Mitotic Checkpoints and Chromosome Instability Are Strong Predictors of Clinical Outcome in Gastrointestinal Stromal Tumors

Gastrointestinal stromal tumors (GIST) are the most frequent mesenchymal tumors of the gastrointestinal tract and account for approximately 25% of soft tissue sarcomas. They are thought to arise from the intestinal cells of Cajal (1) or from a common progenitor cell (2). Most GISTs (80%) have activating mutations in the KIT tyrosine kinase receptor gene, but 8% have platelet-derived growth factor receptor α (PDGFRA) mutations (3, 4) and a few of the remainder have BRAF mutations (5). In addition to these mutations, the most frequently reported genetic changes are 14q, 22q, and 1p losses (6).

Clinical management of GISTs consists mainly of surgical resection and adjuvant targeted therapy with imatinib mesylate (Gleevec, Novartis Pharma AG), which targets mutationally activated KIT or PDGFRA signaling (7). Around 20% to 40% of patients relapse, with distant liver metastasis being the most common manifestation of recurrence. It is mainly patients with these aggressive GISTs who benefit from imatinib therapy. Precise evaluation of metastatic risk is therefore highly desirable.

Many pathologic criteria based on tumor site, tumor size, cell type, degree of necrosis, and mitotic rate have been proposed for predicting the outcome of patients with GISTs. A consensus grading scheme based on tumor size and mitotic count was proposed by the U.S. NIH in 2001 to estimate the prognosis of GIST patients (8). In 2006, the Armed Forces Institute of Pathology (AFIP) proposed an updated system taking into account also tumor location (9). Both systems are based on histopathologic assessment of tumor aggressiveness. Cutoff values defining risk groups have been determined empirically but generate a large intermediate-risk group for which adjuvant imatinib is controversial because the real metastatic risk is poorly defined. Hence, there is a need to better understand GIST biology to identify biomarkers causally linked to poor outcome.

To address this need, multiple DNA copy number and gene expression studies have been carried out but, for a variety of reasons including small sample size and availability of clinical data, the results were generally inconclusive. It has been shown that the number and complexity of genomic rearrangements increase with tumor stage but no threshold has been defined (6, 10–16). At the expression level, Yamaguchi and colleagues reported a gene expression signature based on 32 GISTs that predicts outcome, but only in gastric GISTs (17).

We recently established a 67 gene prognostic signature related to chromosome integrity, mitotic control, and genome complexity in sarcomas (CINSARC for Complexity INdex in SARComa; ref. 18). To assess the effectiveness of this signature in GISTs, we have used it to score 67 fully annotated primary untreated GISTs. To identify the underlying mechanisms leading to high CINSARC scores, we have carried out genome-wide DNA copy number and gene expression analyses of these tumors.

Frozen samples from 67 resected primary GISTs untreated until tumor recurrence were selected from the European GIST database CONTICAGIST (www.conticagist.org). Dates of diagnosis range from June 1995 to February 2009. Information regarding tumors and patients are summarized in Table 1.

DNA was hybridized to 8 × 60 K whole genome Agilent arrays (G4450A) according to the manufacturer's protocol. The ADM-2 algorithm of comparative genomic hybridization (CGH) Analytics v4.0.76 software (Agilent) was used to identify DNA copy number anomalies at the probe level. A low-level copy number gain was defined as a log₂ ratio more than 0.25, and a copy number loss was defined as a log₂ ratio less than −0.25. A high-level gain or amplification was defined as a log₂ ratio more than 1.5, and a homozygous deletion was suspected when the ratio was below −1.

Gene expression analysis was carried out by Agilent Whole human 44K Genome Oligo Array (Agilent Technologies) according to the manufacturer's protocol. All microarrays were simultaneously normalized with the Quantile algorithm. t Tests were carried out using Genespring (Agilent Technologies), and P values were adjusted by the Benjamini–Hochberg procedure. The P value and fold change cutoff for gene selection were 0.001 and 3, respectively. Gene ontology (GO) analysis was conducted to establish statistical enrichment in GO terms using Genespring (Agilent Technologies). MIAME-compliant data have been deposited at Array Express [Experiment name: Prediction of clinical outcome in GISTs (Gastro Intestinal Stromal Tumours); ArrayExpress accession: E-MTAB-373; Reviewer login: E-MTAB-373_Reviewer; Password: Gc3giN7].

The copy number status of p14, p15, and p16 was determined as previously described (19). A normal status was assigned to a ratio of 0.8 or more and 1.2 or less. A ratio of more than 0.1 and less than 0.8 was scored as a hemizygous deletion. When ratio was below 0.1, the deletion was scored as homozygous.

Reverse transcription (RT) and quantitative PCR (qPCR) for p14, p16, AURKA, and RB1 were carried out as previously described (19). A reference C_t (threshold cycle) for each sample was defined as the average measured CT of the 3 reference genes, GAPDH, ACTB, and RPLP0. Relative mRNA level in a sample was defined as: ΔC_t = C_t (gene of interest) − C_t (mean of the 3 reference genes).

Immunohistochemistry experiment was realized on tissue microarrays (TMA) containing 15 cases from the present series and carried out as previously described (20). Antigen retrieval was achieved using the Dako Target Retrieval Solution, pH 9 for 20 minutes at 98°C. Slides were incubated for 1 hour with the AURKA antibody used at a dilution of 1:50 (Novocastra, NCL-L-AK2, clone JLM28). Each case was spotted in triplicate on the TMA, and we used the average value of the 3 spots.

The CINSARC centroids are mean-centered reference profiles for the CINSARC signature genes in the 310 metastatic and nonmetastatic sarcomas from our previous study (18). Each GIST was allocated to the prognostic class with the highest Spearman correlation to the reference centroids.

Metastasis-free survival (MFS) was calculated by the Kaplan–Meier method from the date of initial diagnosis to the date of first metastasis, relapse, last follow-up or death for patients without diagnosis of metastasis. Survival curves were compared with the log-rank test. All survival analyses were conducted using R software (version 2.11.11) and the “survival” package. HRs and multivariate analysis were conducted with the Cox proportional hazard model or Cox regression with the Firth's correction (R software, “coxph” package) depending on occurrence or not of events in the reference group.

To test whether the CINSARC signature has prognostic value in GISTs, we carried out gene expression profiling on a series of 60 of 67 (89.5%) GISTs with mRNA of sufficient quality (Table 1). We assigned these tumors to prognostic groups based on correlation with the published CINSARC centroids from our previous series of 310 sarcomas (18). Survival analysis (Fig. 1) revealed that the CINSARC classification split the tumors into 2 groups with very different MFS (P = 1.4 × 10^–5). No metastasis or other relapse event occurred in the good prognosis group.

The CINSARC signature is based on several different types of sarcoma. To test whether it is possible to derive a better signature that is specific for GISTs, we analyzed the GIST gene expression profiles to identify genes differentially expressed by the metastatic and nonmetastatic tumors. Among the 297 differentially expressed genes (Supplementary Table S1), 70 (86 probe sets) were downregulated and 227 (252 probe sets) were upregulated in metastatic cases (FC > 3 and P < 0.001). GO analysis identified no significantly enriched pathways for the 70 downregulated genes. In contrast, GO analysis revealed that 32 of the 40 (80%) pathways containing upregulated GIST genes were also identified by GO analysis with the CINSARC genes (Supplementary Table S2). Indeed, 45 of the 227 upregulated GIST genes belonged directly to the CINSARC signature. Moreover, GO analysis of the remaining 182 differentially regulated genes not included in CINSARC signature showed enrichment for the same pathways as for the CINSARC genes (Supplementary Table S3).

Among the top-ranked differentially expressed genes identified by t test, AURKA (Aurora kinase A, previously called STK6 or STK15) was the highest ranked gene belonging to the CINSARC signature (Supplementary Table S1). We validated this result by qRT-PCR, which showed that there was a high correlation between the microarray and PCR data (Pearson correlation coefficient = 0.94; P < 1 × 10^–15), and by immunohistochemistry on a TMA containing 15 of the GISTs from the present series (Supplementary Fig. S3). To test the hypothesis that AURKA alone has prognostic value, we stratified samples by AURKA expression. We used the mean AURKA level (9.15) as a cutoff (Table 2). Survival analysis showed that AURKA expression splits the tumors into 2 groups with very different outcomes (MFS: P = 5.31 × 10^–11; Fig. 2A). To validate the result, we studied AURKA expression in the GISTs from the study by Yamaguchi and colleagues (17). This confirmed that AURKA splits GISTs into groups with a large difference in MFS (P = 9.5 × 10^–4; Fig. 2B).

To test the hypothesis that AURKA amplification could account for the AURKA overexpression, we carried out CGH to determine the genomic profile of 66 GISTs for which DNA of sufficient quality was available. No AURKA amplification was detected. We therefore examined the CGH data for other alterations that could potentially explain the increased AURKA expression and poor clinical outcome. The CGH profiles ranged from simple, that is, without any detectable changes, to complex, with multiple full chromosome and segmental gains and losses (Fig. 3A). We compared the frequency of gains and losses for each probe between GISTs with and GISTs without metastatic outcome (Supplementary Fig. S1). No significant difference in gains was observed, whereas several probes showed significant differences in losses. Among the top-ranked losses, the biggest difference was observed for 8 probes on 9p21 deleted in 78.9% and 9.6% of the metastatic and nonmetastatic cases, respectively (Supplementary Fig. S1). All these probes target either the CDKN2A (3 probes), CDKN2B (3 probes), or MTAP (2 probes) loci. 9p21 deletions were observed in 18 patients (18 of 66 = 27%), of whom 13 developed metastases (13 of 18 = 72%). The deletions involved either the whole 9p arms or they were restricted to the CDKN2A/B loci (Supplementary Fig. S2). They were scored as homozygous in 7 cases (6 of 7 with metastatic outcome) because of the very low CGH ratios (Supplementary Fig. S2). These homozygous deletions allowed us to define more precisely the genes of interest because in 2 tumors the homozygous deletion excluded MTAP (GISTs #5 and #17). We checked the CDKN2A and CDKN2B copy number status by genomic qPCR and fully confirmed the exclusion of MTAP from the minimal deleted region. Interestingly, qPCR showed that the minimal deleted region included CDKN2A but not CDKN2B (GIST #5; Table 2). As the Agilent gene expression probes target sequences common to the p14 and p16 mRNA, we carried out RT-qPCR with primers specific for the individual transcripts (Supplementary Table S4). In all 7 tumors lacking both copies of CDKN2A, and in 3 tumors with only 1 copy, both the p14 and p16 transcripts were absent or nearly absent. However, in 2 cases with CDKN2A deletion but without downregulation of p14/p16 expression in the gene expression microarray data, we observed a specific decrease of p16 but not p14 expression, indicating that the target gene of the CDKN2A deletions is likely to be p16. To explain metastatic cases without CDKN2A deletion, we sought other possible genomic alterations that could interfere with Restriction point control. We identified 1 homozygous deletion and 13 hemizygous deletions at the RB1 locus (Table 2). Nine of these deletions occurred in recurrent or metastatic tumors and none of them were observed in cases with CDKN2A homozygous deletions. RT-qPCR analysis confirmed that deleted tumors had significant downregulation of RB1 expression (t test: P = 3.5 × 10^–4; Supplementary Table S4). Comparison of tumors with and without p16/RB1 alterations showed that tumors with p16/RB1 alterations overexpressed 235 probe sets (FC > 3; P < 10^–3) including 42 genes from the CINSARC signature, one of which was AURKA (Supplementary Table S5).

The CGH profiles of the tumors that did not undergo metastasis had no or few losses or gains, generally involving whole chromosomes, whereas the tumors that developed metastases harbored more frequently segmental alterations. We therefore decided to test whether genome complexity could predict metastatic outcome (Fig. 3A). To take into account the number and the type of changes, a Genomic Index (GI) was calculated for each profile as follows: GI = A²/C, where A is the total number of alterations (segmental gains and losses) and C is the number of involved chromosomes. The Genomic Index across the entire series ranged from 0 to 56. The proportion of metastatic cases increased with Genomic Index. Metastatic cases predominated when the Genomic Index was over 10 (Table 2 and Fig. 3B). Stratification by Genomic Index at a cutoff of 10 split the tumors into 2 groups with very different outcomes (Table 2, Fig. 4A). Interestingly, Genomic Index was able to predict metastatic outcome in GISTs in the intermediate-risk group of the AFIP classification (9; Fig. 4B and C).

To assess this issue independency of CINSARC signature, AURKA expression and Genomic Index were evaluated together with AFIP grading system in a multivariate analysis (Table 3). AURKA expression seems to be the stronger prognostic marker (HR = 11.97, 95% CI = (1.60–1406); Table 3, a), and the AFIP is not significant when compared with AURKA expression. As suspected, this analysis showed that CINSARC, AURKA (which belongs to CINSARC), and Genomic Index correlate making the former and the later not significant face to AURKA expression. We thus carried out multivariate analyses comparing each molecular signature with AFIP grading system (Table 3, b) and showed that each signature is superior to AFIP grading system to predict metastasis outcome. Of interest, AFIP intermediate statue against each of the molecular signature has no more significant prognostic value.

The development of a valid and reliable, investigator-independent method of GIST prognostication is essential for the proper clinical management of GIST patients, especially in the context of adjuvant treatment, in which many patients are exposed to imatinib, whereas only a small proportion will likely benefit from such treatment (21).

The main conclusion from this study is that the CINSARC score is a strong and validated predictor of metastasis in patients with GISTs. Remarkably, none of the patients assigned to the good prognosis group developed metastases or relapsed. Prognostic expression signatures have showed their experimental efficacy in several other tumor types, but their clinical application has been complicated by technical issues such as weak reproducibility across array platforms. Importantly, we show here that CINSARC scoring is platform independent: the signature we developed on Affymetrix data was applied and validated here on Agilent data. Furthermore, the CINSARC score was prognostic for both the nontranslocation related sarcomas on which it was originally developed (18) and for the GISTs in this study.

The CINSARC signature comprises 67 genes involved in maintenance of chromosome integrity and mitotic control, indicating that these processes play a crucial role in the development of metastasis in sarcomas (18). Supervised analysis showed that 45 of the 227 genes prognostic in GISTs were common to the CINSARC signature. The top-ranked gene common to both approaches was AURKA. The AURKA protein is a mitotic centrosomal protein kinase amplified in many cancer types (22–25). Increased AURKA expression is associated with poor prognosis in breast carcinoma (26), colon carcinoma (27, 28), neuroblastoma (29), and head and neck squamous cell carcinoma (30). AURKA overexpression induces centrosome duplication and segregation abnormalities leading to aneuploidy and malignant transformation (25). Whole chromosome losses are the most frequently observed alterations in GISTs and are assumed to originate from unequal chromosome segregation, which can be induced by AURKA overexpression (31). Contrary to the mechanism seen in other cancers, AURKA overexpression in GISTs is not explained by gene amplification, but is instead a secondary change we postulate to be caused by defects in Restriction point control. Our results point to AURKA being a very interesting potential therapeutic target in GISTs. With this in mind, it is noteworthy that immunohistochemistry shows that AURKA mRNA overexpression translates into AURKA protein overexpression (Supplementary Fig. S3). AURKA inhibitors have entered clinical trials (32–36) and could be particularly useful for imatinib-resistant GISTs that have not yet disseminated because AURKA could be an essential event leading to acquisition of metastatic potential.

Previous copy number studies identified few aberrations in GISTs, deletions being more common than gains (6, 10–12, 37–39). The authors concluded that chromosome 14, 22, and 1p deletions were the most frequent aberrations. Two studies noted that copy number changes were commoner in high-risk GISTs but did not identify a clear cutoff delineating the high-risk group (6, 10). At the expression level, most studies were designed to facilitate diagnosis (40, 41) or to predict KIT or PDGFRA mutation status (42–44). Yamaguchi and colleagues (17) carried out gene expression profiling on 32 GISTs and identified CD26 as a prognostic marker, but only in GISTs of gastric origin. They concluded that CD26 might not be the cause of malignant progression of gastric GISTs. In contrast, our study shows that CINSARC score, AURKA expression, and Genomic Index are prognostic irrespective of the tumor location (Supplementary Fig. S4). Furthermore, the biological meaning of CINSARC score and its association with genomic changes strongly indicate that CINSARC genes are implicated in malignant progression and are not just a consequence of the process. This hypothesis is supported by the association we observe between CDKN2A deletion, RB1 deletion, AURKA expression, CINSARC score, and metastasis. CDKN2A encodes 2 key tumor suppressor proteins, p16^INK4a and the p14^ARF, which regulate the Restriction point and p53, respectively. Previous studies on GISTs have linked 9p21 alterations to tumor progression (11–16, 45), but the driver gene was not positively identified (CDKN2A, CDKN2B, or MTAP; refs. 37, 39, 46–48). Here, we have shown that homozygous deletions target CDKN2A and more specifically p16^INK4a.

RB1 deletions associated to reduced Rb expression in tumors with high AURKA expression, but normal CDKN2A loci, are consistent with the known cooperation between p16 and RB1 in control of the Restriction point. Most of the CINSARC genes are known to be under the transcriptional control of E2F. RB1 sequesters E2F, which is released from the complex upon RB1 phosphorylation by CDK4. CDK4 is, in turn, inhibited by p16^INK4a. Hence, we hypothesize that alteration of the p16^INK4a or RB1 genes in GISTs is likely to be a causative event that leads to the overexpression of CINSARC genes, which in turn induce chromosome instability and ultimately metastasis. Although this model requires experimental validation in cellular and mouse models of GIST, the strong association between CINSARC gene expression and p16/Rb pathway alteration make it an attractive hypothesis.

Both the AFIP (9) and NIH (8) histologic-based grading systems are widely accepted as “gold standards” in determining metastatic risk and to determine whether a GIST patient should receive adjuvant therapy with imatinib. Adjuvant imatinib is now recommended for localized GISTs of more than 3 cm in the United States or for high-risk or intermediate-risk localized GISTs in Europe. Many patients at AFIP-intermediate risk or with a tumor more than 3 cm will not benefit from imatinib. The ability to select patients likely to benefit from imatinib would be an important advance in the management of GISTs. Here, we show that CINSARC score, AURKA expression, and Genomic Index all outperform the AFIP classification (Table 3b and Fig. 4) and do so independently of tumor location (gastric versus nongastric GISTs; Supplementary Fig. S4). Even if AURKA expression (mRNA level) seems in multivariate analysis as the best predictor of metastasis outcome, its clinical application is limited due to weak quality of mRNA extracted from formalin-fixed paraffin-embedded (FFPE). CGH is a technique applicable to FFPE samples already used in routine diagnostic pathology laboratories, and the Genomic Index is nearly as much effective as AURKA and CINSARC to distinguish good from poor prognosis patients particularly in AFIP-intermediate risk GISTs (which represent around 25% of diagnoses). Genomic Index is therefore potentially the best overall tool to manage imatinib therapy for intermediate risk GIST patients. We recommend carrying out a clinical trial comparing these molecular signatures to the AFIP/NIH methods to validate this hypothesis in a prospective clinical context.

No potential conflicts of interest were disclosed.

R. Sciot, P. Schöffski, M. Debiec-Rychter, A. Neuville, and J.-M. Coindre supplied tumor tissue, did the central pathology review, and collected the clinical follow-up data. V. Dapremont carried out DNA and RNA extractions. F. Chibon supervised the laboratory experiments. P. Lagarde and G. Pérot carried out laboratory experiments. A. Wozniak, M. Debiec-Rychter, and I. Hostein carried out KIT and PDGFRA mutational analysis. A. Kauffmann and C. Brulard calculated centroid scores and conducted survival analysis. P. Lagarde, G. Pérot, and F. Chibon analyzed the data. F. Chibon designed the study. F. Chibon wrote the report. A. Aurias, J.-M. Coindre, and F. Chibon obtained funding for the project. All investigators reviewed and approved the final report.

The authors thank Pippa McKelvie-Sebileau for correction of the English text and Richard Iggo for critical reading of the manuscript.

This work was supported by grants from the French National Cancer Institute (INCa), the European Connective Tissue Cancer Network (CONTICANET, FP6-018806), the French Institut Nationale de la Santé et de la Recherche Médicale (INSERM), the Life Raft Group (M. Debiec-Rychter), and the Fonds voor Wetenschappelijk Onderzoek Vlaanderen (G.0589.09; M. Debiec-Rychter).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.