Carbonic Anhydrase IX, Hypoxia-inducible Factor-1α, Ezrin and Glucose Transporter-1 as Predictors of Disease Outcome in Rectal Cancer: Multivariate Cox Survival Models following Data Reduction by Principal Component Analysis of the Clinicopathological Predictors

Discussion

In the current study we analyzed four hypoxia-related biomarkers: CAIX, HIF-1α, ezrin and GLUT-1 in a cohort of 178 rectal cancer patients treated by preoperative RT or chemoradio-therapy. CAIX intensity was the only independent predictor of DSS (21). In the long-course RT group, negative HIF-1α and negative/weak ezrin expression were associated with favorable DSS (22). Negative/weak ezrin expression in the biopsy was also linked to longer DSS as compared with moderate/strong expression (23). Likewise, a trend for longer DFS in favor of negative/weak GLUT-1 staining was observed among patients in the long-course RT group (24).

HIF-1α, CAIX, ezrin and GLUT-1 are, however, biologically interrelated. CAIX and GLUT-1 are downstream targets of HIF-1α (14, 15, 26). Activation of receptor tyrosine kinases including protein kinase B (AKT), increases mTOR activity (16, 17), which in turn induces HIF-1α (16, 28, 29). Rapamycin, a protein kinase that inhibits mTOR activity, is found to regulate HIF-1α in hypoxia (30). Thus, mTOR is a key regulator of HIF-1α protein synthesis in many cancers (16). Ezrin-related metastatic behavior is shown to be linked to the mTOR pathway (19). When the mTOR pathway is blocked by rapamycin, experimental lung metastasis is inhibited in vivo (19). Thereby, the regulation of mTOR, HIF-1α, CAIX and GLUT-1 are interrelated and univariate analysis of the individual markers is of limited power to disclose these interrelations.

In addition to analyzing individual biomarkers separately, other approaches are needed to explore their interactions and particularly their role as independent predictors of disease outcome together with the classical prognostic factors. The latter include: postoperative tumor stage, tumor size and differentiation, lymphovascular invasion, positive circumferential margin, examination of less than 12 lymph nodes, and type of operation (1, 2, 31-33). In the current cohort, several other clinicopathological variables (e.g. number of metastatic lymph nodes, and disease recurrence) were also recorded and originally analyzed as potential predictors of disease outcome (21, 24). Together with the four biomarkers, however, the number of all these covariates is too large to rend direct multivariate modeling feasible. A useful option is to attempt data reduction using PCA to generate a set of factors that explain as much as possible of the variance by the original covariates (=clinicopathological variables). This set of factors (i.e. their saved scores) can then be directly used in multivariate modeling together with the four markers of interest.

PCA was used in the present study to elucidate the role of these four biomarkers as independent predictors of disease outcome in rectal cancer. The five factors generated by PCA explained almost 70% of the total variance of the original variables subjected to extraction. These five factors also located in the steep slope of the plot, and the rest of the factors located on the shallow slope did not contribute much to the solution. The extracted factors were subjected to varimax and promax rotation to generate uncorrelated and correlated factors, respectively. Although the generation of these five factors by PCA was relatively straightforward, the complexity of the interplay between the original clinicopathological variables is well illustrated by the fact that all these five factors were heavily loaded by more than one original variable. For this reason, these five factors were not given any specific names in order to avoid confusion with the original variables.

By definition, varimax-rotated factors showed no correlation (r=0.000, p=1.000), while promax-rotated factors should be correlated. In our case, this linear correlation (Pearson r) was observed only between F1 and F2 (r=0.451, p=0.0001), implicating a close interrelationship between the original variables loaded into these two first-order factors. In all subsequent analyses, only the five promax-rotated factors were used in order to be able to exploit their complex biological interrelationships. In further analysis, we also generated second-order factors by performing promax rotation for the promax-rotated first-order factors F1-5 (data not shown). Using the same eigen value cut-off (≥1), two second-order factors were generated, showing heavy loadings by the first-order factors F1, F2, F4 and F3, but much less by F5. Not unexpectedly, the two second-order factors were no longer linearly correlated (r=−0.31, p=0.791), reflecting the fact that out of the first-order factors F1-5, only F1 and F2 were significantly correlated.

In multivariate Cox analysis, on the operative samples, F1 was the only independent predictor of DFS and DSS. On the preoperative biopsies, no factor or biomarker predicted DFS, whereas F1 and F3 as well as HIF-1α and ezrin were significantly linked with DSS. When ezrin in the preoperative biopsies and the other biomarkers in the operative samples were analyzed, none of the factors or biomarkers were significantly associated with DFS. However, CAIX and F1 predicted DSS. Deviating results in Cox analyses for the three models would most feasibly be explained by given preoperative treatment. RT may shrink the tumor (34), delivering complete remission in 0-29% of tumors (35-37). Decreased tumor size may alleviate hypoxia and thereby alter the expression of hypoxia-associated biomarkers. RT can also affect the tumor microenvironment (38) or cause direct damage in DNA replication (39). To circumvent this problem, preoperative biopsies as well as tissue samples from patients who had not received preoperative RT were used as controls. However, pre-treatment biopsies may represent only a small portion of the tumor, not enabling the determination of the differentiation grade. Furthermore, preoperative tumor assessment may only be directional due to the limitations of imaging methods (40). Postoperative tumor staging can in turn be affected by RT, which may also influence the results. It would have been advisable to analyze separately the operative samples from the control group of patients (who had not received any treatment prior to surgery). However, the control group was too small to be analyzed alone, and thereby we had to settle for analyzing the pre-RT biopsies.

In the original studies, we were unable to find any significant correlations between the biomarkers and common clinicopathological factors, which may partly be due to the impact of RT. Postoperative tumor stage is known to be the most important clinical prognostic factor of disease outcome (2, 41). In the current PCA analyses, F1 was most heavily loaded by postoperative stage and lymph node status. Thus, it is not unexpected that F1 was singled out as the only significant prognosticator of disease outcome in all three models. F3, most heavily loaded by the number of examined lymph nodes and the year of therapy was significantly associated with DSS only when analyzed in the preoperative biopsies. In our institution, a multidisciplinary team has been active since 2005, in parallel with the introduction of a structured pathology report, bringing about improved quality control and feedback to the radiologists, surgeons, pathologists and oncologists. In fact, at least twelve lymph nodes were examined in 69% of the patients after and including year 2005, whereas this was the case in only 31% operated on before 2005 (unpublished data). The retrieval of lymph nodes after long-course RT may be complex, especially after excellent tumor response (42). Not unexpectedly, the number of examined lymph nodes and the year of therapy seem to be clearly interconnected in this study, reflecting the importance of the acuity in lymph node retrieval.

The present study showed that HIF-1α and ezrin as well as CAIX, but not GLUT-1, significantly predicted DSS in multivariate Cox models, adjusted for the first-order factors 1-5. This is in line with the original studies analyzing each of these markers separately, showing strong expression of these biomarkers to be associated with adverse disease outcome (21-24). The current data implicate their predictive value should not be confounded with the classical prognosticators. Further analyses are ongoing in order to examine the interactions between HIF-1α, CAIX, ezrin and GLUT-1 in full factorial models. It remains to be seen whether any panels can be established and be proven useful as independent predictors of disease outcome.

Taken together, factor analysis using PCA of a multitude of clinicopathological prognosticators generated five first-order factors which proved to be variably loaded by most of the original variables. F1 was the only significant predictor of DFS and DSS in model 1. In model 2, HIF-1α, ezrin, F1 and F3 were significantly associated with DSS. In model 3, DSS was significantly predicted by CAIX and F1. Preoperative RT appears to markedly modify the expression of the studied biomarkers and also interfere with the original clinicopathological prognosticators, emphasizing the complexity of prognostication in rectal cancer.