With the promise of personalized medicine and effective targeted molecular therapeutics on the horizon, developing translational IVDs that can accurately select patients for specific therapies or treatment regimens has become more critical. Two of the best known examples in oncology are the examination of estrogen receptors and HER2 expression using immunohistochemistry (IHC) to predict the response to tamoxifen and Herceptin, respectively (see Figure 1). As a measurement of expression in fixed tissues, IHC preserves tissue architecture and cellular morphology, thus allowing for localization of protein expression. Compared to other gene and protein expression methodologies (i.e., ELISA, RTQ-PCR, gene chip technology), IHC also has the advantage that only a small amount of tissue (one 5-µm section) is needed for testing. As a result, IHC has become a mainstay in clinical pathology laboratories.

However, traditional IHC has disadvantages. In addition to those sources of variation that plague all tissue-based methods (including time to fixation), scoring traditional IHC is qualitative (non-quantitative) in nature and can be subjective. With IHC, pathologists assess how much protein is present through a qualitative, visual determination of the intensity of the brown stain observed in the tissue (see Figure 1). In general, scoring of the tissue is done based on four intensity levels: IHC 0 (negative), IHC +1 (low), IHC +2 (medium), and IHC +3 (high; see Figure 1).

Figure 1. Comparison of traditional IHC with Aqua analysis. Top panel: Images of traditional IHC (brown stain) for indicated markers, ER and HER2, and representative typical data output. Bottom panel: Images of Aqua analysis for indicated markers, ER and HER2, and representative typical data output (click image to enlarge).

As such, scoring reproducibility and thus pathologist-to-pathologist standardization are poor. This problem is best highlighted by recent studies showing poor concordance between local and central laboratories for assessing HER2.^1,2 In addition, traditional chromagen-based IHC lacks the dynamic range that other methods (e.g., immunofluorescence) provide.³ However, despite such disadvantages, traditional IHC remains the standard of care for tissue-based diagnostics in the clinical pathology laboratory.

With the advent of personalized medicine and targeted therapies, AQUA (Automated Quantitative Analysis) technology by HistoRx Inc. (New Haven, CT) provides a solution for quantifying and standardizing biomarker expression analysis that can be adapted in clinical laboratories. AQUA technology was invented to address the disadvantages in IHC-based testing.⁴ AQUA technology is a fluorescence immunohistochemistry-based method that provides objective and continuous protein expression scores for tissue by using automated fluorescence microscopy and advanced image analysis algorithms (see Figure 1). It uses molecular identification of compartments to quantify biomarker expression as a function of pixel intensity in specific tissues or sub-cellular compartments. AQUA scores have been previously demonstrated to be directly proportional to molecules per unit area or protein concentration, and the methodology can be broadly applied to biomarker assessment and clinical characterization.^5-9

Due to the complexity of disease, single biomarker approaches will often fall short of providing a robust level of prediction and prognosis, and therefore multiparametric diagnostic models will be required. In the case of in situ measurements of expression, the limitations of current measurements make them difficult or impossible for building multivariate models. This article discusses using AQUA technology for the potential development of robust clinical diagnostics, specifically examining the development of univariate and multivariate prognostic models for predicting outcomes in glioblastoma multiforme (GBM).

How the Technology Works
Compared to other imaging or analysis systems on the market, AQUA technology is encompassed in a complete platform solution that can take assay applications using fluorescent IHC staining through image acquisition and scoring by AQUA analysis. AQUA analysis technology methodology is based on molecular-based compartmentalization rather than morphometric or adaptive training algorithms. AQUA protocol consists of three main components: tissue staining, image acquisition, and image analysis (compartmentalization and AQUA scoring). Although AQUA technology employs multiple fluorescent stains, its staining methodology is similar to traditional IHC and can be adapted to a pathology laboratoryís workflow.

In a typical AQUA experiment measuring target gene expression in epithelial cancer, 4'-6-Diamidino-2-phenylindole is used for the nuclear mask (UV channel), anti-cytokeratin antibodies are used to identify and differentiate epithelium from stroma (tumor mask) and establish the cytoplasmic mask (Cy2 and/or Cy3 channel), and an antibody directed against the target is used to visualize the target gene protein expression of interest (Cy5 and/or Cy7 channel). However, an AQUA experiment is not limited to these parameters. Depending on the tissue/compartment mask desired, the experiment can employ a number of different masking markers. For example, in the case of neuronal cells for the study of GBM, glial fibrillary acidic protein can be used to define a mask region.

Once the tissues have been stained appropriately, image acquisition can proceed using the PM2000 system by HistoRx. The PM2000 platform is an automated epi-fluorescence microscopy system with accompanying AQUAsition image acquisition software and is designed for high-resolution automated image acquisition for both tissue microarrays and whole tissue sections.

The foundation of the AQUA score is the pixel-based locale assignment for compartmentalization of expression (PLACE) image analysis algorithm.⁴ Tissue is a complex mixture of various tissue components (i.e., epithelium, stroma, blood vessels) and sub-cellular components (i.e., cytoplasm, nuclei, membrane). PLACE enables differential localization of image pixel intensities associated with target gene expression in these different components or masks. In the traditional AQUA software, masks are defined through a series of pixel intensity thresholding steps, producing binary images followed by image exclusion steps (see Figure 2). In the next-generation software, compartmentalization will be performed in a completely unsupervised fashion using simple mathematical algorithms.¹⁰

Figure 2. Aqua image analysis. Schematic representation of the Aqua analysis process for a typical experiment looking at biomarker expression in an epithelial cancer. (1) Tumor mask generation through pixel intensity thresholding (binary gating) and spatial image analysis procedures. (2a and b) Generation of cytoplasmic and nuclear-specific pixel mask through binary gating. (3) Combining cytoplasmic and nuclear-specific pixel masks by 100% mutual exclusion. (4) Identification of coincidental target pixels with compartment pixels (for visualization only) (click image to enlarge).

AQUA analysis is outlined in Figure 2, which provides a step-by-step illustration of how AQUA image analysis is accomplished given a typical experiment with an epithelial tumor using the target estrogen receptor (ER) as an example. Again, this AQUA analysis can be applied to different permutations depending on the application.
AQUA analysis begins by generating a tumor mask through pixel intensity thresholding (binary gating) and spatial image analysis procedures (i.e., fill holes; see Figures 2a and c). Using identified tumor mask pixels as a template, the same type of pixel intensity thresholding is used to identify nuclear and/or cytoplasmic pixels (see Figures 2d and e). Through a process of 100% mutual exclusion, these images are combined to provide pixel assignments for compartmentalization (see Figure 2f). An AQUA score is generated, which represents a concentration of pixels or the summation of the target pixel intensity/compartment area. For purposes of visualization only, step 4 in Figure 2 shows an overlay of the target image (colorized in red) with the compartment image demonstrating colocalization of ER with nuclei (magenta pixels from the overlay of the target in red with nuclei in blue; see Figure 2h).

Standardization
Due to the subjective and qualitative nature of traditional IHC, standardization across laboratories and observers is problematic. Because AQUA technology is objective and strictly quantitative, it can be standardized across instruments, laboratories, and operators. HistoRx has developed instrument calibration methodologies that enable a captured signal to be normalized across multiple instruments.¹¹ This standardization is accomplished through a combination of light source and intrinsic machine calibration methods. These methodologies are automated and use stable non-biological materials, and primary calibration is done in real time with data acquisition.

Software algorithmic methods have been developed that predominantly remove operator decisions from the image acquisition (i.e., auto-exposure) and scoring process (i.e., unsupervised pixel-based clustering).¹⁰ These methodologies can also be applied to any target of interest to achieve CVs of less than 5%, which are comparable with other quantitative immunoassays (i.e., ELISA, flow cytometry). In addition, these standardization methods have been applied to numerous markers including HER2, mTOR, EGFR, ER, and PTEN.^10,11

Of importance for developing clinical diagnostics is the ability to reproducibly classify patients. For example, current ASCO-CAP guidelines suggest that laboratories should achieve a 95% positive/negative concordance for current HER2 assay methodologies.¹² A recent study showed that for HER2 IHC-based scoring, concordance among laboratories ranged from 54% to 85%, falling short of these guidelines.¹³ Positive/negative concordance for AQUA scoring across instruments, operators, and staining days was examined. The overall concordance ranged from 94.5% to 99.3%.¹¹

These analyses include all cases including those that would be considered equivocal by standard IHC methods and would have been reflexed to a different methodology (i.e., fluorescent in situ hybridization). Thus AQUA analysis provides a platform by which a standardized protein expression score can be obtained with a single method for more samples while still allowing for and maintaining typical pathology laboratory workflow.

Quantification
The AQUA score represents a biomarkerís average intensity per unit area. As such, an AQUA score is equivalent to the concentration of protein in a tissue section. This has been demonstrated by using cell line controls whereby matched cell lysates and cell pellets are created for assessment of protein expression with an enzyme-linked immunosorbent assay (ELISA) and for AQUA score assessment, respectively. Although AQUA scores do correlate with traditional IHC scoring, significant overlap is observed (e.g., HER2).¹¹ While this overlap is primarily due to the increased resolution provided by the continuous, quantitative nature of AQUA scores versus traditional IHC, it is also due to how the scoring is accomplished.
With traditional IHC, specifically for HER2, only a small percentage of the tissue (10% for HER2) must appear positive (i.e., IHC +3) in order to be scored positive.2 However, an AQUA score represents average expression over the entire tissue section. This expression translates into enhanced resolution of patient survival, compared to traditional IHC in which low-level expression is differentiated from mid- and high-level. Furthermore, AQUA analysis has demonstrated that an increased percentage of a tumor should be analyzed for accurate assessment of marker expression and that it is not necessarily the highest regions of tissue expression that are predictive.¹⁴

To demonstrate AQUA analysisí quantitative ability, PTEN expression in GBM, was examined. PTEN is a tumor suppressor gene whose expression is frequently lost in numerous cancers. GBM is an aggressive form of brain cancer with median survival times ranging only from 11 to 13 months. At this time, no clinical diagnostics are available to assess prognosis of such patients or their responses to treatment. Although mutations in PTEN have been shown to predict outcomes in GBM, assessing PTEN protein expression has not been demonstrated with respect to prognosis.¹⁵ The difficulty in measuring PTEN protein expression with respect to survival is related to the fact that PTEN expression is inherently low in GBM relative to normal brain tissue. Reliably differentiating expression populations further in the low expression subset of GBM patients is problematic using traditional IHC (see Figure 3a). Measuring PTEN, even in a univariate setting, with AQUA analysis offers potential prognostic utility.

Figure 3. Ability of Aqua analysis to assess and differentiate low level PTEN expressing glioblastoma tumors. (A) Box plot of Aqua scores comparing normal brain to tumor tissue; significant difference in expression by one-way ANOVA (p<0.001). (B) KM survival analysis of low level expressing tumors with indicated cutoff points (inset: frequency histograms) showing a significant (p=0.043) increase in three-year disease specific survival for patients expressing higher levels of PTEN, or low versus very low (click image to enlarge).

In this study comparing normal tissue expression levels to tumor expression for PTEN with AQUA analysis, tumor expression was significantly lower compared to normal levels (see Figure 3a). Nonetheless, differences in low-level expression are resolved with respect to patient survival in which patients who have higher level PTEN expression, or low versus very low levels, have increased survival rates compared to lower level expressing patients (a 25.5% decrease from 45.2%; 19.7% 3-year, disease-specific survival; p=0.043; see Figure 3b). Patients with higher PTEN levels had an 8.3-month improvement in median survival, which is substantial for GBM. Although these findings need to be validated, the data demonstrate the quantitative ability of AQUA analysis to differentiate low-level PTEN expression, which could lead to a prognostic clinical assay for managing patients with GBM.

Localization
Another aspect of AQUA technology is the ability to localize or compartmentalize protein expression and thus quantify protein expression in different subcellular compartments (see Figure 2). Many proteins reside in different cellular compartments depending on their specific function or activation state. For example, thymidylate synthase (TS) resides in cytoplasm for its DNA synthesis role and in nucleus for its translational inhibition function. Another example is cellular signaling molecules such as AKT, STAT, and ERK that move from cytoplasm to nucleus upon their activation.

Although activation of these proteins is mediated by phosphorylation, detection of phosphorylated proteins is problematic in tissue, thus the ability to assess activation by localization is valuable. As discussed later in this article, differential quantitative localization of proteins can be a valuable tool for determining pathway activation in tissue samples. Several studies have been published examining differential subcellular localizations and their association with outcomes by AQUA analysis, including a study examining the effects of TS subcellular localization on colon cancer outcomes.¹⁶

Multiparametric Analysis
Multiparametric-type analyses, or the ability to assess multiple genes and/or proteins simultaneously, have tremendous potential in research and clinical diagnostics. For more than a decade, multiparametric analyses have been applied in gene microarray technology (i.e., mRNA profiling) in which thousands of genes can be assessed concurrently for mRNA expression; the patterns in these gene expression profiles are then examined by statistical analyses such as hierarchical clustering. Because AQUA analysis produces continuous expression scores, multiparametric analyses can be extended to protein expression in tissue.

Figure 4. Unsupervised hierarchical clustering and associated survival outcome for AKT pathway markers. (A) Heat map (green: low level expression; red: high level expression) showing results of unsupervised hierarchical clustering of indicated AKT pathway markers, mTOR, PTEN, phospho-mTOR, and phospho-AKT. The heat map is divided into two main groups: low group and high group. (B) One-year and (C) three-year disease-specific survival in Kaplan-Meier analysis with indicated log-rank P-values for indicated groups from (A) (click image to enlarge).

For example, the AKT pathway in GBM was examined using multiparametric analyses. The AKT pathway is involved in a number of cellular processes including growth, apoptosis, and metabolism, and is being studied as a target for drug treatments, specifically in GBM. Four proteins in the AKT pathway (PTEN, mTOR, phospho-mTOR, phospho-AKT) were assessed with AQUA analysis in a cohort of GBM (n=110). As a first assessment of the population with respect to these four markers, unsupervised hierarchical clustering was performed (see Figure 4). Hierarchical clustering is a valuable tool for assessing multiple parameters (proteins/patients). Two distinct populations of patients were observed: those that in general expressed low levels of all four markers, and those that expressed high levels of all four markers. Although survival comparisons at one and three years were not statistically significant, this type of analysis provides valuable insight into how these proteins are being co-expressed in the population.

Although hierarchical clustering provides valuable information with regard to protein expression patterns in the population, how the expression of each protein for each patient interacts with one another and event-specific outcomes in a continuous multivariate setting provides the most clinical value. Cox proportional hazards modeling is an event-specific linear regression model that enables the development of equations in which prospective data can be entered to determine outcomes.

The interaction of the four markers as continuous variables with survival and with each other was assessed by Cox proportional hazards modeling (see Table I). Although the overall model with all four proteins was significant, only two proteins (PTEN, pAKT) contributed significantly to the model. The optimal model with only these two proteins is highly significant (p=0.009). The data represent a novel finding that continuous expression scores for PTEN and pAKT, as determined by AQUA analysis, can be used to build a prognostic model in GBM.

Conclusion
With the results of the study above, a Cox proportional hazards model equation could be conceived that would be extended to calculate overall risk, thus putting patients on a continuum much like the model established by Genomic Health with its OncotypeDx diagnostic for early-stage breast cancer patients (see Figure 5).¹⁷ The power of this type of analysis comes from the use of AQUA data as a continuous variable in the actual survival analysis, rather than determining a cutoff point. Providing oncologists with prognostic information on a sliding scale could be foreseen, such that when combined with other clinical and pathological factors, this information would provide a more complete prognostic outlook.

Figure 5. Proposed model for prognostic prediction in GBM. (A) A Cox proportional hazards model equation to calculate overall risk was formulated using the optimal model coefficients. (B) Distribution of risk based on indicated model. (C) Schematic representation of possible model diagnostic for determining patient risk based on Aqua scoring of a PTEN/pAKT diagnostic test (click image to enlarge).

As these types of analyses using AQUA technology can be applied to prognostic models, they ultimately can be applied to multiparametric models to predict responses to specific therapies. For example, targeted molecular therapeutics are attempting to inhibit specific cellular pathways in the tumor cell. If these therapies are to succeed, the patients who will have the greatest chance to respond to them must be identified. Responses to therapies will predominantly be determined by whether patients have specific pathways activated in their tumors.^18,19

Therefore, assessing pathway activation will be critical for determining which patients will respond to given treatments. Because assessing pathway activation will most likely involve quantification and association of multiple protein biomarkers and measurements of protein translocations, AQUA analysis is well suited for this task. An example of such a diagnostic would be the assessment of EGFR pathway activation for predicting responses to EGFR-targeted therapeutics by not only quantification of EGFR protein expression but also localization, in conjunction with quantitative association of downstream effector molecules such as ERK, STAT, and AKT.

References
1. EA Perez, et al., "HER2 Testing by Local, Central, and Reference Laboratories in Specimens from the North Central Cancer Treatment Group N9831 Intergroup Adjuvant Trial," Journal of  Clinical Oncology 24 (2006): 3032-3038.

2. S Paik, et al., "Real-World Performance of HER2 Testing-National Surgical Adjuvant Breast and Bowel Project Experience," Journal of the National Cancer Institute 94 (2002): 852-854.

3. DL Rimm, "What Brown Cannot Do for You," Nature Biotechnology 24 (2006): 914-916.

4. RL Camp, GG Chung, and DL Rimm, "Automated Subcellular Localization and Quantification of Protein Expression in Tissue Microarrays," Nature Medicine 8 (2002): 1323-1327.

5. A McCabe, et al., "Automated Quantitative Analysis (AQUA) of In Situ Protein Expression, Antibody Concentration, and Prognosis," Journal of the National Cancer Instute 97 (2005): 1808-1815.

6. Z Zheng, et al., "DNA Synthesis and Repair Genes RRM1 and ERCC1 in Lung Cancer," New England Journal of Medicine 356 (2007): 800-808.

7. RL Camp, et al., "Quantitative Analysis of Breast Cancer Tissue Microarrays Shows that Both High and Normal Levels of HER2 Expression are Associated with Poor Outcome," Cancer Research 63 (2003): 1445-1448.

8. M Dolled-Filhart, et al., "Classification of Breast Cancer Using Genetic Algorithms and Tissue Microarrays," Clinical Cancer Research 12 (2006): 6459-6468.

9. JM Giltnane, et al., "Quantitative Measurement of Epidermal Growth Factor Receptor is a Negative Predictive Factor for Tamoxifen Response in Hormone Receptor Positive Premenopausal Breast Cancer," Journal of Clinical Oncology 25 (2007): 3007-3014.

10. MD Gustavson, et al., "Development of an Unsupervised Pixel-based Clustering Algorithm for Compartmentalization of Immunohistochemical Expression Using Automated Quantitative Analysis," Applied Immunohistochemistry and Molecular Morphology (2009).

11. MD Gustavson, et al., "Standardization of HER2 Immunohistochemistry in Breast Cancer by Automated Quantitative Analysis (AQUA)," Archives of Patholgy and Laboratory Medicine (2009).

12. AC Wolff, et al., "American Society of Clinical Oncology/College of American Pathologists Guideline Recommendations for Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer," Archives of Patholgy and Laboratory Medicine 131 (2007): 18.

13. O Hameed, et al., "Using a Higher Cutoff for the Percentage of HER2+ Cells Decreases Interobserver Variability in the Interpretation of HER2 Immunohistochemical Analysis," American Journal of Clinical Pathology 130 (2008): 425-427.

14. CB Moeder, et al., "Quantitative Justification of the Change from 10% to 30% for Human Epidermal Growth factor Receptor 2 Scoring in the American Society of Clinical Oncology/College of American Pathologists Guidelines: Tumor Heterogeneity in Breast Cancer and its Implications for Tissue Microarray Based Assessment of Outcome," Journal of  Clinical Oncology 25 (2007): 5418-5425.

15. JA Kraus, et al., "Molecular Analysis of the PTEN, TP53, and CDKN2A Tumor Suppressor Genes in Long-Term Survivors of Glioblastoma Multiform," Journal of Neurooncology 48  (2000): 89-94.

16. MD Gustavson, et al., "AQUA Analysis of Thymidylate Synthase Reveals Localization to be a Key Prognostic Biomarker in 2 Large Cohorts of Colorectal Carcinoma," Archives of Patholgy and Laboratory Medicine 132 (2008): 1746-1752.

17. S Paik, "Development and Clinical Utility of a 21-Gene Recurrence Score Prognostic Assay in Patients with Early Breast Cancer Treated with Tamoxifen," Oncologist 12 (2007): 631-635.

18. IB Weinstein and A Joe, "Oncogene Addiction," Cancer Research 68 (2008): 3077-3080.

19. IB Weinstein IB and AK Joe, "Mechanisms of Disease: Oncogene Addiction-a Rationale for Molecular Targeting in Cancer Therapy," National Clinical Practice Oncology 3 (2006): 448-457.

Mark Gustavson is senior research scientist, diagnostics at HistoRx Inc. (New Haven, CT). He can be reached at mgustavson@historx.com.

Wendy Davis is vice president, diagnostics portfolio management at HistoRx Inc. (New Haven, CT). She can be reached at wdavis@historx.com.

Jason Christiansen is senior director, operations at HistoRx Inc. (New Haven, CT). He can be reached at jchristiansen@historx.com.