A benchmarking of deep neural network models for cancer subtyping using single point mutations

Abstract

AbstractIt is now well-known that genetic mutations contribute to development of tumors, in which at least 15% of cancer patients experience a causative genetic abnormality includingDe Novosomatic point mutations. This highlights the importance of identifying responsible mutations and the associated biomarkers (e.g., genes) for early detection in high-risk cancer patients. The next-generation sequencing technologies have provided an excellent opportunity for researchers to study associations betweenDe Novosomatic mutations and cancer progression by identifying cancer subtypes and subtype-specific biomarkers. Simple linear classification models have been used for somatic point mutation-based cancer classification (SMCC); however, because of cancer genetic heterogeneity (ranging from 50% to 80%), high data sparsity, and the small number of cancer samples, the simple linear classifiers resulted in poor cancer subtypes classification. In this study, we have evaluated three advanced deep neural network-based classifiers to find and optimized the best model for cancer subtyping. To address the above-mentioned complexity, we have used pre-processing clustered gene filtering (CGF) and indexed sparsity reduction (ISR), regularization methods, a Global-Max-Pooling layer, and an embedding layer. We have evaluated and optimized the three deep learning models CNN, LSTM, and a hybrid model of CNN+LSTM on publicly available TCGA-DeepGene dataset, a re-formulated subset of The Cancer Genome Atlas (TCGA) dataset and tested the performance measurement of these models is 10-fold-cross-validation accuracy. Evaluating all the three models using a same criterion on the test dataset revealed that the CNN, LSTM, and CNN+LSTM have 66.45% accuracy, 40.89% accuracy, and 41.20% accuracy in somatic point mutation-based cancer classification. Based on our results, we propose the CNN model for further experiments on cancer subtyping based on DNA mutations.