Incomplete Multi-view Inductive Matrix Completion For Gene-disease Associations Prediction

Genetic diseases seriously threaten human health,and deciphering the associations between genes and diseases has become an important goal of biomedical research.Discovering genes closely related to diseases is extremely important for disease prevention,diagnosis and treatment.With the continuous mining of various biological data and the rapid development of computer technology,many methods have proposed for gene-disease associations prediction.However,most existing methods use a two-step strategy with the beginning feature fusion step and the following associations prediction step,ignoring the reciprocal relationship between these two steps.Meanwhile,these methods do not fully excavate the multi-source feature information of genes and diseases,which are often affected by data redundancy and missing.In response to these problems,this article proposes two models from different perspectives,the main research contents are as follows:(1)Aiming at the shortcoming of the existing two-stage models and the problem of multi-source data fusion,a one-step multi-view inductive matrix completion model is proposed.The model employs the multi-view representation learning to fully capture the consistency and complementary information of the multi-view data,and thus obtain the common latent representations for genes/diseases.It is also suitable for incomplete multi-view data.In addition,we also introduce the adaptive weighting scheme into traditional inductive matrix completion model,penalizing the known and the unknown associations differently to adapt large-scale PU(Positive-Unlabeled)learning problem.Multi-view representation learning and weighted inductive matrix completion are integrated into one jointly model to learn latent representations and predictive matrices simultaneously,and promote each other,which can not only improve latent representation learning,but also boost the final prediction performance.Finally,extensive experiments conducted on real gene-disease dataset demonstrate the superior performance of our method compared to other methods.(2)Aiming at the limitations of shallow linear models in extracting nonlinear features and learning complex associations,a deep multi-view inductive matrix completion model is proposed.The model integrates information from multiple views into an intact representation by the nested autoencoder networks.Our model jointly performs view-specific representation learning(with the inner autoencoder networks)and multi-view shared representation learning(with the outer autoencoder networks)in a unified framework,flexibly balancing the complementary and consistency of multi-view data.The multi-modal low-rank bilinear pooling network for associations prediction is used to fully mine complex gene-disease associations.Finally,experimental results on the real-world dataset demonstrate its effectiveness and superiority.