GUIDE: a prospective cohort study for blood-based early detection of gastrointestinal cancers using targeted DNA methylation and fragmentomics sequencing

Abstract

Background: Gastrointestinal (GI) cancers are among the most prevalent and lethal malignancies worldwide. Early, non-invasive detection is essential for timely intervention and improved survival. To address this clinical need, we developed GutSeer, a blood-based assay combining DNA methylation and fragmentomics for multi-GI cancer detection.

Methods: Genome-wide methylome profiling identified 1,656 markers specific to five major GI cancers and their tissue origins. Based on these findings, we designed GutSeer, a targeted bisulfite sequencing panel, which was trained and validated using plasma samples from 1,057 cancer patients and 1,415 non-cancer controls. The locked model was blindly tested in an independent cohort of 846 participants, encompassing both inpatient and outpatient settings across five hospitals.

Results: In the validation cohort, GutSeer achieved an area under the curve (AUC) of 0.950 [95% Confidence Interval (CI): 0.937-0.962] for cancer detection, with 82.8% sensitivity (95% CI: 79.5-86.0) and 95.8% specificity (95% CI: 94.3-97.2). It detected 92.2% of colorectal, 75.5% of esophageal, 65.3% of gastric, 92.9% of liver, and 88.6% of pancreatic cancers. The independent test cohort included 198 early-stage cancers (stage I/II, 66.4%) and 63 advanced precancerous lesions. GutSeer maintained robust performance, with 81.5% sensitivity (95% CI: 77.1-85.9) for GI cancers and 94.4% specificity (95% CI: 92.4-96.5). It also demonstrated the ability to detect advanced precancerous lesions in the colorectum, esophagus, and stomach as a single, non-invasive blood test.

Conclusions: By integrating DNA methylation and fragmentomics into a compact panel, GutSeer outperformed genome-wide sequencing in both accuracy and clinical applicability. Its high sensitivity for early-stage GI cancers and practicality as a non-invasive assay highlights its potential to revolutionize early cancer detection and improve patient outcomes.

Trial registration: ClinicalTrials.gov identifier: NCT05431621.

Keywords: Cell-free DNA (cfDNA); Fragmentomics; Gastrointestinal cancer; Methylation; Multi-cancer early detection (MCED); Multi-dimensional features.

Fig. 1

Overall design of the GutSeer study

Fig. 2

Validity of the GutSeer panel. A The validity of the GutSeer panel was verified by the methylation feature (i.e., average methylation fraction, AMF). Tissue samples of cancer and para-carcinoma, and plasma samples from healthy subjects are well-separated by U-MAP of the AMF data. B The heatmap based on AMF for different cancer types, with top associated biological processes enrichment listed. C The validity of the GutSeer panel was verified by the fragmentomics feature (i.e., tFPKM). Plasma samples from liver cancer patients and plasma from healthy subjects were separated by U-MAP using the tFPKM data

Fig. 3

Evaluation of GutSeer in cancer/non-cancer identification and TOO localization. A,B ROC of cancer/non-cancer prediction, with comparisons among methylation, fragmentomics, and the integrated measurements. A The average ROC based on ten-fold cross-validation in the training set. B The ROC based on the validation set. Methylation: the commonly used methylation measurements, AMF and MHF; fragmentomics: tFPKM and end motif. C,D Confusion matrices of TOO prediction based on the training (C) and validation (D) datasets. Accuracy (column rates) represents the percentage of correct localization within each actual TOO; precision (row rates) represents the percentage of correct localization within each predicted TOO. EsoSto represents esophageal and/or gastric cancers

Fig. 4

Assessment of GutSeer in differentiating non-cancer versus different stages of cancer. Sensitivity (y-axis) is reported by clinical stage (x-axis) in the all five GI cancers (A) and in each of the five cancers (B-F), at the specificity of 95.8% and 94.4% for training/validation and independent test cohorts, respectively. Sensitivities were plotted with 95% confidence interval. Numbers indicate samples in training, validation and independent test sets

Fig. 5

Head-to-head comparison between GutSeer and the WGS-based approach. A ROC of cancer/non-cancer classification in the test dataset for GutSeer and FragD, which is a representative WGS-based approach. The P-value of Delong test between GutSeer and FragD was shown in the plot. B Accuracy of the tissue of origin (TOO) identification in the test dataset, for all cancers and each cancer type. Numbers indicate samples in each data set. Comparison was performed between GutSeer and FragD using χ.² test. N.s., no significance; *, P-values < 0.05; **, P-values < 0.01