Leveraging CpG methylation signatures for robust multi-class cancer classification across platforms

Background: Cancers of unknown primary (CUP) are metastatic malignancies where the primary site cannot be identified, often resulting in empirical chemotherapy and poor outcomes. DNA methylation profiling has emerged as a promising tool for improving tumor identification and classification. By leveraging unique methylation signatures, this approach can enhance diagnostic accuracy and guide personalized treatment strategies. Objective: To develop and validate a prediction model for cancer type classification based on a focused set of CpG sites. Methods: Methylation data (Infinium HumanMethylation450) from 7,476 patients across 21 cancer types were obtained from TCGA and other public datasets. Data were divided into training and test cohorts. A hybrid feature selection approach combining Shapley values and gradient boosting was applied to identify CpG regions, and model performance was evaluated on the test cohort. Louvain clustering based on selected CpG profiles was used to explore tumor phenotypes and assess associations between heterogeneity and prediction performance. Independent validation was performed using Infinium MethylationEPIC v2.0 data from 31 cases representing 17 cancer types from our institution. Results: A total of 1,000 CpG regions were selected. Ridge regression achieved the best performance among tested models, with classification accuracy (CA) of 95.4%, AUC of 0.998, F1 score of 0.953, and Matthews correlation coefficient (MCC) of 0.951 averaged across classes in the training cohort. Performance on the test cohort was 94.7% CA, 0.998 AUC, 0.945 F1, and 0.943 MCC, while independent validation yielded 87.1% CA, 0.9993 AUC, 0.847 F1, and 0.867 MCC. Unsupervised analysis revealed 20 distinct Louvain clusters, highlighting heterogeneity across cancer types. Although heterogeneity and purity correlated with MCC, regression analysis did not confirm an independent predictive effect after adjusting for confounders. Conclusion: A CpG-based methylation signature combined with ridge regression enables highly accurate multi-class cancer classification and demonstrates strong generalizability across platforms. These findings support the clinical utility of methylation profiling for CUP diagnosis and personalized treatment planning.