Machine Learning-Based Detection of Bladder Cancer by Urine cfDNA Fragmentation Hotspots that Capture Cancer-Associated Molecular Features

Background cfDNA fragmentomics-based liquid biopsy is a potential option for noninvasive bladder cancer (BLCA) detection that remains an unmet clinical need. Methods We assessed the diagnostic performance of cfDNA hotspot-driven machine-learning models in a cohort of 55 BLCA patients, 51 subjects with benign conditions, and 11 healthy volunteers. We further performed functional bioinformatics analysis for biological understanding and interpretation of the tool’s diagnostic capability. Results Urinary cfDNA hotspots-based machine-learning model enabled effective BLCA detection, achieving high performance (area under curve 0.96) and an 87% sensitivity at 100% specificity. It outperformed models using other cfDNA-derived features. In stage-stratified analysis, the sensitivity at 100% specificity of the urine hotspots-based model was 71% and 92% for early (low-grade Ta and T1) and advanced (high-grade T1 and muscle-invasive) disease, respectively. Biologically, cfDNA hotspots effectively retrieved regulatory elements and were correlated with the cell of origin. Urine cfDNA hotspots specifically captured BLCA-related molecular features, including key functional pathways, chromosome loci associated with BLCA risk as identified in genome-wide association studies, or presenting frequent somatic alterations in BLCA tumors, and the transcription factor regulatory landscape. Conclusions Our findings support the applicability of urine cfDNA fragmentation hotspots for noninvasive BLCA diagnosis, as well as for future translational study regarding its molecular pathology and heterogeneity.