Amplified centrosomes and mitotic index display poor concordance between patient tumors and cultured cancer cells

Centrosome aberrations (CA) and abnormal mitoses are considered beacons of malignancy. Cancer cell doubling times in patient tumors are longer than in cultures, but differences in CA between tumors and cultured cells are uncharacterized. We compare mitoses and CA in patient tumors, xenografts, and tumor cell lines. We find that mitoses are rare in patient tumors compared with xenografts and cell lines. Contrastingly, CA is more extensive in patient tumors and xenografts (~35–50% cells) than cell lines (~5–15%), although CA declines in patient-derived tumor cells over time. Intratumoral hypoxia may explain elevated CA in vivo because exposure of cultured cells to hypoxia or mimicking hypoxia pharmacologically or genetically increases CA, and HIF-1α and hypoxic gene signature expression correlate with CA and centrosomal gene signature expression in breast tumors. These results highlight the importance of utilizing low-passage-number patient-derived cell lines in studying CA to more faithfully recapitulate in vivo cellular phenotypes.

centrosomes assembled together in interphase (either individually distinguishable or clustered tightly). The third category, termed "PCM" comprised cells with centrosomes whose volumes were above-normal and were represented as only one γ-tubulin spot, not a cluster of γ-tubulin spots.
Centrosome aberrations (CA) was calculated as a percentage by adding percent cells harboring more than two γ-tubulin foci and percent cells harboring γ-tubulin foci with volume greater than upper range of mean centrosomal volumes found in respective normal tissues ( Supplementary   Fig. 1). Since centrosomes pass through a duplication cycle that involves large volume changes, we needed to define a "normal range" for centrosomal volumes using both adjacent uninvolved tissue from cancer patients and normal tissue for disease-free individuals for each cancer type.
To determine the normal range, we analyzed volumes of centrosomes (500 centrosomes for each sample) in adjacent uninvolved tissue from cancer patients (20 samples for each cancer type) and in normal tissues (20 normal tissue samples for breast, pancreas and bladder). Normal tissue samples were obtained from Biomax Inc. in the form of commercial tissue microarrays. We

Enrichment of centrosomal gene expression in tumors with a hypoxia-high gene expression signature
We validated our in vitro findings of a correlation between CA and hypoxia in silico by probing the publicly-available Kao 1 and Jonsdottir 2 microarray datasets using Gene Set Enrichment Analysis (GSEA). 3 Essentially, our goal was to determine whether breast tumors that are enriched in hypoxia-associated transcripts also show a correlational enrichment in centrosomal transcripts.
Publicly available pre-processed gene expression profiles of primary breast tumors (n=327 for the Kao dataset, GSE20685; n=94 for the Jonsdottir dataset, GSE46563) were used for GSEA. Within each dataset, patients were stratified into two groups by a hypoxia score, the reduced hypoxia metagene previously shown to have prognostic ability in multiple cancers. 4,5 As previously defined, hypoxia scores were calculated as the median expression of 26 genes that are upregulated in response to hypoxia. Scores  median were categorized as "hypoxia low" and scores > median were categorized as "hypoxia high." For the Kao dataset, Affymetrix probes with the "x_at" extension were removed unless no other probe was available (e.g., as with ALDOA).
For the Jonsdottir dataset, Illumina probes with the "A" designator were preferentially used. When multiple probes were present, their median expression was used in score calculation. GSEA was performed with 1000 permutations, and false discovery rate q-values<0.05 were considered statistically significant.
Using the Kao dataset, we collapsed features into gene symbols, resulting in 20,606 genes being available for GSEA using curated gene sets from Molecular Signatures Database 6 v5.0, including those from the Gene Ontology (GO) Consortium (for analysis of cellular components and biological processes) and Reactome 7 v53 (for pathway analysis), along with gene sets that we defined based on empirical evidence from the literature. We validated that the hypoxia-high group was differentially enriched in hypoxia-associated genes by performing GSEA with the full hypoxia metagene as shown in Supplementary Fig. 7 (also see Supplementary Table 1 for study details   and Supplementary Table 2 for the ranked gene list; n=44 after filtering). We then performed GSEA to identify gene ontologies associated with the hypoxia-high group, which we found was significantly enriched in microtubule-organizing center and centrosome components, which were among the top-20 enriched cellular components (see Supplementary Table 1 for these and all other enriched gene ontologies). The hypoxia-high group was also enriched in cell cycle-related processes, which constituted the top-ranked gene ontology among biological processes. Cellular pathway analysis using Reactome terms identified mitosis as the third-most enriched pathway, with various other cell cycle-related pathways also significantly enriched. Cellular pathway analysis revealed an enrichment in genes associated with the recruitment of centrosome proteins and complexes. Intriguingly, the hypoxia-high group was also enriched in genes involved in the cellular pathway associated with loss of ninein-like protein (NLP), a γ-tubulin-binding protein, from mitotic centrosomes. It is known that PLK1 and NEK2 phosphorylate NLP at the onset of mitosis, resulting displacement of NLP from the centrosome, which is associated with centrosome maturation (involving the recruitment of γ-tubulin ring complexes and other pericentriolar material components) and a concomitant increase in microtubule-nucleating capacity. PLK1 or NEK2 overexpression results in premature NLP dissociation from centrosomes and also induces CA. 8 Although hypoxia-high breast tumors were clearly found to be enriched in centrosomal components and pathways, we wanted to more specifically test the hypothesis that they are enriched in gene ontologies related to CA per se. No high-throughput screen of CA-associated genes has been performed to inform construction of a CA gene set; nevertheless, the literature reports that CA is associated with hormone receptor-negative and node-positive breast cancer. 9 Thus, we analyzed enrichment of centrosome-associated genes (namely, experimentally identified human centrosomal proteins in the MiCroKiTS 10 database; n=540 genes) in hormone receptor-positive node-negative patients, rationalizing that this gene set has a high likelihood of representing CA. We found that 77 of these genes were enriched in hormone receptor-negative node-positive breast carcinomas. Next, we performed GSEA using these 77 genes as a gene set, which we found was significantly enriched in the hypoxia-high group, as shown in Supplementary   Fig. 7B (also see Supplementary Table 3 for the ranked gene list). Many genes implicated in CA (such as AURKA, CCNA2, CCNE2, CEP152, NEK2, PLK4, and STIL) or amplified centrosome clustering (such as KIFC1, the top-ranked hit, along with BIRC5 and TACC3) from the literature are among the enriched genes from this set. Because CA drives chromosomal instability (CIN), we wondered whether hypoxia-high cases were also enriched in CIN-associated genes. To this end, we performed GSEA with genes from the CIN25 signature, net overexpression of which has prognostic significance in various types of cancer. 11 We found this set was highly enriched in the hypoxia-high group (Supplementary Table 1). Collectively, these results suggest that hypoxic breast tumors are enriched in CA-and CIN-associated genes.

Enrichment of centrosomal gene expression in tumors with a hypoxia-high gene expression signature regardless of mitotic activity
Many CA-associated proteins do not exclusively localize to the centrosome; some also localize to the mitotic spindle. Thus, it could be argued that, rather than having a greater extent of CA, the hypoxia-high group merely has more mitotic cells than the hypoxia-low group. To test this hypothesis, we analyzed the Jonsdottir dataset, which contains gene expression profiles and mitotic activity indices for 94 breast tumor specimens from lymph node-negative patients. To begin, we validated that the hypoxia-high group was enriched in hypoxia-associated genes. We performed GSEA with the full hypoxia metagene and found significant enrichment (Supplementary Table 1), which also underscores the robustness of this 26-gene hypoxia signature across platforms and breast cancer datasets. We then performed GSEA using the 77 potentially CA-associated genes (that is, those that were enriched in the hormone receptornegative node-positive breast carcinomas from the Kao dataset) and found significant enrichment in the Jonsdottir dataset as well (Supplementary Table 4). This is especially interesting because the Jonsdottir patients are also all node-negative, indicating this gene set captures a phenotype that is not wholly dependent on nodal status. There was substantial overlap in the potentially CAassociated genes enriched in the Jonsdottir and Kao hypoxia-high groups. Next, we did find that the hypoxia-high group was associated with a high mitotic activity index (MAI; >10 mitotic figures per 10 fields of vision) based on the Mann-Whitney test (p=0.01). Nonetheless, when we performed GSEA on the MAI-low group (n=60) using hypoxia scores as the phenotype, we still found that the hypoxia-high group was enriched in potentially CA-associated genes (Supplementary Table 5). Thus, even among tumors with relatively low mitotic activity, hypoxiahigh tumors show enrichment in potentially CA-associated genes, minimizing the probability that we are merely capturing proliferation-associated genes with our gene set. Combined with our in vitro data, these in silico data substantiate the hypothesis that hypoxia is associated with CA in patient breast tumors.
Finally, we were interested to determine whether hypoxia-associated CA, as determined by gene expression levels, predicts worse outcomes and, if so, whether its predictive ability depends on mitotic activity. To this end, we created a score based on the top ten CA-associated genes enriched in the hypoxia-high samples of the Jonsdottir dataset (from Supplementary Table 4).
Specifically, we defined the hypoxia-associated CA score as the median expression of those top 10 genes. Kaplan-Meier analysis and Cox regression were performed using SPSS Statistics version 21 (IBM). For multivariate Cox regression analysis, all potential predictors were entered into the full model and then eliminated stepwise based on an α=0.10 elimination criterion. Optimal cut points based on distant-metastasis-free survival (DMFS) were found using X-tile 12 per the highest Χ 2 value following dichotomization. We found that stratifying patients based on a cutpoint of 317 resulted in the CA score having the best predictive ability using the 94 node-negative breast cancer patients of the Jonsdottir dataset (p=0.020; Supplementary Fig. 6C). Univariate Cox regression revealed that a high hypoxia-associated CA score (i.e., >317) was associated with worse DMFS (HR=2.87; p=0.026), which was upheld in multivariate regression adjusting for all available potentially confounding covariates (including tumor size, Nottingham grade, estrogen and human epidermal growth factor receptor 2 statuses, and mitotic activity index). In fact, only this score remained in the final model. When hypoxia score was added to the Cox regression analysis, the effect of CA score on DMFS was more pronounced (HR=3.39, p=0.011). Only the CA score and hypoxia score remained in the final model, though the hypoxia score was no longer significant (HR=2.22, p=0.066). When the analysis was repeated without the CA score in the full model, however, the hypoxia score was a significant predictor of DMFS (HR=2.45, p=0.047), as was mitotic activity (HR=2.88, p=0.0.17), with no other variables in the final model. These results raise the tantalizing possibility that the ability of the hypoxia score to predict DMFS results from its association with CA. Even more intriguing is the idea that hypoxia might upregulate CA to drive metastatic dissemination, an exciting avenue of future research.  Image-J for immunoblot assays provided in main manuscript and supplementary data.