A novel end-to-end learning framework for inferring lncRNA-disease associations based on convolution neural network.

Introduction: In recent years, lots of computational models have been proposed to infer potential lncRNA-disease associations. Methods: In this manuscript, we introduced a novel end-to-end learning framework named CNMCLDA, in which, we first adopted two convolutional neural networks to extract hidden features of diseases and lncRNAs separately. And then, by combining these hidden features of diseases and lncRNAs with known lncRNA-disease associations, we designed five different loss functions. Next, based on errors obtained by these loss functions, we would perform back propagation to fit parameters in CNMCLDA, and complete those missing values in lncRNA-disease relational matrix according to these fitted parameters. In order to demonstrate the prediction performance of CNMCLDA, intensive experiments have been carried out and experimental results show that CNMCLDA can achieve better performances than state-of-the-art competitive predictive models in frameworks of five-fold cross validation, ten-fold cross validation and leave-one-disease-out cross validation respectively. Results and Discussion: Moreover, in case studies of gastric cancer, glioma and breast cancer, there are 19, 17 and 16 out of top 20 candidate lncRNAs inferred by CNMCLDA having been confirmed by recent relevant literatures separately, which demonstrated the outstanding performance of CNMCLDA as well. Hence, it is obvious that CNMCLDA may be an effective tool for prediction of potential lncRNA-disease associations in the future.