To develop a model of acquired resistance to lapatinib, we cultured ErbB2-positive SKBR3 cells with increasing doses of lapatinib for a year, in parallel control parental cell was also cultured in normal conditions with no lapatinib. The resultant cell line variant, SKBR3-R, is almost 100-fold more resistant to lapatinib treatment when compared with the parental SKBR3 (). In line with previous studies, we find that lapatinib induces a significant amount of apoptosis as characterized by sub-G1 DNA content in SKBR3 but not in SKBR3-R cells (). Previous studies have implicated acquired mutations in EGFR/ErbB2 or overexpression of ErbB3 as a possible mechanism of lapatinib resistance. We sequenced EGFR and ErbB2 kinases and found no acquired mutations in the mutation hot spots reported for these genes (Supplementary Figure 1
). We measured ErbB3 protein levels and found no increase in ErbB3 levels between parental and resistant cells (). However, there is a small increase in total ErbB3 levels in response to lapatinib in both parental and resistant cells as previously described (; Garrett et al, 2011
). We next determined the effect of lapatinib treatment on inhibition of EGFR/ ErbB2 phosphorylation in both parental and resistant cells (), and found that EGFR/ErbB2/ErbB3 phosphorylation was equally inhibited in both sensitive and resistant cells ().
Figure 1 Initial characterization of lapatinib-resistant cell line. (A) Percent change in cell numbers in response to increasing doses of lapatinib in parental and resistant SKBR3 cells. (B) Cell-cycle and apoptosis analysis of the parental and resistant cells (more ...)
To determine the effect of lapatinib on downstream signaling pathways from EGFR/ErbB2, we measured changes in phosphorylation of MAPK1,2, AKT, mTOR, TSC2, P70S6K and S6 and found that lapatinib equally inhibited their phosphorylation in both the parental and resistant cells (). We additionally genotyped the parental and resistant cells in the MDACC Characterized Cell Line Core using Sequenome analysis and compared it with the ATCC-derived SKBR3 cells to ensure the acquired resistance cells were in fact SKBR3 cells (Supplementary Figure 2
). Having found no additional mutations in EGFR/ErbB2, no increase in ErbB3 and equal inhibition of EGFR/ErbB2/ErbB3 signaling between the parental and resistant cells we wanted to determine what alterations of cellular processes outside the classical EGFR/ ErbB2 pathway may contribute to acquired lapatinib resistance in breast cancer cells.
To gain insight into the global cellular processes altered during the acquired lapatinib resistance in SKBR3 cells, we measured gene expression changes in SKBR3 and SKBR3-R cells before and after treatment with 0.1 or 1
μM lapatinib for 24
h. Functional analyses of global gene expression data are enhanced by employing network-based approaches within the context of a priori
information (Ideker et al, 2011
). Here, in order to understand the specific network alterations contributing to acquired resistance of SKBR3-R cells to lapatinib, we employed NetWalk, a random walk-based network scoring method for genomic data analyses (Komurov et al, 2010
). In contrast to other similar network analysis methods, NetWalk output is not a collection of networks, but rather a distribution of network-wide scores for each interaction in the network based on the local connectivity as well as the supplied gene expression values. This enables direct comparative analyses of gene expression data between different conditions at a network, rather than at a gene level. To facilitate network analyses of signaling pathways, transcriptional networks, metabolic networks and functional interactions in the gene expression data, we compiled a comprehensive network of binary relationships between genes based on physical, regulatory and neighboring interactions as well as functional similarity as cataloged in various online databases (see Materials and methods). Overall, our network accounts for ~240
000 physical and functional interactions among 15
106 unique genes.
Using the respective microarray gene expression data, we obtained Edge Flux (EF) value distribution for each condition (see Materials and methods). The heatmap of 1000 EF values with highest variance across the six conditions (0, 0.1 and 1
μM lapatinib treatment for 24
h for SKBR3 and SKBR3-R cells) is displayed in . The same analysis using the highest 480 values is shown in Supplementary Figure 3
. Four clusters, K1 through K4, with distinct temporal expression patterns are clearly identifiable. Networks in K1 are specifically upregulated at the basal level in the resistant cells, while those in K2 are specifically downregulated in these cells. K3 contains networks that are upregulated, and K4 contains networks that are downregulated, in response to lapatinib in both cell lines. Heatmap visualization of the data without NetWalk analysis is shown in Supplementary Figure 4
. Plotting of the network in K1 shows extensive upregulation in the resistant cells of the cellular processes involved in glucose uptake, glucagon signaling, unfolded protein response (UPR) and oxidation/reduction (). The network in K2 is primarily composed of signaling pathways involved in TGFβ signaling and inflammatory response (Supplementary Figure 5
), and the network in K3 contains processes involved in oxidation/reduction, cell-cycle arrest and EGF signaling (Supplementary Figure 6
). K4 almost exclusively consists of cell-cycle processes (Supplementary Figure 7
). We performed GSEA analysis of the same data set and the GSEA analysis failed to identify these functionally relevant gene networks (Supplementary Figure 8
). Lapatinib-mediated downregulation of cell-cycle machinery in K4 and upregulation of cell-cycle inhibitory complexes and EGF signaling in K3 in both cell lines are in line with our expectations regarding cellular responses to lapatinib, which involves cell-cycle arrest and attempts at restoring EGF signaling. However, the networks in K1 and K2, which are specifically activated and inhibited, respectively, in resistant cells, and therefore constitute the clusters of highest interest within the context of this study, have not been previously associated with acquired resistance to targeted therapy. Therefore, we chose to analyze these clusters in more detail.
Figure 2 Network analyses of gene expression data. (A) A heatmap of Edge Flux (EF) values with highest variance across the six conditions (see text and Materials and methods). (B) A network plot of the interactions in K1. Nodes are colored by the relative gene (more ...)
Interestingly, processes in the K1 cluster are reminiscent of a classical glucose deprivation response, where endoplasmic reticulum (ER) stress in the form of the UPR, amino-acid catabolism, glucagon signaling, increase in the expression of glucose transporters and glycogen breakdown are common responses (Hotamisligil, 2010
). Indeed, a functional enrichment analysis of the 500 highest EF values in the resistant cells relative to parental cells using hypergeometric probability distribution function shows a specific enrichment of processes associated with the ER (), further suggesting that networks associated with the ER and nutrient stress are upregulated in the resistant cells. Furthermore, several of the previously published markers of glucose deprivation and ER stress response markers are also upregulated in the resistant cells at the basal level relative to parental cells (), indicating that the resistant cells indeed display a nutrient-starved phenotype.
Western blot analysis revealed that key members of the ER stress response pathways, the ER chaperone GRP78 (glucose deprivation response protein of 78
kDa, HSPA5 gene), inositol requiring protein-1 (IRE1) and phospho-JNK, which is activated by IRE1 during ER stress (Hotamisligil, 2010
), are all markedly elevated in the resistant cells (). Additionally, we also found an increase in p38 and PERK and in the phosphorylation of AMPK in the resistant cells ().
Figure 3 Glucose deprivation response phenotype in lapatinib resistance. (A) Immunoblotting of key members of glucose deprivation response in parental and resistant cells HSPA5, HK2, IRE1, pJNK, pAMPK, PERK and p38. (B) Immunoblotting of the AKT and AMPK phosphorylation (more ...)
TSC2 serves as key player in the juxtaposition between signaling and metabolism and is phosphorylated by both AKT and AMPK. Phosphorylation of the TSC2 protein by AMPK is a hallmark of nutrient starvation (Inoki et al, 2003
), and we found increased phosphorylation of TSC2 at the AMPK site (S1387) in the resistant cells at the basal level. Interestingly, lapatinib treatment increased its phosphorylation in both parental and resistant cells; however, the maximal levels were observed in the resistant cells (). In contrast, the AKT phosphorylation site on TSC2 (T1462) exhibited a different behavior with lapatinib blocking the phosphorylation in both parental and resistant cells similarly to the effect on the other signaling proteins ().
Cellular response to nutrient deprivation aims to restore glucose uptake and energy production through upregulation of the glucose transporters as well as alternative pathways of energy production through amino acid and fatty acid break-down; while trying to minimize the damage of ROS and hypoglycosylated proteins in the ER through UPR (Hotamisligil, 2010
). Since inhibition of EGF signaling in ErbB2-positive breast cancers has been associated with glucose deprivation and energetic stress (Weihua et al, 2008
; Schafer et al, 2009
), we hypothesized that lapatinib-mediated cell toxicity may be associated with inhibition of glucose uptake and subsequent energetic stress. Treatment of SKBR3 cells with lapatinib impaired their ability to uptake glucose from the media and perform glycolysis, while the resistant cells were not significantly affected (). Lactate-to-glucose ratio of the fluxes showed that the resistant cells had a higher glycolysis rate as compared with the parental cells (). Lapatinib treatment induced no change in the parental cell but there was decrease in lactate/glucose ratio in the resistant cell, suggesting a switch from glycolysis to the pentose phosphate pathway shunt. Lapatinib treatment leads to ATP depletion in the parental, but not in the resistant cells (). These observations suggest that lapatinib-mediated toxicity is associated with glucose starvation and energy deprivation, and that SKBR3-R cells resist this effect. Since the resistant cells were not phenotypically under glucose starvation and were not energy deprived, the elevated networks of the UPR and other hypoglycemic response pathways in these cells may be an adaptation to prolonged lapatinib treatment, giving these cells the ability to uptake and metabolize glucose independent of the EGFR/ErbB2 pathway. Indeed, forcing parental cells into hypoglycemia by incubating them in a media with no glucose rendered them more resistant to lapatinib-mediated growth inhibition (), indicating that cellular hypoglycemic response can confer resistance to inhibition of ErbB2 signaling. Importantly, lapatinib-induced inhibition of glucose uptake and the protective effect of hypoglycemic response was observed with another ErbB2-positive breast cancer cell line, BT474 ().
Large-scale clinical analyses of mRNA expression profiles of cancer patients provide an invaluable resource for testing the clinical relevance of findings from the in-vitro
cell culture models. We asked whether our findings in the cell culture model of acquired lapatinib resistance are also manifest in the breast cancer patients in vivo
. Unfortunately, ideal sets of data to make such an analysis possible, that is, mRNA expression profiles and clinical data of ErbB2-positive breast cancer patients treated with lapatinib, are not available. Therefore, we asked if the networks associated with lapatinib resistance in our cell culture model correlate with overall survival or relapse rates of patients with breast cancers whose tumors express high ErbB2 levels. To answer this question, we elected to employ a network approach, where we sought to construct a network of gene–gene interactions that most correlate with high relapse rates in ErbB2-positive patients. Using the breast cancer cohort data from Miller et al (2005)
, we calculated COX regression coefficients between the expression of every gene and the relapse status of patients with high ErbB2 levels (see Materials and methods). Here, a high COX coefficient indicates high correlation of the gene's expression level with poor outcome. Using the distribution of COX regression values as input to NetWalk, we obtained the network of highest scoring interactions that best correlate with high relapse rates in ErbB2-positive patients (). In addition to some of the previously well-characterized pathways of cell-cycle progression involved in poor outcome, this network also contains networks involved in glucose/nutrient deprivation response (see ), which are also upregulated in our resistant cells (HK2, SLC2A10, NDRG1; ). In contrast, while ErbB3 by itself shows a significant correlation with poor disease-free survival, most other genes in the ErbB3 network do not show the same association between their expression levels and patient outcome (Supplementary Figure 9
). To identify cell processes that are most significantly associated with high relapse rates in ErbB2-positive patients, we carried out a functional enrichment analysis of the network corresponding to highest EF values calculated using the COX coefficient values above (see Materials and methods). We observed that a GO functional category ‘unfolded protein binding' was one of the most highly enriched processes in this network, along with those involved in cell-cycle progression (), indicating that the pathways involved in UPR similar to those activated in our resistant cells are associated with poor outcome in ErbB2-positive breast cancers. Importantly, several of the markers of glucose deprivation/UPR in the network in were also indicative of poor prognosis in other independent breast cancer patient cohorts (Supplementary Figure 10
). These correlations in the clinical data sets strongly corroborate with our findings in our cell culture models, and implicate the cellular response to glucose deprivation in the form of UPR and/or gluconeogenesis as important biological processes in the relapse of ErbB2 overexpressing breast tumors.
Figure 4 Correlation of glucose deprivation response with clinical relapse rates in ErbB2-positive breast cancers. (A) Network of genes with highest correlation with relapse in ErbB2-positive tumor patients (see Materials and methods). Node coloring reflects the (more ...)
Finding vulnerable intervention points of tumor cells that have progressed on targeted therapies is of critical importance for designing novel therapeutic strategies for such tumors (Haber et al, 2011
). Therefore, we asked if the activated pathways of UPR and glucose deprivation response in the resistant cells can be targeted for therapeutic purposes. To find potential drug candidates for the reversal of the glucose deprivation response phenotype in resistant cells, we employed a chemical genomics bioinformatics approach leveraging the connectivity map data set, which is a useful resource for finding drugs with novel functions based on gene expression (Lamb et al, 2006
). We identified a set of 12 genes that are involved in the glucose deprivation response and UPR that are also upregulated in the resistant cells (see legend), and used this gene set to query the drugs in the CMAP data set that caused their downregulation. We scored each condition in the CMAP data set by summing the ranks (lower means downregulated) of the 12 genes, and identified 5 drugs with the lowest scores (). Interestingly, the drug with the second lowest score was pyrvinium, an anthelmintic drug, that has been previously shown to inhibit UPR associated with hypoglycemia and therefore specifically kill cells that are deprived of glucose (Yu et al, 2008
; Saito et al, 2009
). To test if downregulation of the glucose deprivation response genes by pyrvinium in the CMAP data set is statistically significant, we calculated the average ranks of 10
000 randomly selected 12 genes in the pyrvinium data set and compared with the ranks of the glucose deprivation response genes. Based on this analysis, the downregulation of the glucose deprivation response genes was highly significant relative to what would be expected by chance (). Therefore, we asked if treatment with pyrvinium can be preferentially toxic to resistant cells, as they have a specific upregulation of the hypoglycemic response networks. Indeed, a dose–response survival assay with pyrvinium showed that lapatinib-resistant cells are significantly more sensitive to pyrvinium when compared with parental cells (). Importantly, the parental cells can be sensitized to pyrvinium under low-glucose conditions (Supplementary Figure 11
), indicating that the toxicity of pyrvinium in resistant cells is due to elevated pathways of UPR and hypoglycemic response. We targeted UPR using Metformin which has been previously shown to target the UPR gene program (Saito et al, 2009
) and found that Metformin significantly inhibits lapatinib-resistant cells compared with the parental cells (Supplementary Figure 12
) very similar to the response to pyrvinium.
Figure 5 Identification and treatment of resistant cells with drugs reversing the hypoglycemic response phenotype. (A) Gene expression profiles of 12 genes in the glucose deprivation response gene set that are specifically upregulated in the resistant cells. Tested (more ...)
Similarly to UPR, autophagy is also a survival response to nutrient deprivation (Hotamisligil, 2010
; Kroemer et al, 2010
) and its inhibition can also be selectively toxic in nutrient-limiting conditions (Sato et al, 2007
; Yin et al, 2009
), especially in transformed cells (Sheen et al, 2011
). Autophagy is activated in the resistant cells at the basal level and in parental cells in response to lapatinib (). Therefore, we tested if resistant cells are more sensitive to bafilomycin A, a selective inhibitor of autophagy by the virtue of its ability to inhibit vacuole maturation. Indeed, lapatinib-resistant cells were significantly more sensitive to bafilomycin A doses when compared with parental cells (), and parental cells could be sensitized to bafilomycin A when placed in a low-glucose media (Supplementary Figure 13
). The data above show that acquired resistance of SKBR3 cells to lapatinib is associated with increased expressions of glucose deprivation and ER stress response networks, and that selective targeting of these processes can be an effective therapeutic strategy against these tumors.