At the 2026 American Association for Cancer Research annual meeting, researchers unveiled a machine learning model that identifies the primary origin of metastatic cancers by analyzing DNA methylation patterns, offering new hope for patients with cancers of unknown primary.
Presented by Marco A. De Velasco of Kindai University’s Genome Biology Department, the model focuses on CpG methylation — a chemical modification that acts as a molecular fingerprint for different tissues. By examining these patterns in tumor samples, the tool distinguishes 21 cancer types with high accuracy, addressing a critical gap in oncology where up to 85% of patients with cancers of unknown primary (CUP) receive nonspecific chemotherapy due to undetectable origins.
Patients whose cancer origin is identified can receive site-directed therapies and survive up to 24 months, compared to just six to nine months with standard broad-spectrum treatment. Yet only 15 to 20% of CUP cases currently yield detectable features for targeted therapy, leaving the majority without precision options.
De Velasco noted that while prior molecular profiling methods showed promise, they failed to demonstrate clear survival benefits in clinical trials. The CpG methylation approach, however, leverages tissue-specific epigenetic signatures that persist even after metastasis, providing a more stable biomarker than gene expression alone.
This advance aligns with parallel efforts at McGill University, where researchers developed SIDISH — an AI tool designed to detect rare, high-risk cell populations driving aggressive cancers. Unlike bulk analyses that average signals across millions of cells, SIDISH links single-cell behavior to patient outcomes, identifying dangerous subpopulations in pancreatic, breast, and lung tumors that would otherwise be missed.
Yasmin Jolasun, lead author of the McGill study published in Nature Communications, explained that current tools struggle to integrate detailed cellular data with real-world patient survival metrics. SIDISH bridges this gap by using semi-supervised iterative deep learning to pinpoint cells most strongly associated with poor prognosis, enabling therapies that target the true engines of disease rather than treating all cancer cells uniformly.
The tool’s potential extends beyond cancer, with applications envisioned for other complex diseases where cellular heterogeneity influences outcomes. Still, its preclinical status means clinical validation remains pending.
In breast cancer specifically, AI is already reshaping care pathways, according to a March 2026 review in Biologie et médecine du cancer. The technology enhances mammogram interpretation, refines pathology through whole-image analysis and biomarker evaluation, and supports treatment decisions by integrating clinical, imaging, and pathological data.
Beyond diagnosis, AI is accelerating drug discovery by predicting therapeutic response, identifying biomarkers, and uncovering novel targets — shifting its role from a supplementary aid to a systemic force in oncology.
However, challenges persist: inconsistent data quality and limited accessibility hinder broader deployment, particularly in under-resourced health systems. The review emphasizes that realizing AI’s full potential requires deeper development and equitable access.
Together, these advances reflect a growing convergence in cancer research — where epigenetic profiling, single-cell AI, and integrated diagnostics are moving from isolated innovations toward a unified strategy to reduce diagnostic uncertainty and improve survival.
While the Kindai team’s model predicts cancer origin, McGill’s SIDISH identifies which cells within a tumor are most likely to drive lethality — complementary approaches that together could refine both diagnosis and treatment selection.
For breast cancer, where over 2.3 million new cases occur annually worldwide, AI’s integration across imaging, pathology, and drug development is already improving consistency in diagnosis and enabling more personalized management, though barriers to widespread adoption remain.
The overarching theme across all three efforts is a shift from treating cancer as a uniform entity to recognizing its molecular and cellular diversity — a change that could finally close the gap between diagnostic capability and therapeutic precision for the most elusive cases.
How does the DNA methylation test differ from traditional methods of identifying cancer origin?
Unlike methods relying on gene expression or histology, which can vary and degrade after metastasis, the CpG methylation model analyzes stable epigenetic fingerprints established during tissue development that persist in cancer cells, allowing accurate tracing of origin even in advanced disease.
What makes SIDISH unique compared to other AI tools in cancer research?
SIDISH uniquely connects detailed single-cell behavior to real-world patient survival outcomes by identifying rare, high-risk cell populations that drive aggressive disease — a link most bulk-analysis tools fail to capture due to averaging signals across millions of cells.
Is AI currently being used in routine breast cancer care, or is it still experimental?
AI is already applied in breast cancer care for improving mammogram interpretation, refining pathology analysis, and supporting treatment decisions, though challenges like data inconsistency and limited access prevent broader, equitable implementation across health systems.
