Grading Prediction of Kidney Renal Clear Cell Carcinoma by Deep Learning

The grade of cancer is a way to classify cancer based on certain characteristics of cancer tissue. It is an important issue for the precise diagnosis, treatment, and mechanistic research of cancer. With the rapid development of genome sequencing technology, it has become possible to obtain large amounts of gene expression data, and large-scale genomic data to predict the grade of cancer is a challenging problem. In this study, we used gene expression data to propose a pathway-related deep neural network (K-Net) for predicting the grade of Kidney renal clear cell carcinoma (KIRC) tissues. K-Net provides the capability of model interpretability that most conventional fully-connected neural networks lack, describing which pathways play an important role in the process of predicting grade. The predictive performance of K-Net was evaluated with multiple cross-validation experiments. The K-Net prediction accuracy of 74%. More meaningfully, in contrast to using genes as features, this new classification model using enriched pathways as features can well explain which pathways play an important role in KIRC tissues from highly differentiated to poorly differentiated. Cancer development is a process of degradation of certain functions and enhancement of certain functions of tumor tissue, and understanding which pathways play an important role in cancer development can help explore research directions in cancer treatment.