Embedding of Genes Using Cancer Gene Expression Data: Biological Relevance and Potential Application on Biomarker Discovery

Artificial neural networks (ANNs) have been utilized for classification and prediction task with remarkable accuracy. However, its implications for unsupervised data mining using molecular data is under-explored. We found that embedding can extract biologically relevant information from The Cancer Genome Atlas (TCGA) gene expression dataset by learning a vector representation through gene co-occurrence. Ground truth relationship, such as cancer types of the input sample and semantic meaning of genes, were showed to retain in the resulting entity matrices. We also demonstrated the interpretability and usage of these matrices in shortlisting candidates from a long gene list as in the case of immunotherapy response. 73 related genes are singled out while the relatedness of 55 genes with immune checkpoint proteins (PD-1, PD-L1, and CTLA-4) are supported by literature. 16 novel genes (ACAP1, C11orf45, CD79B, CFP, CLIC2, CMPK2, CXCR2P1, CYTIP, FER, MCTO1, MMP25, RASGEF1B, SLFN12, TBC1D10C, TRAF3IP3, TTC39B) related to immune checkpoint proteins were identified. Thus, this method is feasible to mine big volume of biological data, and embedding would be a valuable tool to discover novel knowledge from omics data. The resulting embedding matrices mined from TCGA gene expression data are interactively explorable online (http://bit.ly/tcga-embedding-cancer) and could serve as an informative reference for gene relatedness in the context of cancer and is readily applicable to biomarker discovery of any molecular targeted therapy.