SEARCH

SEARCH BY CITATION

Keywords:

  • molecular diagnostics;
  • early cancer biomarkers;
  • qPCR;
  • squamous cell carcinoma;
  • FOXM1 diagnostic biomarkers;
  • prognostic biomarkers;
  • clinical translation;
  • personalised medicine;
  • early detection

Abstract

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Histopathological discordance with molecular phenotype of many human cancers poses clinically challenging tasks for accurate cancer diagnosis, which impacts on treatment strategy and patient outcome. Hence, an objective, accurate and quantitative method is needed. A quantitative Malignancy Index Diagnostic System (qMIDS) was developed based on 14 FOXM1 (isoform B)-associated genes implicated in the regulation of the cell cycle, differentiation, ageing, genomic stability, epigenetic and stem cell renewal, and two reference genes. Their mRNA expression levels were translated via a prospectively designed algorithm, into a metric scoring system. Subjects from UK and Norway (n = 299) provided 359 head and neck tissue specimens. Diagnostic test performance was assessed using detection rate (DR) and false-positive rate (FPR). The median qMIDS scores were 1.3, 2.9 and 6.7 in healthy tissue, dysplasia and head and neck squamous cell carcinomas (HNSCC), respectively (UK prospective dataset, p<0.001); 1.4, 2.3 and 7.6 in unaffected, oral lichen planus, or HNSCC, respectively (Norwegian retrospective dataset with up to 19 years survival data, p<0.001). At a qMIDS cut-off of 4.0, DR was 94% and FPR was 3.2% (Norwegian dataset); and DR was 91% and FPR was 1.3% (UK dataset). We further demonstrated the transferability of qMIDS for diagnosing premalignant human vulva (n = 58) and skin (n = 21) SCCs, illustrating its potential clinical use for other cancer types. This study provided evidence that qMIDS was able to quantitatively diagnose and objectively stratify cancer aggressiveness. With further validation, qMIDS could enable early HNSCC detection and guide appropriate treatment. Early treatment intervention can lead to long-term reduction in healthcare costs and improve patient outcome.

Head and neck squamous cell carcinoma (HNSCC) is diagnosed in over half a million individuals worldwide each year, with an expected global incidence of 750,000 by 2015.1 Survival rates are poor (10–30% at 5 years) among patients presenting with advanced disease.2 Early detection of precancer lesions coupled with early intervention could significantly improve patient outcome, reduce mortality and alleviate healthcare costs.2, 3 However, conventional histopathology is currently unable to predict accurately which individual lesions from the oral potentially malignant disorders (OPMD)4 spectrum will transform to squamous cell carcinoma (SCC). Given similar pathogenesis of other epithelial SCCs, the same clinical dilemmas apply to the management of vulva and skin premalignancies.

Recent technological advances in whole-genome research have produced significant detailed biomolecular data across the human cancer transcriptome and cancer genome.5, 6 However, translation of these genome-wide data into clinical use remains a challenge and impractical.7 Currently, diagnostic histopathology and immunohistochemistry are the standard methods for cancer diagnosis, but they rely on a pathologist's subjective interpretation of microscopic structures of cells and tissue morphology as the basis for tumour diagnosis and grading. Because this diagnostic approach may be unreliable for detecting early malignancy,8, 9 there is a need for a practical, reproducible, objective and quantitative molecular method.

FOXM1 cell cycle transcription factor is involved in cancer initiation,10–13 progression14–16 and metastasis,10, 17–19 and ultimately drug resistance.12, 20–23 Here, we exploited the value of FOXM1 (isoform B)-orchestrated transcriptional changes in 14 target genes and 2 reference genes as a ‘molecular gauge’ to quantify malignancy status of tissue specimens using a highly sensitive absolute quantitative reverse transcription PCR (qPCR) method and demonstrate a prototype multibiomarker ‘quantitative molecular index diagnostic system’ (qMIDS) for squamous cell carcinoma (SCC) detection.

Material and Methods

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Clinical samples

The use of human tissue was approved by the relevant Research Ethics Committees at each institution [UK NREC: 06/MRE03/69 (head and neck), P/01/252 (vulva) 08/S1401/69 (skin) and Norway REK Vest: 2010/481-7 (head and neck)]. All clinical samples were collected according to local ethical committee-approved protocols and informed patient consent was obtained from all participants. Clinico-histopathological reports of the tissue samples were obtained from collaborating clinicians at each institution. For the UK head and neck cohorts, fresh biopsy tissues were preserved in RNALater (#AM7022, Ambion, Applied Biosystems, Warrington, UK) and stored short term at 4°C (1–7 days) before transportation and subsequent storage at −20°C until mRNA extraction (Dynabeads mRNA Direct kit, Invitrogen) and cDNA synthesis (Transcriptor cDNA Synthesis kit, Roche). For the Norwegian head and neck and UK vulva and skin cohorts, frozen archival biopsy tissues (embedded in OCT medium) and tissue cryosections (50 μm thick) were preserved in RNALater before mRNA extraction and cDNA synthesis. UK FFPE head and neck tissues (50–100 μm thick) were deparaffinised and rehydrated according to standard protocols. All samples (fresh, frozen or FFPE) were digested with nuclease-free proteinase K at 60°C before mRNA extraction. All samples were pseudo-anonymised and tested blindly to ensure that the qMIDS assays were performed objectively.

Cell culture

All primary normal human oral keratinocytes (OK355, HOKG, OK113, NOK, NOK1, NOK3, NOK16 and NOK376) were extracted from normal oral mucosa tissues donated by healthy disease-free individuals undergoing wisdom tooth extraction and cultured as previously described.14, 16 Dysplastic oral premalignant cell lines [OKF6/T,24 POE9n,25 DOK,26 D1927 and D2027] and HNSCC cell lines [SCC4,28 SCC9,28 SCC15,28 SCC25,28 SqCC/Y1,29 UK1,30 VB6,30 CaLH2,30 CaDec12,30 5PT30 and H35730], SVpgC2a31 and SVFN1-814 were all well characterized lines and were cultured as described previously10, 14, 16.

Real-time absolute quantitative RT-PCR

Real-time absolute quantitative PCR where mRNA (cDNA) transcript quantiation using a standard curve for each gene (Supporting Information Fig. S1) were performed using SYBR Green I Master (#04887352001, Roche Diagnostics, England, UK) in the LightCycler 480 qPCR system (Roche) based on our published protocols,14–16 which are MIQE compliant.32 Thermocycling begins with 95°C for 5 mins before 55 cycles of amplification at 95°C for 10 sec, 60°C for 6 sec, 72°C for 6 sec, 76°C for 1 sec (data acquisition). A ‘touch-down’ annealing temperature intervention (66°C starting temperature with a step-wise reduction of 0.6°C/cycle; 8 cycles) was introduced before the amplification step to maximise primer specificity. Melting analysis (95°C for 30 sec, 65°C for 30 sec, 65–99°C at a ramp rate of 0.11°C/sec with a continuous 5 acquisitions/°C) was performed at the end of qPCR amplification to validate single product amplification in each well. Absolute quantification of mRNA transcripts was calculated based on an objective method using the second derivative maximum algorithm33 (Roche). All qPCR primers and metadata associated with each biomarker panel used in this study are listed in Supporting Information Table ST1. All target gene mRNA (cDNA) transcripts in each sample were normalised to two stable reference genes (YAP1 and POLR2A) previously validated14 to be amongst the most stable reference genes across a wide variety of primary human epithelial cells, dysplastic and squamous carcinoma cell lines, using the GeNorm algorithm.34 Inter-experimental calibrators that give stable qMIDS indexes at 2.0 ± 0.5 or 12 ± 1.0 were included in every qPCR run to monitor assay quality. The qMIDS workflow and detail 384-well assay setup information are provided in Supporting Information Figure S2 and Figure S3, respectively. Assay reproducibility and robustness tests showed that the qMIDS score of a given sample was resistant to >10-fold variation in cDNA concentration (Supporting Information Fig. S4A) and 200% variation in reference standard concentration (Supporting Information Fig. S4B). Furthermore, a longitudinal assay robustness assessment, using data spanning 3 years containing 99 clinical specimens (consisting of 41 normal and 58 tumour samples randomly drawn from both the UK and Norway cohorts), showed no significant difference/shift in gene expression distribution across these different time points, ruling out all interexperimental variations because of reagent batch fluctuations, cDNA sampling techniques, handling and processing protocols and so on.

Gene selection

Target genes were shortlisted based on their statistically significant differential expression and ‘trend-forming’ expression pattern across the 24 cell line panel consisting of 8 primary normal oral keratinocytes, 5 dysplastic and 11 HNSCC cell lines. Statistical t-test and curve-fitting polynomial regression analysis were performed on both raw data points (target gene mRNA transcripts) and Log2 ratio (i.e., Log2 [T(T14m)/Tm(T14)]) data points (see Algorithm [1]). Statistical t-test was performed to determine the significance of differential gene expression between normal (n = 8) and disease cell lines, where the minimum threshold level of p < 0.05 was adopted as our first biomarker selection criterion. Analysis of variance (ANOVA) was performed to determine the significance of differential gene expression across multiple groups. Polynomial curve-fitting (y = a+b1x+b2x2+b3x3+b4x4) regression analysis was performed on both raw and Log2 ratio data of each target gene to survey its ‘trend-forming’ expression pattern during tumour progression. Statistical p-values are shown in Supporting Information Table ST2. The qMIDS diagnostic assay efficiency tests were performed according to the STARD Initiative recommended protocol.35

The qMIDS algorithm

  • equation image(1)

Algorithm [1] was developed for computing and translating multigene expression signatures of each cDNA sample into a qMIDS malignancy index (MI). T represents the target gene mRNA transcripts (normalised against 2 reference genes); Tn represents the sum of n number of target gene mRNA transcripts measured in each cDNA sample; Tm represents a median value of T derived from a control set of 8 independent healthy primary NHOKs; Tnm represents the sum of the n number of Tm values. Q1, Q3 and Q4 represent the first (25%), third (75%) and forth (100%) rank quartile of the x number of target gene absolute fractional Log2 ratio (|Log2 [T(Tnm)/Tm(Tn)]|) distribution values within each cDNA sample. MI values were then proportionately calculated to fit a linear scale.

Statistical analysis

The qMIDS scores were examined in relation to sex, ethnic group and age using non-parametric tests (Kruskal-Wallis ANOVA and Spearman's correlation coefficient). Test performance was estimated at specified qMIDS cut-off values using (i) the detection rate (DR), the percentage of affected samples or subjects who have scores above the cut-off, and (ii) false-positive rate (FPR), the percentage of unaffected samples or subjects who have scores above the cut-off. ‘Affected’ are those with cancer or precancer lesions. ‘Unaffected’ are those with healthy tissue, do not have cancer or have oral lichen planus (OLP) or non-malignant fibro-epithelial polyps (FEP). DR and FPR were obtained from the observed data and also estimated using Gaussian distributions based on the median and standard deviations. Overall HNSCC survival was examined using Kaplan-Meier plots and Cox regression among 97 (out of 105) Norwegian patients for whom data were available on date of death or date last seen alive. Head and neck FFPE, vulva and skin data were analysed in R (version 2.13.1; The R Foundation for Statistical Computing) and plotted using Beeswarm Boxplot software package.36

Results

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Gene selection, functional validation and assay development

Two hundred putative FOXM1-associated genes were initially selected using a bioinformatics meta-analysis based on 28,880 published microarray datasets containing over 791 million data points from 41 different human cancer types available in Oncomine™ Research V3.6 (www.oncomine.org). The main selection criterion was the significance of genes whose mRNA expression levels correlated either positively or negatively with FOXM1 expression. Of the 200 FOXM1-associated genes, 128 genes were found to be differentially expressed when we cross-examined each gene in a number of published Affymetrix microarray datasets comparing normal against several types of SCC tissues (including skin, lung, oral and cervical) using the NCBI's Gene Expression Omnibus (GEO) data-mining research tool as described previously.14, 15 To validate whether the 128 putative genes were indeed transcriptionally associated with the cell-cycle regulated FOXM1 isoform B expression,10 their mRNA transcript levels were measured using absolute reverse transcription qPCR (see Supporting Information) for each gene in primary normal human oral keratinocytes (NHOKs) retrovirally transduced with either a control EGFP or FOXM1 (isoform B) transgene.10, 14, 16, 37

Of the 128 genes, 43 genes were found to be differentially expressed (minimum twofold at p < 0.05) in FOXM1 (isoform B)-transduced cells compared with control cells (Figs. 1a and 1b). To investigate whether the associations of the 43 genes with FOXM1 (isoform B) expression were dose-dependent on tumour progression, we performed absolute qPCR for each of the 43 target genes, 3 FOXM1 isoforms (A, B and C) and 2 reference genes (total 48 genes) across a training panel of 8 independent primary NHOKs (isolated from oral mucosa tissues donated by healthy disease-free individuals), 5 oral precancer (dysplastic lines) and 11 malignant HNSCC cell types (Fig. 1c). Our preliminary data using isoform-nonspecific FOXM1 primers gave less significant results than using FOXM1 isoform-specific primers. All three FOXM1 isoforms, especially isoforms A and B, showed highly significant upregulation in HNSCC (data not shown). This indicates transcriptional upregulation at the gene promoter level. Our previous study10 established that FOXM1 isoform B was the only isoform showing cell-cycle dependent expression patterns, we therefore selected FOXM1 isoform B as a biomarker for its functional significance and to avoid redundancy.

thumbnail image

Figure 1. A systematic framework for the selection of FOXM1-orchestrated genes involved in cancer progression. (a) Two hundred FOXM1-associated genes including both known and novel target genes were initially selected using bioinformatics meta-mining analysis. Hundered and twenty eight genes were found to be putative FOXM1-orchestrated genes in human cancers. Red and green lines represent up- and down-regulated genes between normal and cancer samples. (b) Functional analysis of FOXM1B-induced transcriptional activation and suppression was performed using primary human oral keratinocytes virally transduced with either EGFP control gene or FOXM1B proto-oncogene. The mRNA expression levels of each of the 128 genes in the transduced cells were measured using absolute qPCR. Of the 128 genes, 43 showed statistically significant (P<0.05; >2-fold) differential expression in the FOXM1B-transduced cells compared to controls. (c) A systematic gene expression screening of 46 target genes (including 3 FOXM1 isoforms) and 2 reference genes across a panel of 8 normal primary oral keratinocytes (Normal), 5 oral dysplastic premalignant (Pre-Cancer) and 11 HNSCC cell lines (see Fig 2 for further details). (d) Genes that did not show statistically significant ‘trend-forming’ expression patterns were removed from the study. The remaining panel of trend-forming genes were used to develop the qMIDS algorithm. (e) Differential mRNA expression profile (Log2 Ratio) of 16 genes across the cell line panel. The x-axis represents a progressive ranking of each cell line based on the sum of the differential expression of 14 target to 2 reference gene ratios as indicated in the equation. (f) Individual gene expression profile across the cell line panel with polynomial curve-fitting analysis and R2 values indicated in each figure panel. Each data point indicates the relative gene expression level within each cell type. Twelve genes (red) show progressive upregulation and 2 genes (green) showed progressive downregulation. The two reference genes (black) showed stable expression across all the cell lines.

Download figure to PowerPoint

Student t-tests were performed to determine the statistical significance of differential gene expression between normal (n = 8) and disease cell lines including both oral precancer (n = 5) and HNSCC cell lines (n = 11)] where the minimum threshold level of p < 0.05 was adopted as our first biomarker selection criterion. Genes that showed insignificant differential expression (p > 0.05) were excluded (Fig. 1d), leaving 14 target genes that showed statistically significant ‘trend-forming’ expression pattern across the cell panel (Figs. 1e, and 1f).

Of the 14 target genes (Fig. 2), 12 genes (HOXA7, AURKA, NEK2, FOXM1B, CCNB1, CEP55, CENPA, DNMT3B, DNMT1, HELLS, MAPK8 and BMI1) showed progressive upregulation and two genes (ITGB1 and IVL) showed progressive downregulation across the cell panel. Collectively, these gene expression patterns were consistent with previous findings that upregulation of FOXM1 (isoform B) induces genomic instability14, 16 and perturbs terminal differentiation.10 Interestingly, 6 (FOXM1, CEP55, AURKA, NEK2, CCNB1 and CENPA) of the 12 upregulated genes identified in this study were recently reported to be amongst the 67 prognostic biomarkers that measure genome complexity, which predicts the patient outcome of sarcoma, gastrointestinal stromal tumours, breast carcinomas and lymphoma.19 The two previously validated reference genes YAP1 and POLR2A14–16 showed stable expression across all different cell types used in this study (Fig. 1f). In order to obtain a clinically meaningful quantitative score from the panel of 16 genes (14 target plus 2 reference genes) for cancer diagnosis and tumour grading/ranking, we have prospectively designed an algorithm to compute and translate gene expression signatures into a qMIDS malignancy index (MI) scoring system. The qMIDS algorithm computes the total differential expression of the 14 target genes relative to the median normal control gene expression levels found in a panel of eight healthy primary normal oral keratinocyte strains (Fig. 2).

thumbnail image

Figure 2. Panel of 16 genes and the qMIDS algorithm for translating gene signatures into a qMIDS MI matrix scoring system. The qMIDS assay involves quantification of absolute mRNA levels of 14 target and 2 reference genes from each sample as indicated. The qMIDS algorithm was developed for translating multigene expression signatures within each cDNA sample into a qMIDS MI. T represents the target gene mRNA transcripts (normalised against 2 reference genes); Tn represents the sum of n number of target gene mRNA transcripts measured in each cDNA sample; Tm represents a median value of T derived from a control set of 8 independent healthy primary NHOKs; Tnm represents the sum of the n number of Tm values. Q1, Q3 and Q4 represent the first (25%), third (75%) and forth (100%) rank quartile of the x number of target gene absolute fractional Log2 ratio (|Log2 [T(Tnm)/Tm(Tn)]|) distribution values within each cDNA sample. MI values were proportionately calculated to fit a linear scale. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Download figure to PowerPoint

Proof-of-principle study using clinicopathologically pre-defined head and neck tissues

We tested qMIDS on two geographically independent patient cohorts from the UK and Norway (Fig. 3). Ninety-nine individuals in the UK cohort prospectively provided head and neck tissue biopsies (Nov 2008–July 2011) (36 with precancer, 53 with HNSCC, and 10 who had neither precancer nor HNSCC); 29 patients each donated a single biopsy and 70 each donated ≥2 biopsies (with samples from affected and unaffected areas). When classified according to histopathological features, n = 41 patients had biopsies from normal margins/unaffected, n = 55 with dysplasia, n = 48 with SCC, and n = 15 with lymph node metastasis from primary SCC within the head and neck region. Twenty archival FFPE samples (UK) consisted of n = 5 from each of fibro-epithelial polyps; dysplasia; primary head and neck SCC and lymph node metastasis. The Norwegian cohort consisted of retrospective frozen archival biopsies (Feb 1991–Nov 2010) donated by 200 individuals [histopathological classification: n = 57 unaffected, n = 38 oral lichen planus (OLP), and n = 105 HNSCC patients].

thumbnail image

Figure 3. Patient demography (age, sex, ethnicity and substance habit), tissue anatomical sites and histological classification of head and neck specimens from the UK and Norway subjects.

Download figure to PowerPoint

Sex and age distributions in the UK and Norwegian groups were similar: males (56% UK and 54% Norway), median age of ∼60 years for UK and 59 years for Norway. In the UK group, 61% were Caucasian and 36% were Asians (mainly from Bangladesh, Pakistan and India), whereas the Norwegian group had 98% Caucasian. Tissues from the UK subjects originated mostly from the buccal mucosa (33%), tongue (26%), mandible (8%) and floor of mouth (8%), whilst tissues from the Norwegian patients originated mainly from gingiva (34%), tongue (21%), buccal (19%) and larynx (9%) (Fig. 3).

There was a significant difference between the histopathological groups within both the UK (median qMIDS MI scores 1.3, 2.9 and 6.7 in those with healthy tissue, dysplasia and cancer respectively, p < 0.001, Fig. 4a), and Norway (median score 1.4, 2.3 and 7.6 in those who were unaffected, had oral lichen planus, and cancer respectively, p < 0.001, Fig. 4b). It is interesting to note that two oral lichen planus patients with outlying qMIDS values (4.4 and 5.6 as opposed to a median qMIDS of 2.3; Fig. 4b) were coincidentally diagnosed of breast and colon cancers, respectively.

thumbnail image

Figure 4. Validation of qMIDS in the (a) UK and (b) Norway subjects. Note that in the UK, a subject could provide more than one histological sample type and therefore appear in more than one column on the scatter plot (this is not the case in the Norway group). The mean and 95% confidence intervals are as indicated. (c) Overall survival according to qMIDS values (based on tertiles among patients with HNSCC). The hazard ratios (95% CI) were 1.31 (0.72–2.38) and 1.56 (0.87–2.81) in the middle and highest tertile groups respectively, compared to the lowest category (≤6.39). The median survivals were 5.5, 3.4 and 1.4 years in the lowest, middle and highest tertile groups respectively (p = 0.32). D, Validating qMIDS on archival FFPE head and neck tissues. Data in d and e were plotted as dot-plot with box-and-whisker overlays (median and 25–75% percentiles). An optimum cut-off at four was found for this dataset providing a DR of 83% and FPR of 0%. Statistical t-test between FEP and the disease group (Dysp, HNSCC and LnMets) was found to be significant (p = 4 × 10−3). E, Segregation of disease group according to histopathology: FEP, fibro-epithelial polyps; Dysp, dysplasia; HNSCC, primary head and neck SCC; LnMets, lymph node metastasis.

Download figure to PowerPoint

In the UK group, those with SCC and lymph node metastases had similar qMIDS values, and were therefore subsequently combined for analysis. We saw significant qMIDS variability within the UK dysplasia samples. Although the qMIDS median values appeared to be progressively increased with dysplasia grades (mild, moderate and severe), the overall correlation between qMIDS and histopathological grading of dysplasia was not statistically significant (Supporting Information Fig. S6). As these UK dysplasia samples were prospectively collected, we were unable to investigate the relationship between qMIDS and long-term clinical outcome. Nevertheless, qMIDS appeared to correlate inversely (i.e., poorly differentiated samples tend to have higher qMIDS values, vice versa) with differentiation status of HNSCC samples in both the UK and Norway cohorts (Supporting Information Fig. S7).

The qMIDS scores of normal margins or healthy subjects without cancer were well separated from those with cancer in both the UK and Norway groups. The qMIDS scores did not materially differ according to age, sex or ethnic group, except that males with HNSCC appeared to have even higher qMIDS scores than females (Supporting Information Table ST3). Data on chewing habits in the cohorts were sparse, but among those known to have chewed tobacco, two subjects with healthy tissue had qMIDS values of 0.9 and 1.7, and n = 7 with dysplasia had median 4.8 (not statistically significant from the rest of the group, p = 0.59).

Table 1 shows the observed and modelled detection rate (DR) and false-positive rate (FPR) in each of the histopathological groups (the modelled estimates use parameters in Supporting Information Table ST4). Test performance was good. At a qMIDS cut-off of ≥4.0, 94% of those with HNSCC could be detected compared with only 3.2% of those without cancer (Norway). These were close to the observed values: DR = 97% (95% CI 94–100%) and FPR = 3.5% (95% CI 0–8.3%). The results were consistent with patients who provided both healthy and cancerous tissues (UK); DR = 91% and FPR = 1.3%. Even a lower cut-off of 3.5 might be appropriate. Among those with dysplasia in the UK cohort, 34 and 25% had qMIDS values ≥3.5 or ≥4.0, respectively. The FPR associated with oral lichen planus was relatively low; 15.2 and 7.3% for qMIDS values of ≥3.5 or ≥4.0.

Table 1. Observed and modelled test performance of qMIDS in the UK and Norway patients
inline image

Figure 4c shows overall survival according to qMIDS tertiles, among 97 Norwegian patients with HNSCC, of whom 72 had died. Although not statistically significant (logrank p = 0.32), there was a trend towards a difference, where the median survival was 5.5, 3.4 and 1.4 years in the lowest, middle and highest tertile groups. From a Cox regression (qMIDS as a continuous variable), for every increase of 1 unit the risk of dying increased by 8% (hazard ratio 1.08, 95% CI 0.97 to 1.20, p = 0.17). These data suggest a correlation between qMIDS index and tumour aggressiveness.

In addition, we tested qMIDS on a small cohort of formalin-fixed paraffin-embedded (FFPE) head and neck tissues (n = 20). Of these, three samples were of poor RNA quality and hence were excluded. The remaining 17 samples, consisting of n = 5 nonmalignant fibro-epithelial polyps (FEP), n = 4 oral dysplasia (Dysp), n = 5 primary HNSCC and n = 3 lymph node metastasis (LnMet) provided good qMIDS data. At an optimal qMIDS index cut-off at 4 (derived from both the UK and Norway cohorts), DR and FPR were 83 and 0%, respectively (Fig. 4d). When the FFPE samples were separated according to histopathological classification, qMIDS values showed progressive increase with malignancy status (Fig. 4e). These results showed that molecular signature within FFPE tissue was sufficiently preserved for detection by the qMIDS assay. There is a notable difference in tissue type stratification between data in Figures 4a and 4e. This may be due to the difference in the method of tissue preservation before assay. It is known that nucleic acids in FFPE tissues are significantly degraded on extraction. However, despite the FFPE-induced nucleic acid degradation, we have demonstrated that the qMIDS assay, which was developed using fresh frozen samples, was able to significantly stratify between different sample types, albeit giving more noise in the data. Further study using a larger number of FFPE samples is necessary to provide sufficient training data for customising the qMIDS algorithm to obtain better DRs when using FFPE tissues.

Tissue topological micro-mapping and tumour margin assessment using qMIDS

The high sensitivity and accuracy of the qMIDS prompted us to investigate if the qMIDS could be used as a topological tool to map tumour heterogeneity and/or to locate tumour margin, we performed macro-dissection on two large HNSCCs (HNSCC1, stage-T2 tumour core in Fig. 5a and HNSCC2, stage-T4 tumour core in Fig. 5b) and two premalignant lesions (PML1, white leukoplakia with margin in Fig. 5c and PML2, red leukoplakia with margin in Fig. 5d) into 1–2 mm3 (0.5–1 mg) fragments. Messenger RNA was extracted and qMIDS performed on each fragment independently. qMIDS scores obtained from each fragment were reassembled to provide a malignancy ‘heat map’. Using this method we were able to simultaneously quantify malignancy status and locate tumour margin of a given tissue based on its molecular topology. Although the current technique demonstrated a 2D topology mapping, it is amenable to generate 3D topology by systematic dissection of a given tissue with fixed tissue depth. This study highlights the problem of tumour heterogeneity and we have shown that diagnostic accuracy/sensitivity could be improved by analysing smaller tissue fragment size thereby reducing the ratio of contaminating normal tissues to malignant cells within a given tissue fragment. By titrating the ratio between normal and cancer cells, we have shown that the qMIDS can significantly detect as little as 17% of malignant SCC cells within a given sample containing normal oral keratinocytes (Supporting Information Fig. S5). Given such high level of diagnostic sensitivity, the qMIDS enabled us to study the molecular heterogeneity in tumour tissues that were found to exist in multiple subtypes as illustrated in Figure 5e. The clinical significance of these distinct molecular subtypes is beyond the scope of this study but warrants further investigation into the prognostic value of each molecular subtype.

thumbnail image

Figure 5. Tumour molecular topology and margin analysis using qMIDS. A, stage-T2 HNSCC1 and B, stage-T4 HNSCC2) and two leukoplakia samples (c, white leukoplakia PML1 and d, red leukoplakia PML2). Each tissue was sub-dissected into 1–2 mm3 (0.5–1 mg) fragments and mRNA was extracted, converted to cDNA and performed qMIDS to obtain MI scores for each fragment. qMIDS scores were converted into colour scale to generate a topological heat map displaying areas of advance, moderate and normal tumour margins of the tissue sample. (e) A schematic diagram illustrating that the high resolution diagnostic capability of qMIDS enabled us to study the molecular heterogeneity of tumour tissues.

Download figure to PowerPoint

Adapting qMIDS for vulva and skin SCCs

Given the fact that FOXM1 is upregulated in the majority of human cancers,38, 39 we questioned whether the qMIDS developed for HNSCCs could be used for diagnosing vulva and skin SCCs progression given that these are epithelial squamous cell of origin. We therefore tested qMIDS on 55 vulva (n = 13 normal vulva/NV, n = 5 vulva carcinoma in situ/VIN1, n = 8 VIN2, n = 24 VIN3 and n = 5 VSCC) and 21 skin specimens (n = 6 normal skin/NS, n = 4 actinic keratosis/AK, n = 5 carcinoma in situ/CIS and n = 6 SSCC). For the vulva group, with a qMIDS index optimal cut-off at 1, DR and FPR were 93 and 23%, respectively (Supporting Information Fig. S8A). When the vulva samples were classified according to histopathology, qMIDS values showed progressive increase with malignancy status (Supporting Information Fig. S8B). For skin, with a qMIDS index optimal cut-off at 4, DR and FPR were 80 and 0%, respectively (Supporting Information Fig. S8C). When the skin samples were classified according to histopathology, qMIDS values also showed progressive increase with malignancy status (Supporting Information Fig. S8D). Using the same strategy described earlier for HNSCC, a new panel of biomarkers could be prospectively selected using skin or vulva cell lines or tissue specimens to improve DRs and FPRs for respective tissue types. Nevertheless, these data demonstrated that the qMIDS provides a customisable framework for measuring different forms of cancers.

Discussion

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

By exploiting the aberrant expression of FOXM1 found in many human cancers,38 we have developed a novel molecular diagnostic system that quantify the levels of cell proliferation,38 differentiation,10 ageing,40 genomic instability,2, 14, 16, 17, 19 epigenetic11, 14 and stem cell reprogramming10, 13, 41, 42 as a collective basis for cancer diagnosis. Our proof-of-principal studies tested qMIDS on multiple cohorts of patients with HNSCC and demonstrated that the qMIDS assay is a practical, sensitive, objective, quantitative and customisable method for cancer diagnosis. Our study involving both prospective and retrospective cohorts from two geographically independent subjects demonstrates that the qMIDS assay is a highly efficient system for HNSCC diagnosis. It has high DRs (91–94%) for those with cancer and low FPRs (3.2%) for those without cancer.

Current screening methods for HNSCC in otherwise symptom-free persons include the use of oral cytology (brush biopsy), toludine blue staining and various light-based detection systems.43, 44 More advanced screening methods such as salivary proteomics45 and antibody-based detection46 are under investigation. However, the effectiveness of these oral screening adjuncts in detecting early cancer remains unproven.43, 44, 47 Hence, tissue biopsy remains the most trusted and reliable sample type for informing treatment decision especially when it involves surgical intervention. Histopathology, the standard method of HNSCC diagnosis, has key limitations: it is time-consuming (usually several days for tissue fixing and reporting) and the result is subject to pathologists' interpretation.48 In addition, the presence of dysplasia is often missed because molecular changes indicative of malignant transformation do not necessarily produce clinically or histopathologically detectable changes.8 Analysis of all current methods for oral cancer screening, diagnosis and prognosis (Supporting Information Table ST5) indicates a clinical need for a practical, rapid, quantitative and decisive molecular method to complement histopathology.

A clinically pressing issue is the management of oral potentially malignant disorder (OPMD), which is a notorious clinical dilemma due to the absence of a reliable diagnostic test that can segregate low from high risk subjects with OPMD.47 The majority of HNSCC cases are often preceded by lesions of which leukoplakia is the most common form of OPMD.47 However, some forms of leukoplakia, such as homogenous leukoplakia and oral lichen planus (OLP), have very low risk of malignant transformation, whereas other forms of leukoplakia, such as nonhomogenous leukoplakia, proliferative verrucous leukoplakia, submucous fibrosis and sublingual keratosis, have varying degree of risk of transformation into malignant HNSCC, between 5–30% over 10 years.47, 48 To rule out the possibility that inflammation, often associated with malignancies, may be a confounding factor for qMIDS, we therefore tested OLP which is a benign autoimmune condition. Our data showed that only 7.3% of people with OLP have qMIDS scores ≥4.0, ruling out inflammation as a confounding factor for qMIDS. This data also indicates (or adding further evidence of) that OLP is a relatively low risk lesion for malignant transformation.47, 48

The qMIDS assay objectively measures the ‘malignancy status’ of a biopsy tissue sample using molecular signatures of multiple FOXM1 (isoform B)-orchestrated biomarkers.14, 16, 17, 19 This involves a novel algorithm that produces a quantitative ‘MI’ based on the gene expression signatures of 14 biomarkers and 2 stable reference genes determined by absolute real-time reverse transcription quantitative PCR (qPCR)—an established system now widely available in many diagnostic laboratories. It is highly sensitive, fast, reproducible and amenable to high-throughput automation, without requiring skilled personnel, and therefore, minimising human error/bias in interpreting results. As opposed to gene expression microarray technology, which is at best semiquantitative, absolute qPCR enables accurate quantitative measurement of low abundance mRNAs such as transcription factors. Although a number of clinical studies have already demonstrated the significance of detecting genes involved in genomic stability for cancer prognosis,17–19 to our knowledge the qMIDS index represents the first practical method not only for detecting but also for quantifying the levels of genomic instability generating a ‘MI’ system for SCC diagnosis. Unlike, for example, the microarray-based MammaPrint™ (Agendia BV, The Netherlands) that is non-quantitative and produces dichotomous disease classification (low or high risk),49 but similar to the qPCR-based Oncotype DX™ (Genomic Health, CA)50 and Pinpoint DX Lung™ (Pinpoint Genomics, CA),51 the qMIDS produces a continuous metric scoring system that provides clinicians with a more informative readout of disease risk status.

Furthermore, we have also demonstrated that, by macro-dissecting biopsy samples into smaller fragments (1–2 mm2), the qMIDS could be used to simultaneously quantify malignancy status and provide a molecular tissue topology of the tumour sample, which is important for tumour margin assessment and determination of tumour heterogeneity. Unlike histopathology, qMIDS does not require tissue biopsies to be carefully fixed and embedded in correct orientation by trained personnel thereby minimising sample handling errors and reducing staffing costs. Given the ability to assay very small specimens, multiple biopsies could be performed to identify high risk areas within the oral cavity with field change over the whole oral mucosa. Alternatively, having shown that archival frozen and FFPE tissues were also amenable to this assay, qMIDS could be used as an adjunct to complement histopathology.

While this study validates qMIDS as a diagnostic test for early cancer detection, the next step is to examine qMIDS in a larger prospective study, focussing on the natural history of subjects with precancer lesions, and the relationship between the qMIDS score and progression to cancer. In the currently prospective UK cohort, we saw a significant number of dysplasia samples with qMIDS values discordant with histopathological dysplasia grading—a method known to be highly subjective and no prognostic value. We are currently collecting clinical follow-up data to investigate whether qMIDS could be used to segregate these patients with dysplasia into low and high risk groups. Such a study would enable us to determine the prognostic value of qMIDS for stratifying the cancer risk in patients presenting precancer lesions.

Accumulating evidence shows that FOXM1 is a marker of poor cancer prognosis in oral SCC,52 ovarian carcinomas,53 breast carcinomas,18 pancreatic ductal adenocarcinoma,54 metastatic sarcomas,19 gastrointestinal stromal tumours,19 lymphomas,19 glioblastoma55 and medulloblastoma.56 Furthermore, FOXM1 mediates therapeutic resistance to tamoxifen,22 cisplatin,21 trastuzumab20 and paclitaxel20 in breast cancer cells. This can lead to companion diagnostics whereby cancer therapies could be tailored to a patient's tumour FOXM1 expression status. Given that qMIDS is based on FOXM1-associated genes, it could potentially be used as a personalised companion diagnostic method to improve patient drug treatment response.

In summary, we have developed a multibiomarker molecular diagnostic system, based on FOXM1-orchestrated transcriptional signatures, for objective and quantitative SCC diagnosis. This proof of principle study demonstrates that qMIDS is capable of segregating and also quantifying the malignancy status of clinical tissue biopsies with a high degree of confidence. Such a quantitative diagnostic system will significantly improve the sensitivity and accuracy of HNSCC diagnosis leading to lower iatrogenic morbidity and improve patient outcome. We have also provided evidence that the qMIDS was transferable for diagnosing vulva and skin SCCs, in which there are similar issues such as field change and correct identification of high risk lesions. This study therefore demonstrated a practical FOXM1-based molecular diagnostic system for rapid, highly sensitive, reproducible, quantitative detection and stratification of cancer aggressiveness.

Acknowledgements

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

The authors thank Professor Ian Mackenzie (Centre for Cutaneous Research, Blizard Institute of Cell and Molecular Science) and Professor Ken Parkinson (CDOS) for their critical advice and precious gifts of various oral SCC cell lines used in this study. They also thank Dr. Emilios Gemenetzidis for retroviral transduction, culturing cells and preparing cDNA samples from cell lines. They are indebted to Dr. Tom Vulliamy (Blizard Institute of Cell and Molecular Science) for his critical comments on the manuscript. The authors are extremely grateful to Dr. Karen Gibbon for granting access to her collection of frozen vulval tissue and to Dr Naveena Singh for histopathologic evaluation of the vulval sample series. The also thank Mr. Muy-Lip Teh (Air Operations Software Engineering Department, Thales Australia Centre, Melbourne, Australia) for his advice on data processing and visualisation using R Environment.

References

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Supporting Information

  1. Top of page
  2. Abstract
  3. Material and Methods
  4. Results
  5. Discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Additional Supporting Information may be found in the online version of this article.

FilenameFormatSizeDescription
IJC_27886_sm_SuppInfo.pdf2420KSupporting Information

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.