基于机器学习方法的冷胁迫蛋白识别

发布时间：2018-08-07 09:59

【摘要】：冷胁迫对生物的生长和生活有着极度严重的影响,尤其是对植物来说。对植物的冷胁迫调节机制的研究,对相关的生物技术研究以及提高农作物产量等都有重要意义。当前,对植物冷胁迫蛋白的识别主要依靠人工进行,这种方式费时费力。截止目前,通过前期的资料整理,对整个拟南芥的蛋白数据库进行统计,已发现的与冷胁迫相关的蛋白仅有594条。所以,利用机器学习的方法,通过对已有的数据进行训练和预测,以为生物实验提供数据支持,具有一定的意义和研究价值。在本课题中,仅仅有正例的数据以及拟南芥的全部蛋白质序列,其中在后者中包含很多的未发现的冷胁迫蛋白序列。所以,这里首先考虑了PU Learning方法,从理论上来讲PU Learning是最合适的方法之一。将除正例数据之外的蛋白序列作为未标记的数据。尝试了PUCPI和LibD3C两种当前比较常用的PU Learning算法,但结果仅在50%上下。之后开始采用常用的分类算法进行尝试,将之前的未标记数据作为负例数据处理,并在LibSVM中得到了较好的结果。除了对分类算法的尝试,实验中还尝试了多种特征提取方法,如Pse-One、K-SkipN-Gram、Information Theory等,以及对多种特征提取方法的组合。此时,分类的准确率提高到了80%以上。在最后,一个新的负例集合被构造出来(这样可以有效的降低负例集中包含的未发现的冷胁迫蛋白的数量),并把分类准确率提高到了85%左右,取得了较好的结果。在寻找到了一个较好的冷胁迫蛋白的预测方法后,对现有的数据进行了整理并搭建了冷胁迫蛋白数据库网站。网站采用了Maven、Spring Boot,Mybatis,Mysql、VUE等主流Java Web开发技术,主要使用Java语言进行开发。为使用者提供了对拟南芥冷胁迫蛋白进行序列浏览、全文检索(Lucence)、序列比对(Blast)与分类预测等功能。
[Abstract]:Cold stress has an extremely serious effect on the growth and life of organisms, especially for plants. The study on the regulation mechanism of plant cold stress is of great significance to the study of biotechnology and the improvement of crop yield. At present, the recognition of plant cold stress protein mainly depends on manual, which is time-consuming and laborious. Up to now, only 594 proteins related to cold stress have been found in the whole Arabidopsis protein database. Therefore, it is of significance and research value to provide data support for biological experiments by training and forecasting the existing data by using the method of machine learning. In this study, there are only positive data and all the protein sequences of Arabidopsis thaliana, in which there are many undiscovered cold stress protein sequences. Therefore, the pu Learning method is considered first. In theory, pu Learning is one of the most suitable methods. The protein sequences other than the positive data are taken as unlabeled data. Two popular pu Learning algorithms, PUCPI and LibD3C, are tried, but the results are only about 50%. Then the common classification algorithm was used to process the unlabeled data as negative case data, and a good result was obtained in LibSVM. In addition to the attempt of the classification algorithm, we also try a variety of feature extraction methods, such as Pse-Oneone-K-SkipN-Gram-GramInformation Theory, as well as the combination of various feature extraction methods. At this point, the accuracy of classification improved to more than 80%. Finally, a new negative case set is constructed (which can effectively reduce the number of undiscovered cold stress proteins contained in the negative case set), and the classification accuracy is improved to about 85%, and good results are obtained. After finding a better prediction method of cold stress protein, the existing data were sorted out and the database website of cold stress protein was set up. The website adopts the mainstream Java Web development technology, such as Maven Spring Boot Java Web. It mainly uses the Java language to develop. The users can browse the cold stress proteins of Arabidopsis thaliana and search the functions of (Lucence), sequence alignment (Blast) and classification prediction.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：Q946.1;TP181

【相似文献】