AI Uses DNA Methylation to Identify Primary Sites in Cancers of Unknown Primary

Overview

• Kindai University researchers presented at the AACR Annual Meeting 2026 in San Diego a machine-learning model that reads CpG DNA methylation patterns to classify tumor origin in cancers of unknown primary.
• The team trained the model on methylation data from nearly 7,500 patients across 21 cancer types using The Cancer Genome Atlas and other public datasets, then narrowed the features to about 1,000 CpG markers.
• The classifier reached about 95% accuracy in a held-out test set and about 87% accuracy in an independent 31-case validation cohort that spanned 17 cancer types.
• Because the training used cancers with known primaries, the authors said the next step is a prospective study in true CUP patients and a blood-based version that reads circulating tumor DNA.
• If validated, the tool could shift more patients from broad chemotherapy to site-specific treatments that have been associated with longer survival than standard care in CUP.