Modeling And Analysis On Deep Auto-encoder Based Multi-cancer Molecular Classification

Accurate diagnosis and classification of cancer is very important to cancer treatment, which could contribute to the effective development and analysis of cancer treatment programs. Currently cancer types are defined based on the incidence of its original organ or tissue type. Recently studies have shown that, different types of cancer share same genomic signatures in the levels of genome, transcriptome, proteome and epigenetic molecular; while the patients in same type of cancer have quite a lot of differences in their molecular level. Therefore, in order to complete the deficiency in tissue-of-origin classification, this paper presented a method of cancer molecular classification based on the DAE(Deep AutoEncoder). This method could effectively deal with high-dimensional and sparse cancer data, which could promote the further analyze pathogenesis of cancer within and across tissues of origin.This paper presented a method of cancer molecular classification based on the DAE. Firstly, we reduced the dimension of cancer molecular data from 479 to 30 using DAE. Secondly, we classified the 3199 specimens from 12 cancer types into 11 major categories in the molecular level using optimized K-means Clustering Algorithm. Seven categories were nearly identical to their tissue-of-origin counterparts, but several distinct cancer types were found to converge into common categories. COADREAD and UCEC were divided into several categories separately, due to the subtypes of them. Part of samples from OV, HNSC, LUAD and LUSC converged to one category, due to CDKN2 A mutation and methylation variation. In contrast, KIRC was divided into two categories, based on the states of PBRM1 mutation or not. Finally, we made biological function analysis and prediction of drug targets for those high-frequency mutations in those molecular categories.The result shown that, our method of cancer molecular classification based on the DAE could deal with the high-dimensional and sparse biology data, providing a scientific method for the analysis of cancer within and across tissues of origin. Meanwhile it also provides an effective guidance for the analysis of cancer pathogenesis at the molecular level, and also give a theoretical reference for the personalized medicine in the near future.