BreastMark: An Integrated Approach to Mining Publicly Available Transcriptomic Datasets Relating to Breast Cancer Outcome

Gene expression data sets were downloaded from the Gene Expression Omnibus [18] or authors' websites in the form of raw data files, where possible. Only breast cancer datasets with survival information and at least 48 patients were included. Large datasets were chosen for this analysis so as to avoid the sampling effects associated with small datasets. A cut-off of 48 was chosen as all smaller breast cancer datasets either lacked detailed clinical data or had too few samples (approximately 30 samples or less). In total, 4,738 samples across 26 datasets incorporating 12 different microarray platforms were utilised to develop the BreastMark system (Table 1). Table 2 contains a breakdown of the clinical information available with each dataset. Where raw data were not available, the normalised data as published by the original authors were used. In the case of the raw data for the Affymetrix datasets (.cel files), gene expression values were called using the GeneChip (GC) robust multichip average method [19] and data were quantile normalised using the Bioconductor package, affy [20]. For the dual-channel platforms, data were loess normalised [21] using the Bioconductor package limma. Hybridisation probes were mapped to Entrez gene IDs to gene centre the data [22]. The Entrez gene IDs corresponding to the array probes were obtained using Biomart [23, 24] and the Bioconductor annotation libraries. Probes that hit multiple genes were filtered out. If there were multiple probes for the same gene, the probe values were averaged for that gene. This resulted in expression data for a total of 20,017 Entrez gene IDs across 4,738 samples.

miRNAs are frequently located within the introns of protein coding genes and in exons of non-coding transcripts. miRNA expression can be detected using conventional microarrays through host gene expression for intragenic miRNAs or by direct probe matching for intergenic miRNAs. A total of 1,987 samples were processed on U133A Affymetrix arrays, while 973 were processed on U133plus2 Affymetrix arrays (2,960 in total). U133A and U133plus2 microarrays have 22,277 probe sets in common. Using this information, it is possible to infer the expression of 341 miRNAs across 2,960 samples [25] (based on miRBase version 13.0, Ensembl version 54_36p). As with the gene centred data, this information was also combined with the available clinical data for survival analysis.

The R package genefu [26] was used to classify the 4,739 breast cancer samples into the luminal A, luminal B, HER2, normal-like and basal molecular subtypes using the ssp2003 [5], ssp2006 [27], and pam50 [28], classifiers.

We have combined detailed clinical data from each of the studies used here, including one or more of DFS, DDFS and OS. The software allows for each of these three survival end points to be analysed separately. Median expression was used to dichotomise the data, allowing stratification into high and low groups within each of the 26 individual datasets. Once a sample was assigned to a particular group, the 26 datasets were combined and a global pooled survival analysis was performed in real-time. It is important to treat each dataset separately when determining which group a sample belongs to, as the expression of these genes will vary greatly across the different experiments/platforms. In essence, each dataset is split into high and low in singularity to negate study-specific effects. Survival curves are based on Kaplan-Meier estimates and the log-rank P-value is shown for difference in survival. Cox regression analysis was used to calculate hazard ratios. The R package 'survival' was used to calculate and plot the Kaplan-Meier survival curve. All calculations were carried out in the R statistical environment [29].

The software incorporates all the clinical data made available by the original authors. This allows the data to be analysed based on one or more common clinical parameters including patient age, tumour size, lymph node status, tamoxifen treatment, chemotherapy treatment, ER status, HER2 status, PR status and tumour grade. The software also allows the upper or lower quartiles of the expression of the gene of interest to be used to determine high and low groups within each of the 26 individual datasets.

The interface is available on a publicly accessible web server [17] and is updated quarterly. The software uses CGI to link the web server with the R/perl based algorithm. All calculations are carried out in real-time.

The 21 gene signature used by oncotype DX in predicting patient prognosis was downloaded from the original paper [30]. This panel of prospectively selected genes comprises 16 prognostic genes normalised relative to the expression of five reference genes. The 16 prognostic genes are broken down into five categories: proliferation, invasion, HER2, estrogen and 'other'. The likelihood of breast cancer relapse in patients was based on a recurrence score (RS) algorithm constructed and tested on a cohort of 668 patient samples. The higher the RS, the poorer the patient outcome observed. This algorithm weights each of the five categories based on the influence they have on disease recurrence. For example, the proliferation group is weighted most highly and, therefore, the expression of these genes influences the RS the most. Each of the 16 oncogenes were queried in our dataset to test the effect each gene has on survival using the three above-mentioned survival end-points for prognosis, namely DFS, DDFS and OS. It is expected that the genes with the greatest influence on the RS would have the highest hazard ratios and the lowest P-values. Sample numbers will vary depending on the number of platforms with expression information available for a particular gene.

The 70-gene prognostic signature was downloaded from the original paper along with their correlation with prognosis [31]. It was possible to obtain unique Entrez gene IDs for 61 of these genes (there is more than one copy of PEC1, IGFBP5 and DIAPH3 (three) in the 70 gene list and five others have no Entrez gene ID). As with the oncotype DX signature, each gene was analysed separately within our datasets using the three survival endpoints, DFS, DDFS and OS. Although looking at these genes individually does not represent the full power of this prognostic signature, this dataset should still be enriched for prognostic markers. Additionally, the positive and negative correlation coefficients published by the original authors should be consistent with our observed hazard ratios of less than or greater than 1, respectively. Sample numbers will vary depending on the number of platforms with expression information for a particular Entrez Gene ID.

Oncotype DX is a 21-gene signature (16 oncogenes and five controls) selected using prior knowledge from the literature, which in combination with the developer's algorithm, predicts patient outcome in lymph node-negative (LNN), ER-positive breast cancer [32]. It uses a RS calibrated against approximately 670 patients with known clinical outcome to predict patient survival. Patients with a low score do well, and those with a high score do poorly. The 16 genes are classified into five groups: proliferation, invasion, HER2, ER and other. The algorithm takes gene expression data from 16 oncogenes, normalises the expression against the five controls and weights the 16 oncogenes depending on the effect they have on the RS. The genes are weighted as follows 1.04 × proliferation group score + 0.47 × HER2 group score - 0.34 × ER group score + 0.1 × invasion group score + 0.05 × CD68 score - 0.08 × GSTM1 score - 0.07 × BAG1 score. Genes from the proliferation group, such as Ki67 and Survivin, have the highest weighting and, therefore, the greatest effect on the RS.

Each of the 16 oncogenes were analysed separately within BreastMark using median expression as a cut-off, selecting LNN, ER-positive patients only and using DFS survival as the survival end point to ensure comparability. This information is summarised in Table 3, along with the effect they have on the RS. The 16 genes were also analysed using DDFS and OS as the survival end points [see Additional file 1 Tables S1 and S2] and are consistent with our observations for DFS survival. A hazard ratio (HR) of greater than 1 indicates a negative effect on survival and a HR of less than one has a positive effect. The higher the HR the greater the effect the gene has on survival. As can be seen from Table 3, our results are largely consistent with the weightings calibrated for oncotype DX. The proliferation markers which have the highest weightings, and therefore the largest effect on the RS, have the highest HRs and are highly statistically significant. In contrast, those genes which have only a marginal effect on the RS (CD68, GSTM1 and BAG1) are not significant and have HRs close to one.

Combining the markers (grouping samples where both markers have greater than median expression) identifies patients who will do particularly poorly. The Kaplan-Meier plot for Ki67 in shown in Figure 1(a) (n = 902, HR = 1.68, P = 4.44e-05). The Kaplan-Meier plot for Ki67 and Survivin combined, that is, comparing the survival of patients with greater than median expression of both Ki67 and Survivin against the rest is shown in Figure 1(b). These patients have a worse prognosis than Ki67 alone, that is, they have a higher HR (a HR of 1.99 versus a HR of 1.68). The same occurs when you also combine MYBL2 with Ki67 and Survivin (Figure 1(c)). These patients have an even worse prognosis with an even greater HR (n = 902, HR = 2.00, P = 2.01e-07). However, the same is not true when you combine other markers with the proliferation markers. Figure 1(d) shows Ki67, Survivin and PGR combined (n = 902, HR = 1.537, P = 9.2e-03). The HR is lower and the difference in survival is less significant. In fact, when you combine most of the other oncogenes from the signature, no improvement in prognostic power or decrease in the significance of the HR is observed (data not shown). This suggests that not only are all of the genes from this prognostic signature not necessary, but that potentially our algorithm provides a useful reductionist approach to these more complex prognostic signatures, allowing us to eliminate superfluous markers and highlight those genes that are of the greatest relevance.

Similar to the oncotype DX assay, MammaPrint [33] is a commercially available test for breast cancer recurrence. In contrast, it was developed via a hypothesis-free method from a gene expression profiling study rather than from a prospectively chosen list of known oncogenes. The study used 78 LNN patients specifically to identify a prognostic signature in their gene expression profiles using a supervised classification method. Each of the approximately 25,000 probesets present on those microarrays were correlated with disease outcome and only those genes that were significantly associated with disease outcome were retained to create an optimised list of prognostic markers. Each of the 70 genes had a positive or negative correlation coefficient depending on their association with good or poor prognosis, respectively.

Again, as with oncotype DX, even though the genes from the 70-gene signature are not predicted to act independently, the 70 genes when analysed independently, should correlate with good and poor prognosis based on the correlation coefficients identified in the original MammaPrint study. Genes with positive and negative correlation coefficients should have HRs less than and greater than one, respectively. As we expect, this is what we see with these genes in LNN samples, using a median cut-off and DFS survival as the survival endpoint (DDFS and OS show similar results in Additional file 1 Tables S3 and S4, respectively). Of the 61 genes from the MammaPrint signature for which we had Entrez gene IDs, 53 had HRs consistent with the correlation coefficients from the original study (Table 4). Of the other eight genes, four had HRs close to 1, and were not statistically significant, and the other four were not present in the dataset or present in too few samples. Although not all of the 53 consistent genes were statistically significant, 33 are significantly associated with survival when analysed independently with BreastMark.

Decreased expression of miR-205 has previously been associated with poor prognosis in breast cancer, and miR-93 is highly expressed in high-grade tumours, that is, in tumours of patients who do poorly [10, 34]; however, these studies were relatively small in scope (20 and 93 patients, respectively) [10, 34]. To confirm these observations in a larger dataset and to test our approach, we examined the association of the host genes of these miRNAs with prognosis. The results for miR-205 and miR-93 can be seen in Figures 2(a) and 2(b), respectively. Following BreastMark analysis, high expression of the host gene of miR-205 is indeed associated with good prognosis (HR = 0.768, P-value = 0.02, n = 581) and high expression of host gene of miR-93 is associated with poor prognosis (HR = 1.34, P-value = 1.48e-04, n = 1,563). This confirms that miR-205 and miR-93 are robust markers of good and poor prognosis, respectively.

RTKs are a large family of proteins involved in cell signalling with particular roles in growth, differentiation, adhesion, motility and death of cells [35]. A total of 58 kinases have been classified as receptor type and are listed in Additional file 2. Each of these kinases was assessed in the basal molecular subtype based on the three classifiers (ssp2003, ssp2006 and PAM50). Six of the kinases were significantly associated with poor prognosis in the basal subtype (EPHA5, FGFR1, FGFR3, VEGFR1, PDGFRβ and TIE1). The results are summarised on Table 5. As expected, the RTKs as a group have the potential to act as prognostic markers in this difficult-to-treat subtype of breast cancer. In particular, PDGFRβ would appear to be a strong marker of poor prognosis as it is significant across all three of the survival endpoints. This is not entirely unexpected as elevated levels of PDGFRβ have previously been associated with enhanced cell migration and invasion in breast cancer [36].