A gene signature in histologically normal surgical margins is predictive of oral carcinoma recurrence

This work was performed with the approval of the University Health Network Research Ethics Board. All patients signed their informed consent before sample collection, and had surgery as their primary treatment. Tissue samples were prospectively collected at time of surgery from the Toronto General Hospital, Toronto, Ontario, Canada. The inclusion criteria of patients were based on primary disease, tumor site (oral cavity), histology of the primary tumor (squamous cell carcinoma), and histological status of surgical resection margins. All patients included in this study had a final pathology report of histologically normal resection margins, meaning none of the patients had a positive margin. An experienced head and neck pathologist (BP-O) performed histological evaluation of all surgical margins (frozen section analysis, as per standard of care) to ensure that they were histologically normal. A section from each tumor sample (OSCC) was stained with hematoxylin-eosin, to confirm the presence of at least 80% tumor cells. The primary tumor, histologically normal margins and a distant normal oral tissue were collected from each patient, in both training and validation cohorts. All tissue samples were snap-frozen in liquid nitrogen until RNA extraction and gene expression analysis. We collected a total of 232 tissue samples from 54 patients (divided into training and validation sets), as described below.

96 samples (histologically normal margins, OSCC and adjacent normal oral tissues) from 24 patients were used for oligonucleotide microarray analysis. Patient clinical data for the training set are shown in Table 1.

136 samples (histologically normal margins, OSCC and adjacent normal oral tissues) from an independent cohort of 30 patients were used for RQ-PCR validation analysis. Patient clinical data for the validation set samples are shown in Table 2.

The detailed protocol for RNA isolation and the experimental details for microarrays using the Human Genome HG-U133A 2.0 plus oligonucleotide arrays (Affymetrix, Santa Clara, CA, USA), as well as validation experiments by RQ-PCR are described in Additional file 1, Methods S1.

All bioinformatic analyses of array data were performed in the R language and environment for statistical computing (version 2.10.0) implemented on CentOS 5.1 on an IBM HS21 Linux cluster. We pre-processed each public data set using GCRMA normalization [17] with updated Entrez Gene-based chip definition files [18], using the affy R package (version 1.24.2) [19]. Genes with evidence of tumor-normal differential expression across all public datasets, with a False Discovery Rate (FDR) of 0.01 and fold-change ≥ 2, were identified using the rankprod R package [20]. Gene Ontology enrichment analysis was performed with the GOstats R package (version 2.12.0) [21]. We also used GOEAST (Go Enrichment Analysis Software Toolkit) [22] for graphical representation of GO annotations. Analyses are reproducible using the code (available upon request); data have been deposited in NCBI's Gene Expression Omnibus [23] and are accessible through GEO series accession number GSE31056 http://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​acc=​GSE31056.

Our goal was to identify a prognostic signature for recurrence in OSCC, based on the hypothesis that gene expression deregulation occurring in OSCC would be an early indicator of recurrence if gene expression changes are present in a subset of histologically normal surgical resection margins. We performed gene expression profiling of both resection margins and tumors with the purpose of (1) identifying over-expressed genes in tumors as potential markers of recurrence in histologically normal margins, and (2) finding a subset of those genes predictive of recurrence. In order to generate a very high-confidence gene set, we augmented the analysis of our data with a meta-analysis of five published microarray studies [24‐28] to reliably identify a set of genes significantly deregulated in OSCC compared to normal oral tissues. These five public data sets were selected based on the availability of raw microarray data, as well as for the inclusion of both oral cavity tumors and either adjacent normal tissues or oral tissues from healthy individuals. Although data from Pyeon et al. [28] included HPV positive and HPV negative head and neck carcinomas from different anatomic sites, we selected oral carcinomas only, which are mainly negative for HPV infection, as shown in a recent study performed by our group [29]. This meta-analysis sample set was composed of a total of 199 samples (141 OSCCs, 38 adjacent normal tissues and 20 healthy normal tissues) from 141 oral cancer patients and 20 healthy individuals (without cancer) (Table 3). We pre-processed data from the different array platforms with updated chip definition files, as described above, to correct outdated probe mapping information from older platforms. We used a Rank Product analysis for the public studies, which considered only the ranking of genes by differential expression between pairs of samples within studies [20, 30], avoiding batch and platform-related effects which would occur from directly combining expression values from the different studies. Genes were selected with evidence of up-regulation in tumors with a False Discovery Rate (FDR) of 0.01 and fold-change ≥ 2. We chose to focus on over-expressed genes only, since histologically normal margins may contain only a fraction of genetically altered cells, and the presence of genetically normal cells would likely make down-regulated genes unreliable markers. By using the intersection of genes identified both by meta-analysis and the in-house array training set, we retained only genes that were reproducibly over-expressed compared to normal oral tissues from healthy patients and histologically normal margins. These strict selection criteria for gene signature candidates, based on prior hypothesis, helped to reduce the risk of over-fitting during Cox regression analysis.

Microarray results from our study were normalized as described above, along with normal oral tissue samples from healthy individuals also assayed by Affymetrix Human Genome U133 Plus 2.0 oligonucleotide arrays (downloaded from GEO accession number GSE6791). Probesets with low expression (75^th percentile below log₂(100)) or low variance (IQR on log₂ scale < 0.5) were filtered [31] as well as the quality control probe sets. The treat function from LIMMA: Linear Models for Microarray Analysis (version 3.2.1) [32] was used to identify genes ≥ 2-fold up-regulated in tumors compared to margins from our study, with FDR = 0.01. We took the intersection of genes identified in both the meta-analysis and in-house microarray experiment, and this intersection analysis produced a list of 138 genes, which were used as the feature set for developing the gene signature for recurrence.

LASSO penalized Cox regression was used to train a predictive risk score from the maximum expression of the 138 genes in any margin of each patient (all these 138 genes being found to be over-expressed in OSCC). Normalized expression values were first converted to Z-scores for each gene, to ensure equal effect of the penalty on all coefficients, as is standard for penalized regression. We applied LASSO penalized Cox regression as implemented in the penalized R package (version 0.9-27) [33], to condition a linear risk score with time to recurrence as the event of interest. In keeping with the hypothesis that over-expression of these genes is predictive of recurrence, we constrained the regression coefficients to positive values, and selected the L1 penalty parameter by optimizing 10-fold cross-validated likelihood. Since the cross-validated likelihood depends on the randomized folding of the data, we repeated the regression 50 times and selected the model with the highest cross-validated likelihood.

As shown in Table 1 and Table 2, 9/24 patients (training set) and 7/30 patients (independent validation set) had disease recurrence. Median time to local recurrence (by Kaplan Meier estimate) of patients in the training set was 32 months (range 2-34 months). Similarly, patients from the validation set recurred within 2-36 months. All patients had local recurrence, and some patients also had regional and/or distant failure (detailed data in Tables 1 and 2). Median (by reverse Kaplan-Meier estimate) and range of follow-up times of patients were 20 months (1.4-57 months) in the training set and 23 months (1-81 months) in the validation set.

Meta-analysis of the five public data sets (Table 3) identified 667 up-regulated genes in OSCC; results are provided in Additional file 2, Table S1. Requiring two-fold up-regulation in tumors in both the meta-analysis of public datasets and our microarray experiment, with FDR < 0.01, identified 138 over-expressed genes (Additional file 3, Table S2). The expression patterns of these genes in tumors, margins, and normal oral tissue samples from healthy individuals are shown as a heatmap in Figure 1. As seen in the hierarchical clustering, the 138 genes accurately discriminate between the tumors and margins of the in-house study, as well as normal oral tissues from a publicly available study (as described in Methods). We performed LASSO penalized regression of these 138 genes and selected a total of six genes. Of these six genes, we retained four genes with the largest coefficients (MMP1, COL4A1, P4HA2 and THBS2), and eliminated two genes with very small coefficients (PXDN and PMEPA1), which together contributed less than 5% of the magnitude of the risk score. These four genes constituted a 4-gene signature of recurrence. In the heatmap, Gene cluster "B" contains three of the four genes in the signature (COL4A1, P4HA2 and THBS2), and it shows frequent up-regulation in the surgical margins compared to the normal oral tissues. Strikingly, MMP1, found in gene cluster "A", shows less frequent over-expression in the margins, but has extreme differential expression between margins and OSCCs (400-fold up-regulation as detected by microarrays, and 800-fold up-regulation, validated by RQ-PCR). The proteins encoded by the 138 genes are also shown in a protein interaction network that highlights the most highly inter-connected proteins (Figure 2 and Additional file 1, Methods S1). In the heatmap, the main features of clusters A and B are the large number of interacting MMP proteins in cluster A, which contains MMP1, and collagens plus TGFB1 in cluster B, which also contains P4HA2, THBS2 and COL4A1 genes of the signature. The large number of MMPs and collagen proteins are closely connected; in particular, MMP9 interacts with both THBS2 and COL4A1, and indirectly with MMP1.

Results of Gene Ontology (GO) enrichment analysis of all 138 genes are presented in Additional file 4, Table S3. Graphical representations of GO annotations are shown in Additional file 5, GO biological processes, Additional file 6, GO cellular component and Additional file 7, GO molecular function.

We examined these 138 genes (see above) for univariate association with recurrence in the microarray training data, which, since recurrence was not used in the selection of these genes, constitutes the first validation of our study hypothesis. We found significant enrichment for univariate association with recurrence in these 138 genes (see Additional file 8, Figure S1 for comparison of the distribution of nominal p-values with a null distribution empirically determined by permutation of the outcome labels).

We performed quantitative PCR validation of the 4-gene signature (MMP1, COL4A1, P4HA2 and THBS2) in a separate patient cohort (Table 2). This validation analysis confirmed that all four genes were up-regulated in margins and OSCC samples from patients with disease recurrence compared to margins and OSCCs from patients who did not recur (Figure 3A). The risk score was calculated using the maximum expression in any margin of each patient, and was dichotomized at the median score. These risk groups were predictive of recurrence in an independent test cohort assayed by RQ-PCR (136 samples; N = 30 patients) (HR = 6.8, p = 0.04, logrank test) (Figure 3B). In multivariate Cox analysis, correction for T (tumor size) and N (nodal status) resulted only in a slight increase in significance level of the four-gene risk score (p = 0.06, likelihood ratio test). Clinical variables, alone or in combination, were not predictive of recurrence in either training or validation cohorts. The coefficients of the 4-gene risk score, for use with Z-score scaled expression values, are summarized in Table 4.

In addition, we performed a survival Receiver Operating Characteristic (ROC) analysis to demonstrate the sensitivity and specificity of the model in the test set. This analysis showed the effect of altering the cutoff threshold between high and low-risk patients. Our results showed that the use of a very high threshold to define high-risk patients predicts a majority of recurrences at a low false-positive rate (20%). The area under the ROC curve (AUC) for recurrence within 36 months was 0.73, an improvement over the expected AUC of 0.5 for unpredictive models (Additional file 9, Figure S2). While we maintained the standard median cutoff due to the limited sample size, a larger study in the future may be able to further fine-tune the cutoff threshold to optimize sensitivity and specificity in the context of the relative risks that treatment options informed by this prognostic score entail.

Simulating the selection of only a single margin from each patient, the 4-gene signature maintained a predictive effect in both the training and validation sets (median HR = 2.2 in the training set and 1.8 in the validation set, with 82% and 99% of bootstrapped hazard ratios greater than the no-effect value of HR = 1) (Additional file 10, Figure S3). Results from the bootstrap simulations showed smaller hazard ratios, compared with hazard ratios obtained when using the maximum expression value from several margins. This difference could not be explained by association between the number of margins and recurrence in the training set (HR = 1.17, p = 0.69 logrank test) or the validation set (HR = 1.30, p = 0.24 logrank test). However, a plausible explanation for this difference is that increasing the number of observed margins increases the probability of observing genetic alterations spreading asymmetrically from the tumor, since genetic alterations are not seen in all surgical margins.