Deep Learning-based Identification of Cancer or Normal Tissue using Gene Expression Data

Background: Deep learning has proven to show outstanding performance in resolving recognition and classification problems. As increasing amounts of cancer and normal gene expression data become publicly available, deep learning may become an integral component of efficiently finding specific patterns within massive datasets. Thus, we aim to address the extent to which the machine can learn to recognize cancer. We integrated cancer and normal tissue data from the Gene Expression Omnibus (GEO), The Cancer Gene Atlas (TCGA), Therapeutically Applicable Research To Generate Effective Treatments (TARGET), and Genotype-Tissue Expression (GTEx) databases, including 13,406 cancer and 12,842 normal gene expression data from 24 different tissues. We first trained the deep neural network (DNN) to discriminate between cancer and normal samples using various gene selection strategies and therapeutic target genes from commercial cancer panels and genes in NCI-curated cancer pathways. We also suggest systemic analyzation method to interpret trained deep neural network. We applied the method to find genes mostly contribute to classify cancer in an individual sample. Result: The best trained DNN could classify cancer and normal data with accuracy of 0.997 in the training data set of 13,123 (cancer: 6,703, normal: 6,402) samples. In the independent test set comprising 13,125 (cancer: 6,703, normal: 6,422) samples, the DNN model achieved 0.979 accuracy. Using the same training and test data, our DNN showed better performance than other conventional prediction methods, followed by the support vector machine approach. For interpretation, we propose a method that can extract a gene’s contribution to an individual sample’s cancer probability from the trained DNN. This method distinguished samples dependent on one or a few genes suggesting these samples are possibly}}{{\it “oncogene addicted”. Conclusion: A deep learning approach in conjunction with our interpretation method is not only a useful tool to identify cancer from gene expression data but can also contribute toward understanding the complex nature of cancer based on large public data.