Introduction
Esophageal adenocarcinoma (EAC) is increasingly prevalent, with most cases detected in advanced stages and exhibiting less than a 20% 5-year overall survival rate. Since this cancer develops from Barrett's Esophagus (BE) through dysplastic stages, there is a clear opportunity for early diagnosis. The capsule sponge test, recommended for patients with Gastroesophageal Reflux Disease (GERD) in the AGA, ACG, and ESGE guidelines, facilitates non-endoscopic cell collection for BE detection coupled with biomarker testing. It offers a 10-times higher BE detection rate compared to standard of care, demonstrated with Trefoil Factor 3 (TFF3) as a proteomic biomarker for intestinal metaplasia (IM). However, there is potential for improved molecular biomarkers, particularly DNA methylation, due to their non-subjective, quantitative nature and scalability.
Aims
The primary objective of this study was to assess whether in a cohort with known TFF3 status we can discover and apply methylation biomarkers to increase sensitivity in the context of previously missed diagnoses while retaining specificity.
Methods
We analyzed 190 capsule sponge samples separated into two distinct groups: 144 TFF3-positive cases confirmed through endoscopy (BE+, Prague M: 1-22 cm) with varying extents of IM visible on matched TFF3 pathology (13 out of 144 had no TFF3-positive signal), and 46 TFF3-negative reflux samples without suspected BE (BE-).
We profiled the methylation status of approximately 4 million CpG sites in 190 capsule sponge samples. A generalized linear regression model (GLM) framework was employed to identify differentially methylated regions (DMRs) in samples with and without TFF3-confirmed presence of IM. The top 10% most differential DMRs were used to train a Random Forests classification model for feature selection. Subsequently, these DMRs were employed as predictors in a machine learning model (xgbDART). To select the most robust DMRs, we performed a Kolmogorov-Smirnov test and ranked the 89 DMRs by significance. The lowest-ranking DMRs were sequentially excluded in subsequent model training, using leave-one-out cross-validation for performance assessment.
Results
The optimal xgbDART model with the 9 most significant DMRs following feature elimination achieved an AUC-ROC of 0.96 (CI: 0.92-0.99), an average sensitivity of 96% (CI: 0.92-0.98), and an average specificity of 93% (CI: 0.82-0.98) for diagnosing BE. These results suggest that a selected group of DNA methylation biomarkers using a pragmatic discovery approach can further improve the sensitivity in detecting Barrett's Esophagus in capsule sponge samples compared to the current gold standard TFF3.
Application of a DMR panel to capsule sponge samples has potential as a scalable, quantifiable BE diagnostic test. Additional validation of these results is required in the relevant population.

<b>Figure 1: A</b>. Box-plot showing methylation levels (beta-value, y-axis) of the top 9 selected Differentially Methylated Regions (DMRs) in 144 TFF3-positive cases confirmed through endoscopy (BE+, blue) when compared to 46 TFF3-negative reflux samples without suspected BE (BE-, dark gray). Wilcoxon test p-value < 0.001 (***). <b>B.</b> Area under the receiver operating characteristic curve (AUC-ROC) of training data set for BE detection using top 9 DMRs after performing leave-one-out cross-validation. 95% confidence level used to compute confidence interval.

Figure 1: A. Box-plot showing methylation levels (beta-value, y-axis) of the top 9 selected Differentially Methylated Regions (DMRs) in 144 TFF3-positive cases confirmed through endoscopy (BE+, blue) when compared to 46 TFF3-negative reflux samples without suspected BE (BE-, dark gray). Wilcoxon test p-value < 0.001 (***). B. Area under the receiver operating characteristic curve (AUC-ROC) of training data set for BE detection using top 9 DMRs after performing leave-one-out cross-validation. 95% confidence level used to compute confidence interval.

Alex Bowman

Appearances