Gene Expression Prediction Method Using Deep Learning Based On Landmark Genes

Large scale gene expression analysis has been widely used to characterize the cell state under various conditions of disease and genetic disturbance.With the development of science and technology,although the cost of whole genome expression has been gradually decreased,it is still very expensive to generate a large number of samples.Recognizing that gene expression is usually highly correlated,researchers at the NIH LINCS program have developed an economical and effective way,they picked out about 1000 landmark genes,and they confirmed that these selected genes can capture the information about 80%.Then it is possible to take advantage of these genes and use the machine learning knowledge to perform a genome-wide analysis of gene expression in a relatively short time.There are many methods of machine learning,such as linear regression,decision tree,neural network,support vector machine and some clustering methods.Among them,the LINCS project team used a linear regression method;Yifei Chen,Yi Li et al.used the depth learning method.But the two methods above,there are still some defects,the linear regression can not capture the relationship between gene expression and complex nonlinear,so the accuracy is not high;although the accuracy of deep learning is higher than linear regression,but the cost is too complex and the time we have spent is long.Combining the advantages and disadvantages of the two algorithms of linear regression and depth learning,this paper proposes a new method of gene expression prediction based on deep learning based on convolutional neural network.Through the self-organizing feature map neural network SOM,the original data set is transformed into the similar two-dimensional data in the region.Due to the large amount of data and complex network structure,if we do not use the GPU,it will seriously affect the training speed,so we need to accelerate the deep learning through the GPU,finally verify the results.In order to verify the convolutional neural network prediction algorithm based on gene expression of the feasibility and accuracy of prediction,this paper used GEO,GTEx,1000 G three data sets for experiment,and compared with the above two algorithms.