated data.To optimize the generated sequences to resemble the true distribution and enhance training stability,Sliced Wasserstein Distance is employed in place of JS divergence more closely.On the independent test set of Arabidopsis ubiquitination,the AUC can reach 0.8745.In comparative experiments with existing Arabidopsis ubiquitination site prediction tools,all metrics of Dis-VAE have seen significant improvement,which verified the validity of synthesizing a small number of classes of samples by Dis-VAE.(3)Aiming at the problems of feature dimension explosion and waste of computational resources in the current integrated model,this chapter combines the data augmentation model of Dis-VAE to design a parallel integrated framework(Parallel-Ubi)for Arabidopsis ubiquitination,offering insights for other post-translational modification issues.To resolve the compatibility between feature engineering and classifiers,multi-dimensional representations of amino acid sequences are selected,including sequence features,physicochemical properties,semantic information,and two-dimensional representations.Then the most effective classifiers are selected for each feature project,constituting a parallel integrated classification model with a "feature project-classifier" subnetwork.Subsequently,principal component analysis(PCA)is employed to reduce dimensionality among high-dimensional features,thereby extracting pertinent feature information and reducing noise interference with model efficacy.After that this chapter meticulously evaluates the merits and limitations of subnet ensemble model,feature fusion,and classifier ensemble models,exploring their applicability to a universal model framework for post-translational modification site prediction.The experimental results show that Parallel-Ubi is as high as 0.9688 on AUC,which verifies the validity of its prediction.Meanwhile,the structure of the model running in parallel greatly reduces the computational cost.更多还原