Gastric Cancer Prognostic Signature Development

From September 1994 to December 2005, 1,557 GC patients underwent curative gastrectomy at Samsung Medical Center. Among those, 1,107 patients were selected based on following criteria: histologically confirmed adenocarcinoma of the stomach; surgical resection of tumour without macroscopic or microscopic residual disease; age ≥18; pathology stage IB (T2bN0, T1N1 but not T2aN0) to IV, according to the American Joint Committee on Cancer (AJCC) staging system (6^th Ed); complete surgical record and treatment record, and patients receiving the INT-0116 regimen as adjuvant treatment [7] (link). The study was approved by the institutional review board of the Samsung Medical Center, Seoul, South Korea (IRB approval number: SMC 2010-10-025). All study participants provided written informed consent form recommended by the IRB. In the patients who have deceased at the time of study entry, written informed consent forms were waived by the IRB. Study design and patient cohorts are provided according to REMARK guideline (Figure 1A, 1B, File S1, Section 1). Of the cohort of 1,107 patients, a discovery set of 520 patients and a validation set of 587 patients were randomly assigned and allocated to 6 batches stratified by tumor size and year of surgery for WG-DASL assay.
To avoid false-positive conclusions due to over-fitting, prognostic algorithms and their predefined cut-points were tested in independent cohorts that were not used for prognostic gene discovery and algorithm building. A 4-phase study was designed, with 4 pre-defined independent cohorts recruited from the Samsung Medical Center. The first 3 cohorts include patients with similar clinical and pathological features from chemoradiotherapy-treated study cohorts (File S1, Section 2). The first phase (discovery phase) of the study included GC patients from all clinical stages who were treated with chemo-radiotherapy (N = 520) [8] (link). Tumor blocks from these patients were subjected to prognostic gene discovery using the WG-DASL (Illumina, San Diego, CA), a microarray gene expression profiling method for FFPE [7] (link). An ad-hoc external validation of the gene set was performed to minimize any bias from single institutional cohort. The second phase (algorithm development) was to translate findings from the first phase into a clinically applicable test format. We chose the nCounter platform (Nanostring Technologies, Seattle, WA), because of its ability to interrogate the expression levels of up to 800 genes using total RNA extracted from FFPE in a single-tube reaction [8] (link). We screened stage II patients from the first phase (N = 186) for de novo discovery of prognostic genes, selected ideal combinations of genes using the gradient least absolute shrinkage and selection operator (LASSO) algorithm [10] (link), and then built a first-generation GCPS (GCPS-g1) by adding the products of normalized gene expression and coefficients from the Cox model for DFS. In the third cohort of stage II patients (N = 216). In the fourth phase (testing of clinical utility in a surgery-only setting), we tested the potential clinical utility of GCPS in stage II patients treated with surgery only. A time stamp protocol (Figure S12) was developed before processing of this final cohort. We subsequently developed a refined second-generation GCPS (GCPS-g2) (the final gene set) by analyzing the combined stage II cohorts from the second and third phases of the study.