An Evaluation of DMR-Informed Fine-Tuning of Tissue Array–Pretrained CpGPT for Gastrointestinal Cancer Classification Using cfDNA Targeted Methylation
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Cell-free DNA (cfDNA) methylation profiling is promising for minimally invasive cancer detection, but its translation is limited by high-dimensional data, modest cfDNA cohort sizes, and the difficulty of defining biologically grounded feature sets. CpGPT, a transformer-based DNA methylation foundation model pretrained on large-scale tissue methylation array datasets, may enable transfer of learned methylation representations to data-limited cfDNA applications. We evaluated a differentially methylated region (DMR)-informed framework for fine-tuning CpGPT for gastrointestinal cancer classification using plasma cfDNA methylation data.
Methods
Plasma cfDNA methylation data were obtained from the EpiPanGI Dx cohort, including 254 gastrointestinal cancer samples and 46 non-cancer controls. Tumor and matched normal tissue methylation data included 967 tumor-normal pairs across 18 cancer types from The Cancer Genome Atlas (TCGA). Tissue-derived and cfDNA-derived DMRs were identified independently and intersected to define a candidate set of 20,499 CpG sites. CpGPT was evaluated in a technical sensitivity analysis using the previously published 896-CpG panGI panel and in DMR-informed fine-tuning using the 20,499-CpG candidate set. Performance was compared with radial-kernel support vector machine (SVM radial), elastic net logistic regression, gradient boosting machine (GBM), and Random Forest models using identical data partitions.
Results
In the technical sensitivity analysis using the previously published panGI panel, the CpGPT base configuration achieved a mean test area under the receiver operating characteristic curve (AUROC) of 0.9799 (95% CI, 0.9654–0.9944) across four fixed random seeds. Individual test AUROCs ranged from 0.9597 to 0.9927, while neighboring configurations achieved mean test AUROCs of 0.9661–0.9716. Using the DMR-informed candidate set, CpGPT achieved a mean test AUROC of 0.9904 (95% CI, 0.9715–1.0000) across three split-runs. Mean test AUROCs were 0.9573 for SVM radial, 0.9420 for elastic net, 0.8460 for GBM, and 0.8408 for Random Forest. Cancer type-specific mean test AUROCs ranged from 0.9600 for pancreatic adenocarcinoma to 1.0000 for esophageal squamous cell carcinoma, colorectal cancer, and esophageal adenocarcinoma.
Conclusions
DMR-informed CpGPT fine-tuning achieved high internal discrimination and a higher mean test AUROC than the evaluated classical machine learning models. Integrating tissue and plasma cfDNA methylation evidence provides a biologically constrained feature space for adapting a tissue array-pretrained foundation model to cfDNA classification. Independent external validation, clinically representative control populations, and further feature-set reduction are needed before translation into a targeted cfDNA assay.
