Multi-class Cancer Classification and Biomarker Identification using Deep Learning

Genetic data is important for analysing cellular functions whose disruption gives rise to various kinds of cancer. The intricacies of gene interaction are captured in various kinds of data for cancer detection through sequencing technology, but diagnosis, prognosis and treatment are still hard. Advent of machine learning helped researchers in supervised and unsupervised learning tasks along with gene identification but resourcefulness has not been overtly satisfactory. This research revolves around multi-class cancer classification, feature extraction and relevant gene identification through deep learning methods for 12 different types of cancers using RNA-SEQ from The Cancer Genome Atlas. It has been constrained by hardware resource availability and within them the experiments that have been performed have shown promising results. Stacked De-noising Autoencoders were used for feature extraction and biomarker identification while 1D Convolutional Neural Networks for classification. Classification was performed with extracted features and relevant genes, which gave average performance of around 94% and 95% respectively. We were able to identify generic cancer-related pathways and their associated genes through Stacked De-noising Auto-encoders generated weight matrix and features. The common pathways include WNT Signalling Pathway, Angiogenesis. Moreover, across all pathways some recurrent genes were observed, namely: PIK3C2G, PCDHB8, WNT10A and these genes were found, in literature, to be involved in multiple types of cancer. The proposed approach shows superior performance and promise against traditional techniques used by bioinformatics community, in terms of accuracy and relevant gene identification.