Exploratory integrative machine learning analysis of miRNA datasets identifies a candidate miRNA signature for oral squamous cell carcinoma

Abstract Background Oral squamous cell carcinoma (OSCC) is typically diagnosed at an advanced stage and thus there is a need for biomarkers for OSCC which can be detected with minimum invasiveness. Although circulating microRNAs (miRNAs) have demonstrated great diagnostic promise, there are limited publicly available circulating miRNA datasets associated with OSCC. Objective To exploratively integrate public tissue miRNA datasets using machine learning methods to identify candidate miRNAs as potential OSCC biomarkers. Methods The miRNA profiling of multiple OSCC tissue datasets from GEO and TCGA containing both tumor and control samples were integrated for biomarker discovery. Tumor samples only were excluded prior to the model development. Machine learning models were developed using differentially expressed set of miRNAs. Five classifiers were evaluated and the best model was externally validated by applying it to the only publicly available independent circulating serum miRNA dataset. The signature was subjected to a continuous probability score. The model interpretation was performed using SHAP. Results A 30-miRNA candidate signature associated with OSCC was identified. In internal cross-validation, the panel separated tumor from normal tissue (AUC 0.96). The logistic regression model had the highest AUC, while XGBoost and SVM-RBF had slightly higher accuracy (Table 5) and gave an AUC of 0.67 in external validation. The held-out tissue-level performance was variable and lower (AUC 0.54–0.88, Table 7) and the external serum AUC was modest (0.67). The panel identified included known OSCC-related miRNAs (miR-21, miR-146b-5p) and a number of less studied candidate biomarkers whose function remains to be tested. Conclusion The study demonstrates that a machine learning method can be applied to integrate public tissue miRNA datasets for identification of circulating miRNA biomarkers in OSCC. This is an exploratory study; the signature should not be considered as a validated serum test, but rather a tissue-derived candidate panel that offers limited transferability to serum samples and needs to be validated in more, larger, and independent circulating samples before it can be applied to a diagnostic purpose.