【作者】 李雯；

【导师】 王树林；

【作者基本信息】 湖南大学 ， 计算机科学与技术， 2023， 博士

【摘要】 生物医学和信息学领域一直致力于研究人类疾病相关问题。随着“人类基因组计划”的完成和高通量测序技术的发展,非编码RNA(ncRNA)的生物学功能逐渐被成功鉴定。在转录组分子水平探索疾病致病机制是当前生物医学领域的研究热点之一。深入研究非编码RNA与疾病、基因及蛋白质之间的潜在关联和相互作用,对挖掘疾病的发病机制,发现新的药物靶点并研发疾病的新型治疗药物具有重要意义。传统低通量生物实验周期长、风险高且成本大,且海量的生物数据通常包含的噪声、缺失值和异常值,因此挖掘和提取有价值的信息以预测研究ncRNA的潜在关联,是一个巨大的挑战。本文拟旨在利用计算机技术和机器学习手段,通过生物分子网络,研究调控性miRNA(micro RNA)和具有临床潜力的lncRNA(long non-coding RNA),预测它们与疾病及蛋白质的潜在关联和相互作用。本文主要研究工作如下:(1)为了解决现有miRNA-疾病关联(miRNA-disease association,MDA)预测模型过于依赖已知关联,相似度计算的数据来源单一且存在噪声等问题,本文提出了一种MDA预测的网络模型Tri-HM-RWR。该模型通过构建三方异构且多路复用的miRNA-基因-疾病网络来整合多视图生物数据和关联数据,通过非平衡的多随机游走方法实现关联预测。Tri-HM-RWR在预测潜在MDA的同时还能预测潜在的基因-疾病关联。通过在标准数据集上进行交叉验证和案例分析,实验结果证实了Tri-HM-RWR预测潜在MDA的有效性,同时表明该模型能有效避免对已知MDA的绝对依赖。(2)为了解决可靠的阴性样本无法获取以及矩阵收敛速度普遍较慢的问题,同时提升预测性能,本文提出了一个基于多视图数据融合和谱正则化矩阵补全的lncRNA-疾病关联(lncRNA-disease association,LDA)预测模型(MCS-SRMCLDA)。该模型采用最大余弦相似多核学习(MCS-MKL)将多视图相似度核融合为一个集成核。然后通过构建异构的lncRNA-疾病网络,并对异构网络的邻接矩阵进行基于核范式的谱正则化矩阵补全,以实现关联预测。该模型分别在三个基准数据集上进行训练、对比测试和案例分析,实验结果表明,MCS-SRMCLDA能有效预测潜在的LDA,预测性能优于对比算法。此外,该模型不需要负样本,且能准确预测缺少先验关联的新疾病或新lncRNA的潜在关联。(3)为了进一步提升预测性能,本文在MCS-SRMCLDA模型研究的基础上,提出了一个基于中心核对齐和超图逻辑矩阵分解的LDA预测模型(CKA-HRLMF)。该模型采用中心核对齐来融合多视图相似度基核,并分别为lncRNA和疾病建立超图模型,利用逻辑矩阵补全来实现潜在的关联预测。CKA-HRLMF在三个基准数据集上的实验结果表明,其运行速度和预测准确度均优于MCS-SRMCLDA,且能成功预测没有已知关联的全新lncRNA或新疾病的潜在关联。扰动测试结果表明CKA-HRLMF抗干扰能力强,具有很好的鲁棒性。案例分析结果再次验证了CKA-HRLMF预测LDA的有效性。(4)为了进一步了解lncRNA的功能和调控机制,挖掘疾病的发病机制,推动药物研发,本文针对lncRNA-蛋白质的相互作用(lncRNA-protein interaction,LPI),提出了一个基于深度学习框架的预测模型。为了突破已鉴定的LPI数据量较少,且正负样本不平衡,图节点提取的特征不够丰富的瓶颈,LPI-FKLGCN利用多视图生物数据提取多个特征及相似度基核,然后采用快速核学习(FKL)对其进行融合,并通过带注意力机制的多层图卷积编码器来获取lncRNA和蛋白质的嵌入表示向量,最后通过解码器将两组嵌入表示向量转换为一个相互作用概率矩阵,从而实现相互作用预测。实验结果表明,LPI-FKLGCN比其他基线模型的性能更佳,并且可以预测全新的关联。案例分析也进一步验证了模型预测的有效性和准确性。更多还原

【Abstract】 The fields of biomedicine and informatics have been persistently committed to researching issues related to human diseases.With the completion of the ”Human Genome Project”and the development of high-throughput sequencing technologies,the biological functions of non-coding RNA(ncRNA)have gradually been successfully identified.Exploring the pathogenic mechanisms of diseases at the transcriptomic molecular level is currently one of the research hotspots in the field of biomedicine.In-depth studies on the potential associations and interactions between non-coding RNA and diseases,genes,and proteins are of great significance for revealing the pathogenesis of diseases,discovering new drug targets,and developing new therapeutic drugs for diseases.Traditional low-throughput biological experiments have long cycles,high risks,and large costs,and the massive biological data usually contains noise,missing values,and outliers,so mining and extracting valuable information to predict the potential associations of ncRNA research is a significant challenge.This paper intends to use computer technology and machine learning methods,through the biological molecular network,to study regulatory micro RNA(miRNA)and clinically potential long non-coding RNA(lncRNA),and predict their potential associations and interactions with diseases and proteins.The main research work of this paper is as follows:(1)To address the issues that the existing miRNA-disease association(MDA)prediction models overly rely on known associations and that the data sources for similarity calculations are singular and noisy,this paper proposes a network model for MDA prediction,Tri-HM-RWR.This model integrates multi-view biological data and association data by constructing a trilateral heterogeneous and multiplexed miRNA-gene-disease network and realizes association prediction through an unbalanced multi-random walk method.TriHM-RWR can predict potential gene-disease associations while predicting potential MDAs.Through cross-validation and case analysis on standard datasets,experimental results have confirmed the effectiveness of Tri-HM-RWR in predicting potential MDAs,and demonstrated that this model can effectively avoid absolute dependence on known MDAs.(2)To address the problems of unavailability of reliable negative samples and the generally slow convergence speed of matrices,and to improve prediction performance,this paper proposes an lncRNA-disease association(LDA)prediction model based on multi-view data fusion and spectral regularization matrix completion(MCS-SRMCLDA).This model uses the maximum cosine similarity multi-kernel learning(MCS-MKL)to fuse multiple similarity kernels into an integrated kernel.Then,by constructing a heterogeneous lncRNA-disease network and completing the spectral regularization matrix based on the kernel norm of the adjacency matrix of the heterogeneous network,the model realizes association prediction.The model is trained,contrast tested,and case analyzed on three benchmark datasets respectively.Experimental results show that MCS-SRMCLDA can effectively predict potential LDAs,and its predictive performance is superior to comparative algorithms.In addition,this model does not require negative samples and can accurately predict potential associations of new diseases or new lncRNAs that lack prior associations.(3)In order to further improve prediction performance,building upon the research of the MCS-SRMCLDA model,this paper proposes an LDA prediction model based on center kernel alignment and hypergraph regularized logic matrix factorization(CKA-HRLMF).This model uses center kernel alignment to fuse multi-view similarity base kernels,establishes hypergraph models for lncRNA and diseases separately,and utilizes logic matrix completion to achieve potential association prediction.The experimental results of CKAHRLMF on three benchmark datasets demonstrate that its running speed and prediction accuracy are superior to MCS-SRMCLDA,and it can successfully predict potential associations of entirely new lncRNAs or new diseases with no known associations.The disturbance test results show that CKA-HRLMF has strong anti-interference ability and good robustness.The case analysis results re-validate the effectiveness of CKA-HRLMF in predicting LDAs.(4)In order to further understand the functions and regulatory mechanisms of lncRNA,unravel the pathogenesis of diseases,and promote drug development,this paper proposes a prediction model based on a deep learning framework targeting lncRNA-protein interactions(LPI).To overcome the bottleneck of a limited amount of identified LPI data,imbalance of positive and negative samples,and insufficiently rich features extracted from graph nodes,LPI-FKLGCN utilizes multi-view biological data to extract multiple features and similarity base kernels.It then uses fast kernel learning(FKL)to fuse them and applies a multi-layer graph convolutional encoder with an attention mechanism to obtain the embedded representation vectors of lncRNA and proteins.Finally,the decoder converts the two sets of embedded representation vectors into an interaction probability matrix,thereby realizing interaction prediction.Experimental results show that LPI-FKLGCN outperforms other baseline models and can predict entirely new associations.Case analyses also further validate the effectiveness and accuracy of the model’s predictions.更多还原