When cancer metastasizes, malignant cells break away from the original tumor site and travel through the vascular or lymphatic systems to colonize distant organs. In approximately 2% to 5% of all cancer diagnoses, clinicians detect these secondary tumors but are entirely unable to locate the original source. This clinical dilemma is known as Cancer of Unknown Primary (CUP).
The inability to find the primary site creates a critical therapeutic bottleneck. Modern oncology relies heavily on site-specific treatment protocols; a breast cancer cell that has migrated to the liver retains the biological vulnerabilities of breast tissue and responds far better to breast cancer therapies than to standard liver cancer regimens. Without a known origin, clinicians must resort to broad, empirical chemotherapies, which frequently result in poor patient outcomes.
Unusual clinical presentations underscore this diagnostic challenge. For instance, male breast cancer can occasionally manifest solely as an axillary lymph node mass without any palpable breast tumor, as detailed by Guo-Li Gu et al. in Axillary metastasis as the first manifestation of male breast cancer: a case report and by Lang Chen et al. in Misdiagnosed male breast cancer with an unknown primary tumor: A case report. In such scenarios, the primary tumor remains hidden, mimicking an anonymous metastatic cancer from an unidentified source.
To solve this riddle, a major research direction has emerged: using computational models to infer a tumor's tissue of origin by decoding its molecular and morphological signatures.
The Core Intuition: Lineage and Molecular Memory
Why should we be able to predict the primary site from a metastatic biopsy? The fundamental insight is that tumors retain a molecular memory of their tissue of origin.
During embryonic development, cells undergo strict epigenetic and transcriptional programming to differentiate into specific lineages—becoming a lung cell, a colon cell, or a breast cell. When a cell undergoes malignant transformation and metastasizes, it experiences massive genetic mutations and structural chaos. However, it rarely erases its entire developmental blueprint. The cell's baseline gene expression, epigenetic markings, and even certain structural mutation patterns remain anchored to its tissue of origin. By training machine learning algorithms on large datasets of known primary tumors, researchers can build classifiers that recognize these subtle, underlying blueprints in metastatic tissue.
[Healthy Tissue] ──(Developmental Lineage)──> [Unique Epigenetic/RNA Blueprint]
│
(Malignant Transformation)
▼
[Metastatic Tumor] ◄──(Computational Model)─── [Retained Molecular Memory]
Transcriptomic Fingerprints
The most direct way to read a tumor's molecular memory is through its transcriptome—the complete set of RNA transcripts produced by the genome.
In a foundational study, Sridhar Ramaswamy et al. demonstrated in Multiclass cancer diagnosis using tumor gene expression signatures that multiclass cancer diagnosis using gene expression signatures could scale to handle the reality of clinical oncology. This shifted the paradigm from simple binary classifications to tracking complex, multi-site tissue profiles.
But which molecular layer holds the clearest signal for tracing a tumor's origin? To answer this, Haiyan Liu et al. systematically evaluated different biomarkers in Evaluating DNA Methylation, Gene Expression, Somatic Mutation, and Their Combinations in Inferring Tumor Tissue-of-Origin. The authors discovered that gene expression serves as the single superior predictor, achieving over 94% accuracy in identifying the tissue of origin. Interestingly, they noted that combining multiple types of biomarkers did not improve performance, confirming that the transcriptome alone captures the necessary lineage signals.
To translate this insight into a scalable clinical tool, Jackson Michuda et al. developed a machine learning classifier trained on a massive dataset of more than 43,000 tumor biopsies. In Validation of a Transcriptome-Based Assay for Classifying Cancers of Unknown Primary Origin, the authors demonstrated that their model achieved 91% accuracy across 68 distinct cancer subtypes, proving that transcriptomic profiling can reliably resolve diagnostic uncertainty in CUP cases.
Decoding the Genomic Blueprint
While RNA expression provides a dynamic snapshot of cellular identity, the underlying DNA also holds vital structural clues. However, reading DNA signatures requires navigating a landscape altered by severe genomic instability.
As shown by Jonas Vikeså et al. in Cancers of unknown primary origin (CUP) are characterized by chromosomal instability (CIN) compared to metastasis of know origin, CUPs are characterized by exceptionally high levels of chromosomal instability and DNA damage repair signatures. This biological volatility makes them highly aggressive and difficult to track.
Despite this genomic chaos, machine learning models can isolate predictive structural patterns from DNA. Kee Pang Soh et al. combined somatic mutations and copy number changes using a Support Vector Machine framework in Predicting cancer type from tumour DNA signatures, achieving high prediction accuracy with a minimal gene set.
Similarly, Yulin Zhang et al. developed an XGBoost-based method in A Novel XGBoost Method to Identify Cancer Tissue-of-Origin Based on Copy Number Variations. By focusing exclusively on Copy Number Variations (CNVs) and applying dimensionality reduction, the authors demonstrated that large-scale structural DNA alterations retain spatial, tissue-specific patterns capable of pinpointing the primary tumor site.
Morphological Inference via Deep Learning
While molecular profiling is highly accurate, it can be resource-intensive, costly, and slow. An alternative approach asks: can we infer the primary site using standard, widely available clinical data, such as hematoxylin and eosin (H&E) stained histology slides?
This is the domain of computational pathology. Drew FK Williamson et al. introduced a deep learning model called TOAD in AI-based pathology predicts origins for cancers of unknown primary. The model uses multi-task learning on whole-slide images to predict whether a tumor is primary or metastatic, while simultaneously identifying its site of origin.
[H&E Histology Slide] ──> [Deep Neural Network] ──> [Feature Extraction] ──> [Predicted Origin]
(Cellular Architecture,
Nuclear Pleomorphism,
Stromal Interactions)
This approach works because deep neural networks can extract multi-scale spatial features that elude human eyes. While a human pathologist might struggle to differentiate highly de-differentiated metastatic cells, a neural network can recognize subtle geometric arrangements, nuclear pleomorphism, and stromal interactions that correlate strongly with the parent tissue. This method democratizes primary site identification, offering a rapid, low-cost alternative to genomic sequencing.
From Diagnostics to Precision Therapeutics
The ultimate goal of identifying the primary site is to transition patients from broad empirical treatments to targeted precision therapies. Comprehensive genomic profiling can reveal actionable therapeutic targets in up to 75% of CUP patients.
However, a major clinical hurdle remains. As noted by Lena Weiss et al. in Cancer of unknown primary (CUP) through the lens of precision oncology: a single institution perspective, the primary barrier to precision medicine in CUP is often the rapid clinical decline of the patients. Because these cancers are highly aggressive, patients frequently deteriorate before traditional, multi-step diagnostic workflows can identify a primary source.
This reality underscores the true value of computational tissue-of-origin tools. By deploying machine learning classifiers—whether transcriptomic, genomic, or image-based—as frontline diagnostic protocols rather than measures of last resort, clinicians can rapidly trace a tumor's lineage. This speed allows oncologists to deploy site-specific therapies and targeted inhibitors early enough to alter the clinical trajectory of the disease.
Key Foundational Work
Axillary metastasis as the first manifestation of male breast cancer: a case report (2008) — contributed 4 concept elements to this direction.
Multiclass cancer diagnosis using tumor gene expression signatures (2001) — contributed 1 concept elements to this direction.