Use of Multiple Machine Learning Approaches for Selecting Urothelial Cancer-Specific DNA Methylation Biomarkers in Urine

• Published in: International journal of molecular sciences Vol. 25; no. 2; p. 738
• Main Authors: Köhler, Christina U, Schork, Karin, Turewicz, Michael, Eisenacher, Martin, Roghmann, Florian, Noldus, Joachim, Marcus, Katrin, Brüning, Thomas, Käfferlein, Heiko U
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 01-01-2024
• Subjects: Accuracy; Analysis; Biological markers; Biomarkers; Biomarkers, Tumor - genetics; Body Fluids; boosted trees; Cancer; Decision making; Disease; DNA; DNA Methylation; DNA methylation biomarker; DNA, Neoplasm; Fingers - abnormalities; Genetic aspects; Hair Diseases; Health aspects; Humans; Langer-Giedion Syndrome; LASSO; Learning strategies; Machine Learning; Male; Methylation; Neoplasms; Nose - abnormalities; random forest; Repressor Proteins; Tumors; Urine; urothelial cancer; United Kingdom; Germany; Taiwan; urothelial cancer; random forest; DNA methylation biomarker; boosted trees; LASSO
• ISSN: 1422-0067; 1661-6596; 1422-0067
• DOI: 10.3390/ijms25020738

Diagnosing urothelial cancer (UCa) via invasive cystoscopy is painful, specifically in men, and can cause infection and bleeding. Because the UCa risk is higher for male patients, urinary non-invasive UCa biomarkers are highly desired to stratify men for invasive cystoscopy. We previously identified multiple DNA methylation sites in urine samples that detect UCa with a high sensitivity and specificity in men. Here, we identified the most relevant markers by employing multiple statistical approaches and machine learning (random forest, boosted trees, LASSO) using a dataset of 251 male UCa patients and 111 controls. Three CpG sites located in
,
and an intergenic region on chromosome 16 have been concordantly selected by all approaches, and their combination in a single decision matrix for clinical use was tested based on their respective thresholds of the individual CpGs. The combination of
and
yielded the best overall sensitivity (61%) at a pre-set specificity of 95%. This combination exceeded both the diagnostic performance of the most sensitive bioinformatic approach and that of the best single CpG. In summary, we showed that overlap analysis of multiple statistical approaches identifies the most reliable biomarkers for UCa in a male collective. The results may assist in stratifying men for cystoscopy.