LSTM neural network for pan-cancer classification & marker gene discovery by symmetrical occlusion method

Abstract

Abstract RNA-Seq has been widely used for capturing the transcriptome profile of clinical samples. The quantitative measurement of RNA expression level provided by RNA-Seq is an ideal replacement for conventional cancer diagnosis of microscope examination with a more sensitive & automated approach. Accurate classification of the status and the tissue of origin of a clinical sample is crucial for the cancer diagnosis and helps clinicians to determine the appropriate treatment strategy. However, existing studies often use maker genes that exhibits statistical difference between healthy and cancer samples, neglecting genes with low expression level differences. To conduct pan-cancer classification, this paper employed normalized read count for gene expression level normalization. A baseline LSTM neural network was trained using RNA-Seq data containing a complete list of genes to distinguish 28 classes of samples with different origins and statuses. Most importantly, this paper introduces a novel maker gene discovery method named “Symmetrical Occlusion” (SO), which works in conjunction with the trained LSTM network by mimicking the “gain of function” and “loss of function” of genes to evaluate the importance of the gene in pan-cancer classification. This is achieved by calculating the change in the prediction score of the LSTM network. Furthermore, a new neural network would be trained using dataset containing only genes of high importance to achieve better classification performance with fewer genes. The baseline LSTM neural network achieves a validation accuracy of 96.59% in pan-cancer classification. After employing occlusion and selecting the top 33% of genes ranked by their importance, the accuracy of the second LSTM neural network with the same architecture is later improved to 98.30% with 67% fewer genes than the baseline. Our method successfully discovered many marker genes which are not differentially expressed. Compared with the existing methods, our neural network has more comprehensive prediction classes, and outperformed others in terms of classification performance, including metastasized cancer classification. In addition, our method may also be applied to marker gene discovery as well as novel cell type detection using single-cell RNA-Seq.