Cancer remains one of humanity's most challenging diseases, with over 19 million new cases and 10 million deaths annually. Most cancers are detected only at advanced stages, by which point metastatic spread accounts for approximately 90% of cancer deaths. Early detection -- before spread occurs -- is one of the most powerful ways to improve survival outcomes.
While genetic mutations are widely studied as cancer drivers, changes to the epigenome -- the layer of chemical modifications that control how genes are read without changing DNA sequence -- occur even earlier. DNA methylation, which involves the addition of chemical tags called methyl groups to specific sites in the genome, is one of the most important and well-studied epigenetic modifications.
In cancer, methylation patterns go awry in two characteristic directions. Hypermethylation of gene promoters silences tumor suppressor genes, removing normal brakes on cell growth. Simultaneously, global hypomethylation destabilizes chromosomes and can activate cancer-promoting genes. Both patterns serve as molecular fingerprints of cancer.
Crucially, methylation changes appear very early in the process of cancer development, often preceding the tumor's clinical detectability. This makes methylation an attractive target for early detection tools. Computational methods capable of reading these complex methylation signatures across the genome could form the basis of new cancer screening approaches.
The researchers used DNA methylation microarray data from The Cancer Genome Atlas (TCGA), covering 13 cancer types and corresponding normal tissue samples, totaling 6,224 samples. The data measured methylation at approximately 277,000 genomic positions called CpG sites, capturing a near-complete picture of the methylome.
Four types of machine learning models were trained and compared: logistic regression, support vector machines (SVM), XGBoost (a gradient-boosted decision tree ensemble), and a deep neural network (DNN). Both binary models (cancer vs. normal for one cancer type at a time) and multiclass models (distinguishing all 13 cancer types simultaneously plus normal tissue) were developed.
The key innovation was EMethylNET, a two-stage architecture. First, an XGBoost model processed all 277,000 methylation features and identified 3,388 sites that were informative for cancer classification -- termed probes contributing to classification (PCCs). These selected features were then used as inputs to a deep neural network, creating a more robust and generalizable model.
An important design choice was allowing the XGBoost model to select features from the raw data without pre-filtering, unlike many prior approaches that first narrowed the feature set using statistical tests. This unbiased feature selection allowed the model to discover informative methylation sites that might not be identified by conventional differential methylation analysis.
The binary XGBoost models -- one for each cancer type distinguishing cancer from matched normal tissue -- achieved an average test set accuracy of 0.987. Five of the thirteen models (for colorectal, kidney, lung adenocarcinoma, lung squamous, and uterine cancer) achieved perfect performance. The Matthews correlation coefficient (MCC), a metric robust to class imbalance, averaged 0.919 across all binary models.
The multiclass XGBoost model, which simultaneously classified all 13 cancer types and normal tissue into 14 categories, achieved an overall accuracy of 0.982 and an MCC of 0.980. This demonstrates that a single model trained on methylation data can reliably distinguish between diverse cancer types -- a task comparable to what a molecular pathologist performs with conventional tissue analysis.
A particularly striking finding was that the model correctly classified adenoma samples (benign but precancerous polyps) as cancer-like in colorectal and thyroid datasets, even though no adenomas were in the training data. The colorectal model classified all adenoma samples as cancerous, consistent with their known biological position on the pathway to cancer. This suggests the model may detect early neoplastic methylation changes before frank malignancy develops.
The final EMethylNET model achieved an accuracy of 0.976 on the TCGA test set and an accuracy of 0.867 with an MCC of 0.844 on nine fully independent external datasets from seven countries, totaling 940 samples. Half of the independent datasets achieved an F1 score above 0.9, demonstrating real-world generalizability.
A key advantage of XGBoost over black-box neural networks is interpretability: the model explicitly assigns an importance score to each of its input features. The 3,388 PCCs selected by the multiclass model could therefore be mapped to nearby genes, enabling a biological investigation of what the AI had learned to detect.
Enrichment analysis revealed that the genes near the model's chosen methylation sites are significantly enriched in well-established cancer processes. There was striking overlap with the COSMIC Cancer Gene Census (140 genes, 19% of all known cancer genes) and the OncoKB Cancer Gene List (217 genes, 19.7%), both at extremely high statistical significance (p-values below 10^-17 and 10^-27 respectively).
The model's features mapped to genes involved in virtually all known cancer hallmarks: cell cycle regulation, apoptosis, DNA damage response, and metabolic reprogramming. Signaling pathways known to drive cancer -- including Wnt, TGF-beta, PI3K-AKT, RAS-MAPK, and Hedgehog signaling -- were all significantly enriched. Genes involved in cancer invasion and metastasis, immune evasion, and angiogenesis were also represented.
Text mining of over four million cancer-related PubMed abstracts confirmed that 65.6% of the model's identified genes appear in published cancer research, and 7,831 publications support cancer-associated methylation changes in 892 of these genes. The top genes identified include well-known cancer drivers such as STAT3, BRCA1, AR, MYC, NOTCH1, and TP53, providing strong biological validation for the model's choices.
An unexpected discovery was the prevalence of long non-coding RNA (lncRNA) genes among the model's selected features. While most genes encode proteins, lncRNAs produce functional RNA molecules without making proteins, and their role in cancer has only recently come to be appreciated.
Across all 13 cancer-specific gene lists, lncRNA genes made up 14-26% of the identified genes -- a much higher proportion than would be expected from the overall genomic background. The multiclass model identified 596 lncRNA-associated methylation sites.
Literature review and cross-referencing with two cancer lncRNA databases confirmed evidence for 142 of these lncRNAs in cancer. The most evidenced included HOTAIR, NEAT1, and HOTTIP -- lncRNAs with established roles in cancer proliferation, invasion, and epigenetic regulation. The most common cancer hallmarks associated with the identified lncRNAs were proliferation, invasion, and migration.
Statistical comparison with the Cancer LncRNA Census 2 confirmed that the model's lncRNAs share properties of validated cancer lncRNAs, including longer gene lengths, higher expression levels, greater genomic conservation, and differential expression in tumors. This validates them as genuinely cancer-relevant, while also suggesting that epigenetic regulation of lncRNAs may be an underappreciated aspect of carcinogenesis.
Beyond classification, the researchers tested whether the genes identified by the model could predict patient survival outcomes. For every cancer type tested, the expression of the model's identified genes was able to significantly differentiate between patients with high and low survival risk.
The most striking results were seen in head and neck squamous cell carcinoma (HNSC, p = 3.15 x 10^-16) and kidney renal clear cell carcinoma (KIRC, p = 3.06 x 10^-15), where gene expression data based on the model's methylation-identified genes powerfully separated patient survival curves.
Predictive survival models were then tested on held-out data for 5-year survival prediction. For kidney (KIRC) and colorectal (COAD) cancer, these gene lists achieved strong predictive performance with best ROC AUC values of 0.817 and 0.895 respectively. Survival prediction was less reliable for cancer types with fewer samples (such as kidney papillary and esophageal cancer).
These findings suggest that the methylation-identified genes are not just diagnostic markers but may also carry prognostic information, opening the possibility that the same genomic analysis used for cancer detection could also stratify patients by disease severity and guide treatment intensity decisions.
This study demonstrates that DNA methylation patterns alone carry sufficient information to classify 13 different cancer types with accuracy approaching 98% on held-out test data, and to generalize to heterogeneous independent datasets from multiple countries. The model's performance is competitive with or superior to related approaches in the literature.
A key advance is the interpretable design: unlike many deep learning systems that function as black boxes, EMethylNET's XGBoost backbone provides an explicit list of the methylation sites it uses. These sites map to biologically meaningful genes and pathways, meaning the model's decisions can be investigated, validated, and built upon by cancer biologists.
The most transformative future application would be extending this approach to cell-free DNA in blood -- also called a liquid biopsy. Tumor cells shed DNA fragments into the bloodstream, and these circulating DNA fragments carry the same aberrant methylation patterns as the tumor. A blood test that reads methylation patterns could potentially detect cancer early without any invasive tissue sampling.
The researchers note that with sufficient training data, this framework could be extended to hundreds of cancer types. Near-term clinical applications include screening for specific cancers in high-risk populations and determining the tissue of origin for cancers of unknown primary -- a clinically important problem where methylation's tissue-specificity is particularly valuable.