Patients and sample collection
This study was approved by the Institutional Review Board (IRB) of Kanagawa Cancer Center, the central institute for this study (approval number: 26 - 42), as well as the IRBs of all institutions that participated in this study. Representative blocks from formalin-fixed, paraffin-embedded (FFPE) gastrectomy specimens were collected retrospectively from participating institutions according to the following inclusion criteria: (1) patients were participants in the SAMIT, (2) FFPE blocks or unstained cut sections were available, and (3) the translational study protocol was approved by the IRB of the respective institution. Samples were collected by the data center of Kanagawa Cancer Center and shipped to Yokohama City University for RNA extraction and analysis. Sections (each 10-μm thick) were cut from the FFPE blocks and stored at 4°C until microdissection.
RNA extraction and complementary DNA (cDNA) synthesis
Hematoxylin and eosin-stained slides were reviewed, and the area with the highest tumor content was outlined manually. After manual microdissection, total RNA was isolated using NucleoSpin® FFPE RNA XS (MACHEREY-NAGEL GmbH & Co. KG, Düren, Germany). For RNA quality control, the OD260/OD280 ratio was measured using a NanoDrop 2000 (Thermo Fisher Scientific Inc., MA, USA; RRID:SCR_018042). The total RNA integrity number was measured using an Agilent 2100 Bioanalyzer (Agilent Technologies Inc., Waldbronn, Germany, RRID:SCR_018043). To confirm that the total RNA samples were not contaminated with DNA, RNA18S1 expression was evaluated by quantitative real-time PCR (qRT-PCR) in each sample before cDNA preparation. cDNA was prepared from samples that passed all the quality control checks. cDNA was synthesized from 0.4 µg of total RNA using an iScript cDNA Synthesis Kit (Bio-Rad Laboratories, Inc., CA, USA), diluted to 0.2 µg/µL with distilled water, and stored at −20°C until use.
qRT-PCR was performed using the QuantiFastTM Probe Assay (QIAGEN, Venlo, Netherlands) and QuantiFastTM Probe PCR (QIAGEN) according to the manufacturer’s instructions. The expression of each gene was quantified in triplicate. A standard curve was plotted for each run using three fixed concentrations of human control cDNA synthesized using Xpress Ref Universal Total RNA (QIAGEN) with an iScript cDNA Synthesis Kit (Bio-Rad Laboratories, Inc.) to measure the mRNA expression levels in all samples. The concentration of each sample was determined based on the point of intersection of the sample value with the standard curve. β-actin and RNA18S1 were used as the internal controls.
The 105 selected genes included 63 genes analyzed in an exploratory biomarker study of ACTS-GC trial participants10. Of these, 57 genes have been previously reported as biomarkers of paclitaxel resistance or sensitivity. The functional annotation of each gene using DAVID 6.7, is outlined in Supplementary Table S4 (Online Resource 1).
Defining the predictive value of the biomarkers
The mRNA expression level of each gene was classified as low versus high using the median mRNA expression level as a cut-off point, as described previously40. If the mRNA expression level of a particular gene was below 1.0×10-8 ng/μL, the expression level was set to ‘0.00’. The value of a biomarker in predicting the benefit of sequential paclitaxel based on the OS, DFS, and cumulative incidence of relapse was determined by examining the p-values of the interaction between the dichotomized gene expression level and the treatment group (sequential paclitaxel versus monotherapy) after adjusting for clinical and pathological factors using Cox regression or Fine-Gray models41,42. The genes were ranked according to treatment interaction-related p-values. Values were considered significant at p<0.05. Additionally, we combined the expression levels of selected genes to identify sensitive and non-sensitive patient subsets.
We adopted an internal validation strategy, as proposed by Wahl et al.43, to address the potential overestimation of the standard error owing to multiple imputations and optimism in the predictive performance. We used Harrell’s C statistics to analyze the predictive performance of the survival data and addressed the optimistic bias by Harrell’s C statistics using the bootstrap 0.632+ method with 20 bootstrap samples from the original dataset with replacement, followed by multiple imputations.
The pre-defined statistical analysis plan for this study has been reported previously40. The primary and secondary endpoints were the OS and DFS, respectively. The OS and DFS curves were constructed using the Kaplan-Meier method, and the cumulative incidence curves of relapse were constructed using the Aalen-Johansen method44 to compare sequential paclitaxel and monotherapy, considering the expression levels of the selected genes either individually or in combination. The adjusted hazard ratios (HRs), 95% confidence intervals (CIs), and p-values of the major treatment effects and interactions were estimated for the entire patient population and subgroups according to the UICC TNM 8th ed stage2. We used multiple imputations to handle missing clinical and pathological factor data and generated 20 multiply imputed datasets for parameter estimates. The reported p-values were two-tailed, and the major effects and interactions were considered statistically significant at p<0.05. Statistical analyses were performed using SAS version 9.4 (SAS Institute, Inc., Cary, NC, USA).