Use of Machine Learning for Diagnosis of Cancer in Ovarian Tissues with a Selected mRNA Panel

Ovarian cancer (OC) is a leading cause of death among women in the United States. Diagnostic molecular biomarkers for OC have been suggested as a prominent approach towards reducing the mortality rate. Standard single-molecule biomarkers do not provide sufficient sensitivity and specificity for detection of OC, while panels of biomarkers may potentially do. We previously have reported a panel of 26 mRNAs, each having a distinct expression pattern between OC and non-cancerous tissues. In this study, we use the 26-gene panel as a candidate biomarker set for training machine learning predictive models. We have used the selected feature set for classifying an integrated expression data of 530 ovarian tissues. After preprocessing samples and batch effect removal, we used the integrated dataset to train and evaluate multiple classification methods. Our analysis finds highly specific and sensitive Random Forest and Support Vector Machine pipelines for predicting cancerous tissues. We also verified the quality of the selected 26 mRNAs by reevaluating the pipeline on randomly selected mRNA expressions. These results suggest that the presented mRNA panel is a small candidate set of gene expression signatures for designing molecular-based OC biomarkers.