Large-scale analysis reveals distinct molecular subtypes in real-world gastric cancer data

Introduction Gastric cancer (GC) remains a global health concern, with limited outcome improvement despite targeted therapy. Here, a large real-world dataset of diverse disease stages and tumor sites was used to profile the molecular and clinical landscape of GC. Methods De-identified clinico-genomic records from GC patients profiled with Tempus xT (DNA) and xR (RNA)-seq assays were analyzed. Molecular subtyping used non-negative matrix factorization on xR data from primary tumors of the stomach and esophagus (n=1,385; 89.9% Stage III/IV). A classifier trained on tumor-intrinsic features was applied to metastatic sites (n=640; 99% Stage III/IV) and samples with low subtype probability were excluded (n=159). Real-world overall survival (rwOS), from first-line therapy to death was evaluated in patients with available data (31.5%). Sensitivity analysis for delayed entry included prospective-like patients with samples received/sequenced before treatment. Only subtypes with ≥20 patients contributing rwOS data were reported. Results To enhance interpretability and biological relevance, nine clusters were identified and merged into seven molecular subtypes with prognostic differences (p=0.01) based on centroid distance. The G1 subtype (25.6%; median OS 12.2 months, 95% CI 6.4-16.1) was associated with chromosomal instability (CIN) (78% CIN-High), Stage IV (91%), liver site samples (26%), microsatellite stability (MSS; 97%), and low tumor mutation burden (TMB<10; 97%). G1 exhibited TP53 mutations (81.9%), CCNE1 amplifications (16.7%), and CDKN2A / B - MTAP deletions (~15%). The G2 subtype (21.8%; median OS 20.6 months, 95% CI 14-22.2) had the highest TMB (TMB≥10; 21%), microsatellite instability (MSI; 20%), frequent ERBB2 (14.3%) and FGFR2 (5.9%) amplifications, and lower TP53 mutations (62.6%). The G3 subtype (21.8%; median OS 18.1 months, 95% CI 11.6-27.4) had diffuse histology (58%), CDH1 mutations (21.2%), low TMB (94%), and stomach site enrichment (54%). The G4 subtype (5.6%) was enriched for KRAS mutations (11.4%) and amplifications (9.6%), TP53 mutations (75.4%), and CIN (62%, trending significance). The G5 subtype (6.4%) had the highest PD-L1 (CPS≥10; 40%), EBV positivity (16%), high TMB (17%), and low CIN (55%). G5 exhibited ARID1A / B (41.9%/9.3%), PIK3CA (21.7%), and KRAS mutations (18.6%), and had the lowest TP53 mutation proportion (46.5%). The G6 subtype (4.9%) was enriched for Stage III (22%) and CDKN2A / B - MTAP deletions (~13%). The G7 subtype (17.1%; median OS 20.6 months, 95% CI 7.6-15.5%) was associated with diffuse histology (67%), CDH1 mutations (23.9%), MSS (97%), low TMB (97%), and low CIN (53%). Conclusions This study shows that real-world data integration enables identification of clinically relevant, biologically distinct GC subtypes with prognostic differences. These findings provide a foundation for tailoring therapeutic strategies and improving patient outcomes.