Machine Learning Model Uses DNA Methylation to Predict Tumor Origin in Cancers of Unknown Primary

Many patients consequently receive broad, nonspecific chemotherapy associated with poorer outcomes than site-directed care. Molecular profiling has been explored to infer tissue of origin but has not shown clear survival gains in trials. Researchers now present a machine learning approach that classifies CUP using CpG DNA methylation patterns.

Researchers at Kindai University (Osaka, Japan) developed a computational model that analyzes cytosine–phosphate–guanine (CpG) DNA methylation to infer tumor origin, and presented the work at the American Association for Cancer Research (AACR) Annual Meeting 2026. The method treats CpG methylation as a tissue-specific molecular “fingerprint” that can persist after metastasis. The classifier was designed to distinguish among 21 cancer types.

The approach applies machine learning to tumor methylation profiles to identify informative loci and build class-specific signatures. The investigators sought to replace large, complex feature sets with a smaller, practical marker set while maintaining performance. They selected about 1,000 CpG regions from hundreds of thousands across the genome to constitute the predictive panel.

The model was trained and tested using methylation data from nearly 7,500 patients with 21 cancer types obtained from The Cancer Genome Atlas (TCGA) and other public datasets. Data were partitioned into training and test cohorts to develop and evaluate the classifier. In held-out testing, it correctly identified cancer type in about 95% of cases, and it achieved approximately 87% accuracy on an independent validation cohort from the investigators’ institution comprising 31 cases representing 17 cancer types.

The researchers noted that in CUP, only 15% to 20% of patients have features that enable site-specific therapy, while 80% to 85% receive more generalized chemotherapy; reported survival can reach up to 24 months with site-directed treatment compared with six to nine months with standard care.

A key limitation of the current study is that the model was developed using cancers of known origin rather than true CUP, underscoring the need for prospective evaluation in this population. The team also highlighted challenges in obtaining tissue for genetic testing and identified blood-based biopsy using circulating tumor DNA as a priority for future adaptation.

"Our findings suggest that DNA-based approaches can help identify where a cancer may have started, even when the original tumor is not visible. By using a much smaller and more focused set of markers, this approach could make these types of tests more practical and accessible in the future," said Marco A. De Velasco, Ph.D., a faculty member in the Department of Genome Biology at Kindai University in Japan.

“Overall, we see this research as part of a broader effort to better understand cancer using molecular information, with the goal of supporting more informed and personalized care in the future. However, this work is still in the research stage. We next have to evaluate how well this approach performs in a prospective analysis of patients with true cancers of unknown primary,” added De Velasco.

Related Links
Kindai University