IDDF2024-ABS-0107 In-depth plasma proteomics integrating with genetics identify protein biomarkers for gastric cancer

<h3>Background</h3> Gastric cancer is the leading cause of cancer-related deaths worldwide. This study aims to identify plasma proteomic biomarkers underlying the etiology of gastric cancer and their potential for early detection. <h3>Methods</h3> A comprehensive and unbiased plasma proteomic characterization of gastric cancer was conducted through LC-MS/MS in a case-control study involving 100 GCs and 94 non-cancerous controls (IDDF2024-ABS-0107-Figure 1. Study design and proteomic profiling through LC-MS/MS). This study also utilized two GWAS summary datasets, including independent FinnGen and UK Biobank, for proteome-wide Mendelian randomization analysis (IDDF2024-ABS-0107-Figure 3. Mendelian randomization analysis). Summary-level statistics of genetic associations with levels of 4,907 circulating proteins were extracted from a large-scale protein quantitative trait loci study in 35,559 Icelanders. MR analysis was then conducted in cis-regions to identify genes that are causally linked with gastric cancer. Furthermore, biomarker panels for detecting gastric cancer were developed using five machine-learning methods and their performance was compared to traditional risk factors and tumor biomarkers. <h3>Results</h3> A total of 4652 highly reliable plasma proteins were identified in our case-control study, of which 918 proteins were significantly associated with GC after adjusting for age and sex (IDDF2024-ABS-0107-Figure 2. Proteomic signatures significantly associated with GC). Collectively, genetically predicted levels of 12 proteins were associated with GC risk. Elevated levels of 9 proteins (APOC1, S100A7, ADH7, LYZ, NPNT, C9, MMP7, TPI1, SAR1A) and decreased levels of 3 proteins (REG4, CRISPLD2, ENO2) were associated with an increased risk of GC (IDDF2024-ABS-0107-Figure 4. Summary results from case-control study and mendelian randomization). Based on the importance of variables and area under curve (AUC), APOC1, TPI1 and C9 were selected for constructing the diagnostic model. Our machine learning-derived diagnostic model demonstrated superior performance (AUC=0.97) compared to the traditional model that used clinical risk factors (AUC=0.75). The diagnostic model showed a sensitivity of 95%, which was significantly higher than that of seven clinical blood-based biomarker tests (CA724, CA199, CA50, CA242, CEA, CA125, AFP), which had sensitivities ranging from 5% to 30%. The findings remained consistent when analyses were restricted to early GC (IDDF2024-ABS-0107-Figure 5. Diagnostic model construction and evaluation for GC). <h3>Conclusions</h3> Our study highlights the potential of plasma proteomics for accurate screening and early detection of GC. Furthermore, it provides insights into the etiology of GC.