Quantitative imaging for development of companion diagnostics to drugs targeting HGF/MET.

The limited clinical success of anti-HGF/MET drugs can be attributed to the lack of predictive biomarkers that adequately select patients for treatment. We demonstrate here that quantitative digital imaging of formalin fixed paraffin embedded tissues stained by immunohistochemistry can be used to measure signals from weakly staining antibodies and provides new opportunities to develop assays for detection of MET receptor activity. To establish a biomarker panel of MET activation, we employed seven antibodies measuring protein expression in the HGF/MET pathway in 20 cases and up to 80 cores from 18 human cancer types. The antibodies bind to epitopes in the extra (EC)- and intracellular (IC) domains of MET (MET4EC, SP44_METIC, D1C2_METIC), to MET-pY1234/pY1235, a marker of MET kinase activation, as well as to HGF, pSFK or pMAPK. Expression of HGF was determined in tumour cells (T_HGF) as well as in stroma surrounding cancer (St_HGF). Remarkably, MET4EC correlated more strongly with pMET (r = 0.47) than SP44_METIC (r = 0.21) or D1C2_METIC (r = 0.08) across 18 cancer types. In addition, correlation coefficients of pMET and T_HGF (r = 0.38) and pMET and pSFK (r = 0.56) were high. Prediction models of MET activation reveal cancer-type specific differences in performance of MET4EC, SP44_METIC and anti-HGF antibodies. Thus, we conclude that assays to predict the response to HGF/MET inhibitors require a cancer-type specific antibody selection and should be developed in those cancer types in which they are employed clinically.

Introduction

Immunohistochemistry (IHC) is widely used clinically to aid with challenging diagnostic decisions that require formalin‐fixed and paraffin embedded (FFPE) tissues. In the era of precision medicine, the ease associated with detecting proteins and phosphoproteins in single cancer cells via IHC is ideal for development of companion diagnostics needed to stratify patients for treatment with targeted therapies. However, technical problems diminishing the accuracy of quantifying the expression of drug targets by IHC hinder this biomarker development. The recent emergence of whole slide imaging and digital image analysis generates quantitative numerical data associated with protein expression in FFPE tissues 1, 2. In theory, this approach could provide new opportunities for measuring the expression of drug targets in cancer tissues obtained from patients during routine medical care. However, the use of digital image analysis for biomarker development and companion diagnostics is still in its infancy 1 because it requires multidisciplinary interactions between pathologists, experts in image processing and software developers.

The MET receptor tyrosine kinase is a documented oncoprotein in human cancer (visual database available here: http://www.vai.org/met/). Our and other groups linked MET to epithelial‐to‐mesenchymal transition, invasion and metastasis in human cancer cells and in mouse models 3. Germline activating mutations of MET cause hereditary papillary renal carcinoma and MET expression is associated with adverse outcome in practically all solid tumours 4, 5. In addition, the human MET gene lies in an unstable genomic region on chromosome 7 that is prone to amplification 6, 7. Hence, clonal expansion of cancer cells with multiple copies of the MET gene or with autocrine production of the MET ligand, Hepatocyte Growth Factor/Scatter Factor (HGF), can cause resistance to drugs that inhibit other receptor tyrosine kinases 8. The key role of MET in oncogenesis and drug resistance provides a strong rationale for extensive drug development 9. While most clinical trials with MET kinase inhibitors have not demonstrated an overall survival advantage, a subset of patients clearly experienced cancer control 10.

The lack of appropriate companion diagnostics compromises the success of clinical trials with MET inhibitors 11 and the limited success may stem from inadequate selection of patients. Measuring the active MET kinase, ie, the drug target, relies on detection of protein phosphorylation, which constitutes an inherently unstable posttranslational modification. In addition, biomarker development is complicated by the sequestration of MET in signalling endosomes and shedding from the cell surface 12, 13. Companion diagnostic assays for MET inhibitors employed MET FISH and MET IHC 14. However, association of these measurements with activation of the MET kinase varies among cancer types 15. Because of extensive post‐translational regulation of MET protein expression, MET mRNA expression has not been explored as a biomarker of MET activation. Furthermore IHC to measure MET protein expression has not been stringently validated in FFPE tissues 16. Collectively, the different mechanisms of MET kinase activities, ie, genomic mutations, binding of HGF and overexpression and/or crosstalk with other cell surface receptors 3, 17, have so far not resulted in the development of a uniform assay of MET receptor activation.

To analyze the activation of the MET pathway, we generated a tissue microarray (TMA) of 18 different cancer types. We employed seven antibodies that react with proteins in the MET pathway. SP44_METIC and D1C2_METIC bind epitopes within the cytoplasmic C‐terminus of MET and MET4EC recognizes a motif in the MET receptor alpha‐chain, which is extracellular 18. We also utilized an antibody reactive with HGF. To determine the activation of cytoplasmic MET signalling pathways, we selected antibodies that bind phospho‐MAPK and phospho‐SRC‐family kinases (SFK). Using quantitative digital imaging of slides stained by IHC, we generated a dataset consisting of 9139 individual measurements. The analysis of the data revealed that antibodies most predictive of MET activation differ amongst cancer types. This approach highlights the importance of cancer‐type specific development of companion diagnostic assays for drugs against the active MET receptor kinase.

Materials and methods

Selection of cases and TMA construction

For TMA construction, 20 cases from 18 cancer types, which were randomly retrieved from the pathology archive, were displayed and confirmed by a subspecialty pathologist (Dataset S1 and Table S1). The MET genomic status was presumed to be identical to germline 19. The core diameter was 0.6 mm except for sarcoma, melanoma, lymphoma, glioblastoma, head and neck cancer, and colon cancer, where it was 1 mm. Normal control blocks included kidney, tonsil, liver and colon. TMAs were constructed with the TMArrayer™ (Pathology Devices, Inc.).

Immunohistochemistry

Antibodies were obtained from Cell Signaling Technology (CST) or from other sources as indicated: MET‐pTyr1234/pTyr1235 (clone D26), MET4 18, c‐Met (clone SP44, Spring Bioscience), c‐Met (clone D1C2), HGF (LifeSpan, BioSciences), Src‐family‐pTyr416, pErk1/2 (clone 137F5). Where needed, the specificity of antibodies was confirmed by western blotting before and after treatment with inhibitors (Figures S1A, S1B) 20. The pSRC family‐Tyr416 has not been optimized for IHC by the company (CST) and we therefore followed a published protocol 21, which we also validated in a neoadjuvant bladder cancer study 22. The HGF antibody has been characterized in the Vande Woude laboratory 23 and our protocol provides a staining pattern that is consistent with staining in mesenchymal cells. The specificity of the MET4 antibody was published in 18. For each antibody, dilution, signal amplification and incubation conditions are listed in Table S2. Staining was performed as described previously 24. Secondary antibodies were purchased from CST. The tyramide signal amplification system (TSA) was purchased from Perkin Elmer and the polymer amplification system was obtained from VECTOR labs. The immunofluorescent staining was performed with the Opal multiplex immunofluorescent system (Perkin Elmer) and fluorescent staining with MET4EC and SP44_METIC was followed by IHC with the pMET antibody.

For every core, the percent coefficient of variation (%CV) for between‐days (five slides) or within‐day (three slides) assays was calculated by dividing the mean core intensity by the standard deviation. The total %CV was calculated by first averaging the squares of individual %CV values followed by taking the square root this value.

The IHC staining variability amongst cores within a case was compared to the variability across cases using F‐statistics. Details of calculations are provided in the material.

Digital image acquisition, image processing and pathology annotation of TMA cores

TMA slides were digitized on the high‐resolution Leica SCN400NF whole‐slide scanner and images were sequentially analyzed with the Tissue IA Optmiser (Leica) software installed on the Leica Digital Image Hub as described in 25. Cancer and stromal areas were annotated by one of three pathologists and confirmed by second pathologist. The mean signal intensity (staining concentration) per unit area was calculated within annotated regions and exported for statistical analysis (Dataset S2).

Data acquisition and normalization

All statistical analyses were conducted in R v3.1.3 26. Plots were generated with the use of the following packages in R, beeswarm 27, extrafont 28, plyr 29, RColorBrewer 30, reshape2 31, scales 32. TMA data were subjected to background intensity subtraction (lowest core intensity of individual slide) and log2 transformation. Intensity values were normalized to the average signal intensities of kidney and tonsil control cores within respective slides. The final data represent the log2 transformed fold change from control. As a second normalization method, control adjusted TMA data values (without background subtraction) were quantile normalized across slides and antibodies using limma 33 package in R.

Determination of distributions and correlation coefficients

Histogram density profiles of each antibody were constructed using ggplot2 34 package in R. Unsupervised hierarchical clustering of pairwise Pearson's correlation matrix was calculated using hclust function based on Euclidean distance and complete linkage method. p‐values of correlations generated for each antibody pair using data from all cancer types combined were adjusted using the stringent Bonferroni correction. p‐values of pairwise correlations performed in cancer types individually were adjusted using the less stringent FDR correction 35. Heatmaps were generated using heatmap.2() in gplots package 36. The Dendrogram in Figure S5A of the correlation matrix was generated using cluster package 37. The Wilcoxon rank‐sum test was used to determine the significance of the difference between of T_HGF and St_HGF signal intensities.

pMET z‐score

In order to compare average levels of pMET expressions across cancer types, z‐scores of pMET staining intensities were calculated as the difference of the cancer‐type mean and grand mean, divided by the standard deviation of all samples, and then multiplied by the square root of number of cases in that particular cancer type. See supplementary material for details of the calculations.

Random forest classification

The expression of pMET was dichotomized at the 7th decile to classify cases as positive or negative. Levels of pMET were predicted by fitting a Random Forest model to all data with the following features: cancer‐type plus all antibodies, or cancer‐type plus either 1 of the antibodies (MET4, SP44_MET, D1C2_MET, T_HGF, St_HGF). Random forest classification analyses were conducted using the randomForest function 38. 10‐fold cross validation was performed and the area under the curve (AUC) of the receiver operating characteristic (ROC) curves was determined for each model using the ROCR package 39. Variable importance for each model was calculated using the cforest function in the party package 40, 41, 42.

Groups of five cancer‐types (n = 100 cases) were constructed for all possible cancer‐type combinations (n = 8568). Random Forest models were constructed in each group to predict pMET positivity with individual antibodies as predicting variables. Five‐fold cross validation was performed for each model and the AUC was calculated.

Results

Differences in tissue staining between antibodies to the extra‐ and intracellular domain of MET

To identify potential differences in staining patterns of antibodies that react with the intracellular and extracellular domains of MET and with MET‐pY1234/pY1235 (pMET), we stained tissue sections of renal cortex with α‐pMET, SP44_METIC (antibody against MET intracellular domain) and MET4EC (antibody against MET extracellular domain) (Figure 1A). Surprisingly, we observed marked heterogeneity of staining in renal tubules. In particular, individual cells in collecting duct epithelium preferentially reacted either with MET4EC or SP44_METIC. Tubules that stained with α‐pMET reacted with MET4EC, but not with SP44_METIC. These data highlight cell specific differences in availability of the MET4EC and SP44_METIC epitopes and suggest that antibodies against total MET protein might differ in their correlations with pMET levels. To take advantage of this result for development of companion diagnostics, we assembled a biomarker panel to investigate the relationship between MET phosphorylation, total MET protein expression, expression of HGF and activation of MET signalling pathways through SFK and MAPK.

Figure 1

Antibody‐based analysis of the MET pathway in 18 cancer types. (A) SP44_METIC and MET4EC staining in renal tubules. Normal kidney tissue was stained with SP44_METIC (green) and MET4EC (red) by immunofluorescence and with α‐pMET by immunohistochemistry (purple) and a representative case is shown. The low power image reveals heterogeneous staining patterns of the MET antibodies in multiple renal tubules. The arrow in the magnified panels points to the overlap between MET4IC and α‐pMET staining. (B) Data acquisition. 18 cancer types are listed and the numbers of cores with data are indicated in parentheses next to each cancer‐type. Each cancer type was stained with seven antibodies. The total number of TMA cores providing data for each antibody is indicated in the legend. Number of cores analyzed per antibody in individual cancer types is plotted on the y‐axis. (C) Specificity of antibody staining and annotation of cancer and stromal regions. Tissue cores were stained with antibodies indicated above each image and consist of areas of cancer and non‐cancer tissue. The insert for each core shows an enlarged area of the cancer‐to‐stroma interface. For measurement of staining intensities, multiple areas in the cancer are circled in each core (red outline) and for cores stained with the HGF antibody, additional areas of stroma (blue) are circled for analysis.

Quantitative imaging of IHC with 7 antibodies in 18 cancer types

In order to test the MET‐associated biomarker panel, we constructed a multi‐cancer TMA consisting of 18 cancer types (lung adeno‐ and squamous carcinoma, mesothelioma, endometrium, ovary, breast, prostate, kidney clear cell, bladder, sarcoma, melanoma, pancreas, stomach, liver, lymphoma, glioblastoma, head and neck, and colon), 20 cases per cancer type and four cores per case (Figure S3, Dataset S1). In total, 1440 cores from 360 cases were displayed on seven glass slides. The average core loss after IHC was equally distributed across cancer types, such that 16–20 cases per cancer type provided data for statistical analysis. In addition, normal tissue cores were used for normalization and comparison of staining intensities across the seven slides.

Each TMA was stained with seven antibodies (Figures 1B, S3, Table S1). In order to overcome the problem of low sensitivity of measuring MET autophosphorylation at Y1234/Y1235 43, we combined polymer and tyramide amplification systems without compromising the dynamic range (Figures S1C, S2B). The improved amplification method applied to α‐pMET IHC resulted in TMA cores above the limit of quantification in all 18 cancer types. Control rabbit and mouse antibodies were negative (data not shown). We acquired 1180–1230 measurements per antibody in areas of cancer and 726 measurements in the stroma (Figure 1B). For all antibodies and cancer types together, we obtained 9139 measurements (Dataset S2).

To compare the staining intensities from each antibody across seven slides, we subtracted the background staining and calculated the fold change relative to control cores in the same slide (Figure S4). The resulting distributions of control‐adjusted, log2‐transformed staining intensities were normally distributed, ie, ‘bell‐shaped’ (S2A). The expression of pMET levels in 18 cancer types is shown in Figure 2A. Phospho‐MET staining was highest in bladder cancer and lymphoma. In addition, outlier cases with high pMET levels were identified in lung adenocarcinoma, endometrial, prostate, stomach and liver cancer, and in melanoma. In order to assess the biological variability, we compared expression values between cases to those within a case. Between‐case variability was significantly higher than within‐case variability (F‐test, adjusted p < 0.05) (Figure 2B, Dataset S2), indicating that the within‐tumour heterogeneity was less than the heterogeneity exhibited across cases. Furthermore, the α‐pMET within‐day coefficient of variation (%CV) = 8.5% and the between‐day %CV = 26.7%, both acceptable for IHC assays (Figure 2C) 44.

Figure 2

Protein expression levels. (A) Quantative imaging of pMET levels. Each TMA with its cancer types is listed on the x‐axis and the number of cases per cancer type is listed in parentheses. Control‐adjusted staining intensities are plotted on the y‐axis. Horizontal lines in boxes represent the 1st, 2nd and 3rd quartiles. Whiskers are indicated outside the box with limits of 1.5× the inter‐quartile range (IQR). (B) Inter‐ and intra‐tumour variance. One‐way analysis of variance was used to calculate the variance within cores of a case (within‐case variability) and across all cases within a cancer type (between‐case variability). Each colour indicates one of 18 cancer‐types stained with a single antibody. (C) Within‐ and between‐assay concordances. To determine within day (WD) technical variability of the assay, three slides of TMA 7 (WD1–3) were stained with the pMET antibody on the same day. To determine day‐to‐day variability, a single TMA 7 slide was stained on five different days (BD 1–5). Pairwise correlations for WD and BD slides are shown and the coefficients of variation (CV) are indicated for within day and between day variability. (D) HGF expression. The boxplots depict the control adjusted log2 expression of St_HGF and T_HGF in each cancer type. St_HGF expression could not be measured in cancer types that did not display sufficient regions of stroma with peritumoural mesenchymal cells (lymphoma, glioblastoma, sarcoma). A significant difference in expression between T_HGF and St_HGF is indicated by **p < 0.01, *p < 0.05.

A switch from paracrine to autocrine HGF has been demonstrated during progression of AML and in the clonal selection of lung adenocarcinoma 8, 45. However, the prevalence of autocrine HGF in other primary solid tumours is not known. To determine the relationship between HGF expression and MET activation, we profiled tumour HGF levels (T_HGF) in 1188 cores from 18 different cancer types. In addition, we measured HGF levels in the stroma adjacent to tumour (St_HGF) in 726 cores. The highest stromal HGF levels were detected in prostate and stomach cancer (Figure 2D). In addition, tumour and stromal HGF were correlated (r = 0.56, p < 0.001) (Figure 3A).

Figure 3

Correlations of protein expression levels determined by quantitative imaging (QI). (A) Correlation matrix for all cancer types. Pairwise correlations were calculated across all cases from 18 cancer types stained with the antibodies listed on the x‐ and y‐axes. Pairwise correlation matrix was used in an unsupervised cluster analysis. The three main clusters are labelled above the matrix. (B–E) Scatter plots of normalized, relative staining intensities. A regression line is shown for each scatter plot and Pearson's correlation coefficients are indicated on the bottom right. (F) Correlations with pMET in individual cancer types. Pearson's correlation coefficients between levels of pMET and the MET and HGF antibodies listed next to the heatmap. The last row shows the pMET z‐score (normalized average level), which was calculated by normalizing the average pMET level of an individual cancer type to the mean of pMET levels across all cancer types. Values in the heatmap are matched to the color bar.

Correlations between antibodies in the MET pathway and MET kinase activation

We reasoned that a better understanding of associations between antibodies that report the activation state of the MET pathway might allow for selection of surrogate biomarkers of MET activation. Thus, a correlation matrix was generated through unsupervised clustering of pairwise correlations between the seven antibodies. The clustering revealed three major groups (Figures 3A, S5A). The first group consists of the SP44_METIC and D1C2_METIC, which both bind the C‐terminus of the MET receptor. While measurements with SP44_METIC and D1C2_METIC were significantly correlated (r = 0.45, p < 0.001), correlations of the SP44_METIC and D1C2_METIC antibodies with MET4EC were lower (r = 0.11, p = 1 and r = 0.11, p = 1, respectively) (Figures 3A, 3B, S5B). The second group contains all the phospho‐antibodies. The highest correlation in this group is observed between α‐pSFK and α‐pMET (r = 0.56, p < 0.001). In contrast, the correlation between α‐pMAPK and α‐pMET is only r = 0.20 (p = 0.0053), suggesting greater pMET to pSFK than pMET to pMAPK signalling (Figure 3A, C). MET4EC demonstrated the highest correlation with α‐pMET (r = 0.47, p < 0.001), while correlation coefficients of SP44_METIC and D1C2_METIC with α‐pMET were lower (r = 0.21 and r = 0.08, respectively) (Figure 3A, D). The third group consists of MET4EC, T_HGF and St_HGF. T_HGF and St_HGF are significantly correlated (r = 0.56, p < 0.001) and, in addition, T_HGF is significantly correlated with MET4EC (Figure 3E, r = 0.65, p < 0.001). Next, we performed an unsupervised hierarchical clustering analysis of pairwise correlation coefficients from 28 antibody pairs across the 18 cancer types (S6). Two major clusters were observed. The smaller cluster of five antibody pairs corresponds to pairs with the highest correlations in the correlation matrix in Figure 3A. To exclude a bias from the normalization method, we performed control‐adjusted and quantile normalization method (S7) and observed similar correlation matrices (S8 and Figure 3A). In summary, the main observation from the correlation matrix is the stronger correlation between MET4EC and α‐pMET compared to either SP44_METIC or D1C2_METIC and α‐pMET. In addition, MET4EC is strongly correlated (p < 0.001) with T_HGF and α‐pSFK.

MET protein expression and MET phosphorylation in individual cancer types

Since MET can be activated through overexpression, we determined the relationship between MET protein expression and MET receptor phosphorylation in individual cancer types. MET protein expression levels measured with MET4EC, SP44_METIC and D1C2_METIC differed between cancer types (Figure S4A, one‐way ANOVA, pMET4 < 0.001, pSP44 < 0.001, pD1C2 < 0.001). We calculated the correlation coefficients between pMET levels and total MET protein expression measured with MET4EC, SP44_METIC or D1C2_METIC in individual cancer types (Figure 3F). Correlation coefficients r > 0.5 are twice as frequent for SP44_METIC/pMET than for D1C2_METIC/pMET (Figure S9A). Two cancer types (glioblastoma and endometrial) revealed high correlation coefficients (r > 0.4, FDR adjusted p < 0.1) for all three α‐MET antibodies with α‐pMET. The correlations with α‐pMET in individual cancer types were not related to the z‐score (normalized average level) of pMET expression (p > 0.1) (Figure S10). Collectively, the results demonstrate that across cancer types, MET4EC, which reacts with an epitope in the extracellular domain, reports pMET more faithfully than the antibodies that bind the cytoplasmic C‐terminus of MET and that MET activation and protein levels are not related.

Next, we determined the correlations between HGF and pMET expression in individual cancer types. Correlations r ≥ 0.5 (FDR adjusted p < 0.1) between pMET and T_HGF were identified in four cancer types (head and neck, lymphoma, lung squamous and bladder carcinoma). While T_HGF (r = 0.38) and St_HGF (r = 0.29) were significantly correlated with pMET levels overall (Figures 3A, S5B), no correlation was observed with the z‐scores of pMET levels (p > 0.1) (Figure S10). The results suggest that in some cancer types, either autocrine or paracrine HGF expression may contribute to the activation state of MET.

Predictive models of MET activation

The ultimate goal of this multi‐cancer investigation is to compare the accuracy of antibodies for predicting MET activation. We applied a random forest (RF) classification approach to determine which antibodies singly or in combination best predict MET receptor phosphorylation. The activation state of the MET receptor was set at the 70th percentile cutoff of signal intensities of α‐pMET. We constructed ROC curves using 10‐fold cross validation for RF models that consisted either of a combination of all α‐MET and α‐HGF antibodies or of single antibodies. The accuracy of predicting pMET levels was based on the area under curve (AUC) (Figure 4A). In addition, a variable importance measurement is provided by the RF model for the contribution of each antibody and of the cancer type to the prediction. As previously noted, data normalized by both background with control adjustment or by quantile normalization methods generated similar results (Figure S11). The results from the RF prediction model underscore the important role of the cancer type in the prediction of MET activation status.

Figure 4

Random Forest (RF) models predicting MET activation status. (A) Variable importance. The y‐axis of the top panel shows the AUC as determined by a RF model with 10‐fold cross validation. Each bar depicts the total AUC from a model fitted with cancer type and antibodies indicated on the x‐axis. The bottom panel shows the contributions of cancer type and antibody (RF variable importance) to the prediction by the model. The RF variable importance is shown on a scale from 0 to 1. (B) Prediction of pMET status in groups of five cancer types. Cases were combined from groups of five cancer types listed at the top. ROC curves were generated to determine the accuracy of the model predicting pMET activity status for each antibody (MET4EC, SP44_METIC or T_HGF) with fivefold cross‐validation. The AUC with the highest value across all possible combinations of five cancer types is shown for MET4EC, SP44_METIC or T_HGF. ROC curves of St_HGF and D1C2_METIC are included for comparison.

To further examine the role of the cancer type in the prediction model, we combined data from five cancer types to increase the sample size. We also excluded D1C2_METIC and St_HGF from further evaluation because of the low variable importance of D1C2_METIC and the absence of St_HGF in some cancer types. Figure 4B shows the combinations of five cancer types with the greatest AUC for MET4EC, SP44_METIC and T_HGF. These cancer type combinations were obtained from a total of 25,704 ROC curves for the three antibodies in all possible 5‐cancer type combinations. Since a different combination of cancer types is obtained for each of the antibodies, the ROC curves confirm the important role of the cancer type in the prediction of pMET activation by MET4EC, SP44_METIC and T_HGF.

Discussion

This study demonstrates for the first time that digital image analysis of IHC stained slides can be used as a general approach to improve the selection of biomarkers for companion diagnostics to drugs targeting HGF/MET. We demonstrate that quantitative digital imaging (QI) can overcome difficulties of accurately measuring the inherently unstable and low abundance phosphorylation of the active MET receptor kinase in patient tissues by IHC. Replacing traditional categorical measurements with numerical measurements on a continuous scale improves the statistical data analysis in IHC assays. While a system for automated quantification of fluorescent antibody signals (AQUA) has been available for a long time 46, the method is not transferrable to chromogenic detection of antibody staining. In particular for IHC, QI lowers the limit of detection, reduces observer bias, lowers the technical variability and permits the normalization of experimental data to internal controls.

Using a robust pipeline we demonstrate that (1) it is feasible to measure levels of MET receptor activation and signalling in archival tissues from 18 cancer types; (2) in agreement with pre‐clinical systems (reviewed in 3), the magnitude of MET activation correlates with HGF and total MET protein expression levels; (3) an antibody reactive with an extracellular epitope in MET (MET4EC) 18 more faithfully reports MET phosphorylation than antibodies that react with intracellular epitopes (SP44_METIC, D1C2_METIC); (4) intratumoural HGF expression levels predict MET activation more accurately than HGF expression in the tumour microenvironment 8, 45, and (5) the prediction of MET activation is cancer type‐specific. Collectively, the data provide novel insights into the choice of antibodies to measure the active MET receptor kinase and are consistent with cancer‐type specific mechanisms of MET activation.

The comparison of MET antibodies that react with epitopes in the intracellular and extracellular domain of the receptor was prompted by the surprising differences of staining patterns in renal tubules (Figure 1A). The nature of MET receptor proteins that preferentially react with MET4EC versus SP44_METIC or D1C2_METIC is not known. MET4EC binds to an epitope in the 25 kDa alpha‐chain of MET 18 and not to the transmembrane 140 kDa MET receptor beta‐chain that interacts with SP44‐METIC and D1C2_METIC at its C‐terminus. The MET C‐terminal domain contains docking sites for multiple proteins that form a large signalling complex in response to MET kinase activation and autophosphorylation 47. It is conceivable that this protein complex hinders the binding of SP44_METIC and D1C2_METIC to active MET, particularly after tissues are fixed by formalin treatment. The steric hindrance is also consistent with the weaker correlation of SP44_METIC and D1C2_METIC with α‐pMET. In addition, MET4EC and α‐pMET staining intensities correlate strongly with intratumoural HGF and with α‐pSFK, suggesting that the heightened MET activation is associated with an HGF–MET–pSFK signalling axis (Figures 3A, S12). Previous measurements of tissue and plasma HGF in clinical trials with AMG‐102 utilized HGF‐specific ELISA assays and did not distinguish between intra‐ and extracellular HGF 48. As confirmed by our prediction model, intratumoural HGF qualifies as a novel biomarker of MET activation status that might improve the prediction of pMET activation status.

There are limitations of this study that remain to be addressed. One is the true clinical predictive power of biomarkers in the MET pathway. While we have improved the reproducibility and sensitivity of detection of pMET through a combination of signal amplification and QI, the clinical significance of measuring pMET using this approach remains to be determined. It is also unclear whether pMET is sufficient as a treatment response biomarker and whether α‐MET protein and α‐HGF measurements are required to improve the prediction. Since phosphorylation is unstable and rapid fixation of biopsies is required to preserve the phosphoproteome 49, it is a reasonable assumption that antibodies that are not dependent on phosphorylation will be needed to determine the MET activation status in clinical samples. An answer to these questions can be obtained in the future by analysis of tissues from clinical trials with MET inhibitors. Another limitation is the small sample size of individual cancer types. Based on the marked differences of biomarker expression among cancer types, expanding cohorts of individual cancer types will be necessary to obtain conclusive results for each of the cancer types included in the study. Our attempts to build a prediction model by using 5‐cancer type combinations is the most detailed analysis that our sample size permitted and the prediction of MET activation by MET4 and T_HGF expression levels is consistent with the correlations between pMET and MET4 or T‐HGF.

The results in our study explain conclusions by others demonstrating the lack of strong correlation between MET expression and response to MET inhibitory antibodies 50, 51, 52, 53 and highlight the importance of a cancer type specific development of companion diagnostics to MET inhibitory drugs.

Author Contributions

F.H., M.B.A., B.S.K designed study; Z.M., S.P., S.S., A.G., J.M., S.B., M.R., S.M.F. generated data; F.H., W.G.A., B.B. analyzed data; F.H., M.R.F., G.VW wrote the paper; X. Y., M.G., D.D., A.E.W., M. dP. V., A.M.M., D.J.L. created the Tissue Microarray.

SUPPLEMENTARY MATERIAL ONLINE

Supplementary materials and methods

Supplementary figure legends

Table S1. Tissue Microarray construction profiles

Table S2. Antibodies and staining conditions for immunohistochemistry

Click here for additional data file.

Figure S1A. Specificity of antibodies. Specificity of MET antibodies. A cell lysate from the DU145 cell line was generated in RIPA buffer supplemented with protease and phosphatase inhibitors and equal amounts of protein (100 micrograms) were probed directly in the first and second panel with the SP44_METEC and D1C2_METIC antibodies. Cells were treated with HGF + 10μM SGX523, a specific MET inhibitor [1], or with + 200U/ml HGF as indicated below each lane in panels 3 – 6. The WB membranes were probed with α‐pMET

Figure S1B. Specificity of pSRC and pMAPK antibodies. The DU145 prostate cancer cell line was pretreated with SRC (10μM Saracatinib) and MEK (50μM PD90859) inhibitors before treatment with HGF. Western blots were probed for pSRK and pMAPK as indicated and reprobed with the non‐phosphoantibodies as indicated. In addition, cells were encased in HISTOGEL® and cell pellets were processed for IHC with pSRC antibodies

Click here for additional data file.

Figure S1C. Scatterplot of pMET intensities obtained by horseradish peroxidase polymer amplification (Polymer) versus tyramide signal amplification (TSA). The normalized staining intensity values of pMET acquired from signal amplification method TSA were plotted against that from polymer amplification. There is a high correlation between the two amplification methods (r = 0.7, p < 0.001). Individual cancer types are indicated as different colors and shapes. The regression line is shown and Pearson's correlation coefficient is indicated on the top right of the figure

Click here for additional data file.

Figure S2A. TMA data distribution and dynamic range. Distribution of signal intensities for each antibody. For each antibody, the frequencies of the normalized signal intensities across all cancer types are plotted on the y‐axis

Click here for additional data file.

Figure S2B. Dynamic ranges of antibody measurements. For each antibody and cancer type, the dynamic range, plotted on the y‐axis, was determined as the difference of the lowest and highest intensity value. The intensity was calculated by subtraction of the background, log2 transformation and calculation of fold‐change to control cores

Click here for additional data file.

Figure S3. Workflow of case selection, tissue microarray (TMA), and digital image data acquisition. 18 cancer types were selected for analysis and are listed below their slide designation. Schematic representation of 20 cases per cancer type, 4 cylinder cores per case for construction of 7 TMA blocks. Each block was stained with 7 antibodies shown in the “Immunohistochemistry” section. Slides were imaged for collection of expression values by digital image analysis. Areas for analysis were annotated under the supervision of a pathologist. Both cancer and stromal regions in each core were annotated separately. Values of signal intensity measurements for each antibody were subjected to statistical analysis

Click here for additional data file.

Figure S4A. Boxplots of protein expression in 18 cancer types and kidney and tonsil controls. Boxplots of anti‐MET antibodies. Background adjusted protein expression data for pMET, MET4EC, SP44_METIC and D1C2_METIC are shown as box and whisker plots. The TMA slide number is listed at the top of each column and the cancer types, kidney and tonsil controls are listed on the x‐axis at the bottom of page. Background‐adjusted staining intensities are plotted after log2 conversion on the y‐axis. Horizontal lines in boxes represent the 1st, 2nd and 3rd quartiles. Whiskers are indicated outside the box with limits of 1.5x the inter‐quartile range (IQR). Case numbers for which expression values could be obtained are indicated in each column at the bottom of the graph

Click here for additional data file.

Figure S4B. Boxplots of T_HGF, St_HGF, pMAPK, pSFK. The same format was used as described for panel A

Click here for additional data file.

Figure S4C. Examples of IHC for pMET, MET4 and SP44_MET.

Click here for additional data file.

Click here for additional data file.

Click here for additional data file.

Click here for additional data file.

Figure S5A. Dendrogram and p‐values for correlation matrix Figure 3A. Dendrogram of hierarchical clustering. The distance between the branches is shown as the height on the y‐axis

Figure S5B. p‐values of pairwise Pearson correlation coefficients. The p‐values corresponding to the correlation coefficients in Figure 3A are shown

Click here for additional data file.

Figure S6. Unsupervised cluster analysis of pairwise Pearson correlation coefficients. Antibody pairs are shown at the bottom of each column and constitute all possible pairwise combinations. The magnitude of the calculated Pearson correlation coefficient is matched to the color bar. The clustering of antibody pairs is shown above the heatmap

Click here for additional data file.

Figure S7A. Comparison of 2 normalization methods for protein expression measurements. Comparison of 2 normalization methods across slides. Upper panel – “Control Adjustment”: Control adjusted normalization method. Boxplots of normalized expression measurements are shown for each slide. After background subtraction, expression measurements were normalized to control cores in the same TMA slide. Lower panel – “Control Adjustment and Quantile Normalization”: Quantile normalization. Control adjusted values were normalized by quantile normalization. Expression measurements were adjusted to control cores without background subtraction. The antibody and slide designation is indicated below the box and whisker plot. Horizontal lines in boxes represent the 1st, 2nd and 3rd quartiles. Whiskers are indicated outside the box with limits of 1.5x the inter‐quartile range (IQR)

Figure S7B. Boxplots of normalized expression measurements for individual antibodies. The same methods as described in panel A were employed. Data from all cores are used to generate the box plot for a given antibody, indicated on the x‐axis

Click here for additional data file.

Figure S8A. Analysis of quantile normalized data. Correlation matrix. Pairwise correlations for all antibodies were used in an unsuperwised cluster analysis. 3 main clusters are labeled above the matrix

Click here for additional data file.

Figure S8B. Scatterplots of pMET, 3 MET antibodies and HGF expression in all cancer types for quantile normalized data. The expression values of MET4EC, SP44_METIC, D1C2_METIC, T_HGF, and St_HGF were plotted against those of pMET in all cancer types. Individual cancer types are indicated as different colors and shapes. A regression line is shown for each scatter plot and Pearson's correlation coefficients are indicated on the top right

Click here for additional data file.

Figure S9A. Scatterplots of pMET, MET and HGF expression levels. Scatterplots of pMET and antibodies reactive with MET protein. For each cancer type, the expression of MET protein was measured by either MET4EC, SP44_METIC or D1C2_METIC and values are plotted against the pMET expression values. The line depicts the linear regression of the data and the Pearson correlation coefficient and its adjusted p‐value are shown in the lower right corner

Figure S9B. Scatter plot of pMET and HGF expression levels. For each cancer type shown at the bottom of the panel, the expression values of stromal (St_HGF) and tumor HGF (T_HGF) (y‐axis) are plotted against the pMET (x‐ axis) expression values. The line depicts the linear regression of the data and the Pearson correlation coefficient and its adjusted p‐value are shown in the lower right corner

Click here for additional data file.

Figure S10. Scatterplots of pMET z‐scores and pMET correlation coefficients. The average level of pMET calculated as the pMET z‐score is plotted on the y‐axis. The x‐axis shows the correlation coefficients between pMET and T_HGF (T_HGF/pMETCor), ST_HGF (St_ HGF/pMETCor), MET4 (MET4/pMET), SP44_MET (SP44_MET/pMET) and D1C2_MET (D1C2_MET /pMET). The Pearson correlation coefficient (black) and its p‐value (blue) are listed on bottom right.

Click here for additional data file.

Figure S11. Model performance and variable importance of quantile normalized data. The y‐axis of the top panel shows the area under the curve (AUC) as determined by a RF model with 10‐fold cross validation. Each bar depicts the total AUC from a model fitted with cancer type and antibodies indicated on the x‐axis. The bottom panel shows the contributions of cancer type and antibody (RF variable importance) to the prediction by the model

Click here for additional data file.

Figure S12. Schematic diagram of MET receptor pathway and the binding sites of antibodies used in the study. The binding sites of pMET, MET4EC, SP44_METIC, D1C2_METIC, HGF, pSRC and pMAPK are shown

Click here for additional data file.

Dataset S1

Click here for additional data file.

Dataset S2

Click here for additional data file.