A Comparative Study of Deep Learning and Machine Learning Approaches for Viral Transcriptomic Classification and Alcohol-Associated Hepatocellular Carcinoma

Although distinct molecular profiles are found based on the etiology of hepatocellular carcinoma (HCC), classifying cases as viral or alcohol-based remains difficult due to the high-dimensionality nature of transcriptomic data. In this analysis, transcriptomic data from the TCGA GDC portal, which contains data for 134 viral-associated HCC samples and 119 alcohol-associated HCC, were used to evaluate classification. After eliminating overlapping samples found in both groups, 190 samples were contained in the final dataset-including a total of 112 viral-associated samples and 78 alcohol-associated samples. Following processing an initial collection of 34,686 genes, 3,000 of the most variable genes were selected for an autoencoder to perform dimensionality reduction to 128 latent dimensions (R² = 0.918). Classifiers including CNN, Random Forest, Support Vector Machine, and Logistic Regression were evaluated using features created through the dimensionality-reduced autoencoder. Among the classification methods used, Random Forest produced the best results (AUC-ROC = 0.817, accuracy = 73.7%) and outperformed the CNN (AUC-ROC = 0.747, accuracy = 68.4%) methods. Through gene-latent feature correlation analysis, the following genes are representative of the classifiers: BRCA1, PRKCZ, UBE2I and ATG16L1, all of which have been associated with pathways in DNA repair, cell signaling, and autophagy. Machine learning techniques that utilize classical machine-learning techniques and effective dimensionality reductions can produce results that are superior to deep-learning techniques when dealing with limited samples of transcriptomic data. The approach proposed in this study provides a foundation for defining an etiology-based classification system for HCC.

Authors:
Muhammad Rezki Rasyak, Mahmud Isnan, and Bens Pardamean.

The 11th International Conference on Computer Science and Computational Intelligence (ICCSCI)

Read Full Text