Proteomic maps of breast cancer subtypes

Abstract

Systems-wide profiling of breast cancer has almost always entailed RNA and DNA analysis by microarray and sequencing techniques. Marked developments in proteomic technologies now enable very deep profiling of clinical samples, with high identification and quantification accuracy. We analysed 40 oestrogen receptor positive (luminal), Her2 positive and triple negative breast tumours and reached a quantitative depth of >10,000 proteins. These proteomic profiles identified functional differences between breast cancer subtypes, related to energy metabolism, cell growth, mRNA translation and cell-cell communication. Furthermore, we derived a signature of 19 proteins, which differ between the breast cancer subtypes, through support vector machine (SVM)-based classification and feature selection. Remarkably, only three proteins of the signature were associated with gene copy number variations and eleven were also reflected on the mRNA level. These breast cancer features revealed by our work provide novel insights that may ultimately translate to development of subtype-specific therapeutics.

Figure 1. Super-SILAC-based quantitative proteomics of breast cancer clinical samples.

(a) Proteomics workflow involved combination of the super-SILAC mix with FFPE tumour samples, followed by FASP digestion with trypsin, peptide fractionation and analysis on the Q Exactive MS. (b) Ratio distribution between the tumour proteome (of one representative tumour) and the super-SILAC mix showed overall narrow distribution that enables accurate ratio determination. (c) Plot shows the number of identified and quantified proteins in each tumour sample. (d) Hierarchical clustering of Pearson correlations of breast cancer samples shows high diversity between tumour samples, with only partial co-clustering of samples of the same classical subtype. Two triple-positive tumours are marked with *.

Figure 2. Molecular signatures database (MSigDB) analysis.

Enrichment analysis of molecular signatures from the MSigDB database was performed on the average protein fold changes between the ER (a), Her2 (b) and TN (c) subtypes. Molecular signatures that differentiate one subtype with respect to both other subtypes are shown in orange, whereas categories that differ only between two subtypes are shown in blue or green. Complete list is given in Supplementary Data 3.

Figure 3. Tree-maps of subtype-specific KEGG categories.

Enrichment of breast cancer subtype-specific KEGG pathways is shown for ERPR (a), Her2 (b) and TN (c) breast cancer subtypes. Each Tree-map includes all KEGG pathways mapped to a high hierarchical level. The size of the boxes corresponds to the number of proteins in that category and the color—to the enrichment score S, computed with the one dimensional enrichment test. Categories in red are characterized by higher average expression in the corresponding subtype, and in dark blue—by lower average expression. The categories that are significantly enriched at FDR 5% are indicated by ***. Category ‘1' stands for ‘global and overview maps' and category ‘2'—for ‘biosynthesis of other secondary metabolites'.

Figure 4. Energy metabolism protein–protein interaction network in ERPR tumours.

Protein–protein interaction network of all proteins in our data set that mapped to the ‘energy metabolism' KEGG category was constructed in String (www.string-db.org). The color of the nodes correspond to the protein expression fold change between the ERPR and the two other subtypes; red indicates higher expression and blue—lower expression in the ERPR subtype. The size of the nodes corresponds to the absolute protein expression fold change.

Figure 5. Computational workflow of feature selection in breast cancer patients.

(a) To establish a predictive classifier of breast cancer subtypes, we developed a workflow that embeds classification and feature selection in a random sampling cross-validation procedure. In this framework, the user can define the fraction of samples to be randomly selected for the test set, as well as the number of such random samplings. The features in the training set are then ranked based on a pre-defined scoring scheme. Next, the predictive power of sets of ranked features of different sizes is computed in the test set and recorded for each cross-validation run. (b) Receiver operating characteristics (ROC) curves are given for each of the three predictors built using the optimal number of selected features. The performance of the predictors is shown by the area under curve (AUC) and ranges from 0.87 for the ERPR subtype to 0.94 for the Her2 positive subtype.

Figure 6. Comparison of SVM selected proteins on different expression levels.

(a) Mapping of the 19 signature proteins to RNA and copy number variation (CNV) data from Curtis et al.. Genes with statistically significant difference in expression at a given molecular level are marked by ‘+' and ‘−' if their expression at that level is higher or lower, respectively, in the particular subtype. Comparison of protein expression, mRNA expression and CNV distributions in each subtype are given for three selected proteins: Her2 (b), MAPK3 (c) and FOXA1 (d). All other proteins are presented in Supplementary Fig. 6. Quantitative information and local FDR values (q values) are given in Supplementary Data 4.