WEDNESDAY, April 22, 2026 (HealthDay News) -- A CpG-based methylation signature combined with ridge regression allows highly accurate multicancer classification in patients with cancers of unknown primary, according to a study presented at the annual meeting of the American Association for Cancer Research, held from April 17 to 22 in San Diego.
Marco A. De Velasco, Ph.D., from the Kindai University Faculty of Medicine in Sakai City, Japan, and colleagues developed and validated a prediction model for cancer type classification based on a focused set of CpG sites using methylation data from 7,476 patients across 21 cancer types. Data were divided into training and test cohorts. CpG regions were identified by applying a hybrid feature selection approach combining Shapley values and gradient boosting; model performance was assessed on the test cohort. Independent validation was performed using data from 31 cases representing 17 cancer types.
The researchers selected 1,000 CpG regions. Among tested models, the best performance was achieved with ridge regression, with classification accuracy (CA) of 95.4 percent, an area under the receiver operating characteristic curve (AUC) of 0.998, F1 score of 0.953, and Matthews correlation coefficient (MCC) of 0.951 averaged across classes in the training cohort. In the test cohort, performance was 94.7 percent CA, 0.998 AUC, 0.945 F1, and 0.943 MCC, and performance in independent validation was 87.1 percent CA, 0.9993 AUC, 0.847 F1, and 0.867 MCC. Twenty distinct Louvain clusters were identified in an unsupervised analysis, highlighting heterogeneity across cancer types.
"Our findings suggest that DNA-based approaches can help identify where a cancer may have started, even when the original tumor is not visible," De Velasco said in a statement.