School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, 221116, China.
Artificial Intelligence Research Institute, China University of Mining and Technology, Xuzhou, 221116, China.
Brief Bioinform. 2022 Jan 17;23(1). doi: 10.1093/bib/bbab500.
MicroRNAs (miRNAs) play crucial roles in multiple biological processes and human diseases and can be considered as therapeutic targets of small molecules (SMs). Because biological experiments used to verify SM-miRNA associations are time-consuming and expensive, it is urgent to propose new computational models to predict new SM-miRNA associations. Here, we proposed a novel method called Dual-network Collaborative Matrix Factorization (DCMF) for predicting the potential SM-miRNA associations. Firstly, we utilized the Weighted K Nearest Known Neighbors (WKNKN) method to preprocess SM-miRNA association matrix. Then, we constructed matrix factorization model to obtain two feature matrices containing latent features of SM and miRNA, respectively. Finally, the predicted SM-miRNA association score matrix was obtained by calculating the inner product of two feature matrices. The main innovations of this method were that the use of WKNKN method can preprocess the missing values of association matrix and the introduction of dual network can integrate more diverse similarity information into DCMF. For evaluating the validity of DCMF, we implemented four different cross validations (CVs) based on two distinct datasets and two different case studies. Finally, based on dataset 1 (dataset 2), DCMF achieved Area Under receiver operating characteristic Curves (AUC) of 0.9868 (0.8770), 0.9833 (0.8836), 0.8377 (0.7591) and 0.9836 ± 0.0030 (0.8632 ± 0.0042) in global Leave-One-Out Cross Validation (LOOCV), miRNA-fixed local LOOCV, SM-fixed local LOOCV and 5-fold CV, respectively. For case studies, plenty of predicted associations have been confirmed by published experimental literature. Therefore, DCMF is an effective tool to predict potential SM-miRNA associations.
微小 RNA（miRNAs）在多种生物过程和人类疾病中发挥着关键作用，可以被视为小分子（SMs）的治疗靶点。由于验证 SM-miRNA 关联的生物学实验既耗时又昂贵，因此迫切需要提出新的计算模型来预测新的 SM-miRNA 关联。在这里，我们提出了一种名为双网络协同矩阵分解（DCMF）的新方法来预测潜在的 SM-miRNA 关联。首先，我们利用加权 K 最近已知邻居（WKNKN）方法对 SM-miRNA 关联矩阵进行预处理。然后，我们构建矩阵分解模型，分别获得包含 SM 和 miRNA 潜在特征的两个特征矩阵。最后，通过计算两个特征矩阵的内积获得预测的 SM-miRNA 关联评分矩阵。该方法的主要创新之处在于，WKNKN 方法的使用可以预处理关联矩阵中的缺失值，并且引入双网络可以将更多不同的相似性信息集成到 DCMF 中。为了评估 DCMF 的有效性，我们基于两个不同的数据集和两个不同的案例研究，实施了四种不同的交叉验证（CV）。最后，基于数据集 1（数据集 2），DCMF 在全局留一法交叉验证（LOOCV）、miRNA 固定局部 LOOCV、SM 固定局部 LOOCV 和 5 折 CV 中分别实现了 0.9868（0.8770）、0.9833（0.8836）、0.8377（0.7591）和 0.9836±0.0030（0.8632±0.0042）的接收者操作特征曲线下面积（AUC）。对于案例研究，大量的预测关联已被已发表的实验文献所证实。因此，DCMF 是一种预测潜在 SM-miRNA 关联的有效工具。
