Prognostic role of Ki-67 in glioblastomas excluding contribution from non-neoplastic cells

Survival of glioblastoma patients varies and prognostic markers are important in the clinical setting. With digital pathology and improved immunohistochemical multiplexing becoming a part of daily diagnostics, we investigated the prognostic value of the Ki-67 labelling index (LI) in glioblastomas more precisely than previously by excluding proliferation in non-tumor cells from the analysis. We investigated the Ki-67 LI in a well-annotated population-based glioblastoma patient cohort (178 IDH-wildtype, 3 IDH-mutated). Ki-67 was identified in full tumor sections with automated digital image analysis and the contribution from non-tumor cells was excluded using quantitative double-immunohistochemistry. For comparison of the Ki-67 LI between WHO grades (II-IV), 9 IDH-mutated diffuse astrocytomas and 9 IDH-mutated anaplastic astrocytomas were stained. Median Ki-67 LI increased with increasing WHO grade (median 2.7%, 6.4% and 27.5%). There was no difference in median Ki-67 LI between IDH-mutated and IDH-wildtype glioblastomas (p = 0.9) and Ki-67 LI was not associated with survival in glioblastomas in neither univariate (p = 0.9) nor multivariate analysis including MGMT promoter methylation status and excluding IDH-mutated glioblastomas (p = 0.2). Ki-67 may be of value in the differential diagnostic setting, but it must not be over-interpreted in the clinico-pathological context.

www.nature.com/scientificreports/ lower grade gliomas or ependymomas [24][25][26] . In the glioblastoma patients, the prognostic role of Ki-67 LI is less clear. Some groups showed that a high Ki-67 LI was associated with improved OS [27][28][29] , whereas other groups reported that high levels of Ki-67 were associated with poor OS 20,21,23,30 . Yet other groups reported that Ki-67 LI was not associated with OS 3,19,[31][32][33][34] . These previous studies on Ki-67 did not include MGMT promoter methylation status and post-surgical treatment in the survival analysis although these parameters have significant prognostic impact on the outcome in the glioblastoma patients.
The reported results may also be influenced by observer-based scoring, which is prone to intra-and interobserver variation, e.g. due to different counting methods and subjective assessment of staining positivity. Vörös et al. 35 reported that a semi-quantitative assessment of the reproducibility of the Ki-67 LI in breast cancer is an acceptable and reproducible method. Similar approaches based on automated quantification have also been used to assess Ki-67 LI in tumors in the brain; indicating that automated digital quantification may be more robust than semi-quantitative scoring performed by pathologists [36][37][38] .
Another aspect influencing the results in previous studies is the Ki-67-positive non-neoplastic cells within the tumor tissue. A considerable number of microglia and macrophages, with a certain proliferative potential, can be found in brain tumors, and especially high-grade gliomas are characterized by dense infiltration with activated microglial cells and macrophages, lymphocytic infiltrates, reactive astrocytic gliosis and vascular proliferation 39,40 . Thus, non-neoplastic cells and in particular activated microglia and macrophages contribute to the overall proliferative activity and Ki-67 labeling in these tumors and may influence results obtained in earlier studies 41 . We expected that we, by exclusion of the Ki-67 contribution from proliferating non-tumor cells would be able to assess the prognostic value of the true Ki-67 level in tumor cells in glioblastomas more precisely than previously. Therefore, we used a double immunohistochemical staining approach combining Ki-67 immunostaining with a cocktail of antibodies against CD45 (immune cells), CD31 (vascular structures), Iba-1 (microglia and macrophages) and smooth muscle actin (ASMA) (vascular structures), which made it possible to identify non-neoplastic Ki-67 positive cells in the investigated glioma tissues. This allowed us to obtain a more precise evaluation of the Ki-67 LI in the actual tumor cell fraction of the tumor tissue. Because we investigated central tumor tissue only, it was decided not to stain for glial cells and neurons. To obtain a representative Ki-67 LI for each tumor we used full sections-like in daily diagnostics-instead of tissue microarrays consisting of tissue cores being only 1-3 mm in diameter.
The aim of this study was to investigate the prognostic value of Ki-67 LI restricted to glial tumor cells using a digital computer-based quantification method and a double immunohistochemical staining approach for exclusion of non-neoplastic cells. This approach was used on a well-annotated population-based cohort of astrocytic tumors. Importantly, critical prognostic parameters such as IDH mutation, MGMT promoter methylation as well as post-surgical therapy were taken into account.

Materials and methods
Patients. From a population-based cohort of 433 patients in the Region of Southern Denmark we identified 181 glioblastoma patients (178 IDH-wildtype, 3 IDH-mutated) with a sufficient amount of viable tumor tissue for immunohistochemical analyses. For the purpose of comparison, tissue from 9 IDH-mutated diffuse astrocytomas and 9 IDH-mutated anaplastic astrocytomas were stained. Tissue were obtained from the local pathological department, where it was ruinously stored. All patients underwent initial surgery between 01.01.2005 and 31.12.2009, and no treatment was received prior to surgery. This cohort has been described thoroughly and has been used in previous biomarker studies 40,[42][43][44][45][46][47][48] . Patient characteristics are listed in Table 1.
Immunohistochemistry. Immunohistochemical staining was carried out using the BenchMark Ultra IHC/ISH staining system (Ventana Medical Systems, Inc, AZ, USA). The primary antibody was Ki-67/ MIB-1 (monoclonal mouse antibody, Dako, no. M7240). To detect Ki-67 labeling in non-neoplastic cells a cocktail of antibodies against CD45 (monoclonal mouse antibody, Dako, no. M0701), CD31 (monoclonal mouse antibody, Dako, no. M0823), Alpha-smooth muscle actin (ASMA) (monoclonal rabbit antibody, Spring Bioscience no. M4710) and ionized calcium-binding adaptor molecule-1 (Iba1) (polyclonal rabbit antibody, Wako) was used for double immunohistochemical staining. For Ki-67 a detection system UVDAB-HRP plus amplification followed by application of the uDAB-kit (ref. no. 760-500) was used. For the other antibodies, we used the detection system UVRed-AP followed by application of the uRed-kit (ref. no. 760-501). Nuclear counter staining was performed using Hematoxylin II (Ventana Medical Systems, Tucson, AZ) at the BenchMark Ultra instrument. Finally, slides were washed, dehydrated, and coverslipped using a Tissue-Tek Film coverslipper (Sakura, Alphen aan den Rijn, The Netherlands). The slides were scanned on the Hamamatsu whole-slide scanner (Hamamatsu, Hamamatsu City, Japan).
MGMT assessment. The MGMT promoter status was determined using pyrosequencing (MGMT Pyro kit; Qiagen, Hilden, Germany) as described by the manufacturer. Briefly; DNA was purified from 10 lm paraffin slides using QIAamp DNA FFPE Tissue kit (Qiagen), and MGMT pyrosequencing was performed according to the kit instructions. Methylation percentages at four CpG sites were measured, if one or more sites had a methylation of 10% or higher, the tumor was considered as MGMT methylated. Image analysis. The digital images were analyzed using the Visiopharm software module (Visiopharm, Hørsholm, Denmark). The regions of interest were manually outlined, excluding areas containing staining-artifacts, brain parenchyma, necrosis or large vessels. Sampling was performed using systematic uniform random sampling at 20 × magnification; all images were reviewed and non-tumor areas were manually outlined. Images were analyzed using an algorithm developed in the Tissuemorph APP-Control system. The algorithm identi- www.nature.com/scientificreports/ fied all nuclei expressing Ki-67 and a 2.5 μm perimeter was subsequently grown around all detected nuclei. The exclusion cocktail identified non-tumor cells by a red-labelled cytoplasm, and nuclei surrounded by red-labelled cytoplasm were subsequently excluded from further analyses. It was pre-defined that only images including more than 50% viable tumor tissue were included in further analysis and that more than 5 useable images should be obtained for each tumor. Based on a sample fraction study it was shown that sampling in 10% or more of the tissue did not provide further information than sampling in 5% of the tissue (data not shown). As sampling in 10% or more was time-consuming, sampling in 5% of the tissue was chosen. Due to small amount of tissue, 27 tumors were re-sampled at a 20% fraction.

Results
Staining patterns and software classifier. The immunohistochemical double staining identified Ki-67 positive nuclei in both neoplastic and non-neoplastic cells (Fig. 1A-F). IDH-mutated diffuse astrocytomas (WHO grade II) with low cellularity had only a few scattered Ki-67 positive tumor cells and moderate expression of exclusion-marker (Fig. 1A), whereas IDH-mutated anaplastic astrocytomas (WHO grade III) with moderate cellularity showed moderate expression of the exclusion-marker (Fig. 1B). In glioblastomas (WHO grade IV) the frequency of non-neoplastic cells varied from moderate (Fig. 1C) to high (Fig. 1D).
Morphologically and immunohistochemically most non-neoplastic cells resembled activated microglia and tumor-infiltrating macrophages (Fig. 1A-D), but both small and large vessels were also identified ( Fig. 1A-D). The trained software-based classifier easily detected Ki-67 positive and negative nuclei in both neoplastic and non-neoplastic cells (Fig. 1E-F).
Ki-67 LI. In astrocytic IDH-mutated tumors median Ki-67 LI was 2.7%, 6.4% and 27.5% in WHO grade II, III and IV tumors. There was a significant difference between WHO grade II and IV tumors (p = 0.004) and between WHO grade III and IV tumors (p = 0.04). In IDH-wildtype glioblastomas median Ki-67 was 24.4%. This was not significantly different from IDH-mutated glioblastomas (p = 0.9) (Fig. 2).
An optimal cut-point analysis was performed. It was not possible to identify a cut-point that provided more information than the median (data not shown).

Discussion
In this study we investigated the prognostic value of the Ki-67 LI in glioblastoma patients from a populationbased cohort including MGMT promoter methylation and post-surgical treatment in the survival analysis. As an important methodological aspect we excluded the contribution of non-neoplastic cells to the Ki-67 LI and used digital quantification of full sections in order to minimize intra-observer bias and bias from evaluating small tumor areas, both being known pitfalls in biomarker studies 38,49,50 . Moreover, the use of full sections reflects daily diagnostics for possible glioblastomas, where most biomarkers are evaluated on full sections.
In accordance with previous results, we found that the Ki-67 LI increases with increasing WHO grade [19][20][21][22][23]51 . As expected, we also found that cells with microglial cell and macrophage morphology expressed Ki-67 and that the fraction of these cells increased with increasing WHO grade. This is in line with a study by Klein el al. 41 , who reported on a immunohistochemical double-labeling study with Ki-67 as proliferation-marker and Ki-M1P (CD68) as microglia marker in 40 astrocytomas WHO grade I-IV. The authors showed that proliferating microglia cells are present in all WHO grades and that the proliferative activity in microglia increased with increasing WHO grade. As the degree of microglial cell infiltration may differ between individual gliomas, we speculate that the Ki-67 contribution from microglial cells may have contributed to the divergent results reported for Ki-67 LI and outcome of glioma patients in previous studies. Therefore, we excluded the proliferative microglial cells, macrophages and other non-neoplastic cells in the tumor tissue in the present study, and specifically focused on the evaluation of proliferative tumor cells only. This is supposed to provide a more accurate estimate of the proliferative capacity of the tumor cells per se and thereby also the patient outcome.
In patients with glioblastomas, we found no correlation between Ki-67 LI and OS. This is in accordance with several other reports 3,19,31-34 , but in contrast to other studies 20,21,23,[27][28][29][30] . We divided our cohort of glioblastoma patients at the median Ki-67 LI (24%), which is similar to other groups who included the contribution of Ki-67 from non-tumour cells 21,23 . Interestingly; Moskowitz et al. 31 used the Ki-67 LI as a continuous parameter and Kuriyama et al. 34 used quartiles, but neither of these groups identified an association between the Ki-67 LI and OS. To address a potential bias due to different cutoff values, we performed an explorative optimal cutoff analysis in our cohort; however, we were unable to identify an association between Ki-67 LI and outcome for glioblastoma patients despite the use of different cutoffs.
A strength in our study is the inclusion of MGMT promoter methylation status and information on postsurgical treatment in the multivariate analyses, as both parameters are known to influence survival. Other groups investigating the prognostic value of Ki-67 LI in glioblastomas included only MGMT status 29 , adjuvant chemotherapy 27,30,33 or none of these important parameters in their analyses 3,20,21,23,28,31,32 . Yang et al. was the only group to 33 include both MGMT status and post-surgical treatment in the multivariate analyses. They reported on 254 glioblastoma patients, however; no information regarding IDH-status was reported. Further; although the Ki-67 immunohistochemical stainings were performed on whole slides-scoring was performed by pathologists using a 5-point scale. Despite these methodological differences compared to our study, the authors reported that Ki-76 LI was not associated with OS, a result similar to ours. It should be noticed that all studies reporting that Ki-67 LI is associated with OS, a finding that we could not validate in our population-based patient cohort, lack information of MGMT promoter methylation status and post-surgical treatment in the multivariate analyses. Therefore, the prognostic value of Ki-67 LI in glioblastomas is most likely very limited. www.nature.com/scientificreports/  www.nature.com/scientificreports/ Several groups, including our group 38 , have reported that the reproducibility of Ki-67 LI varies between studies. This may be due to inter-and intra-observer variability when detecting Ki-67 LI in tumors 35   www.nature.com/scientificreports/ that manual scoring has limited reproducibility in glioma patients 38 , and we thus decided to use digital quantification in the present study.

Conclusion
We found no association of Ki-67 LI with overall survival in IDH-wildtype glioblastomas, independent of the employed cutoff values and even when MGMT promoter methylation status and postsurgical treatment were included in the survival analysis. Due to the small number of IDH-mutated glioblastomas it is not possible to assess the prognostic value of Ki-67 in these patients. For the first time the prognostic value of Ki-67 was addressed using digital quantification of full sections excluding Ki-67 positive nuclei from non-neoplastic cells.
Over-interpretation of the prognostic potential of Ki-67 LI in the clinico-pathological setting should therefore be avoided.

Data availability
Data is available from the corresponding author upon reasonable request.