Integrated proteomic, transcriptomic, and biological network analysis of breast carcinoma reveals molecular features of tumorigenesis and clinical relapse

Abstract

Gene and protein expression changes observed with tumorigenesis are often interpreted independently of each other and out of context of biological networks. To address these limitations, this study examined several approaches to integrate transcriptomic and proteomic data with known protein-protein and signaling interactions in estrogen receptor positive (ER+) breast cancer tumors. An approach that built networks from differentially expressed proteins and identified among them networks enriched in differentially expressed genes yielded the greatest success. This method identified a set of genes and proteins linking pathways of cellular stress response, cancer metabolism, and tumor microenvironment. The proposed network underscores several biologically intriguing events not previously studied in the context of ER+ breast cancer, including the overexpression of p38 mitogen-activated protein kinase and the overexpression of poly(ADP-ribose) polymerase 1. A gene-based expression signature biomarker built from this network was significantly predictive of clinical relapse in multiple independent cohorts of ER+ breast cancer patients, even after correcting for standard clinicopathological variables. The results of this study demonstrate the utility and power of an integrated quantitative proteomic, transcriptomic, and network analysis approach to discover robust and clinically meaningful molecular changes in tumors.

Fig. 1.

Panels a–c depict the three approaches we took to integrate our proteomic and transcriptomic differential expression data sets with annotated biological networks. a, Approach 1 applied network analysis on the list of significantly differentially expressed genes whose corresponding proteins were also differentially expressed in NBE and MBE. b and c, Approaches 2 (gene-centric NSEA) and 3 (protein-centric NSEA) built networks in either the transcriptomic or proteomic data set, respectively, and chose among these the subset enriched in differentially expressed proteins or genes, respectively. Black and white arrows in the histological tissue sections denote NBE and IBC epithelium, respectively. d, shows the clinical validation approach we used to test gene expression clinical classifiers based on networks generated by Approaches 1–3 using independent training and validation cohorts of ER+ breast cancer patients. The numbers alongside each test and training data set indicate the number of ER+ patients used in the analysis from each in each cohort.

Fig. 2.

Concordance of differential gene and protein expression in NBE and MBE epithelial samples. a, scatter plot of differential protein expression (SpI) and differential gene expression (t statistic) for 1236 genes/proteins measured on both transcriptomic and proteomic platforms. Each point represents a gene/protein. Points are colored according to whether the corresponding gene/protein is significant in none, one, or both of the genomic and proteomic data sets. b, gene-protein set concordance analysis results for 1015 MsigDB gene/protein sets plotted as a histogram of observed normalized concordance enrichment score (black) against the background distribution obtained through shufflings of tumor and normal sample labels (red). NES scores for gene-protein sets showing significant concordance (q < 0.25, NES > 1.36) are highlighted in blue. See “Materials and Methods” for details of the analysis.

Fig. 3.

The single gene and protein network generated by Approach 3 (Network 1) that demonstrated significance as a clinical relapse predictor following a stringent permutation based test. Nodes correspond to proteins in the Ingenuity Pathway Database. a and b show the network with protein and gene expression overlaid, respectively. In each figure, red and green coloring represents proteins/genes that are up- and down-regulated, respectively, in the laser capture microdissected cancer cells relative to normal epithelium. Nodes in a and b are annotated with the spectral index and log fold change for that protein and gene, respectively.

Fig. 4.

Kaplan-Meier curves for a clinical classifier built from Network 1 from Approach 3 that was trained in Ref. and tested across four independent cohorts. Panels a–d show results for the four independent test cohorts in Refs. , , , and , respectively. Panel e shows the results of meta-analysis across all four cohorts. Kaplan-Meier curves depict the percentage of patients without disease (a, b, and d) or relapse (c) as a function of time from diagnosis, stratified by computed prognostic class. Disease and relapse events were treated equivalently in the meta-analysis results shown in e. Notches on each curve (marked with ×) represent censure times. The significance of outcome differences between the two prognostic classes in each analysis was computed using the log rank test and is displayed alongside the respective Kaplan-Meier curve.