利用深度稀疏自动编码器预测miRNA与疾病的关联关系

Predicting miRNA-Disease Associations through Deep Sparse Autoencoder

摘要: microRNA (miRNA)是一类具有调控功能的内源性非编码RNA，在各种生物的生命发展过程中发挥着关键作用。许多生物学实验研究证明，miRNA与人类疾病密切相关，包括疾病的发生、流行、传播、诊断和治疗。但是生物实验既昂贵又耗时。因此，有效的计算模型变得越来越重要。在这项研究中，我们将稀疏性嵌入到现有的自动编码器中，形成一个新的计算框架(DSAEMDA)。首先，通过两种方式计算疾病语义相似度得到两个疾病语义相似度矩阵，计算miRNA功能相似度得到miRNA功能相似度矩阵，分别同疾病和miRNA的高斯相互作用谱核相似度矩阵融合。然后嵌入高维空间提取疾病和miRNA高维表达，利用已被证明的miRNA-疾病关联数据训练我们的深度稀疏自动编码器。此外，通过计算未知关系对的重建误差可用于预测某些疾病相关miRNA的相关值。实验结果表明，DSAEMDA可以有效地预测疾病相关的miRNA且准确率高。

Abstract: MicroRNA (miRNA) is a series of endogenous non-coding RNAs with regulatory functions that take a key part in the life development of various organisms. Many biological experimental studies have proved that miRNA is closely allied to human diseases, including the occurrence, prevalence, transmission, diagnosis and treatment of diseases. But biological experiments are expensive as well as time-consuming. Therefore, efficient computational models to avoid the above problems are be-coming increasingly necessary to identify potential miRNA-disease associations. In this study, we add sparsity to the existing autoencoder to form a new computational framework named DSAEMDA (deep sparse autoencoder miRNA-disease association). First, two disease semantic similarity matrices were obtained by computing disease semantic similarity in two ways, and miRNA functional similarity matrices were obtained by computing miRNA functional similarity, which were fused with the kernel similarity matrix of Gaussian interaction spectrum of disease and miRNA respectively. Then, disease and miRNA high-dimensional expressions were extracted by embedding in high-dimensional space, and our deep sparse autoencoder was trained by using proven miR-NA-disease association data. In addition, the reconstruction errors of unknown relationship pairs can be used to predict the correlation values of some disease-related miRNAs. The experimental results showed that DSAEMDA could effectively predict disease-related miRNA with high accuracy.