Urine IRF4/PENK/PXDN methylation signatures enable machine learning-driven bladder cancer detection

Diagnosing and monitoring bladder cancer (BC) currently relies heavily on invasive cystoscopy, which is uncomfortable, costly, and associated with high recurrence rates. There is a critical need for non-invasive, highly sensitive, and specific methods for early detection and surveillance. DNA methylation, a key epigenetic modification, offers a promising biomarker for cancer, as changes can be detected in bodily fluids like urine, reflecting tumor presence and characteristics without invasive procedures.

Researchers conducted a prospective cohort study from May 2022 to November 2023, enrolling 155 individuals (68 BC, 87 non-BC) whose BC diagnosis was confirmed by cystoscopy-guided biopsy and histopathology. Urine DNA samples were analyzed using targeted next-generation sequencing to quantify methylation at 44 CpG sites within the IRF4, PENK, and PXDN genes. Supervised machine learning classifiers, specifically a random forest model, were trained on these methylation features to distinguish tumor from non-tumor urine samples. Additionally, systems analyses were performed to map signaling pathways and tumor-microenvironment contexts.

Significantly increased methylation of IRF4, PENK, and PXDN was observed in BC urine samples compared to non-BC samples (P < 0.0001). Corresponding mRNA levels of IRF4 and PENK were significantly downregulated in tumor tissues, with PXDN showing a declining trend. Methylation of IRF4 and PXDN negatively correlated with their expression (P < 0.0001). A 41-CpG-site random forest classifier, termed the BladderCando model, demonstrated excellent performance in distinguishing BC from non-BC individuals.

The BladderCando model achieved an Area Under the Curve (AUC) of 0.9783 and an F1 score of 0.9773, significantly outperforming urine cytology for low-grade BC detection. Co-expression and enrichment analyses identified DCN as a central hub gene primarily linked to ECM functions. High expression of PXDN, IRF4, and DCN correlated with upregulated immune checkpoint genes and increased immune cell infiltration. Single-cell sequencing further localized PXDN to fibroblasts and endothelial cells, DCN to fibroblasts, and IRF4 to T cells, with urine expression patterns mirroring tumor tissue profiles.