GaEsSeer: Early detection of gastric and esophageal cancer by integrating methylation and fragmentomics signatures in cfDNA.

Abstract

4069 Background: Esophageal and gastric cancer (EC and GC) are two common cancer types that severely impact patients’ health. The 5-year survival rate for EC and GC is as low as 19% and 31%, respectively. However, early detection will significantly increase the survival rate: stage-1 EC has a 5-year survival rate of 51%, while for stage-1 GC it’s 69%. Invasive screening methods, such as endoscopy and biopsy, caused low compliance. Computational tomography and carcinoembryonic antigen were limited by low sensitivity. To address this problem, we developed GaEsSeer, a non-invasive targeted-sequencing-based assay that utilizes multiple methylation and fragmentomics features of cell-free DNA (cfDNA) to accurately detect EC and GC signals in blood. Methods: cfDNA was tested using the GaEsSeer panel, which was developed using in-house genome-wide sequencing data on EC and GC samples, and public datasets from databases and literature. Methylation features, which was quantified as methylation haplotypes or methylation encoding score, and fragmentomics features including copy number and end motif ratio were taken for modeling. Separate sub-models were trained utilizing each type of feature, which were eventually combined via logistic regression to establish the final predicting model. Results: A total of 1770 participants were recruited from multiple centers. This included 787 healthy individuals, 448 cancers (209 EC, 239 GC; stage I:156, -II:120, -III:78, and -IV:58), 174 benign esophageal diseases, and 361 benign gastric diseases. For cancer detection, the methylation-only model had an AUC of 0.909 and 0.897 in training (618 total) and test sets (617 total), respectively; while the AUC of the fragmentomics-based model was 0.885 and 0.911, respectively. The combinatorial model further improved performances, which achieves an AUC of 0.940 and 0.931 in the training and test cohorts, respectively. While the specificity remained at 96.7%, GaEsSeer detected 81.1% EC and 70.3% GC cases in the test cohort. It had a sensitivity of 74.2% and 48.9% for stage-I EC and GC, respectively. GaEsSeer also has high specificities of 87.9% and 89.8% for benign esophageal and gastric diseases, respectively. Additionally, the performance of GaEsSeer was compared with known serum cancer markers such as CEA, CA19-9, and CA72-4; and the results show that it had significantly higher sensitivity than any of these serum markers (54.8% vs 6.4% when against CEA; 53.5% vs 7.1% when against CA19-9; 50% and 16.7% when against CA72-4). Conclusions: In this pilot study, we developed the blood-based GaEsSeer assay and a model for EC and GC detection with high accuracy by stacking multiple methylation- and fragmentomics-based submodules together. Further optimization and validation of GaEsSeer using larger prospective cohorts are needed to validate its potentials for clinical application.