Deep Collaborative Filtering For Disease Genes Prediction

It is of great medical importance to explore new genes which are associated with the disease.Disease genes provide effective clues to elucidate disease mechanisms and predict potential risks of disease.With the development of bioinformatics technologies,an ocean of biological data have been exploded in recent years,such as genomes,genotype,and phenotypes.The computational method based on multiple biological networks has played a key role in disease gene prediction.However,high noise and high dimensionality lead to unreliable results of networks.This dissertation mainly studies the relationships between genes and disease,which includes two important processes of feature extraction of genes and diseases and matrix completion.Specific research work is as follows:1)In order to obtain deeper understanding of biological data,we introduced a deep learning model,the Stacking Denoising Auto-Encoder(SDAE)to process gene datasets and combined them with traditional collaborative filtering(CF)method to build a Deep Collaborative Filtering(DCF)model.This model is then taken as the benchmark of the research work.2)Due to the implicit feedback of gene-disease association data,we use Positive-Unlabeled(PU)learning method which punishes the misclassification of positive and unlabeled samples differently.The experimental results show that the performance of the PU model is slightly higher than that of the benchmark model,and thus verify the validity of the biased method.3)In order to further uncover the available information of gene datasets,the DCF uses the node2vec method of network representation learning to analyze the structure of gene interaction network.Each gene node is mapped into a low-dimensional vector.The experimental results show that applying gene embedding can effectively improve the precision and recall of DCF model prediction.