Blood Test for Stomach Cancer: cfDNA Methylation Markers Reach 94% Sensitivity

Gastric cancer remains one of the most lethal malignancies worldwide, largely because it is often detected at advanced stages when curative treatment options have narrowed. Current diagnostic approaches, such as gastroscopy with biopsy, are invasive, resource-intensive, and uncomfortable for patients, while existing blood-based markers lack the sensitivity needed for reliable population screening. A new study published in PLOS Genetics offers a potential way forward: a machine learning model that reads the methylation patterns of cell-free DNA circulating in plasma and distinguishes gastric cancer patients from healthy individuals with remarkable accuracy.

The research, led by Hailing Ou and colleagues, took advantage of a fundamental feature of cancer biology. Tumor cells leave behind a distinctive epigenetic signature, and fragments of their DNA released into the bloodstream carry traces of that signature. DNA methylation, the addition of chemical tags to cytosine bases that can silence genes without altering the genetic code, is one of the most stable and informative of these marks. Because methylation patterns are preserved in cell-free DNA fragments, they can serve as a molecular fingerprint of tumor presence detectable through a simple blood draw.

To identify such fingerprints, the team first performed whole-genome bisulfite sequencing on 29 pairs of gastric tumor tissue and adjacent normal tissue. Bisulfite treatment converts unmethylated cytosines into uracil while leaving methylated cytosines intact, allowing researchers to map methylation across the entire genome at single-base resolution. This analysis revealed two hallmark features of the gastric cancer epigenome: widespread global hypomethylation, in which large stretches of the genome lose their methylation marks, and focal hypermethylation at CpG islands, the gene-regulatory regions where dense clusters of cytosine-guanine dinucleotides reside. These opposing patterns are well-documented in cancer and can disrupt gene expression in ways that promote tumor growth.

The critical innovation of the study lay in connecting tissue-level findings to what can actually be measured in blood. The researchers mapped the differentially methylated regions discovered in tumor tissue onto methylation data from plasma cell-free DNA obtained from 61 gastric cancer patients and 29 healthy controls. This tissue-plasma concordance analysis identified 32 regions whose methylation changes were consistent in both compartments, suggesting that the tumor’s epigenetic alterations are faithfully reflected in the circulating DNA fragments that reach the bloodstream.

From these 32 candidate regions, the team applied two complementary feature-selection algorithms, LASSO regression and Random Forest, to narrow the panel to eight core diagnostic markers. LASSO, or least absolute shrinkage and selection operator, shrinks the coefficients of less informative features toward zero, effectively eliminating them, while Random Forest ranks features by how much they improve the purity of decision trees built on random subsets of the data. The convergence of both methods on the same eight markers strengthened confidence that these regions carry genuine diagnostic signal rather than statistical noise. One of the markers maps to the ATP8A2 gene, which encodes a phospholipid-transporting ATPase involved in membrane lipid trafficking.

With the marker panel established, the researchers built diagnostic classifiers using four machine learning algorithms: Support Vector Machine, Random Forest, K-Nearest Neighbor, and Naïve Bayes. Each algorithm approaches classification differently. Support Vector Machines find the hyperplane that best separates cancer from control samples in high-dimensional feature space; K-Nearest Neighbor classifies samples by the majority label of their closest training examples; Naïve Bayes applies probabilistic reasoning under an assumption of feature independence; and Random Forest aggregates hundreds of decision trees trained on bootstrapped samples. When the models were evaluated on training and independent test sets, the Random Forest classifier built on the eight cfDNA methylation markers performed best.

The performance figures are striking. The Random Forest model achieved an area under the receiver operating characteristic curve, or AUC, of 0.978 in the training set and 0.938 in the independent test set. An AUC of 1.0 represents perfect discrimination, so values approaching unity indicate that the model separates cancer patients from healthy controls with very few errors. In the test set, the model reached an accuracy of 88.46 percent, a sensitivity of 94.44 percent, and a specificity of 75.00 percent. Sensitivity measures how reliably the test catches true cancer cases, a crucial metric for screening, while specificity reflects how well it avoids false alarms. Because age can influence methylation patterns independently of disease, the researchers also adjusted for age and found the model retained an AUC of 0.924 on the test set, confirming that its discriminatory power is not merely an artifact of demographic differences between groups.

Beyond diagnosis, the study explored what the eight markers might reveal about tumor biology. Functional enrichment analysis linked the differentially methylated regions to biological pathways including phagocytosis, the process by which cells engulf and digest particles, and axon guidance, the signaling system that steers growing nerve fibers. The markers also showed significant correlation with immune cell infiltration patterns in the tumor microenvironment, hinting that methylation changes at these loci may reflect or influence the interplay between gastric tumors and the immune system. Such connections could eventually inform not only detection but also prognostic stratification and immunotherapy decisions.

Perhaps the most consequential technical finding concerns the limitations of existing methylation screening platforms. The Illumina methylation arrays known as 450K and EPIC v2, which are widely used in epigenetic research, do not cover the probes corresponding to any of the eight core DMRs identified in this study. In other words, researchers relying solely on array-based methylation profiling would never have found these biomarkers. This underscores a key advantage of whole-genome bisulfite sequencing: by surveying the entire methylome without probe constraints, it can uncover regulatory regions that fall outside the curated content of commercial arrays, opening discovery space that standardized platforms systematically miss.

The study’s design, matching tissue and plasma methylation profiles from the same patients before applying machine learning, offers a template for biomarker development in other cancers where non-invasive detection remains an unmet need. Liquid biopsy approaches based on circulating tumor DNA have already transformed management of several malignancies, and methylation-based cfDNA analysis is emerging as a particularly informative modality because methylation changes occur early in carcinogenesis and are tissue-specific. For gastric cancer, where endoscopic screening is impractical at scale in many regions and symptoms often appear late, a blood test achieving over 90 percent sensitivity could meaningfully shift the detection curve toward earlier, more treatable stages. The authors conclude that their tissue-plasma concordance strategy has produced a panel of plasma cfDNA methylation biomarkers with strong diagnostic potential, and that the resulting Random Forest model represents a promising non-invasive strategy for gastric cancer screening. Larger validation cohorts across diverse populations and clinical settings will be the necessary next step before such a test can move from research bench to clinical practice, but the evidence presented here demonstrates that the epigenetic whispers of a stomach tumor can indeed be heard in a tube of blood.

Subject of Research: Genome-wide cfDNA methylation biomarkers for non-invasive gastric cancer diagnosis

Article Title: Genome-Wide DNA methylation profiling of tissue-plasma pairs identifies Novel cfDNA biomarkers for the diagnosis of gastric cancer

Article References: Ou, H., Miao, J., Ling, F., Gao, J., Wang, T., Ye, Z., Zhang, C., Huang, M., & Qin, Y. (2026). Genome-Wide DNA methylation profiling of tissue-plasma pairs identifies Novel cfDNA biomarkers for the diagnosis of gastric cancer. PLOS Genetics, 22(10), e1012322. https://doi.org/10.1371/journal.pgen.1012322

Image Credits: AI Generated

DOI: 10.1371/journal.pgen.1012322

Keywords: gastric cancer, cell-free DNA, DNA methylation, liquid biopsy, machine learning, Random Forest, whole-genome bisulfite sequencing, biomarkers, epigenetics, cancer screening, PLOS Genetics, differentially methylated regions