A pan-cancer perspective of matrix metalloproteases (MMP) gene expression profile and their diagnostic/prognostic potential

The TCGA gene expression (RNAseq) data (IlluminaHiSeq: log2-normalized_count+ 1) was downloaded from Xena browser (https://​xenabrowser.​net/​datapages/​). The fifteen cancer types selected had at least ten patients with adjacent normal samples. The cancer types are denoted by their TCGA four-letter abbreviation (See Abbreviations). Statistical analyses were performed to compare and contrast the expression levels of 24 MMP genes.

All statistical analyses were performed using the R language and environment for statistical computing (R version 3.2.2; R Foundation for Statistical Computing; www.​r-project.​org). The normalized counts were log2 transformed prior to all statistical analyses to achieve a normal distribution. Expression differences between cancer patients and adjacent normals were initially examined using a t-test where the amount and significance of change were depicted using fold change (FC: Cancer vs Adjacent Normal) and p-value (pval). Fold change was calculated as the median gene expression level. Trends in expression differences were identified using unsupervised clustering. Diagnostic power of expression differences (Cancer vs. Adjacent Normal) of individual, or combinations of, MMPs was assessed using the area under the curve (AUC) of the receiver operating characteristic (ROC). Sensitivity values were calculated at various specificity thresholds (90%; 95%; 99%; 100%). Cox proportional hazards models were used to evaluate the association of gene expression levels on overall survival (diagnosis to date of death). Survival data was obtained from the TCGA patient phenotype files. Patients who at the time of analysis were alive with no evidence of disease were censored at the date of last follow-up visit. Kaplan-Meier survival analysis and the log-rank test were used to compare survival differences in groups separated by median expression level at multiple thresholds. Each cancer was considered individually. Multivariate analysis was performed to identify combinations of proteins that had significant findings in the univariate analysis. Data was visualized using R software and Tableau 10.4 (Tableau Software, www.​tableau.​com).

Gene expression differences for the twenty-four MMP family members were analyzed in fifteen different cancer types in the Cancer Genome Atlas (TCGA) (Fig. 1). The most prevalent gene expression changes were upregulation as opposed to downregulation in tumor tissue versus control tissue. MMP subtypes had similarities in expression patterns, with dramatic upregulation in cancer for the majority of collagenases, matrilysins, metalloelastase, and stromelysins. Other MMPs that showed significant differences were strong upregulation of MMP14 and downregulation of MMP23B, MMP27, and MMP28. The most significant dysregulations were among collagenases and stromelysins. MMP11, a stromelysin, was significantly upregulated in all cancer types except KICH and UCEC (p = 1E-07 to 1E-122). The other stromelysins, MMP3 and MMP10, were significantly upregulated in 7 cancer types and 10 cancer types respectively. MMP9, a gelatinase, was significantly upregulated in twelve of the fifteen analyzed cancer types (p = 1E-05 to 4E-27). Collagenases, MMP1 and MMP13, were significantly altered in 11 cancer types and 12 cancer types, respectively (p = 5E-05 to 1E-77).

Across cancer types, lung squamous (LUSC) and uterine corpus endometrial (UCEC) were distinct from the other cancer types with several significantly downregulated MMPs (Fig. 1). Additionally, renal cancers comparatively lacked significant upregulation of MMPs compared to adjacent normal tissue. This pattern is most noticeable among collagenases and stromelysins other than MMP11.

Expression of MMPs has a large degree of heterogeneity across cancers (Fig. 2). Some MMPs had very high numbers of transcripts present (gene cluster A), while six MMPs in gene cluster C (MMP27, MMP21, MMP20, MMP26, MMP23A, and MMP8) featured scant numbers of transcript copies across any cancer tissue (Fig. 2a and b). The cancer types with the most significant changes in MMP gene expression were LUSC and HNSC (19 and 15 MMPs, respectively), and patients from these cancers generally clustered together upon unsupervised hierarchical clustering as seen in cluster 2 of Fig. 2a. This grouping in cluster 2 features high expression of gene cluster B genes (MMP3, MMP10, MMP13, MMP1, and MMP12) in a pattern that is relatively unique among cancer types.

Among the high abundance MMPs, MMP9, MMP11, MMP14, and MMP2 had marked differences in expression among the three patient clusters. These genes were extremely abundant in breast, esophageal, head & neck, lung, and stomach cancers compared to the other tumor types.

MMP11 and MMP13 expression was dramatically higher in most cancer types compared to tissue matched controls (Fig. 3). Aside from KIRP and PRAD for MMP13, and KICH and UCEC for MMP11, the upregulation was consistent across all cancer types. MMP11 expression was comparatively high in the normal UCEC tissue, while the other non-significant pairs retained relatively low expression within the cancer samples. Both genes typically featured a very high fold change difference between groups and were among the highest expression differences observed within the dataset (Fig. 1).

The area under the curve (AUC) values for ROC analysis of MMP expression in each cancer are illustrated in Fig. 4a. Values for the range of the AUC as well as the sensitivity values at 90, 95, 99, and 100% specificity are included in Additional file 1: Table S1 for gene-cancer pairs with the highest AUC values. Each cancer type featured at least one MMP with an AUC greater than 0.9 except for PRAD. Six cancer types (BRCA, COAD, ESCA, HNSC, LUSC, and UCEC) showed 4 or more MMPs with AUC greater than 0.9. MMP11 most frequently had significant predictive values, with AUC greater than 0.9 in twelve cancers and greater than 0.95 in eight. The highest AUC observed for MMP11 was in LUSC with an AUC value of 0.991 (Fig. 4b). Seven MMPs (MMP7, MMP11, MMP12, MMP13, MMP24, MMP27, and MMP28) had an AUC of greater than 0.95 for at least one cancer type. For example, MMP12 had an AUC of 0.983 in lung squamous cancer (LUSC) (Fig. 4b).

Using a multivariate analysis, ROC curves were made analyzing expression data using combinations of up to five MMPs for each cancer type. Additional file 1: Figure S1 displays AUC graphs for the top two individual MMPs and combinations of MMPs. The AUC for a combination was usually higher than with a single gene, however was not necessarily superior, as many of the confidence intervals for individual MMPs overlapped with the AUC value of MMPs in combination (Additional file 1: Table S1). For example, for thyroid cancer (THCA), MMP11 showed an AUC of 0.928 alone, but when combined with MMP19, multivariate analysis showed an AUC of 0.961 (Fig. 4c). In colon cancer (COAD), an AUC of 1 was achieved after multivariate analysis of a combination of MMP7, MMP13, and MMP28. The cancer type with the next highest AUC after multivariate analysis was lung squamous (LUSC). Individually, MMP11 and MMP12 were the MMPs with the best diagnostic value with an AUC of 0.991 and 0.983 (Fig. 4b), respectively for LUSC. After multivariate analysis two combinations of MMP11 and MMP12 yielded an AUC of 0.996.

Twenty-four proteins from the MMP family were independently analyzed for their ability to predict overall survival in 15 different cancer types using tumor samples from TCGA. Each cancer type was considered independently when examining MMP prediction of overall survival. Hazard ratios, expression thresholds, and p-values for the thirty-seven curves that had (p < 0.01 with HR or 1/HR > 1.5) are shown in Additional file 1: Table S2 and summarized in Fig. 5a. Of these relationships, sixteen were in a type of renal cell carcinoma (KICH, KIRC, or KIRP). Liver cancer (LIHC) was the only other cancer type with three MMPs associated with a survival change. Genes frequently associated with survival differences included MMP14, MMP17, and MMP23B which were associated with exclusively poorer prognosis for higher expression while MMP15, MMP20, and MMP24 were associated exclusively with poorer prognosis for lower expression. Interestingly, several MMPs featured inconsistent survival trends across cancer types with MMP19 showing opposite trends when comparing renal papillary (KIRP) with renal clear cell (KIRC) (Fig. 5b-c). The strongest hazard ratios were observed for MMP15 within renal papillary and renal chromophobe (KICH) subtypes (Fig. 5d-e).

The observation of MMP11 upregulation across cancer types is consistent with prior knowledge of gene function facilitating tumor invasion by degrading collagen, fibronectin, and laminin. Upregulated in 13 of 15 cancer types (Fig. 1), MMP11 expression differences yielded particularly high fold changes and high AUCs. For example, in lung cancers there was a 43-fold increase in squamous (LUSC) and 36-fold increase in adenocarcinoma (LUAD) over control. This upregulation of MMP11 was contrasted by other MMPs where up-regulation was more cancer-type specific. A reason for this pan-cancer upregulation might be related to the functional ability of MMP-11 in helping evade immune surveillance by desensitizing cancer cells to NK-cells [38]. MMP-11 differs from other MMPs because it is not secreted as a pro-MMP, but is already active upon secretion through intracellular activation by furin, within the constitutive secretory pathway of the trans-Golgi network [39]. Our analysis supports a theory that MMP-11 expression is critical in cancer development and progression.

MMP-7, a matrilysin protein, had a gene expression fold change of 141.85 in COAD, the highest fold change we found in our analysis. This is consistent with prior studies of colon cancer, that also have indicated that MMP-7 could predict a more aggressive phenotype of colon cancer and correlate negatively with patient survival [40‐43]. However, TCGA data did not show any significant associations between survival and MMP7 expression in colon cancer. MMP-7, a stromelysin, degrades collagens, proteoglycans, elastin, laminin, fibronectin, and casein. It also activates other MMPs, including gelatinase. Upregulation of MMP7 was primarily found within GI cancers (esophageal, stomach, colon) as well as thyroid and head/neck. Though upregulated in other cancer types, the degree of upregulation for colon cancer, is truly remarkable with a fold change greater than 140. Part of the reason for this level of upregulation could be ascribed to the intrinsic expression of MMP7 in colon versus other tissues. Additionally, MMP-7 is very closely associated with some of the prime factors in the tumorigenesis of colon cancer. MMP-7 activation is directly associated with APC, and it is well established that APC mutations are frequently implicated as among the first mutations that occurs in the disease history of colon polyps as they progress to cancer [44]. APC mutations often lead to increased β-catenin and thus overexpression of MMP-7. If these changes occur in the earliest stages of colon cancer or precancer as we suspect, MMP-7 may be an excellent early diagnostic marker for colon cancer.

Several MMPs that were largely downregulated in tumor versus control were membrane-type or unclassified by their primary degradation target. MMP-28, also known as epilysin, is downregulated in nine cancer types in our analysis (Fig. 1). This downregulation was highly pronounced in both lung squamous and adenocarcinoma subtypes. The finding is especially intriguing given that stable expression of MMP-28 in lung adenocarcinoma cells results in epithelial mesenchymal transition with accompanying loss of E-cadherin as reported by Illman et al. [37, 45]. Our analysis and prior reports do not capture a clear mechanistic or clinical picture how MMP-28 downregulation interplays with disease progression or metastasis.

We believe that MMPs have tremendous potential as diagnostic biomarkers. Survival in cancer can be improved by diagnosis at an early stage. Screening tests such as colonoscopy plus biopsy, the gold standard for colon cancer diagnosis, are often invasive, costly, or uncomfortable while others such as a fecal occult blood test have a lower sensitivity. Identifying novel biomarkers for cancer screening hopefully will lead to highly sensitive, cost-effective, and non-invasive diagnostics. MMPs offer theoretical potential for these properties. Six cancer types featured four or more MMPs with AUC > 0.9 (Fig. 4a). Cancer-by-cancer, lung squamous (LUSC) has two MMPs with sensitivities over 95% (MMP11 and MMP12), colon cancer (COAD) has three over 94% (MMP11, MMP28, and MMP7), and esophageal cancer (ESCA) has two over 94% (MMP11 and MMP12). MMP11 was 100% specific and over 93% sensitive in colon and esophageal cancer. While, this data indicates that MMPs have the potential to serve as quality biomarkers, confirmation of upregulation in more easily accessed biofluids remains to be determined as increased gene expression does not necessarily mean a protein will be overexpressed at a high enough level to be detected within a biofluid, such as serum, at a high enough level to be acted upon. Will be needed to fulfill the qualities of an excellent, minimally-invasive biomarker. If serum detection is achievable, ubiquitously upregulated MMPs such as MMP11 or MMP13 could serve as a pan-cancer marker, capable of identifying the presence of cancer but not the specific location of the cancer.

Several MMPs interact with each other or rely on other MMPs for their activation. For example, the activation pathway of pro-MMP-2 is cleaved by MMP-14 [2]. It is likely that the dysregulation of one MMP alters the MMP ecosystem and means that MMPs are better predictors when analyzed in combination as opposed to individually. Simultaneous upregulation of MMPs that cleave the same substrates were frequently observed within our data. The combined signal of MMP11 and MMP19, both of which cleave aggrecan and gelatin [46], resulted in a higher AUC value than individually in thyroid cancer. Additionally, stromelysins and collagenases both showed similar patterns of upregulation across tumor types. The inter-dependence and similar functions of MMPs give physiological rationale for performing multivariate analysis when observing MMP dysregulation.

MMPs are generally not great prognostic biomarkers for most cancer types. However, MMPs have very good prognostic value for kidney clear cell carcinoma (KIRC) (Fig. 5). High expression of MMP7, MMP9, MMP14, MMP17 and MMP19 each individually correlated with a significantly poorer survival in patients with KIRC. Pathologic spread of these tumors from the renal cortex to the renal vein is facilitated by the lack of connective tissue between the renal columns and the vasculature in the renal sinus. Further degradation of the extracellular matrix from higher expression of MMPs could allow invasion of the renal vein and therefore more metastases earlier in the disease course. MMP-2 and -9 are reported in literature to be associated with poor prognosis in kidney clear cell carcinoma [47]. Both MMP-2 and MMP-9 are implicated in angiogenesis which is critical for highly vascularized malignancies such as renal cell carcinoma. Higher expression of MMP-2 and MMP-9 are found in patients who have kidney clear cell carcinoma metastases when compared to patients without distant metastases [48]. Other MMPs, especially those less studied or of lower abundance, may also have critical roles in the promotion of invasion, angiogenesis, and metastasis within this disease state.

The MMP dysregulation in renal cancer might be linked to VHL. VHL mutation, whether inherited or sporadic is frequently implicated in the tumorigenesis of renal clear cell carcinoma. With the loss of VHL, HIF is no longer properly degraded and free to induce the expression of proteins needed in hypoxic conditions. This leads to the upregulation of MMP14 in kidney clear cell carcinoma in patients with deletion of VHL [49]. We found that patients with higher expression of MMP14 had generally poorer survival, consistent with the more aggressive and invasive malignancy associated with VHL functional loss. Additionally, VHL regulates the assembly of collagen IV, a major constitute of the basement membrane and target of MMP-3, MMP-8, and MMP-9 [50]. This profile likely results in increased destruction and invasion through the basement membrane and may explain why higher expression of MMP-3 correlates with a poorer prognosis in kidney clear cell carcinoma patients [51]. Our evidence corroborates the literature knowledge of the significant role of MMPs in the disease history of kidney cancer.