Exploration of the Relevance of MicroRNA Signatures for Cancer Detection and Multiclass Cancer Classification

miRNA expression profiles are heterogeneously expressed among cancer types, with miRNAs serving as highly tissue specific tumor suppressors and oncogenes. Machine learning methodologies have been used to develop high performance pan-cancer classification models and identify potentially novel miRNA biomarkers for clinical investigation. However, it is important to understand how such data science techniques correlate to established biological processes to advance integration into clinical environments. This research aims to assess how the top miRNA features selected by machine learning models relate to clinically and biologically verified miRNA biomarkers. We developed Support Vector Machine and Random Forest machine learning models for cancer classification, iteratively adding cancer classes to the multiclass models. The relationship between the selected top features (miRNAs) and clinically verified miRNA biomarkers was assessed through percent relevance, i.e., the number of verified miRNAs vs the number of selected features. We found that as the number of cancer classes increased, the performance metrics decreased, yet the percentage relevance of the miRNA feature selection signature slightly increased before stabilizing. Additionally, after conducting principal component analysis, the non-cancer tissues from all samples had very similar expression visualizations, while all cancerous tissues had unique profiles. The results indicated that models with a greater number of cancer classes shift towards focusing on cancer-diverse miRNAs of greater relevance with characterized functionality. This work suggests that miRNAs may be highly unique to specific cancerous tissues and can be strong biomarkers for detection and classification, but current verified biomarkers fall toward more cancer-wide miRNAs when detecting cancer.