LMTRDA: Using logistic model tree to predict MiRNA-disease associations by fusing multi-source information of sequences and similarities

Emerging evidence has shown microRNAs (miRNAs) play an important role in human disease research. Identifying potential association among them is significant for the development of pathology, diagnose and therapy. However, only a tiny portion of all miRNA-disease pairs in the current datasets are experimentally validated. This prompts the development of high-precision computational methods to predict real interaction pairs. In this paper, we propose a new model of Logistic Model Tree for predicting miRNA-Disease Association (LMTRDA) by fusing multi-source information including miRNA sequences, miRNA functional similarity, disease semantic similarity, and known miRNA-disease associations. In particular, we introduce miRNA sequence information and extract its features using natural language processing technique for the first time in the miRNA-disease prediction model. In the cross-validation experiment, LMTRDA obtained 90.51% prediction accuracy with 92.55% sensitivity at the AUC of 90.54% on the HMDD V3.0 dataset. To further evaluate the performance of LMTRDA, we compared it with different classifier and feature descriptor models. In addition, we also validate the predictive ability of LMTRDA in human diseases including Breast Neoplasms, Breast Neoplasms and Lymphoma. As a result, 28, 27 and 26 out of the top 30 miRNAs associated with these diseases were verified by experiments in different kinds of case studies. These experimental results demonstrate that LMTRDA is a reliable model for predicting the association among miRNAs and diseases.

越来越多的研究证据表明，微小RNA（microRNAs, miRNAs）在人类疾病研究中发挥着关键作用。鉴定二者之间的潜在关联，对于病理学、诊断学与治疗学的发展具有重要意义。然而，当前数据集内的所有miRNA-疾病配对中，仅有极小一部分经过了实验验证，这推动了高精度计算方法的发展，用以预测真实存在的互作配对。本文提出一种用于预测miRNA-疾病关联的全新模型——逻辑斯蒂模型树（Logistic Model Tree, LMTRDA），该模型融合了多源信息，包括miRNA序列、miRNA功能相似性、疾病语义相似性以及已知的miRNA-疾病关联数据。尤为关键的是，本研究首次在miRNA-疾病预测模型中引入miRNA序列信息，并借助自然语言处理（Natural Language Processing, NLP）技术提取其特征。在HMDD V3.0数据集的交叉验证实验中，LMTRDA取得了90.51%的预测准确率、92.55%的灵敏度，其受试者工作特征曲线下面积（Area Under Curve, AUC）达90.54%。为进一步评估LMTRDA的性能，我们将其与多种分类器及特征描述模型进行了对比。此外，我们还针对人类疾病验证了LMTRDA的预测能力，所涉及的疾病包括乳腺肿瘤（Breast Neoplasms）、乳腺肿瘤（Breast Neoplasms）以及淋巴瘤（Lymphoma）。实验结果显示，在与上述疾病相关的排名前30位miRNA中，分别有28、27和26个通过各类案例研究得到了实验验证。上述实验结果证明，LMTRDA是一款用于预测miRNA与疾病之间关联的可靠模型。