Machine Learning DNA Methylation Model Shows Promise as Early Ovarian Cancer Detection Tool

By: Lisa Astor
Posted: Monday, September 22, 2025

Researchers have developed and validated a prediction model for high-grade serous ovarian cancer using step-wise artificial intelligence (AI) methodology and methylated DNA probes that showed an area under the curve of 100%, and 84% by external validation. The report of the model and pilot study was published in Scientific Reports.  

The study authors sought to create a more effective screening tool for epithelial ovarian cancer in order to detect cancer cases before they progress, become symptomatic, and become harder to treat. They believed that a combination of adequate markers, including DNA methylation, and artificial intelligence could help to improve the predictive ability of ovarian diagnostic tools.  

The researchers conducted a pilot case-control study of 99 samples of patients with high-grade serous ovarian cancer and 12 normal fallopian tube samples for controls, all of which were obtained from the Gynecologic Oncology Bank of the University of Iowa.  

Using the Illumina Infinium MethylationEPIC BeadChip Array, they examined over 850,000 methylation sites. Variable reduction began with the deep-learning tool MethylNet, followed by univariate analysis of variance and multivariate lasso regression, so that the model could be verified more easily. This step-wise process identified nine highly informative probes capable of distinguishing high-grade serous ovarian cancer with an area under the curve of 100% in the training set. 

External validation in an independent dataset confirmed strong performance, with areas under the curve of 98% (95% confidence interval [CI] = 95%–100%) for a larger probe set and 84% (95% CI = 76%–93%) for the simplified nine-probe model. Importantly, results held across multiple platforms, including the machine learning analytic platform TensorFlow. Although the study was limited by its small and relatively homogeneous patient cohort, the findings do provide proof-of-concept for a clinically useful biomarker panel. 

“For this model to be effective as early diagnosis or screening, we need to optimize its performance in blood, with a diverse population of patients [with epithelial ovarian cancer] from all stages and appropriate controls, with and without benign pelvic masses,” the study authors, led by Jesus Gonzalez Bosquet, MD, PhD, Associate Professor of Obstetrics and Gynecology-Gynecologic Oncology, University of Iowa, noted. “Ideally, for this model to be generalizable, the validation cohort should have similar composition of these conditions that the general population. Only then we would have an idea of the discriminatory potential of this model. That is our future goal,” they concluded.  

Disclosure: For full disclosures of the study authors, visit nature.com.

Scientific Reports

---