Development and Validation of a Novel Four Gene-Pairs Signature for Predicting Prognosis in DLBCL Patients

Abstract

Diffuse large B-cell lymphoma (DLBCL) is the most common subtype of non-Hodgkin's lymphoma. Because individual clinical outcomes of DLBCL in response to standard therapy differ widely, new treatment strategies are being investigated to improve therapeutic efficacy. In this study, we identified a novel signature for stratification of DLBCL useful for prognosis prediction and treatment selection. First, 408 prognostic gene sets were selected from approximately 2500 DLBCL samples in public databases, from which four gene-pair signatures consisting of seven prognostic genes were identified by Cox regression analysis. Then, the risk score was calculated based on these gene-pairs and we validated the risk score as a prognostic predictor for DLBCL patient outcomes. This risk score demonstrated independent predictive performance even when combined with other clinical parameters and molecular subtypes. Evaluating external DLBCL cohorts, we demonstrated that the risk-scoring model based the four gene-pair signatures leads to stable predictive performance, compared with nine existing predictive models. Finally, high-risk DLBCL showed high resistance to DNA damage caused by anticancer drugs, suggesting that this characteristic is responsible for the unfavorable prognosis of high-risk DLBCL patients. These results provide a novel index for classifying the biological characteristics of DLBCL and clearly indicate the importance of genetic analyses in the treatment of DLBCL.

Keywords: diffuse large B-cell lymphoma; drug sensitivity; gene signature; prognosis; public database.

Figure 1

Workflow from training data integration to development of risk-scoring model. The training dataset 1 includes five Affymetrix microarray datasets and the training dataset 2 includes three Illumina microarray datasets were used for identification of prognostic genes. The risk-scoring model based on four prognostic gene-pairs was developed in the manner shown in Section 4.

Figure 2

Screening of 408 prognostic genes and prognostic four gene-pairs. Venn diagrams summarizing the overlap of FPGs (A) or UPGs (B) or prognostic gene-pairs (C) between two training datasets. (D–K) KM survival curves showing OS in binary score groups of four gene-pairs. p values obtained using log-rank test.

Figure 3

The predictive performance of the risk score in the training and validation datasets. (A–D) KM survival curves showing OS of five risk groups. p values obtained using log-rank test.

Figure 4

Survival analysis of the risk score in different clinical characteristics. (A) KM survival curves showing OS of patient with low-, intermediate-, and high-IPI-score. (B–D) KM survival curves showing OS of low- and high-risk patients in the three different IPI-score groups. (E) KM survival curves showing OS of patients with GCB, UNC, ABC, and MHG molecular subtypes. (F–I) KM survival curves showing OS of low- and high-risk patients in the four different molecular subtypes. (J) KM survival curves showing OS of patients with Myc-normal and double-hit DLBCL. (K,L) KM survival curves showing OS of low- and high-risk patients in the Myc-normal or double-hit DLBCL.

Figure 5

Comparison of AUC score between four gene-pair signature and nine other gene signatures. The numbers on the right side of the bar graph indicate the average AUC score at the 1-, 2-, and 3-year time points. The “Ranked AUC” is a relative ranking in predictive accuracy among 10 gene signatures [13,14,15,16,17,18,19,20,21].

Figure 6

Differential signaling pathways in low- and high-risk DLBCL. DLBCL patients in two training datasets were stratified by risk score. GSVA scores of 85 KEGG pathways in each DLBCL patients in two training datasets were illustrated by heat map.

Figure 7

Differential drug sensitivity in low- and high-risk DLBCL cell lines. (A) 17 DLBCL cell lines were stratified into low- and high-risk groups by risk score (B–G) Boxplots showing log-normalized (LN) IC50 values of anti-cancer drugs in low- (blue) and high-risk (red) DLBCL cell lines. Each circle represents LN_IC50 values in an individual cell line. p values obtained using Weltch t test.