br Technologies Samples were diluted fold in
Technologies). Samples were diluted 100-fold in this step to minimize the matrix effect and the chance of extension generated by unpaired ol-igonucleotides. Next, 2·8 μl of each first round PCR product was mixed with 7·2 μl detection reagent from which 5 μl was loaded into the sam-ple wells, while specific pairs of PCR primers were loaded into the primer wells of a microfluid chip (Fluidigm 96.96 Dynamic Array, Fluidigm, CA, USA). The microfluid chip was then primed in Fluidigm IFC controller and loaded for real-time qPCR in Fluidigm Biomark™ thermocycler. Quantification cycle (Cq) value was converted to Normal-ized Protein eXpression (NPX) by normalizing with extension- and neg-ative controls spiked in each sample. NPX is a unit on log2 scale, where one NPX increase corresponding to a two-fold increase in concentration of the protein. LOD was determined as the NPX value three times the standard deviation beyond its background. Multiplex PEA detects rela-tive protein levels but not absolute concentration. Correlations between NPX values and protein concentrations in mass unit (pg/ml) are avail-able at Olink's website (www.olink.com/products/complete-biomarker-list). The specificity for each panel was determined by carry-ing out the whole assay in which the test samples were pools of full length recombinant Minocycline HCl corresponding to every block of 8 proteins from all the 92 proteins in the panel, resulting in signals generated only from the present proteins but not the others (detailed in https://www. olink.com/data-you-can-trust/validation/).
For proteins present in more than one panel, only one was chosen for further analysis. Thus, seven proteins from commercial panels and 18 from experimental panels were removed. The assay reproducibility for the same proteins is displayed in Supplementary Fig. 2, while the re-producibility for sample duplicates is shown in Supplementary Fig. 3.
Data analyses were performed using R software (www.r-project.
NPX values were used for analysis since the values tended towards a normal distribution. To minimize the inter-plate variation, samples from different disease groups were evenly distributed throughout the plates, and the inter-plate variation was further normalized for each protein in each plate by adding the Z-score factor calculated as follows: factor = (actual value – median of all samples)/standard deviation. For comparison between GC and controls, NPX values were adjusted if a sig-nificant effect (corrected p value b0·05) of age or gender on the protein levels was found by linear regression in both the control and cancer groups. Therefore, 13 proteins measured in serum (CDCP1, CTSV, CXCL9, EPHA1, KIT, OPG, RET, RSPO3, TGFBR2, TNFRSF10B, TRANCE, VEGFR2, WFDC2) were adjusted for age, while no protein was found significantly associated with gender in either group. Principal compo-nent analysis (PCA) was applied for an overview of the relationships be-tween variables and the presence of outliers.
Differences between two groups for continuous variables were analysed by non-parametric Mann-Whitney-Wilcoxon test, while for category variables, Chi-square or Fisher's exact test was performed, and ANOVA was applied for comparisons of more than two groups. Dif-ferences between before and after surgery were tested using paired Mann-Whitney-Wilcoxon test. Correlation coefficients or co-linearity between each two protein markers were tested by Spearman's rank rho. In order to manage multiple tests errors, P-values were adjusted using the Benjamini-Hochberg procedure using 5% as an acceptable false discovery rate .
To evaluate the diagnostic performance, receiver operating charac-teristic (ROC) curves were constructed, and areas under ROC curve (AUC), optimal cutoff, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated through R packages pROC  and ROCR . The optimum cutoff value was de-fined by maximizing the Yoden's index (sensitivity+specificity-1).