Survival analysis and external validation

Survival analysis was performed using Survival (http://cran.r-project.org/package=survival) and Survcomp [18] packages in R environment. The Cox proportional-hazards analysis was used for constructing a model for the prediction of survival. In this analysis, the association between a group of covariates (genes) and the response variable (DFS) was evaluated. Two datasets were employed as training and validation (test) groups, where important prognostic gene(s) was identified in a group (training group) and then validated in the other dataset (validation group). We used an external validation instead of internal validation, as the former is generally more robust to the overfitting problem [19].

First, the univariate Cox analysis was performed and genes with a z score greater than 1.5 or less than -1.5 [13, 20] were selected for the multivariate Cox analysis, where a negative score and a positive score associated with longer and shorter survival respectivley. In the multivariate Cox analysis, statistically significant genes were entered into the analysis and significant covariate(s) was detected at a P-value lower than 0.05. Survival curves were depicted by Kaplan–Meier method and compared using the log-rank test. Furthermore, some clinical prognosis parameters such as animal age, sex, and tumor grade (high or low) (Additional file 1: Table S2) were assessed in the Cox analysis to determine their roles in the prediction model.

Next, the external validation of the resulted prognostic genes was determined. The prognostic gene(s) in each group was tested in the other group via the Kaplan-Meier method and the log-rank test. In addition, the expression of the prognostic genes were compared in human ABC-like (activated B-cell like) and human GCB-like (germinal center B like) groups, because GCB-like and ABC-like cases are associated with better and poorer prognoses, correspondingly [21]. For this analysis, the patients were categorized as GCB-like and ABC-like groups based on 1,180 canine-specific differentially expressed probe sets proposed by Richards et al. (2013) [16]. Grouping was carried out using the hierarchal clustering analysis provided in geWorkbench 2.5.1 package [22]. Subsequently, the expressions of the prognostic genes were compared between the two groups using the Student's t-test analysis provided in geWorkbench 2.5.1 package.