Analytic Variability in Immunohistochemistry Biomarker Studies Free

Immunohistochemistry (IHC) is frequently used in clinical diagnosis and the classification of neoplasms (1). Despite the widespread use of IHC in routine diagnostic practice, no universally accepted standardization guidelines have been developed. IHC is subject to variable preanalytic, analytic, and postanalytic factors that decrease its reproducibility, including tissue preparation and fixation, diverse reagents, different staining methods, scoring systems, and the definition of a “positive” result (2). Considerable variability has been attributed to the selection and insufficient validation of the primary antibodies, poor quality controls, and lack of rigorous assay optimization (1, 3).

Epidermal growth factor receptors (EGFR) are among the most studied cancer biomarkers because of their oncogenic activity in diverse tumor types (4). EGFR in particular was the first receptor to be proposed as a target for cancer therapy, and several EGFR-targeted therapies are now available in the clinic (5). EGFR IHC results highly depend on the type of antibody, protocols, scoring system, and cutoffs used (6), rendering the prognostic role of EGFR protein expression as assessed by IHC highly controversial (7). This lack of reproducibility may be responsible for its failure as a companion diagnostic test, and studies done to date, largely nonstandardized, have shown that detection of high EGFR expression by IHC does not reliably predict the clinical outcome of EGFR-targeted treatment (8). Therefore, IHC is not routinely used to determine EGFR status.

In contrast, IHC on formalin-fixed paraffin-embedded (FFPE) tissue is the current standard for evaluating estrogen receptor (ER) status in breast cancer, and IHC measurements directly affect the management of breast cancer patients (9). Standardized and validated IHC assays for ER include those that use monoclonal antibody (mAb) clones 1D5 and SP1 (10, 11), and efforts to establish best practice guidelines for ER IHC testing are ongoing (12). Nevertheless, high false negative rates of tissue samples evaluated by IHC have been reported when laboratory quality control was inadequate and proficiency testing requirements were not met (3); this highlights the critical need for the development of standardized IHC assays.

Here we analyze, validate, and compare ER, EGFR, and human epidermal growth factor receptor (HER) 3 protein expression detected by commonly used ER, EGFR, and HER3 antibodies in breast and lung cancer. Ultimately, we show the effects of variable EGFR antibody selection and validation on outcome in a breast cancer cohort of 642 patients.

FFPE primary breast cancer specimens from 642 patients that underwent surgery at Yale-New Haven Hospital (New Haven, CT) between 1962 and 1983 were obtained from the archives of the Pathology Department of Yale University (New Haven, CT). A smaller cohort was retrospectively collected from 42 non–small cell lung cancer patients from Yale-New Haven Hospital between January 1995 and May 2003. The demographics, tumor characteristics, and routine IHC scoring for ER, progesterone receptor, and HER2 in the breast cancer cohort are shown in Supplementary Table S1. The study was approved by the Institutional Review Board.

Tissue specimens were prepared in a tissue microarray format containing one 0.6-mm representative FFPE tumor core for each primary tumor. FFPE cell line pellets were used as controls: A431, H1666, H1355, MCF-7, MB468, BT474, CHO, BaF3, MB436, and HER3-transfected BaF3 cells were purchased from the American Type Culture Collection (Manassas, VA) or donated by other laboratories. FFPE EGFR–stable transfected CHO cell pellets were generated as described in detail elsewhere (13).

Equivalent amounts of protein (25 μg) were resolved by SDS-PAGE in 4% to 12% bis-TRIS gels (150V for 1h) and transferred at 45V for 2 h to a nitrocellulose membrane. Immunoblots were probed with primary antibodies diluted 1:1,000, except for ER 1D5 and SP1 clones, which were diluted 1:500, and ER60c, which was diluted 1:2,000, followed by anti-rabbit or anti-mouse horseradish peroxidase–conjugated secondary antibodies (Santa Cruz Biotechnology, Santa Cruz, CA) diluted 1:4,000 and detected with the use of enhanced chemiluminescence (SuperSignal West Pico, Thermo Scientific, Rockford, IL). β-Tubulin (rabbit polyclonal, Cell Signaling Technology, Danvers, MA) was used to visualize total protein loading. Band intensity was quantified with the use of the ImageJ software (14). The scores generated were normalized to the maximum β-tubulin signal.

Heat-induced antigen retrieval with the use of sodium citrate (pH 6) resulted in reproducible staining when tissue microarray sections were incubated with mAb clones H11 and 15F8, as well as polyclonal 2232; however, heat-induced antigen retrieval did not yield specific EGFR staining with mAb clones 31G7 and 2-18C9 in breast cancer tissue, nor in EGFR–stable transfected CHO cells or A431 cells (data not shown). Digestion with proteinase K was done instead at room temperature and 37°C at four time points (5, 10, 15, and 45 min) to identify the optimal temperature and incubation period. Incubation for >5 min at room temperature or any incubation at 37°C resulted in the destruction of the tissue architecture and appearance of an edge artifact; moreover, cytokeratin multiplexed on the same slides as a positive control was similarly affected (Fig. 1). Based on these results, we selected pretreatment with proteinase K for 5 min at room temperature as the optimal antigen retrieval method for staining tissue microarray sections with mAb clones 31G7 and 2-18C9.

Sections (5 μm thick) were deparaffinized, rehydrated, and antigen retrieved by pressure cooking in 10 mmol/L citrate (pH 6) or 10 mmol/L Tris/1 mmol/L EDTA buffer (pH 9) for all primary antibodies except anti-EGFR clones 2-18C9 (pharmDx; Dako, Carpinteria, CA) and 31G7 (Zymed/Invitrogen, Carlsbad, CA), for which proteinase K (Dako) digestion was done instead. Slides were incubated with a cocktail of the primary antibody and a mouse monoclonal cytokeratin (clone AE1/AE3, M3515; Dako) or a polyclonal rabbit cytokeratin antibody (Z0622; Dako) overnight at 4°C for all antibodies except ER clone 1D5 and EGFR clone 2-18C9, which were incubated for 1 h at room temperature (Table 1). This was followed by 1-h incubation with Alexa 546–conjugated goat anti-mouse (A11003; Molecular Probes, Eugene, OR) or Alexa 546–conjugated goat anti-rabbit secondary antibody (A11010; Molecular Probes) diluted 1:100 in rabbit (K4003; Dako) or mouse EnVision (K4001; Dako). Cyanine 5 tyramide (FP1117; Perkin-Elmer, Boston, MA) diluted 1:50 was used for target detection. ProLong mounting medium (ProLong Gold, P36931; Molecular Probes) containing 4′,6-diamidino-2-phenylindole was used to identify nuclei. The protocol specifications for all antibodies used are described in detail in Table 1.

Serial sections of a smaller specialized tissue microarray (control array) were stained alongside all cohorts to confirm assay reproducibility; this allows for the construction of a normalization standard curve to adjust for run-to-run variability. Tumor heterogeneity for each marker was assessed in duplicate cores of the same tumor. MCF-7 and A431, EGFR-transfected CHO cells, and HER3-transfected BaF3 cells were used as positive controls for ER, EGFR, and HER3 assays, respectively (Table 1). Negative control sections, in which the primary antibody was omitted, were used for each immunostaining run. In addition MB436, CHO, and BaF3 FFPE cell pellets that are known to not detect ER, human EGFR, and HER3, respectively, were analyzed by quantitative immunofluorescence as negative controls.

Automated quantitative analysis (AQUA) allows the exact measurement of the protein concentration within subcellular compartments (15). In brief, high resolution 4′,6-diamidino-2-phenylindole, cytokeratin–Alexa 546 and target–cyanine 5 monochromatic images were captured by a PM-2000 microscope (HistoRx, New Haven, CT). Tumor was distinguished by creating an epithelial tumor “mask” from the cytokeratin signal on the basis of an intensity threshold set by visual inspection of histospots. The AQUA scores of target proteins in the tumor mask for EGFR and HER3, and the nuclear compartment for ER were calculated by dividing the sum of the target compartment pixel intensities by the area of the compartment in which they were measured. Specimens with <5% tumor area per histospot were not included in the AQUA for not being representative of the corresponding tumor specimen.

AQUA scores were log normalized for run-to-run variability. Pearson's correlation coefficient (r) was used to assess correlations between ER, EGFR, and HER3 measurements with different antibodies, as well as AQUA and western blot readings for the same ER and EGFR antibodies on cell line controls. Assay reproducibility was assessed by comparing AQUA scores on serial sections of the control array; a Pearson's r > 0.9 was indicative of minimal assay variability. Survival curves were constructed with the use of the Kaplan-Meier method, and survival differences were analyzed by the log rank test. All P-values were based on two-sided testing, and differences were considered significant at P < 0.05. All statistical analyses were done with the use of the SPSS software program (version 13.0 for Windows; SPSS Inc., Chicago, IL).

ER antibody specificity was confirmed by western blotting on MCF-7 and BT474 cells; each of the primary antibodies tested yielded a single 67-kDa band, and the band intensity was comparable with protein measurements by AQUA in FFPE cell pellets (Supplementary Fig. S1). MB436 cells that do not express any ER were negative by AQUA (data not shown).

We used AQUA to measure ER protein expression detected by clones 1D5, SP1, F10, and ER60c on a tissue microarray of 642 breast cancer specimens. Clones 1D5 and ER60c recognize epitopes in the N terminal domain of ER-α, whereas SP-1 and F10 bind epitopes at the C terminus (Fig. 2). ER AQUA scores ranged from 21.5 to 2,727.3 (n = 261; mean ± SE, 438.98 ± 31.23), 5.7 to 1,865.3 (n = 366; mean ± SE, 347.71 ± 24.98), 11 to 3,503 (n = 389; mean ± SE, 490.9 ± 35.7), and 2.6 to 1,016.7 (n = 396; mean ± SE, 135.74 ± 9.8) for ER 1D5, SP1, F10, and ER60c, respectively. Parametric comparisons between different ER antibodies revealed a tight correlation with a Pearson's r² of 0.98 (P < 0.0001) between SP1 and F10, F10 and ER60c, and SP1 and ER60c; 0.97 (P < 0.0001) between 1D5 and ER60c or SP1; and 0.96 (P < 0.0001) between F10 and 1D5 clones (Fig. 2). Quantitative immunofluorescence ER measurements were strongly correlated with semiquantitative pathologist-based ER scores (Spearman's rho = 0.68, ρ = 0.69, ρ = 0.69, and ρ = 0.68; P < 0.001 between pathologist-based ER and AQUA F10, SP1, 1D5, and ER60c, respectively). Evaluation of the assay reproducibility did not reveal significant differences between redundant tumor cores or serial sections of the control array stained in each IHC run (Pearson's r² > 0.9 for all runs; P < 0.0001; Supplementary Fig. S2).

Western blotting was done for EGFR antibodies on A431, MB468, H1355, and H1666 cell lines that express different levels of EGFR. MAbs 31G7 and 15F8 detected a single 170-kDa band. MAb H11 also detected a 170-kDa band in A431 and MB-468 cells; however, it detected a 140-kDa band in H1666 (Supplementary Fig. S1). The rabbit polyclonal 2232 revealed multiple bands, in addition to the expected 170-kDa band, at various molecular weights (Supplementary Fig. S1). The mAb 2-18C9 recognizes a conformation-specific epitope dependent on disulfide bonds and therefore was not useful for western blotting analysis. We also did AQUA on matched FFPE cell pellets as well as FFPE EGFR-transfected CHO cell pellets that revealed the same range of expression shown in western blot analysis. Comparison of quantified EGFR bands and AQUA scores of the same cell lines revealed an excellent correlation for 31G7 (Pearson's r² = 0.9), a good correlation for the 15F8 clone (r² = 0.62), and a moderate correlation for clone H11 (r² = 0.34). Polyclonal 2232 bands were not quantified because the blotting was not apparently specific for 170-kDa EGFR. CHO cells that do not express any human EGFR were negative by AQUA (data not shown).

EGFR protein levels were measured by AQUA, and ranged from 1.44 to 146 (n = 554; mean ± SE, 5.77 ± 0.41), 1.41 to 146 (n = 608; mean ± SE, 13.31 ± 0.61), 7.25 to 121.74 (n = 478; mean ± SE, 32.28 ± 0.71), 4.43 to 165.66 (n = 549; mean ± SE, 9.64 ± 0.41), and 0.01 to 167.5 (n = 395; mean ± SE, 39.22 ± 1.16) for clones 31G7, 2-18C9, H11, 15F8, and polyclonal 2232, respectively. Clones 31G7, 2-18C9, and H11 recognize epitopes on the extracellular domain of EGFR protein as well as the 145-kDa variant III form of EGFR. The 15F8 clone binds to the C terminus, whereas polyclonal 2232 recognizes residues around tyrosine 1068 of the EGFR protein (Fig. 3). In contrast to the strong correlation observed among the ER antibodies, we found that the EGFR antibodies resulted in higher analytic variability. The highest correlation was identified between EGFR measurements by the 2-18C9 and 31G7 clones (r² = 0.61; P < 0.00001; Fig. 4), whereas a weak correlation was observed between 15F8 and 2-18C9 (r² = 0.21; P = 0.01) or 31G7 (r² = 0.17; P = 0.01), and no association was found between 15F8 measurements and H11 (Pearson's r² = 0.006; P = 0.5) or polyclonal 2232 (r² = 0.05; P = 0.1). H11 was not associated with any other EGFR antibody (r² = 0.03 and P = 0.1; r² = 0.0007 and P = 0.59; and r² = 0.003; P = 0.25 for comparisons with polyclonal 2232, 2-18C9, and 31G7 clones, respectively). There was also no association between EGFR protein levels assessed by the polyclonal 2232 antibody and mAb 2-18C9 (r² = 0.0005; P = 0.71) or 31G7 (r² = 0.008; P = 0.2). A significant correlation was found between EGFR 31G7 and 2-18C6 scores in redundant tumor cores (Pearson's r = 0.64 and r = 0.77; P < 0.0001; Supplementary Fig. S2). No significant differences were observed between serial sections of the control array stained in each run (Pearson's r² > 0.9 for all runs; P < 0.0001; Supplementary Fig. S2).

HER3 expression was measured by six antibodies in a smaller lung cancer tissue microarray, and AQUA scores ranged from 12 to 52.9 (n = 42; mean ± SE, 26.9 ± 1.8), 111 to 244 (n = 42; mean ± SE, 168.3 ± 6.21), 58.9 to 200 (n = 42; mean ± SE, 145.4 ± 7.8), and 15 to 50 (n = 42; mean ± SE, 29.3 ± 0.8), to 40 to 552 (n = 42; mean ± SE, 309 ± 25.8) and 104 to 338 (n = 42; mean ± SE, 157.8 ± 20) for clones RTJ1, RTJ2, SGP1, and M7297, and polyclonals C17 and RB-9211, respectively. MAbs RTJ1, RTJ2, and M7297, and polyclonal C17 recognize epitopes on the intracellular domain of HER3, whereas the SGP1 clone and RB-9211 bind to the extracellular domain. All antibodies except for C17 failed to detect the expected membranous and/or cytoplasmic localization of HER3, and comparisons between HER3 measurements did not reveal strong correlations between C17 and RTJ1 (r² = 0.02), C17 and M7297 (r² = 0.02), C17 and RB-9211 (r² = 0.28), C17 and RTJ2 (r² = 0.18), RB-9211 and RTJ1 (r² = 0.08), RB-9211 and SGP1 (r² = 0.26), RB-9211 and M7297 (r² = 0.38), SGP1 and RTJ1 (r² = 0.01), SGP1 and M7297 (r² = 0.1), or RTJ1 and M7297 (r² = 0.00001; Fig. 4). A tight correlation was observed between C17 and SGP1 (r² = 0.61). However, C17 yielded membranous and cytoplasmic staining (Fig. 4), whereas SGP1 yielded nuclear staining without any cytoplasmic localization (data not shown).

To investigate whether the prognostic value of EGFR expression in breast tumors was dependent on the antibody used for IHC detection, we applied different cut points, including tertiles, quartiles, and the top deciles of the continuous EGFR AQUA scores to reproduce the cutoff points most commonly used in the EGFR literature. The P-values shown were not normalized for multiple cut-point testing because this was an exploratory analysis to illustrate the effect of cut-point selection. The top 66.66% EGFR-expressing tumors had a worse prognosis compared with the bottom 33.33% expressing group when EGFR protein levels were assessed by mAb H11 (n = 290 versus n = 135, respectively; log rank P = 0.015; Fig. 5). The same trend was observed when EGFR was detected by mAb 31G7 (top 66.66% n = 333 versus bottom 33.33% n = 167; log rank P = 0.126; Fig. 5), and the prognostic stratification was more significant when the median AQUA score was used as a cutoff point (top 50% n = 250 versus bottom 50% n = 250; log rank P = 0.059; Fig. 5C). Interestingly, survival analysis with the use of the median AQUA score for H11 failed to reach significance (top 50% n = 212 versus bottom 50% n = 213; log rank P = 0.63; Fig. 5). Survival analysis with the use of all cutoff points described above for EGFR protein levels assessed by the 15F8 clone, the 2-18C9 clone, and the 2232 polyclonal did not reveal any statistically significant association between EGFR and disease-specific survival (Fig. 5 and data not shown). Thus, the statistical significance, direction of trend, and optimal cut point seem to be highly dependent on antibody selection.

IHC is among the most commonly used research tools in biomedical research. A 3-fold increase in immunohistochemical studies over the last 20 years is indicative of the increased popularity of IHC tests (16). Efforts toward standardization and quality controls have been made (1, 17-20); however, there are no universally accepted guidelines that primarily address IHC laboratory practice (21). Antibody specificity and validation, use of appropriate controls, type of antigen retrieval, and automation are critical analytic factors that affect IHC results (22, 23). Here we measured three commonly studied cancer-related biomarkers, ER, EGFR, and HER3, with the use of different primary antibodies to assess the extent of analytic variability in IHC biomarker studies.

Specificity has been considered the criterion standard for antibody selection, and antibodies are often validated by western blotting when the antibody recognizes the denatured antigen. We and others have noted that most antibodies produced by immunization with synthetic peptides fall into this category (24). Antibodies recognizing ≥2 bands might still identify the same target protein, which might be present as a splice variant, deletion mutant, degradation product, or one that contains different post-translational modifications in the cells of interest. However, additional cross-validation studies with antibodies recognizing different epitopes would then be required. We identified a single band at the expected molecular weight (67 kDa for ER and 170 kDa for EGFR) with each of the antibodies tested for ER and EGFR, except for the polyclonal 2232 and mAb clone H11 anti-EGFR antibodies. Nevertheless, comparisons between all EGFR antibodies recognizing different epitopes on the full-length 170-kDa protein did not reveal strong associations between any of the EGFR antibodies, with the exception of clones 31G7 and 2-18C9. These results are consistent with previous studies that reported strong correlations between mAb clones 2-18C9 and 31G7 with the use of conventional chromagenic IHC (7), as well as between AQUA and IHC-based EGFR detection (25).

Validation of the antibodies by a second independent method in a panel of cell lines with known target expression has been used to confirm specificity (23, 26-29). We used quantitative immunofluorescence to confirm western blotting results and found an excellent correlation between the two methods. MCF-7 and A431, which overexpress endogenous ER and EGFR, respectively, and EGFR-transfected CHO cells were used as positive controls. MB436 breast cancer cells and nontransfected CHO cells were used as negative controls for the ER and EGFR IHC runs. IHC negative controls including no primary antibody or use of an unimmunized antibody from the same species are valuable but insufficient. Here we used in addition to conventional negative controls cell lines that are known to not express the protein of interest.

Antibody specificity has also been evaluated in the context of biologically relevant subcellular localization. We have assessed the importance of subcellular localization and its correlation with analytic variability in a series of experiments with the use of different HER3 antibodies in a small lung cancer tissue microarray. We found that 50% of the antibodies used yielded a nuclear pattern for HER3, and no association in terms of specific staining was found among the clones tested. In contrast, all EGFR and ER antibodies yielded membranous and nuclear localization, respectively, in both cell-line positive controls, and EGFR-positive and ER-positive clinical cases, consistent with the biological relevance of these proteins.

One of the critical analytic factors is antigen retrieval, which is used to unmask antigens in FFPE tissue sections. Most antibodies show increased staining when heat-induced antigen retrieval is done (30, 31); however, heat-induced antigen retrieval is not effective for all antibodies, and loss of specific immunostaining (1) as well as “retrieval” of misleading cross-reacting sites (24) have been reported. Enzymatic digestion yields superior results for specific proteins (32-34); however, it has been shown to destroy the targeted epitopes (35, 36), and the lack of a standardized protocol in terms of incubation time and temperature may result in tissue breakdown, loss of morphology, and high levels of background (37). We used heat-induced antigen retrieval for all ER, EGFR H11, 15F8, and 2232 antibody detection, and an optimized protocol for proteinase K digestion for EGFR 31G7 and 2-18C9 clones. Not surprisingly, we observed the destruction of the tissue morphology and loss of EGFR-specific staining with prolonged proteinase K incubation, consistent with previous findings. Heat-induced antigen retrieval for the same clones did not yield any specific EGFR immunoreactivity; this is not consistent with the study of Derecskei et al. showing that microwave antigen retrieval in citrate buffer (pH 6) converted EGFR-negative into EGFR-positive tumors (38).

EGFR-related analytic variability precludes firm conclusions on the prognostic and predictive potential of EGFR. Lung cancer studies failed to consistently show a prognostic (39-42) or predictive (43, 44) role for EGFR protein levels, and the same heterogeneity in study results was also observed in breast cancer studies in which the clinical application of an EGFR IHC test is highly controversial (25, 45-47). Interestingly, only H11 IHC analysis yielded a statistically significant prognostic classification, and a trend toward prolonged survival was observed for the low EGFR–expressing group when 31G7 was used to detect EGFR protein expression. We believe that our results reflect the heterogeneity seen in the EGFR literature attributed to the variation in the study methodology and interpretation of the results. Quantitative immunofluorescence has been shown to have accuracy comparable with an ELISA (48), and has been broadly used for tissue microarray and whole tissue section studies (49). The reproducibility and comparison of AQUA-derived clinical cutoff points with conventional IHC for ER and EGFR have been described in two previous studies (25, 50). We assessed outcome as a function of tumor heterogeneity on different constructs of the same cohort and observed similar prognostic classifications for both ER and EGFR (data not shown).

In conclusion, our work emphasizes on the impact of analytic processing on the validity of biomarker measurements by IHC. Evaluation of IHC proficiency by external agencies is a key first step toward IHC standardization. In the future, we believe rigorous standardization methods will be adopted in the leading laboratories that will include a series of index cases or some other method to show robust reproducibility in the assessment of protein expression on FFPE tissue.

M. Gustavson: Employment, HistoRx, Inc. D. Rimm: Consultant/Advisory Board, HistoRx, Inc.

Grant Support: Department of Public Health, State of Connecticut #2009-0097 and NIH R33 CA 110511 (D.L. Rimm); and the Susan G. Komen Breast Cancer Foundation, the Breast Cancer Alliance, and the Milheim Foundation for Cancer Research (J.L. Reiter).