Abstract
This study aimed to identify candidate diagnostic miRNAs from the serum of colorectal cancer (CRC) patients using Boruta, a wrapper-based feature selection technique, in combination with decision tree-based machine learning methods. We analyzed three serum miRNA expression profile datasets from the gene expression omnibus (GEO) database to identify differentially expressed miRNAs common to both cancerous and non-cancerous samples. The GSE106817 dataset, comprising 2568 miRNAs, was used to train our models. The Boruta machine learning feature selection method was applied to identify robust and significant miRNAs associated with CRC in the training cohort. Next, random forest and XGBoost models were trained using the selected miRNAs. To validate the predictive efficacy of the identified candidate miRNAs, we tested them against two independent datasets (GSE113486 and GSE113740). Finally, we performed ontology analysis and constructed a regulatory network to explore the potential links between the selected miRNAs and CRC development. The GSE106817 dataset included 115 CRC patients and 2759 non-cancerous samples. Using Boruta, we identified 146 miRNAs as potential biomarkers for CRC diagnosis. Among these, the highest-scoring miRNAs were: hsa-miR-1228-5p, hsa-miR-6787-5p, hsa-miR-1343-3p, hsa-miR-6717-5p, hsa-miR-3184-5p, hsa-miR-1246, hsa-miR-4706, hsa-miR-8073, hsa-miR-5100. The machine learning models achieved an AUC of 100% when tested on the internal dataset. Additionally, the external validation datasets showed an AUC exceeding 95%, confirming the robustness and reliability of our findings. Furthermore, functional annotation analysis revealed the involvement of several miRNA-mediated pathways in the pathogenesis of CRC.