Machine learning-guided identification of metastasis-associated miRNAs and their integration into a PFI-based cox risk score and nomogram.

BACKGROUND

Metastasis drives mortality in breast invasive carcinoma. We sought miRNA biomarkers that (i) discriminate metastatic potential, (ii) stratify prognosis, and (iii) translate into a clinically useful PFI predictor.

METHODS

We analyzed 858 TCGA-BRCA primary tumors (20 M1, 838 M0). After filtering low-expression miRNAs, DESeq2 identified 10 miRNAs downregulated in M1. Class imbalance was addressed with ADASYN; Random Forest and XGBoost feature importance over 50 iterations converged on four candidates (hsa-miR-150, -5694, -6510, -7156). Ten ML models (single learners and ensembles) were trained with nested tuning and evaluated on balanced test sets and the original cohort. Prognostic value was tested by Kaplan-Meier and Cox regression across OS, DSS, DFI, and PFI. Endpoint-specific Cox β-coefficients yielded miRNA risk scores; a PFI nomogram combined the PFI score with N and M stage. We profiled miR-150, its isoforms (3p/5p), and the three additional candidates in cell lines (MCF-7, MDA-MB-231) and in a 4T1 murine model with histologic confirmation of lung metastasis. Circulating/metastasis-related biomarkers (LDH/PDH ratio, VEGF, Angiopoietin-2, MMP-2) were assayed in serum.

RESULTS

Ensembles showed near-perfect discrimination on balanced data and strong transfer to the original cohort (test AUCs: Bagging ≥0.979, Random Forest 0.981; original-cohort XGBoost 0.973 ± 0.008). High expression of miR-150 and miR-6510 associated with longer OS and DSS; for PFI, miR-150, miR-6510, and miR-5694 were favorable. The three-miRNA PFI score independently predicted progression (multivariable HR = 1.85; 95 % CI: 1.14-3.01) and, integrated with N and M stage, improved 3- and 5-year PFI discrimination (AUC 0.68 and 0.70) with robust calibration. Experimentally, miR-150 (3p/5p) declined in metastatic tissues and blood, while metastatic mice showed elevated LDH/PDH, VEGF, Ang-2, and MMP-2, supporting a mechanistic axis linking miRNA suppression, metabolic rewiring, angiogenesis, and matrix remodeling.

CONCLUSIONS

An integrative pipeline identifies a four-miRNA signature associated with lung metastasis and delivers a translational PFI nomogram. Concordant experimental data and serum biomarkers reinforce biological plausibility and clinical potential, including liquid-biopsy applications.