【作者】 于涵；

【导师】 杜朴风；

【作者基本信息】 天津大学 ， 计算机科学与技术， 2021， 硕士

【摘要】 非编码RNA(non-coding RNA,nc RNA)曾经被认为是基因组中的“噪音”或“暗物质”。随着越来越多不同类型的nc RNA被发现,它们的分子功能也得到了进一步研究。nc RNA在多种生物过程中扮演着关键的角色,通过与其他生物大分子之间的相互作用来实现其分子功能。其中,蛋白质是最重要的相互作用对象。发现新的nc RNA-蛋白质相互作用对于研究nc RNA的功能有着重要意义。采用实验方法鉴定nc RNA-蛋白质相互作用的成本高且耗时长,因此越来越多的计算方法作为替代方法被提出来。本文提出了一个新的预测nc RNA-蛋白质相互作用的方法:基于关系型图卷积网络自编码器的nc RNA-蛋白质相互作用预测方法(predicting nc RNA-protein interactions using the Relational Graph Convolutional Network Auto-Encoder,NPIRGCNAE)。首先,这一方法基于蛋白质序列相似度为每对nc RNA-蛋白质对计算了相互作用得分,在此基础上筛选可靠的负样本。其次,在分别构建了正负样本对应的二部图的基础上,使用关系型图卷积网络(Relational Graph Convolutional Network,R-GCN)作为编码器,同时从正负关系的二部图中学习节点嵌入。最后,NPI-RGCNAE使用Dist Mult作为解码器,将成对节点嵌入输入到解码器中来重构节点之间的关系。NPI-RGCNAE可以在较短的时间内同时聚合来自正样本网络和负样本网络中的拓扑信息用于相互作用预测。因此,NPI-RGCNAE仅需要不到其他方法10%的训练时间就获得了与其他已有方法相近的性能。NPI-RGCNAE还提出了更为可靠的负样本筛选方法,这一负样本筛选方法不仅能提高NPI-RGCNAE方法的性能,还能提高已有其他方法的预测准确率。实验结果表明,NPI-RGCNAE是一个高效、准确且稳健的nc RNA-蛋白质相互作用预测方法。本文研究工作中涉及到的数据集和编写的源代码已经存档到Github的公开仓库中(https://github.com/Angelia0hh/NPI-RGCNAE)。更多还原

【Abstract】 Non-coding RNA(nc RNA)used to be considered as "noise" or "dark matters" in the genome.However,as more and more different types of nc RNAs were discovered,molecular functions of nc RNAs have also been further studied.nc RNA plays a key role in a variety of biological processes.nc RNA achieves its molecular functions by interacting with other biomacromolecules.Among them,the most important interaction object is protein.Therefore,the discovery of new nc RNA-protein interactions is of great significance to the study of nc RNA functions.Identifying nc RNA-protein interactions by experimental methods is costly and time-consuming,so more and more computational methods are proposed as alternative methods.This paper proposed a new method for predicting nc RNA-protein interactions:NPI-RGCNAE(predicting nc RNA-protein interactions using the Relational Graph Convolutional Network Auto-Encoder).Firstly,this method calculated the interaction score for each nc RNA-protein pair based on protein sequence similarities.Reliable negative samples were screened according to the score.Secondly,on the basis of constructing the bipartite graphs corresponding to the positive and negative samples respectively,the Relational Graph Convolutional Network(R-GCN)was used as the encoder.The node embeddings were learned from the bipartite graphs of the positive and negative relations at the same time.Finally,NPI-RGCNAE used Dist Mult as the decoder.Each pair of node embeddings were input into the decoder to reconstruct the relationship between them.NPI-RGCNAE can simultaneously aggregate the topological information from the network of positive samples and negative samples for interaction prediction in a short time.Therefore,NPI-RGCNAE only needs less than 10% of the training time of other methods to obtain performance similar to other existing methods.NPI-RGCNAE also introduced a more reliable negative sample screening method.This method can not only improve the performance of NPI-RGCNAE,but also improve the prediction accuracy of other existing methods.The experimental results show that NPI-RGCNAE is an efficient,accurate and robust method for predicting nc RNA-protein interactions.All datasets and source codes involved in the research of this paper have been archived in a public Github repository(https://github.com/Angelia0hh/NPI-RGCNAE).更多还原