一种预测miRNA与疾病关联关系的矩阵分解算法

PDF下载HTML

Title:

A matrix factorization method for predicting miRNA-disease association

作者:

刘晓燕¹, 陈希¹, 郭茂祖^1,2, 车凯¹, 王春宇¹

1. 哈尔滨工业大学 计算机科学与技术学院,黑龙江 哈尔滨 150001;
2. 北京建筑大学 电气与信息工程学院, 北京 100044

Author(s):

LIU Xiaoyan¹, CHEN Xi¹, GUO Maozu^1,2, CHE Kai¹, WANG Chunyu¹

1. School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China;
2. School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture, Beijing 100044, China

关键词:

microRNAs; 疾病; 关联预测; 矩阵分解; 迭代最小二乘

Keywords:

microRNAs; disease; association prediction; matrix factorization; iterative least squares

分类号:

TP391

DOI:

10.11992/tis.201805043

摘要:

越来越多的证据表明microRNAs（miRNAs）在生命进程中发挥着重要作用。近年来，预测miRNAs与疾病的关联关系成为一个研究热点。然而，现有的方法大多数是基于已知的miRNA-疾病关联，对没有任何关联信息的miRNA或疾病的效果是很不理想的。本文提出了一种矩阵分解的方法LMFMDA（least squares optimization matrix factorization method for mirna-disease association）对miRNAs和疾病的关联关系进行预测。LMFMDA基于miRNAs相似度矩阵、疾病相似度矩阵和miRNAs-疾病关联关系矩阵，用迭代最小二乘法求解miRNAs和疾病的表达向量，最终利用miRNAs和疾病的表达向量完成对miRNA与疾病关联关系的预测。与常规做法不同的是，我们引入了辅助的miRNAs和疾病变量，来保证在优化时能够收敛到最优解。实验结果表明，采用留一交叉验证法得到的AUC值可达0.820 6，明显优于当前其他方法，尤其在没有任何关联信息的miRNA和疾病上，LMFMDA算法比最新的算法有了极大的提升。

Abstract:

There are increasing evidences that microRNAs (miRNAs) play an important role in life processes. In recent years, predicting the association between miRNAs and diseases has become an active topic. However, most of the existing methods are based on known miRNA-disease associations and are not ideal for miRNAs and diseases without any known associations. This paper presents a least squares optimization matrix factorization method for miRNA-disease association (LMFMDA) prediction. The LMFMDA, which is based on miRNAs similarity matrix, disease similarity matrix, and miRNAs-disease relationship, uses the iterative least squares method to solve the expression vectors of miRNAs and disease and approximates the existing associations between miRNAs and diseases by the expression vector of miRNA and disease. Different from the conventional approach, we introduce auxiliary miRNAs and disease variables to ensure that these variables converge to the optimal solution during optimization. The experiments show that the AUC obtained by applying the leave-one-out cross-validation method is 0.820 6, which is obviously better than other current methods. Especially in the miRNA and disease without any associated information, the LMFMDA algorithm significantly outperforms the latest algorithm.