Accurate Detection of Esophageal Squamous Cell Carcinoma (ESCC) Using Machine Learning with Methylated DNA Biomarkers

Esophageal cancer, with 544,076 deaths in 2020,1 includes esophageal adenocarcinoma and esophageal squamous cell carcinoma (ESCC). ESCC comprises 90% of esophageal cancers globally. The 5-year survival rate remains poor (15%–25% worldwide) because patients usually present with late disease.2 Although ESCC can be diagnosed by means of esophagogastroduodenoscopy (EGD) with biopsy, EGD is not widely available in low-income countries.2 Inexpensive, safe, accessible diagnostic alternatives will likely improve diagnosis and outcome.

DNA methylation often silences gene expression, inducing loss-of-function of tumor suppressor genes. Hypermethylation occurs at high levels and frequencies in ESCC. However, these potential biomarkers have not been tested extensively in minimally invasively–obtained cytologic samples from patients with ESCC, where they could increase detection of ESCC.

Less invasive, nonendoscopic sampling techniques have examined Barrett’s esophagus, a precursor of esophageal adenocarcinoma.3, 4, 5, 6 A recent article described samples of mostly squamous esophageal dysplasia obtained via endoscopic biopsy, brushings, or older balloons in Iran and China.7 A study by Middleton et al8 was the first to use a sponge-based approach to screen for ESCC in Tanzania. Our approach uses quantitative methylation-specific polymerase chain reaction instead of histology, which requires pathologists to interpret findings of paraffin-embedded cell clots. Our study combines sponge-based sampling with a predictive biomarker panel to develop an accessible ESCC screening method in low-income countries, where EGD is not widely available. We used epigenetic changes (DNA methylation levels) as a basis for this strategy.

To discover and select epigenetic biomarkers, we first searched PubMed and The Cancer Genome Atlas for ESCC markers described previously (Supplementary Figure 1A). We then selected CpG sites using minimum and maximum β-value (percent methylation) thresholds: high methylation levels in ESCC, but low levels in all types of normal tissues. Markers emerging from this step were tested in fully methylated DNA, unmethylated DNA, and finally paired normal–tumor tissues. Markers passing these steps were used to construct a multivariable model on a training set composed of EsophaCap sponge samples. Finally, this model was applied to a separate independent test set.