BMP2 stimulation of lung stromal fibroblasts induces regulatory genes of the BMP signaling pathway
The aim of this study was to uncover the transcriptional responses to BMP signaling in lung fibroblastic stroma. To characterize the effects of BMP2 on stromal cells, we stimulated pre-starved primary human fibroblasts (CCL-171) with BMP2 in a physiological concentration of 200ng/ml for 24 h. Biologically independent duplicate samples were profiled for gene expression changes using human exonic evidence-based oligonucleotide (HEEBO) microarrays. After filtering for data quality as described in the methods section, filtering for data distribution (SD > =0.7) and zero-transformation by the mean of the duplicate mock control average linkage unsupervised hierarchical clustering (Pearson correlation) was performed and the data were presented in a heatmap (Figure 1). Following BMP2 stimulation, we observed a remarkable change in the gene expression profile. 171 genes exhibited a greater than 1.5-fold increase (mean: 2.68, standard deviation: 0.86 ) and 206 genes showed a more than 1.5-fold decrease in the expression level (mean decrease: 2.99, standard deviation: 1.604) (Additional file 1: Table S1).
A notable feature of the transcriptional response was the induction of genes with direct roles in the regulation of BMP signaling either as inhibitors (e.g., BAMBI) or positive regulators (e.g., BMPR1A). Co-expression of both negative and positive regulators of BMP signaling provides a network of opposing control mechanisms that are likely to contribute to the robust and precise regulation of BMP target gene expression. Furthermore, we detected an up-regulation of BMP target genes such as SMAD7, ID1, ID2 and ID3. Other genes such as BMP4 and CXCL12 were systematically down-regulated. The down-regulation of CXCL12 indicates a tie to stem cell regulation [17]. The “CCL-171-derived BMP2” signature as marked in Figure 1 (Additional file 2: Table S2), also contains genes that are known to be involved in developmental processes, including GATA6, DKK1, ID2, BAMBI, SERPINE1, and PTGIS.
The identification of transcription factors and other regulatory genes involved in the transcriptional response (e.g. HEY1, PHTF2, GATA6, and LMCD1) suggests that a more extensive reprogramming of transcription occurs downstream of the primary target genes; the mechanism of target gene activation is likely to be far more complex than direct activation through SMADs, involving a cascade of downstream transcription factors and regulatory proteins that play a role in further regulation of secondary BMP2 target genes. Among the transcription factors regulated by BMP signal transduction, it was interesting to find HEY1 and GATA6, which play important roles in pattern formation, morphogenesis and body-axis specification. The increased expression of TPM1, TPM2 and TNC [18] indicates a fundamental reprogramming of normal lung fibroblasts such that they exhibited a gene expression pattern that is typically found in carcinoma-associated fibroblasts.
For an unbiased assessment of which features are shared by members of the “CCL-171-derived BMP2” signature and to verify the significance of enrichment of a specific gene ontology, we used the GO::TermFinder tool [16]. The analysis revealed that the “CCL-171-derived BMP2” signature as shown in Figure 1 (Additional file 2: Table S2) is significantly enriched for genes involved in biological processes such as the cellular response to chemical stimuli and the BMP signaling pathway with Bonferoni corrected p-values of 0.00017 and 0.006, respectively (Additional file 3: Table S3).
Furthermore to identify genes with a significant change in expression levels we performed a two class unpaired SAM analysis using a false discovery rate <1%. 76 significantly induced and 151 significantly repressed genes after BMP2 stimulation are shown in Additional file 4: Table S4. According to our expectations, all the significantly up-regulated genes are contained within the list of Additional file 1: Table S1. In addition 10 genes not comprised within Additional file 1: Table S1 were found to be significantly down-regulated with SAM analysis, namely: GNG11, ABHD5, PHF17, GAS1, ARSI, JUN, KLHL29, CHEK2, ST6GAL1, PDGFRA, H1F0, BTG1 (fold changes of these genes: 2.15-2.58 fold decrease).
A global picture of genes that are differentially expressed in response to BMP stimulation or inhibition in lung stromal fibroblasts
To obtain a more general picture of the transcriptional responses to BMP signaling, we examined the gene expression changes in response to physiological concentrations of BMP2, BMP4, and BMP7, as well as their antagonists, Noggin and Gremlin, in CCL-171 cells. CCL-171 cells were starved, and the medium was subsequently replaced with fresh low-serum D-MEM with or without 200 ng/ml BMP2 (human recombinant in Escherichia coli; Sigma Aldrich), 24 ng/ml BMP4, 200 ng/ml BMP7, 240 ng/ml Noggin, or 1 μg/ml Gremlin. To search for genes with a significant change in expression upon exposure to at least one of the agents, we performed a multiclass SAM analysis on BMP2, BMP4, BMP7, Noggin, Gremlin and mock stimulated biologically independent duplicate samples of CCL-171 with a false discovery rate of 1% [11] After zero-transformation with the mean expression levels of the mock stimulated samples we performed unsupervised hierarchical clustering of the genes identified by SAM and displayed the data as a heatmap (Figure 2A). A set of 115 genes was identified that were commonly induced by BMP2, BMP4 and BMP7 (Additional file 5: Table S5). To correlate the responses of all genes derived by SAM upon BMP2 stimulation with the responses to BMP4, BMP7, Noggin and Gremlin we calculated the Pearson correlation coefficient r and displayed their correlation as a scatter plot (Figure 2B). Thus the response to BMP2 observed in CCL-171 cells was highly significantly correlated to that observed in response to BMP4 (r = 0.892, p < 0.001) and BMP7 (r = 0.892, p < 0.001). An expression pattern opposing the BMP responses were observed in response to Noggin (r = −0.940, p < 0.001) and Gremlin (r = −0.924, p < 0.001), as would be expected for antagonists.
A comparison of the effect of BMP2 stimulation in lung stromal fibroblasts and breast cancer cells
To determine whether the response to BMP2 is a general effect or specific for mesenchymal cells such as the lung fibroblasts (CCL-171), we compared the gene expression profile of BMP2-stimulated CCL-171 cells with the expression profiles of the breast cancer cells MDA-MB-231 and T47D treated with BMP2. First, to systematically define a gene signature reflecting the common response to BMP2 shared by CCL-171, MDA-MB-231 and T47D cells, we used a two class SAM with block permutation to identify genes with a significant change using a false discovery rate of less than 0.8%. A common set of 80 genes that were significantly induced, and 200 genes that were significantly repressed in all three cell types were identified (Additional file 6: Table S6).
Second, we intended to determine the cell type specific gene expression changes upon BMP2 stimulation. The gene expression levels of CCL-171, MDA-MB-231 and T47D upon stimulation with BMP2 were subtracted by the mean expression levels of duplicate samples of the same cell types with mock stimulation. Genes with a difference in expression of at least 4 fold were selected and after unsupervised hierarchical clustering displayed in a heatmap (Figure 3). We observed a cell type dependent induction or repression of specific genes. We observed genes that were specifically up- or down-regulated in CCL-171 cells, including POSTN, a gene that was recently described to be essential for the formation of metastatic stem cell niches [19]. The fibroblast-specific up-regulation of topoisomerases and cyclins indicates an induction of fibroblast proliferation in response to BMP2 (Figure 3). A set of 67 genes, which are specifically induced in CCL-171 compared to MDA-MB-231 and T47D is referred as the “Fibroblast specific BMP2 induced gene list” (Additional file 7: Table S7). To support the up-regulation of this cluster in BMP2 stimulated CCL171 cells compared to non-stimulated CCL-171 cells and compared to MDA-MB-231 and T47D we calculated the average expression levels of the genes in the cluster (centroid) for each cell line and condition. A graphic representation of the average expression value is shown as centroid above the heatmap of the magnified area in Figure 3.
BMP4 stimulation of lung stromal fibroblasts induces a time-dependent gene expression signature
To exemplarily show the development of the gene expression changes in response to BMP stimulation over time we performed a 48 h time course analysis after BMP4 stimulation in CCL-171 cells. CCL-171 cells were first placed under replicative quiescence in DMEM media with 0.1% serum for 48 h and then exposed to fully supplemented media, DMEM with 10% serum, and incubated with BMP4 for 1, 3, 6, 12, 24 or 48 h. The temporal effect of BMP4 on gene expression was assessed using DNA microarrays. The BMP4 dependent changes in the gene expression profiles over time were assessed by normalization against the time dependent changes with mock stimulation and normalization against the gene expression profile at time 0. Some genes were up-regulated as early as 3 hours, whereas others were only expressed after 24 hours of stimulation with BMP4 (Figure 4). The strong and highly significant correlation (r = 0.892, p < 0.001) of the gene expression changes upon BMP4 and BMP2 stimulation over 24h (Figure 2B) suggests a concordant and reliable pattern of gene expression changes in CCL-171 upon 24h of stimulation by BMP4 and BMP2. This suggests the response to BMP2 to be suitable for further analysis.
Prognostic significance of the “Fibroblast specific BMP2 induced gene list” in human lung adenocarcinomas
To verify the relevance of our in vitro experiments, we checked the expression of genes in the “Fibroblast specific BMP2 induced gene list” (Additional file 7: Table S7) in publicly available microarray data from lung cancer biopsies [13]. Garber et al. published global gene expression profiles of 41 human lung adenocarcinomas; patient survival data were available for 24 of these adenocarcinomas [13] (GEO: GSE3398). A subset of the genes that overlapped between the “Fibroblast specific BMP2 induced gene list” and the Garber dataset (n = 37), was coherent to provide a basis for segregation of the tumors into two groups (Figure 5A). As visualized in the Kaplan-Meier plots, the lung cancer patients with a high expression level of the “Fibroblast specific BMP2 induced gene list” showed a strong trend to a higher risk of death than the patients with a low expression level (76% versus 38% after 12 months, hazard ratio (HR): 0.57, cox-p = 0.07). Thus, the “Fibroblast specific BMP2 induced gene list” has a potential to be a prognostic marker in lung cancer.
To further support this finding we analyzed the stage I lung adenocarcinomas published by Lee et al. [14]. From 67 genes building “Fibroblast specific BMP2 induced gene list” 48 were present in Lee dataset. Patients stratification based on unsupervised hierarchical clustering divided patients into two groups. Patients carrying tumors with high expression levels of the “Fibroblast specific BMP2 induced gene list” showed a significantly higher risk of death than the patients with a low expression level (76% versus 38% after 24 months, hazard ratio (HR): 0.49, cox-p =0.002) (Figure 5B).
We further validated our findings using a larger and better-annotated dataset published by Bhattacharjee [15], which contains microarray profiles of 151 stage I lung adenocarcinomas from patients who had undergone surgery. Consistent with our hypothesis, based on 66 genes present in the dataset, the expression of genes of the “Fibroblast specific BMP2 induced gene list” provided a basis for segregation of the tumors into two groups (Figure 6). Compared with patients with low expression levels of genes in the “Fibroblast specific BMP2 induced gene list” (underlined in red), patients with high expression levels of these genes (underlined in black) had a significantly shorter disease-specific survival (87% versus 46% after 5 years, HR: 0.57, cox-p = 0.002) and overall survival (68% versus 34% after 5 years, HR: 0.51, cox-p = 0.0002). Both Kaplan-Meier curves are shown in Figure 6.
Because the classification of data based on hierarchical clustering has been suggested to be unstable and codependent on many factors such as the presence of missing values [47], we validated the results using continuous scoring and stratified the patients in the Lee and the Bhattacharjee datasets based on a score derived from the average expression level of genes in the “Fibroblast specific BMP2 induced gene list”. The continuous scoring approach consistently divided the lung cancer patients into two groups with significantly different outcomes (Lee dataset: overall survival: HR 0.38, cox-p = 0.00028, Bhattacharjee dataset: disease free survival: HR 0.69, cox-p = 0.0023, overall survival: HR 0.72, cox-p =0.027) (Figure 7). Taken together, these findings indicate that genes included in the “Fibroblast specific BMP2 induced gene list” are of importance in vivo in human lung carcinomas and helpful in predicting outcomes for patients with lung cancer.