Automated Quantification of CD44, c-MET, EGFR, MTOR, and GLUT1 Expression in Head and Neck Squamous Cell Carcinoma: The Impact of p16 Status and Response to Chemoradiation

This study assessed automated quantification of CD44, c-MET, MTOR, EGFR, and GLUT1 protein expression in a tissue microarray of 109 Stage II-IV p16 positive and negative head and neck squamous cell carcinomas (HNSCC) treated with definitive chemoradiation. Immunohistochemistry-based


Introduction
Tissue microarrays (TMAs) can interrogate protein expression levels across many samples simultaneously and are therefore powerful tools to discover biomarkers with potential predictive or prognostic utility. Compared to the conventional immunohistochemistry (IHC) using individual tissue sections, TMAs facilitate high-throughput IHC analysis through their ability to combine many samples in a single slide, improving efficiency and eliminating possible variability in processing [1,2]. Though they are subject to limitations, including loss of areas of heterogeneity from larger tissue blocks and the potential for variations in fixation and processing, their ability to facilitate high throughput makes them ideal for screening large numbers of specimens.
However, the ability to efficiently query multiple samples is still limited by quantification. Manual quantification through visual inspection is labor-intensive and also prone to issues with reproducibility due to its inherent subjectivity, limitations which are reduced or eliminated with automated analysis [1]. Therefore, efforts to establish a method for automated quantification have been the subject of multiple studies [3][4][5]. Several commercial platforms for automated analysis have also been developed [6]. While there are minor variations in these methods and commercial platforms, in general, they have the common benefits of improved efficiency and improved inter-experimental normalization.
The prognostic utility of biomarkers has played a major role in recent changes to staging and treatment in head and neck cancer (HNSCC). The prognostic value of HPV status in oropharyngeal cancer has led to multiple trials investigating the possibility of therapy de-escalation in this setting [7,8]. Improved understanding of the prognostic value of additional biomarkers in this disease site may lead to further improvements in treatments and clinical outcomes.
Among previous research making use of automated quantification of biomarker expression levels in TMAs, there is limited literature on the validation of this technique using biomarkers with known clinical prognostic significance as defined by traditional, manual scoring of biomarker expression levels. Therefore, we analyzed the expression of several biomarkers in a TMA constructed from tissue samples from HNSCC patients treated with chemoradiation. We performed automated quantification of biomarkers as we had previously assessed these biomarkers for clinical prognostic significance using semi-quantitative visual inspection [9][10][11].

I Patient Selection and Generation of TMA
After approval from the Institutional Review Board, a tissue microarray was generated using tissue from pre-treatment biopsies of stage II-IV HNSCC patients as previously described [9]. Briefly, the representative areas of invasive carcinoma were marked by a pathologist, and two 1.0mm tumor tissue samples were taken from each block and inserted onto the TMA. All patients had archived tissue available at our institution and were treated with chemoradiation from 2003 through 2011. A total of 109 patients were included.

III Quantification of IHC protein expression
High-resolution digital images were created by scanning slides for each biomarker using an Aperio Scan-Scope AT Turbo (Leica, Buffalo Grove, Illinois) at ×20 magnification and organized using Image Scope eSlide Manager (Leica) and TMA images were imported into Definiens Tissue Studio software (Definiens Inc., Cambridge, MA). Definiens uses a proprietary de-arraying algorithm consisting of three phases: TMA core segmentation, gridding, and mapping, which overcomes processing and sectioning artifacts and identifies and isolates each core. Within each core, a software algorithm was trained to automatically detect regions of interest (ROI), which were defined as viable tumor, non-tumor, white space, and artifacts (histological deposits, tissue folds etc.). Within the viable tumor ROI of each core, individual nuclei were detected based on hematoxylin staining. Cells were simulated around each nucleus based on the presence of a membrane. Algorithm IHC thresholds were then established to classify each cell based on its membrane, cytoplasmic, or nuclear stain intensity as negative, low, medium, or high. Once completed, the algorithm was applied objectively to each core. Data generated included the number of negative, low, medium, or high classified cells in each core. The histologic score (H-score) for each core was calculated as follows: The H-score is a combined parameter that corresponds to the percent positivity and intensity of expression within areas of tumor and has a range of 0-300. For p16 staining, positive expression was defined as diffuse 3+ staining (all 3+ cases demonstrated diffuse staining) or 2+ staining in 50% or more of tumor cells.

IV Statistics
H-scores were stratified by high and low expression using cut-points determined with ROC analysis. Biomarker expression levels measured using the H-score generally showed a normal distribution and were correlated with one another using the Pearson rank correlation. Biomarker expression levels were correlated with p16 status using Spearman's rank correlation. Effects of biomarker and p16 status on locoregional failure (LRF), disease-free survival (DFS), and overall survival (OS) were analyzed using Kaplan Meier and Cox proportional hazard modeling for uni-and multivariable analyses (UVA, MVA), respectively.

I Baseline Patient Characteristics
The baseline characteristics of the patients are shown in (Table 1). Of the 109 patients included in this analysis, the majority had oropharynx primaries (62%), while larynx (28%), hypopharynx (6%) and oral cavity (4%) were also represented. The majority were clinical stage IVA (64%), according to AJCC version 7 staging. The median follow-up was 6.19 years, with a maximum of 13.46 years.

II Biomarker Analysis
Biomarker expression levels were correlated with one another and with p16 status. Using the H-score as a continuous variable, there was a significant negative correlation between CD44 expression and p16 status and a significant positive correlation between CD44 and both MTOR and GLUT1, c-MET and MTOR, and MTOR and GLUT1 (Table 2). When patients were stratified by p16 positive or negative status, the significant positive correlation between CD44 expression and MTOR remained for both the p16 positive and negative subsets, while correlations between CD44 and GLUT1 and c-MET and MTOR were seen in the p16-negative subset only. A significant correlation between MTOR and GLUT was seen overall and for the p16 positive subset. There were no significant correlations with EGFR. Figure 1 presents the IHC data as a heat map sorted by p16 status and CD44 expression, clearly showing the correlation between the two biomarkers. Figure 1 presents the IHC data as a hierarchical cluster analysis for all the biomarkers and as a loading plot from principal component analysis (PCA) for p16-and p16+ tumors. The clustering and color map indicates that there are discernible associations between the biomarkers without a strong phenotype that emerges encompassing several biomarkers. The loading plot clearly highlights the different contributions of these biomarkers in p16-and p16+ patients. When correlations were assessed by disease site (oropharynx, larynx, or other), a significant correlation between CD44 and MTOR remained for both oropharynx and larynx; the correlations between CD44 and GLUT1 remained for oropharynx primaries only; the correlations between c-MET and MTOR and MTOR and GLUT remained for other primary sites only. There were significant correlations between EGFR and MTOR for oropharynx primaries and between c-MET and GLUT1 for other primary sites (Supplementary Table 1).

III Biomarkers and Clinical Outcomes
To examine the effects of biomarker status on clinical outcomes, ROC analysis was performed to stratify H-scores by high and low expression.
Representative images of IHC staining for CD44, with corresponding Hscores stratified by the cut-point of 163, are shown in (Figure 2). The cut-off values for each biomarker as determined by ROC analysis comprised of an H-score greater than or equal to 163 for CD44, greater than or equal to 187 for c-MET, greater than or equal to 86 for mTOR, greater than or equal to 32 for EGFR, and greater than or equal to 202 for GLUT1. Fifty-one percent of the patients were p16 positive. The ROC cut-off values do not necessarily signify high or low expression, as was the case in our previous publications; it is the optimal value for statistical significance with regard to overall survival.

Figure 2:
Examples of cores stained for CD44 displaying different staining characteristics. The numbers below each represent the H-score using definiens tissue studio analysis. The prognostic impact of each biomarker expression levels on clinical outcomes are also presented in (Table 3) Due to the fact that p16 positive HNSCC is known to have a more favorable clinical course and response to treatment as compared to p16 negative tumors, we also assessed the influence of biomarker expression levels on our clinical endpoints as a function of p16 status (Figure 3) [12]. With the exception of CD44, which was significantly predictive for LRC and DFS in the p16 negative subset and for OS in the p16 positive subset, and GLUT1, which was significantly predictive for OS in the p16 positive subset, the majority of biomarkers were no longer significantly predictive for LRC, DFS, or OS. This is possibly explained by the reduced number of patients in the p16 positive and negative subsets.

Discussion
In this study, we have developed a quantitative image analysis approach to analyze known biomarkers in HNSCC that allowed us to more accurately study their correlation with one another and investigate their association with clinical outcomes in p16+ and p16-HNSCCs. We had previously analyzed these biomarkers using visual analysis carried out by two or three independent observers [9][10][11]13]. The results of this present analysis generally correspond with our previous findings, although differences did arise. The differences were attributed to the different methods in scoring and in determining the cut-points for analysis. Using visual scoring methods, the data was generally classified as high or low expression based on semi-quantitative scoring and grouping of these scores. The grouping of scores varied depending on the marker, i.e., in some cases it was 1 and 2 versus 3 whilst in others it 0 and 1 versus 2 and 3. The grouping then determined the distribution of data for correlation analysis (using a Chi-squared analysis for ordinal data) and for analysis of clinical outcome parameters. The methodology in this study allowed a more robust correlation between biomarker expression as the values basically became a continuous variable.
An important point to note is that the use of the terms high or low expression in this current study is not relevant. The cut-offs were determined statistically, providing the optimal segregation of the data with reference to overall survival. In some cases, the cut-off was close to the median value (e.g., CMET 187 by ROC versus a median of 170), while in other cases it was not (e.g., CD44 163 by ROC versus a median of 103). Supplementary Table 2 presents the clinical outcome data using the median value for each biomarker and highlights the dependence of clinical significance on the choice of cut-off. However, using the more rigorous cut-offs for the biomarkers based on ROC analysis rather than a semi-quantitative classification, we were able to confirm the prognostic significance of high expression levels of CD44 and CMET for worse LRC, DFS and OS, as well as high EGFR or GLUT1 expression predicting for worse DFS and OS but not LRC.
The correlation between the biomarkers and p16 status yielded new data. As p16 serves as a surrogate for HPV status, this added to knowledge concerning the difference between HPV-driven tumors compared to HPV negative tumors. These two HNSCC subtypes have distinct etiologies, molecular properties, clinical features, and prognostic outcomes [14][15][16]. Differences in the molecular landscapes of HPVdriven and HPV-negative HNSCC occur genome-wide and encompass changes in genomic, epigenetic, and transcriptional landscapes [17,18]. The mutational load and spectrum differ between these two subtypes, with HPV+ cancers characterized by a lower mutational burden, with the most common mutations in PIK3CA, IGFR1, EPHA2, DDR2, KEAP1, NOTCH1, FGFR2, and FGFR3 [19,20]. Mutations TP53, FADD, CCDN1, CDKN2A, FAT1, NOTCH1, MYC, and PIK3CA dominate in HPV-tumors [19,20]. Interestingly, there was no difference in expression levels of MTOR, CMET or EGFR between p16+ and p16tumors.
CD44 was also strongly correlated with MTOR (Table 2), and this correlation was significant irrespective of p16 status. The link between CD44 and MTOR has been reported in several recent studies in the context of cancer stem cells (CSCs) [23][24][25][26][27][28]. Rapamycin inhibition of the MTOR signaling pathway downregulated cancer stem cell properties of nasopharyngeal carcinoma, and anti-CD44 antibodies inhibit both mTORC1 and mTORC2 in acute myeloid leukemia [23,24]. A molecular link was discovered between the PI3K/AKT/MTOR signaling pathway and CSCs, characterized by high CD44 and aldehyde dehydrogenase (ALDH) activity, in patient-derived xenografts from patients with both HPV positive and negative HNSCC [29].
Given the evidence that CD44 is consistently associated with a poor prognosis in HNSCC and its established role in treatment-resistant cancer stem cells in HNSCC, it is somewhat surprising that therapeutic strategies targeting this molecule are lacking in solid tumors [30][31][32][33][34][35][36][37].
Multiple methods for automated quantification of biomarker expression within TMAs have been established, with the common objective to reduce the time-consuming nature and potential biases of semiquantification by visual inspection. While some have been developed for general quantification, others have been created for more specific purposes, such as measuring changes in the subcellular location of biomarkers, where the ability to eliminate the bias inherent to visual inspection may be of particular importance [3][4][5]. While automated quantification has been used previously to correlate biomarker expression levels with clinical prognostic factors in human tumor specimens, to our knowledge, this is the first study that used automated analysis to interrogate the potential prognostic significance of biomarkers using a TMA and biomarkers which had previously been shown to have significant prognostic value using semi-quantitative visual inspection [42]. Specifically, we used Definiens Tissue Studio, which uses feature-based or contextual compartmentalization to train the software to recognize areas of interest for analysis and which is one of several commercially available image analysis software with the potential to allow high-throughput analysis in IHC [6].
In conclusion, we performed automated quantification of CD44, c-MET, MTOR, EGFR, and GLUT1 expression in HNSCC samples from patients who underwent definitive chemoradiation. Our analysis shows that higher expression of CD44, c-MET, EGFR, and GLUT1 were associated with inferior DFS and OS, while MTOR expression levels did not correlate with any clinical endpoints. The trend for OS and DFS remained significant for CD44 and EGFR but not for c-MET or GLUT1 on MVA. This confirms findings reported in our previously published studies using this same TMA and highlights the importance of CD44 expression in both p16 positive and negative subtypes of HNSCC.

Conflicts of Interest
None.

Acknowledgments
Funding was provided by the Beaumont Health System Department of Radiation Oncology. The authors thank Barbara Pruetz and the Beaumont BioBank for specimen cutting and processing.

Author Contributions
A