Genome-wide DNA copy number predictors of lapatinib sensitivity in tumor-derived cell lines

Lapatinib is a potent, orally active dual kinase inhibitor of the epidermal growth factor receptor (EGFR) and HER2 tyrosine kinase receptors. In tumor cells that overexpress either EGFR or HER2, lapatinib treatment results in growth arrest and cell death (1), suggesting that it may be possible to predict clinical response based on a tumor's genetic profile. Although larger in vitro screens confirm strong associations between EGFR and HER2 amplification with lapatinib responsiveness, some tumor models seem to respond in the absence of these markers (2, 3). This observation indicates that additional proteins may cooperate with HER2 in promoting lapatinib responsiveness, thereby forming the basis of a more inclusive means of predicting clinical response. Such proteins may include the linear cascades that constitute the mitogen-activated cascade, the stress-activated protein kinase cascade, protein kinase C, and the Akt pathway (4).

Altered DNA copy number is a hallmark feature of the cancer genome. A number of tumor suppressor genes have been mapped to homozygous deletions, including RB1 and SMAD4/DPC4 (5, 6). Similarly, oncogenic copy number amplifications have been described in such loci as PDGFRA, EGFR, and HER2 (7–9). Until recent technological advances, the assembly of more complete lists of genes whose copy number is altered in cancer has been limited by nonsystematic techniques (10). However, the development of high-resolution microarray-based genome-wide DNA copy number has remedied this limitation. Although earlier BAC clone–based platforms provided a resolution of ∼1 Mb (11), more recent oligonucleotide-based platforms such as Representational Oligonucleotide Microarray Analysis decreased probe spacing considerably (12). Single-nucleotide polymorphism arrays (SNP chips) represent a type of high-density oligonucleotide-based microarrays that have been adapted for DNA copy number analyses (13) and perform comparably with traditional two-channel arrays (14).

The increasing role of pharmacogenomics in oncology compound development and treatment has closely paralleled the advent of high-throughput genomics platforms. A major attraction of using genome-wide assays is the increased ability to more comprehensively characterize genomic mechanisms of drug sensitivity. For example, gross patterns in gene expression effectively show that in vitro response to a farnesyltransferase inhibitor of the Ras pathway is directly correlated with Ras pathway activity (15). Translational research focuses on the discovery of rational predictive markers derived from such genomic assays that could then be used for response enrichment in a clinical setting. A growing number of these studies include an effective predictor of bortezomib response in myelomas that was retrospectively calculated and validated from gene expression microarrays run on patient tissue collected from multiple independent clinical sites (16). Although genomewide expression profiling is currently the most widespread platform for pharmacogenomic studies, the mapping of specific DNA copy number alterations to multiple pathways is also effective in delineating markers associated with the responsiveness to targeted therapy (reviewed in ref. 17). Fixed genetic alterations such as DNA copy number gains and losses are appealing for clinical study because (a) they are relatively stable in vivo over time, and (b) they are faithfully modeled in cell lines in vitro (18, 19). For example, copy number studies have associated specific DNA copy number alterations in ovarian carcinomas with carboplatin response (20), whereas amplifications of the multidrug resistance gene MDR1 are correlated with therapeutic resistance (21, 22).

The growing use of targeted therapies in cancer treatment reflects an increasing knowledge of key oncogenic pathways. Difficulties in predicting response to targeted therapies are likely a consequence of the limited global knowledge of altered genetic pathways in each cancer. Also, unique genomic profiles observed in each cancer suggests the existence of many more contributing pathways than previously thought. For example, contrasting sets of recurring alterations found to differentiate between certain types of lung cancers indicate that common oncogenic pathways may be subject to unique alterations (23). Collectively, these results suggest that developing genome-wide strategies to discover predictive models of therapeutic response can be more sensitive than traditional single-gene approaches. In this study, we present a model that takes advantage of global characterizations of DNA copy number to accurately predict in vitro responses to lapatinib. The model, which emphasizes high-level copy number alterations such as homozygous deletions and oncogenic amplifications, rediscovers the HER2 locus, which is already a known predictor of clinical activity. Furthermore, this model discovers novel markers of response that are capable of classifying a panel of non–HER2-amplified head and neck cancers. Eventually, these DNA alterations may provide clinically relevant markers that allow the selection of patients most likely to benefit from therapy.

A panel of 34 cell lines of mixed histology were obtained from American Type Culture Collection and were cultured in RPMI 1640 supplemented with 10% heat-inactivated fetal bovine serum, 2 mmol/L of glutamine. Similarly, for the purpose of characterizing a diploid reference in the copy number analysis of the tumor-derived cell lines, DNA from a panel of 20 cell lines composing a human genome diversity panel were obtained (Coriell Institute, Camden, NJ; ref. 24). For tumor-derived cell lines, DNA was extracted before any compound treatment (i.e., untreated cells).

Cells were plated into 24-well plates at a density of 2 × 10⁵ to 5 × 10⁵ and grown in cell line–specific medium without or with increasing concentrations of lapatinib (ranging between 0.008 and 10 μmol/L). Cells were harvested by trypsinization on day 7 and counted using a particle counter (Z1, Beckman Coulter, Inc.). Growth inhibition was calculated as a percentage of the untreated controls. Experiments were carried out twice or thrice in duplicate for each cell line. The log of the fractional growth inhibition was plotted against the log of the drug concentration, and the IC₅₀s were interpolated from the resulting linear regression curve fit (Calcusyn, Biosoft). IC₅₀ values <1.0 μmol/L were considered as sensitive, 1.0 to 2.5 μmol/L as intermediate, and values >2.5 μmol/L were considered resistant. Proliferation data for several of these lines have been previously presented in ref. (2).

DNA was extracted from each line using Mini DNeasy kit (Qiagen, Inc.) and subsequently purified. For the 24 cell lines of mixed tumor type composing the gtrainingh panel, two aliquots (250 ng each) were digested with the restriction enzyme XbaI or HindIII (New England Biolabs), whereas a similar quantity of material was digested with Sty and Nsp enzymes for the 10 head and neck lines in the validation set. Digested DNA was subsequently ligated to an adaptor and amplified by PCR using Platinum Pfx DNA Polymerase (Invitrogen), yielding a product of ∼250 to 2,000 bp. For each enzyme digest, PCR was carried out in four 100-μL aliquots, pooled, purified, quantified, normalized to 40 μg/45 μL, and fragmented with DNase I to yield a size range of ∼25 to 200 bp. The fragmented products of the training set and test set were then labeled, denatured, and hybridized to the Affymetrix 100K and 500K chip, respectively (Affymetrix Inc.). Upon completion of hybridization, each assay was washed and stained using Affymetrix fluidics stations. Image data were acquired using the GeneChip Scanner 3000. For all 34 cell lines, transcript abundance was quantified in triplicate by using the Affymetrix U133 Plus2 GeneChips. First, cell lines were plated in triplicate and lysed in TRIzol. Lysates were captured with chloroform and purified using QIAGEN RNeasy Mini Kit (QIAgen, Inc.). cDNA was prepared from 5 μg of total RNA using the Invitrogen SuperScript Double-Stranded cDNA Synthesis Kit (Invitrogen, Inc.) and amplified using the ENZO BioArray High-Yield RNA Transcript Labeling Kit (Enzo Biochem, Inc.). Finally, the samples were fragmented and hybridized to the HG-U133Plus2 GeneChips, stained, and scanned according to the manufacturer's protocols.

All SNP chip images (“CEL files”), were extracted using the Affymetrix Genotype software, and analyzed using a modification of the RLMM implementation in the R programming language (25). Specifically, a quantile intensity normalization was carried out for the CEL file–extracted intensity scores. Next, SNP-wise copy number ratios (log₂ scale) were calculated for all cancer cell lines by dividing the SNP intensity score by the respective median intensity score for the lymphoblastic reference panel. Subsequently, data were adjusted under the assumption that the median copy number for all samples was diploid. Finally, copy number inferences were made by circular binary segmentation to reduce noise (e.g., from unmasked complex sequences in the target) and provide a consensus score for all regions of the genome based on at least two underlying SNPs (26). For all subsequent analysis, the genome was divided into 100 kb units in which each was assigned the circular binary segmentation–extracted value for every cancer cell line. Transcript abundance was estimated by normalizing all probe signal intensities to a value of 150 using the mas5 algorithm in the Affymetrix Microarray Analysis Suite 5.0. For subsequent analyses, the average intensity of each probe was used.

An externally cross-validated prediction model, using methods described in ref. (27), was constructed using 24 tumor-derived cell lines of mixed histologies. The predictive model framework is solely determined by the three variables in the feature selection step. The feature selection step included (a) converting log₂ segment scores to zero if they were less than the data threshold t; (b) calculating the mean difference between phenotypic groups for each feature (100 kb in this case) and filtering out features with a small mean difference by applying a mean difference threshold d, and also removing neighboring redundant features of the included features; and (c) estimating the P values of the mean difference of selected features using 10,000 random permutations and subsequently filtering out insignificant features by applying a P value threshold. The remaining features were then used to build a weighted-voting model as described in ref. (28). To find the optimal model, a grid search was carried out in the following search space: data threshold t between 0.2 and 0.5; mean difference threshold d between 0.2 and 0.5; P value threshold between 0.01 and 0.05.

To evaluate the performance of a model, leave-one-out cross-validation was carried out (with a new feature selection for each cross-validation step) and area under the receiver operating curve (AUROC) was calculated. Cross-validation AUROC showed that model performance was generally good when the variables were in the following ranges: t between 0.3 and 0.5, d between 0.35 and 0.45, and P between 0.03 and 0.05. The optimal model had the variables t = 0.45, d = 0.4, and P = 0.03. Leave-one-out cross-validation showed that the performance of the optimal model had an AUROC score of 0.85. To establish confidence intervals, 200 runs of bootstrapping were conducted and an AUROC of each run was calculated using leave-one-out cross-validation. The mean of the AUROC was 0.85, with an 80% confidence interval (0.70–0.98). To estimate the statistical significance of the optimal model, the phenotypic grouping information of the 24 cell lines was randomly permuted 500 times. Less than 1% of the random data sets had an AUROC score of 0.85 or above. This showed that the optimal model was significant at P < 0.01. Finally, an independent panel of 10 head and neck cancer cell lines were classified using the loci and variables in the optimal model derived from the training data. All lines in this test set lack amplification of the HER2 locus (the most significant predictor in the training set), which addresses whether additional predictors of sensitivity were identified.

The effects of lapatinib on cell growth were studied in 24 human cancer cell lines of mixed histology. These cell lines exhibited a wide range of sensitivities based on their IC₅₀ values measured in the proliferation assays (range, 0.025–7.71 μmol/L; Table 1). Several showed hypersensitive phenotypes, having IC₅₀ scores of <0.1 μmol/L.

Many cancer cell lines in this panel are subject to complex rearrangements. For example, 1q in the breast cancer cell line BT474 contains more than 20 calculated copy number transitions, and includes an amplification of the AKT3 gene. Other recurring high-level changes are apparent, including homozygous losses of the CDKN2A locus, as well as high-level amplifications of the oncogene c-MYC. Furthermore, copy number aberrations seen in this study seem to be concordant with those noted in earlier studies. For instance, the breast cancer cell line SK-BR-3 shows a balanced loss of 8p and an amplification of the HER2 locus, which are alterations noted in previous studies of this cell line (11, 12, 33). Two additional cell lines harbored DNA amplifications of the HER2 locus (NCI-N87 and BT-474), whereas a total of six lines were EGFR-amplified (HN5, A431, MDA-MB-468, SCC-15, Calu-27, and SCC-4). In all these cases, elevated protein abundance was confirmed by Western blot analysis (ref. 3; Supplemental Tables S1 and S3).⁴

A weighted voting predictive model for lapatinib sensitivity was constructed using genome-wide DNA copy number measurements derived from 100K Affymetrix SNP Chips from 24 tumor-derived cell lines of mixed histology. Model performance, evaluated using leave-one-out cross-validation, was found to be significant (P < 0.01). This significance is reflected in the strong relationship between cross-validation scores for each cell line and its relative sensitivity (Fig. 1A; ROC cross-validation analysis shown in Supplemental Fig. S1; Supplemental Table S1).⁴ Notably, cell lines with intermediate sensitivity to lapatinib spanned the boundary between resistant and sensitive.

Nine distinct genomic regions were identified as contributing to the optimal model (Fig. 2). Regions encompassing the EGFR and HER2 genes (7p11.2 and 17q12, respectively) were among these loci (Fig. 2). Several other cancer-related genes map to the loci in the optimal model. Included in the model are deletions of the familial melanoma locus CDKN2A (9p21.3), DNA gains of the candidate oncogene ZNF217(20q13.2) as well as losses of the candidate tumor suppressor gene PTPRD (9p23; Table 2; additional details for all genes available in Supplemental Table S2).⁴ There seemed to be multiple contributing rearrangements in each cell line, in which those with the highest sensitivity had rearrangements in all nine genomic regions of the optimal model, whereas less sensitive lines harbored fewer alterations.

To independently assess the optimal model, an additional data set consisting of 10 head and neck cell lines was analyzed for sensitivity to lapatinib and genome-wide copy number alterations. The head and neck cell lines had a range of sensitivities to lapatinib, including six sensitive cell lines (IC₅₀ < 0.10 μmol/L) and two resistant cell lines (IC₅₀ > 2.5 μmol/L; Fig. 3; Supplemental Table S3).⁴ Although, high EGFR amplification is prevalent in head and neck tumors and is a poor prognostic indicator (29), HER2 amplifications have not been previously reported. As expected, several lines in this panel (3 of 10; 30%) harbored EGFR amplification (more than five copies), whereas none seemed to have amplification of HER2. Predictive model scores that were generated for each head and neck cell line seemed to be generally related to its sensitivity to lapatinib (Fig. 1B; Supplemental Table S3).⁴ Given score boundaries derived from cross-validation, the optimal model successfully predicted two of two of the lapatinib-resistant cell lines, and four of six of the lapatinib-sensitive cell lines, whereas both of the intermediate-sensitivity cell lines had scores close to the boundary between resistant and sensitive (all intermediate to those correctly classified as most sensitive or resistant). Specifically, sensitive lines scored <0.15 whereas resistant cell lines scored >0.3. Only two cell lines, CAL27 and SCC9, had scores discordant with their measured sensitivities; this may be due to the presence of additional alterations that confer resistance.

In order to determine the effects that DNA copy number alterations may have on transcript levels, a comparison was made between sensitive cell lines versus the resistant cell lines (including both training and test panels) for the genes mapping to the regions composing the optimal performing predictive model. Statistical significance was assessed by simple t test, where a P value threshold of 0.05 was applied. For those probes expressed at a minimum of 50 units in at least one sensitivity group, a total of 55 probes (67.1%; n = 82) representing 31 unique genes (70.5%; n = 44) were differentially expressed (Table 2). The fold change (mean of sensitive/mean of resistant) of 54 of 55 (98.2%) differentially expressed probe sets was concordant with the DNA copy number changes exhibited by the regions to which they map (30 of 31 genes; 96.8%; Supplemental Table S4).⁴ This high concordance between DNA copy number and transcript abundance is highly significant for the genes mapping to 17q12, in which 20 of 22 (90.9%) probe sets showed differential expression and fold changes concordant with the DNA copy number gains seen in the sensitive group.

By using full genome copy number profiles and associated sensitivities to lapatinib in a panel of 24 cell lines of mixed histologic origins, a predictive model was constructed for the purpose of identifying the likelihood of response in untested cell lines. Both the sensitivity categorizations and model variables were subsequently applied to an independent panel of non–HER2-amplified head and neck cancer cell lines (n = 10). Both leave-one-out cross-validation and the independent validation set suggest that copy number profiles can efficiently predict cell line sensitivity to lapatinib even in the absence of HER2 amplification.

The most significant finding of this study is the unbiased rediscovery, from genome-wide copy number data, of the EGFR and HER2 amplicons as predictors of lapatinib sensitivity. This is consistent with previous studies demonstrating that lapatinib treatment leads to the inhibition of EGFR and HER2, and consequently results in growth arrest and/or apoptosis in a variety of EGFR and HER2-dependent tumor cell lines (1, 30). This is also concordant with the correlation between HER2 expression and the ability of lapatinib to inhibit HER2, Raf, Akt, and Erk phosphorylation in a panel of breast cancer cell lines (2). Notably, other known members of the HER2 signaling pathway, such as PTEN and AKT, were not identified as useful predictors.

Although it can be expected that copy number increases of target genes are competent predictors of drug sensitivity in cell lines, their recognition from an unbiased genome–wide association study has several important implications. First, it suggests that molecular profiling using genome-wide assays can be a powerful and valid means of identifying biomarkers predictive of clinical response. This has been shown even for less selective agents, such as molecular profiles that predict for taxane activity in the treatment of breast cancer (31). Second, genome-wide profiles could point to additional genetic aberrations that further influence the efficacy of an agent in the absence of (or in conjunction with) primary markers; the fact that clinical responses may occur in only ∼25% of an already-enriched population (32) highlights the potential existence of additional predictive markers. Finally, given that genome-wide copy number aberrations in cell lines have been shown to correlate with primary tumor tissues in individual tumor types (e.g., breast cancer; refs. 33, 34) as well as across histologies (18), finding such loci in vitro can yield clinically relevant testable hypotheses. In the current study, discerning both target genes of lapatinib (EGFR and HER2) from a relatively limited number of cell lines of mixed histologies infers the potential to delineate molecular markers that span tumor types. The clinical activity of lapatinib against HER2+ breast cancer has recently been proven (35), and there are ongoing studies in patients with HER2+ tumors of other histologies (i.e., gastric, bladder, and ovarian cancers). Although substantial activity has not been seen in unselected patients in phase II studies of lung and colorectal cancers (36), a retrospective analysis of a phase III study in renal cell cancer suggests benefit in patients whose tumors markedly overexpress EGFR (amplification of HER2 is extremely rare in renal cell cancer; ref. 37).

The novel in vitro predictors observed in this study are partially composed of genes with established roles in oncogenesis, apoptosis, and tumor suppression. The CDKN2A locus (9p21.3) contains three related genes (ARF, p15INK4b, and p16INK4a) that encode distinct tumor suppressor proteins, all of which have well-established roles in tumor suppression (38). Both the loss of heterozygosity and complete DNA deletion of CDKN2A are common occurrences in a wide range of tumors including melanoma (39), leukemias (40), and non–small cell lung cancer (41). The recurring amplification of ZNF217 (20q13.2) has also been identified in studies of breast cancer and ovarian carcinomas. Specifically, ZNF217, a Kruppel-like zinc-finger protein, which acts as an oncogene targeting E-cadherin, Akt phosphorylation, and cell survival (42–45). The FADD gene (11q13.3), the amplification of which was associated with lapatinib sensitivity in this study, is frequently mutated in human malignancies. FADD seems to play a central role in the control of programmed cell death, as loss of expression in tumor cells confers survival and growth advantages (46). PTPRD (9p23), another candidate tumor-suppressor gene, has been identified by high-resolution genomic analysis in both cell lines and primary tumors for neuroblastomas (47) and non–small cell lung cancer (41, 48), suggesting an important contribution to carcinogenesis in these tumor types.

A high degree of concordance (98.2%) was found between DNA copy number and transcript abundance of probes mapping to DNA model regions. Although the direct effects of DNA copy number have been well documented (e.g., ref. 49), this result suggests that computational models derived from transcript data may be reverse-engineered in a way that allows clinical testing by DNA aberrations (and vice versa). Furthermore, this has functional implications for the genes that reside in the copy number regions, suggesting that the proteins encoded by these genes, as is the case with EGFR and HER2, may affect the clinical responsiveness to lapatinib.

Recent studies have shown the in vitro sensitivity of EGFR-amplified head and neck cancers to lapatinib (3). As expected from previous studies, HER2 amplification was absent in this panel of head and neck cell lines, and provided no prognostic significance (3). Thus, the nine-locus predictive model highlights the utility of additional genomic alterations in predicting lapatinib response in head and neck cancer cell lines. Limited array-CGH studies of head and neck cancers corroborate many of these loci as candidate genes for driving the development and progression of head and neck cancer (50, 51). Collectively, the sensitivity predicted from a limited set of fixed genetic alterations in a subset of head and neck cell lines suggests the potential for clinical benefit. A recent phase II study of the combination of lapatinib with chemotherapy and radiation in head and neck cancer showed promising results (52), and phase III studies are ongoing.

Patient enrichment strategies have guided the development of several cancer therapies. In the cases of trastuzumab with HER2 and of imatinib with BCR-ABL, PDGFRA, or KIT, clinical activity has correlated best with the presence of specific fixed genetic aberrations (e.g., ref. 53). An immunohistochemical assay for detecting EGFR positivity has been included in the development of EGFR inhibitors, at least in CRC; however, to date, this assay has not been an adequate predictor of clinical activity (54). Furthermore, more recent data suggest that the best predictor is the absence of a downstream mutation in KRAS. Whether this is due to a biological reason (i.e., the presence of a genetic aberration encodes for oncogenic addiction; ref. 55) or a technical one (whereby quantifying the expression of a specific protein is less reliable than quantifying the relative amounts of a given gene) is unclear. Given this and the technical limitations in proteomic profiling, the strategy of many recent pharmacogenomic studies has been to quantify relationships between transcript abundance and clinical response to cancer therapy (15, 16, 56). To a lesser extent, DNA alterations have been used for this purpose (e.g., ref. 20). As with other genome-wide profiling platforms, microarray-based DNA copy number studies provide an unbiased, parallel, high-throughput interrogation of many genes; however, pharmacogenomic studies of DNA copy number may be advantageous for several reasons. First, genetic aberrations seem to be relatively stable in vivo over time and place (i.e., between primary tumors and metastasis), likely more so than RNA or protein changes. Second, genetic aberrations are uniquely derived from the tumor cells (whereas gene transcripts likely represent stromal cells as well). Finally, as noted above, genetic aberrations are faithfully represented in cell lines, allowing for hypotheses to be generated and tested in vitro. Limitations of array-based copy number analysis include the requirement of sufficient tissue material, the relative expense of the technique, and the lack of availability of array equipment in the clinical setting (17, 57, 58). In addition, copy number analyses will obviously not discover specific mutations (e.g., EGFR and KRAS) relevant to drug responsiveness.

In summary, we show here that an unbiased model derived from high-resolution genome-wide DNA copy number profiles estimates the in vitro efficacy of lapatinib. This model correctly identified increases in copy number of the two known target receptors, EGFR and HER2, as well as additional regions harboring several cancer-related genes. Collectively, these alterations provide more robust estimations of cell line sensitivity to lapatinib compared with target genes alone. These analyses illustrate for the first time that genome-wide copy number analyses can be used to discover clinically relevant predictors of sensitivity to anticancer therapeutics.

We thank the Discovery Technology Group at GlaxoSmithKline for all gene expression assay work.