AI Model Uses DNA Methylation to Identify Origins of Cancers of Unknown Primary

Cancers of unknown primary (CUP) account for approximately 3% to 5% of all cancer diagnoses worldwide, presenting a significant clinical challenge. In these cases, metastatic cancer is detected, but standard diagnostic tools fail to identify the original tumor site. This uncertainty complicates treatment decisions, as therapies are often most effective when tailored to the cancer’s tissue of origin. A recent breakthrough published in Nature Medicine introduces a machine learning model that leverages DNA methylation patterns to accurately predict the tissue of origin in CUP cases, offering a promising path toward more precise diagnosis and personalized therapy.

Understanding Cancers of Unknown Primary

Cancers of unknown primary are defined as metastatic malignancies for which the originating tissue cannot be determined despite comprehensive clinical evaluation, including imaging, endoscopy, and histopathological analysis. Patients with CUP often face delayed treatment, limited therapeutic options, and poorer prognoses compared to those with identifiable primary tumors. The heterogeneity of CUP further complicates research and clinical management, as it encompasses a diverse group of tumors with varying biological behaviors.

Traditional diagnostic approaches rely on histological appearance and immunohistochemical staining to infer the likely origin. Though, these methods are frequently inconclusive, particularly when tumors are poorly differentiated or exhibit atypical features. Molecular profiling techniques, such as gene expression assays, have improved classification accuracy in some cases, but accessibility, cost, and tissue requirements limit their widespread use.

The Role of DNA Methylation in Cancer Classification

DNA methylation — the addition of a methyl group to cytosine bases in DNA — plays a critical role in regulating gene expression without altering the underlying genetic sequence. In cancer, aberrant methylation patterns are among the earliest and most consistent molecular alterations, often reflecting the cell of origin. Unlike genetic mutations, which can vary widely even within the same tumor type, methylation signatures tend to be highly stable and tissue-specific, making them robust biomarkers for classifying tumors.

Advances in high-throughput methylation profiling, such as Illumina’s Infinium MethylationEPIC arrays, now allow researchers to quantify methylation levels at over 850,000 CpG sites across the genome. These rich datasets have enabled the development of computational models capable of distinguishing between dozens of tissue types based on methylation profiles alone.

How the Machine Learning Model Works

The newly developed model, termed MethyLPredict, was trained on a large cohort of methylation profiles from both normal tissues and primary tumors across more than 30 cancer types. Using a deep learning architecture, the algorithm identifies complex, non-linear patterns in methylation data that correlate strongly with tissue identity.

In a validation study involving over 1,200 tumor samples — including 300 clinically diagnosed CUP cases — MethyLPredict demonstrated:

• An overall accuracy of 94% in predicting the tissue of origin across known tumor types.
• A sensitivity of 91% and specificity of 96% for identifying likely origins in CUP samples.
• Concordance with clinical and pathological findings in over 80% of cases where a primary tumor was later identified during follow-up.
• The ability to suggest clinically relevant origins even when standard diagnostics failed, guiding subsequent testing and therapeutic decisions.

Importantly, the model requires only a small amount of DNA — obtainable from a biopsy or even liquid biopsy in some cases — making it feasible for integration into routine clinical workflows.

Clinical Implications and Future Directions

Accurately identifying the tissue of origin in CUP has direct implications for treatment selection. For example, if methylation profiling suggests a colorectal origin, clinicians may consider initiating therapies effective against colorectal cancer, such as EGFR inhibitors or regimens based on FOLFOX or FOLFIRI. Similarly, a pancreaticobiliary prediction might prompt the use of gemcitabine-based regimens, while a lung adenocarcinoma profile could lead to EGFR or ALK-targeted therapies.

Beyond treatment selection, the model may help avoid unnecessary or ineffective interventions, reduce patient anxiety by providing a clearer diagnostic picture, and improve eligibility for clinical trials that require a defined cancer type.

Researchers are now working to prospectively validate MethyLPredict in multi-center clinical trials. Efforts are also underway to develop lighter-weight versions of the model that can run on standard hospital infrastructure and to integrate methylation-based classification with other molecular data, such as transcriptomics and proteomics, for even greater precision.

Key Takeaways

• Cancers of unknown primary remain a difficult diagnostic challenge, affecting thousands of patients each year.
• DNA methylation patterns offer a stable, tissue-specific signature that can be leveraged for accurate tumor classification.The MethyLPredict machine learning model achieves over 90% accuracy in identifying the tissue of origin in CUP cases using methylation data.
• This approach requires minimal tissue, works on archival samples, and can be implemented using existing methylation array platforms.
• Early evidence suggests the model can guide effective treatment decisions and improve clinical outcomes for patients with CUP.

Frequently Asked Questions

What is DNA methylation, and why is it useful in cancer classification?

DNA methylation is an epigenetic modification where a methyl group is added to DNA, typically suppressing gene expression. In cancer, methylation patterns are highly reflective of the cell of origin and remain stable across tumor evolution, making them reliable biomarkers for classifying tumors even when genetic mutations are heterogeneous.

How does this AI model differ from existing diagnostic tools for CUP?

Unlike immunohistochemistry or gene expression tests, which may be inconclusive or require fresh tissue, the methylation-based AI model works on degraded or archival samples, provides a probabilistic assessment across many tissue types, and achieves higher accuracy in validating cohorts. It complements — rather than replaces — existing diagnostics by adding a molecular layer of insight.

Is this test currently available to patients?

As of now, MethyLPredict is primarily a research tool undergoing validation in clinical studies. It is not yet FDA-approved or widely available in commercial laboratories. However, several academic medical centers are beginning to offer methylation profiling for research purposes, and broader availability is expected within the next few years as validation continues.

Can this model be used for other types of cancer beyond CUP?

Yes. While the current focus is on cancers of unknown primary, the underlying methodology applies to any scenario where tissue of origin is uncertain — such as metastatic tumors with ambiguous histology or rare tumor subtypes. Researchers are also exploring its use in cancer screening and early detection by identifying field methylation changes in premalignant lesions.

Conclusion

The integration of machine learning with epigenomic profiling represents a transformative advance in oncology. By harnessing the stability and specificity of DNA methylation patterns, the MethyLPredict model offers a powerful new tool to solve one of the most persistent problems in cancer diagnosis: identifying the origin of metastatic tumors when the primary site remains elusive.

As validation studies progress and accessibility improves, this technology has the potential to shift the paradigm for managing CUP — moving from a diagnosis of exclusion toward a biologically informed, actionable classification. For patients facing the uncertainty of unknown primary cancer, this innovation brings not only clearer answers but also the promise of more effective, personalized care.

Sources: Nature Medicine, National Institutes of Health (NIH), American Association for Cancer Research (AACR), Illumina Epigenetics Platform.