| According to the annal report by World Health Organization(WHO)in 2018,cancer is the second leading cause of death globally,and is responsible for an estimated 9.6 million deaths that year.In recent years,both the number of new cancer cases and deaths are increasing.At the same time,deep learning technologies,as the newest trends in machine learning and artificial intelligence,have brought revolutionary advances in many research domains,including the field of bioinformatics.People can't wait to use computer technology to study the phenomena and laws of life.Cancer is a genetic disease initiated by genetic mutations and progressed by an accumulation of genomic aberrations.Therefore,cancer genomics can provide important insights into not only carcinogenesis and cancer progression,but also cancer detection,diagnosis and treatment.In this paper,we exploit three deconvolution methods: NNLS,CIBERSORT and XGBoost to perform deconvolution on the latest RNA-seq human gene expression datasets DS389 and DS488,released by the Dream Challenges community to obtain the proportion of different cell types.By comparing the correlation between the deconvolution results and the real data,we reveal that the deconvolution results of the CIBERSORT method have the highest accuracy,the XGBoost method is the second,and the NNLS method is the lowest.It is worth mentioning that this is the first time that the XGBoost is used for gene-expression-based deconvolution.Based on deep learning technology,we developed a novel neural network evaluation model that can recommend deconvolution methods for each gene expression input profile.The results suggest that this deep-learning-based evaluation model can significantly improve the accuracy of the deconvolution results.Finally,we developed an open-source R package,named cellsorter,which consists of three deconvolution methods and the novel evaluation mode.The R package is shared at the Github platform for other researchers to use and to further imporve. |
PDF Full Text Request