Noninvasive detection of early gastric cancer via red blood cell DNA features associated with tumor-induced hematopoietic stress.

4093 Background: Early detection of gastric cancer (GC) is limited by invasive endoscopy and the low sensitivity of current blood-based biomarkers for early-stage disease. Emerging evidence indicates that GC engages in complex crosstalk with the bone marrow microenvironment, imposing stress on hematopoiesis. Tumor-derived exosomal and inflammatory mediators reprogram transcriptional and epigenetic profiles in hematopoietic progenitors, driving abnormal hematopoiesis and inducing genomic instability. Here, we report a noninvasive strategy for GC detection based on genome-wide profiling of DNA remnants in mature red blood cells (rbcDNA) and validate its performance for early-stage disease across independent multicenter cohorts. Methods: We analyzed rbcDNA from 1–2 mL peripheral blood of 865 individuals, including 434 treatment-naive patients with stage I–III GC (~60% stage I) and 431 non-GC controls with benign gastric conditions. Shallow whole-genome sequencing revealed reproducible rbcDNA read-depth features that distinguished GC from non-cancer states. A machine-learning classifier was trained on dimensionally reduced rbcDNA features in a discovery cohort (n = 435), with a cutoff fixed at 90% specificity. The locked model was then evaluated in a test cohort (n = 109) and three independent validation cohorts from distinct medical centers (n = 321). Results: In a discovery cohort, we identified GC-associated rbcDNA features that were significantly enriched in hematopoietic regulation and cell-cycle pathways, reflecting systemic hematopoietic stress rather than tumor-derived copy-number alterations. The resulting classifier achieved AUCs of 93% in the discovery cohort and 95% in the test cohort. At the predefined cutoff, the classifier yielded 83% overall sensitivity and 79% sensitivity for stage I GC in the test cohort. Across three independent medical center cohorts, the model achieved sensitivities of 83%, 90%, and 86%, with specificities ranging from 89% to 94%, consistent with the discovery and test cohorts. The assay effectively distinguished early-stage GC from precancerous lesions, including atrophic gastritis and intestinal metaplasia, while showing limited detection in non-gastric solid tumors. Importantly, rbcDNA identified ~80% of GC cases that were negative for conventional serum tumor markers. Conclusions: These findings establish rbcDNA as an effective biomarker for accurate, noninvasive detection of early-stage GC.