Serum PlGF and EGF are independent prognostic markers in non-metastatic colorectal cancer

The aim of this study was to determine the prognostic value of circulating angiogenic cytokines in non-metastatic colorectal cancer (CRC) patients. Preoperative serum samples of a training (TC) (n = 219) and a validation cohort (VC) (n = 168) were analyzed via ELISA to determine PlGF, EGF, VEGF, Ang1, PDGF-A, PDGF-B, IL-8 and bFGF levels. In addition, survival was correlated with PlGF and EGF expression measured by microarray and RNAseq in two publicly available, independent cohorts (n = 550 and n = 463, respectively). Prognostic values for overall (OS) and disease-free survival (DFS) were determined using uni- and multivariate Cox proportional hazard analyses. Elevated PlGF is predictive for impaired OS (TC: HR 1.056; p = 0.046; VC: HR 1.093; p = 0.001) and DFS (TC: HR 1.052; p = 0.029; VC: HR 1.091; p = 0.009). Conversely, elevated EGF is associated with favorable DFS (TC: HR 0.998; p = 0.045; VC: HR 0.998; p = 0.018) but not OS (TC: p = 0.201; VC: p = 0.453). None of the other angiogenic cytokines correlated with prognosis. The prognostic value of PlGF (OS + DFS) and EGF (DFS) was confirmed in both independent retrospective cohorts. Serum PlGF and EGF may serve as prognostic markers in non-metastatic CRC.

Affymetrix CRC cohort. A database of publicly available CRC patient samples measured by Affymetrix gene chips was set up as described previously 24 . In brief, gene chip datasets with transcriptome-wide gene expression data generated by Affymetrix gene arrays and available survival data were identified in the Gene Expression Omnibus repository (www.ncbi.nlm.nih.gov/geo/). Samples were MAS5 normalized and a second scaling normalization was performed to set the mean expression across all probes to 1000. We selected the probe sets 209652_s_at for PlGF and 206254_at for EGF. TCGA CRC cohort. CRC patients measured by RNA-seq were published in The Cancer Genome Atlas (TCGA) 25 . Pre-processed level 3 data generated using Illumina HiSeq. 2000 RNA Sequencing V2 and probe IDs 5228 (PlGF) and 1950 (EGF) were used. For each sample, the expression level was determined using a combination of MapSplice and RSEM. The individual sample files were merged in R using the plyr package 26 . statistical analysis. All values are reported with standard deviation (SD). For univariate analyses, categorical variables were compared using the χ 2 -test. Continuous variables were expressed as arithmetic mean with standard deviation and compared using students' t-test. All variables with p < 0.05 were included in a stepwise backward, multivariate logistic regression model using the median as a cutoff. Survival analysis was performed by employing Cox proportional hazard regression in the R statistical environment (www.r-project.org) as described previously 27 . Hazard ratios with 95% confidence intervals and log-rank P values were calculated using the library "survival". Overall survival, defined as time to death, and disease-free survival, defined as time to recurrence, were determined. Kaplan-Meier curves were drawn to visualize the survival differences. All variables with p < 0.05 on univariate analysis were added to a multivariate Cox regression model adjusting for age, sex, site of disease (SOD) and UICC stage and addressed for multiple testing controlling the false discovery rate by using the Benjamini-Hochberg procedure 28 . Statistical analyses were carried out using IBM SPSS Statistics v23 (SPSS Inc., Chicago, IL), R (www.r-project.org) and GraphPad Prism v7 (Graph Pad Software Inc., La Jolla, CA).   Table 2. Univariate analysis and arithmetic means for VEGF, Ang1, PDGF-A, PDGF-B, IL-8 and bFGF are given in Supplementary Tables 1-3. The multivariate logistic regression models were adjusted for age, sex, SOD and UICC stage (    Disease-free survival. A total of 31 patients (14.2%) in the TC and 37 (22.0%) in the VC developed recurrent disease within the follow-up period. Univariate analysis identified age >70 (TC: p = 0.018; VC: p < 0.001), UICC stage (TC: p = 0.009; VC: p = 0.028) and neoadjuvant therapy in the VC (p = 0.046), but not in the TC, as prognostic factors for DFS. Of the measured angiogenic factors, serum PlGF levels significantly correlated with DFS in the VC cohort (TC: p = 0.324; VC p = 0.024) when using the median as cutoff (Fig. 1), when using the 75 th percentile as cutoff, significant effects were seen in both cohorts ( Supplementary Fig. S3). Similarly, serum EGF levels correlated with DFS significantly in the VC (TC: p = 0.299; VC p < 0.01, Fig. 2); these effects were even stronger when using the 75 th percentile as cutoff ( Supplementary Fig. S4). Multivariate analysis showed elevated PlGF to be associated with worse DFS in both cohorts (TC: HR 1.052; 95% CI 1.005-1.107; p = 0.029; VC: HR 1.091; 95% CI 1.033-1.153; p = 0.009). Conversely, elevated serum EGF significantly predicted longer DFS in both cohorts (TC: HR 0.998; 95% CI 0.996-1.000; p = 0.045; VC: HR 0.998; 95% CI 0.996-1.000; p = 0.018). Serum VEGF, Ang1, PDGF-A/B, IL-8 and bFGF levels did not show any influence on DFS (Table 4). Again, the 75 th percentile as cutoff exhibited the strongest effects. The results for other cutoffs are given in Supplementary Figs S1-S4, significance levels for all analyses are reported in Supplementary Table S5. Of note, no synergistic effects of PIGF and EGF  Table S7).
prognostic impact of intratumoral plGF and eGF expression. Given the prognostic relevance of serum PlGF (OS + DFS) and EGF (DFS), we hypothesized that these cytokines may be produced by tumor cells and thus analyzed their prognostic role in two publically available, independent cohorts: The Affymetrix cohort with 550 CRC patients of all stages and expression data from Affymetrix microarrays 24 , and the TCGA cohort with 463 CRC patients of all stages and RNA expression data generated via RNAseq 25 . In both cohorts, our findings were independently confirmed (Fig. 3).
There was no influence of EGFR expression levels on the prognostic value of EGF in both cohorts ( Supplementary  Fig. S5). In patients with non-metastatic disease in the Affymetrix cohort, PlGF expression higher than the median showed a tendency towards being a negative prognostic factor for overall survival, but failed to reach statistical significance (p = 0.193; HR 1.45; 95% CI 0.83-2.55). In the same patients, EGF proved to be a positive prognostic factor for overall survival (p = 0.019; HR 0.51; 95% CI 0.29-0.9). For disease-free survival, both factors were confirmed as negative (PlGF, p = 0.036; HR 2.14; 95% CI 1.03-4.45) and positive (EGF, p = 0.041; HR 0.48; 95% CI 0.23-0.99) prognostic factors. There was no data on disease recurrence available for the TCGA cohort.

Discussion
This study demonstrates elevated serum PlGF levels to be a negative prognostic factor for OS and DFS in non-metastatic CRC patients. This prognostic association was demonstrated in two independent patient cohorts. Additionally, elevated serum EGF levels are associated with favorable DFS, but not OS, in both cohorts. Serum VEGF, Ang1, PDGF-A/B, IL-8 and bFGF levels did not predict OS or DFS on multivariate analysis. Our findings suggest that quantification of serum PlGF and EGF may be useful for preoperative risk stratification of patients with non-metastatic CRC. On tissue RNA level, identical prognostic effects could be demonstrated for both cytokines in two other independent cohorts with expression data measured via microarray and RNA next generation sequencing, respectively, thus strongly confirming the data from our own cohorts.   www.nature.com/scientificreports www.nature.com/scientificreports/ The currently available body of evidence on PlGF and EGF in CRC is limited. Gomceli et al. reported high PlGF to be prognostic for local recurrence but not OS in non-metastatic CRC, thus supporting our findings 29 . Contradictory to our present finding of PlGF as a negative prognostic marker, previously published data from our own group demonstrated PlGF to be of positive prognostic value for recurrence-free survival in metastatic CRC 30 . However, different cutoffs were chosen in the metastatic cohort previously published and only patients with PlGF levels >90 th percentile experienced favorable survival. Additionally, PlGF indicated only favorable disease-free survival and was not associated with overall survival. Moreover, the previously published data are from metastatic patients undergoing liver resection for CRC metastases, representing a different patient cohort as well as a different microenvironment in hepatic metastases. These differences in the angiogenic microenvironment between hepatic metastases and the primary colorectal tumor have resulted in anti-VEGF treatment to be used only in metastatic disease and may thus also explain the different effects of PlGF in different disease stages. This disparity in tumor biology is emphasized by data indicating that primary tumor cell lines produced more PlGF than cell lines from metastatic CRC 31 .
PlGF stimulates the chemotactic migration of human mesenchymal progenitor cells and stimulates neovascularization via the VEGF pathway, especially under hypoxic conditions 32,33 , and facilitates vessel growth and maturation 34 . PlGF is able to independently activate endothelial cells and synergizes with VEGF in driving angiogenesis 12,35 . In addition to angiogenesis, PlGF influences intratumoral macrophage polarization towards an immunosuppressive phenotype via the VEGF1 pathway 36 . Furthermore, PlGF, but not VEGF-A or VEGF-B, is increased in obesity 15,36 . Our data confirmed a correlation of PlGF levels with obesity and may partly explain the clinically well-known unfavorable outcome of obese cancer patients 15,36 . While the data presented here is insufficient to confirm PlGF as the reason for the more malignant biology of cancer in obese patients, they clearly warrant further studies on this matter.
An interesting finding is the markedly increased PlGF level in patients after neoadjuvant therapy. The addition of bevacizumab, a humanized, monoclonal antibody targeting VEGF-A, to chemoradiation in the neoadjuvant setting for non-metastatic rectal cancer has limited benefit, which may indicate escape mechanisms to bevacizumab treatment 37,38 . As PlGF is elevated during neoadjuvant therapy and able to activate the VEGF signaling axis independent of VEGF-A or -B 39 , it may act as a mediator of such escape mechanisms after bevacizumab-mediated VEGF depletion. This theory is supported by the activity of PlGF inhibitors in anti-VEGF-resistant tumors 40,41 . www.nature.com/scientificreports www.nature.com/scientificreports/ Consequently, elevated PlGF may be a negative predictive factor in this setting and if so, agents targeting VEGF signaling downstream of PlGF may be of more benefit than bevacizumab. This question needs to be addressed in future clinical trials investigating neoadjuvant anti-angiogenic therapy in rectal cancer.
In addition to PlGF, our data identified elevated serum EGF levels as a positive prognostic factor in CRC. The biological role of EGF in CRC is currently unclear. Although intuitively it should act as a negative prognostic factor due to its activation of the MAPK/RAS/RAF signaling cascade, previous studies were unable to demonstrate a correlation between serum EGF levels and prognosis in CRC 21 and the here presented study even observed a positive correlation. In breast cancer, the prognostic value of EGF is also a matter of debate while most studies point toward a positive correlation between serum EGF levels and prognosis 42,43 . The EGF receptor (EGFR, HER1, ErbB1) is a receptor for ligands of the epidermal growth factor family. Other receptors in this family include HER2 (ErbB2), HER3 (ErbB3) and HER4 (ErbB4) 44 . Upon ligand binding, receptors of the HER family form homo-or heterodimers, thus potentiating the MAPK activation as compared to their activity as monomers 44 . Ligands of the EGF family include EGF, Heparin-binding EGF-like growth factor (HB-EGF), transforming growth factor-α (TGF-α), Amphiregulin (AR), Epiregulin (EPR), Epigen, Betacellulin (BTC), neuregulin-1 (NRG1), neuregulin-2 (NRG2), neuregulin-3 (NRG3) and neuregulin-4 (NRG4) 45 . While EGFR is activated by EGF, HB-EGF, TFG-α, AR, EPR, BTC and Epigen, neuregulins activate HER3 and HER4 and no ligand is known for HER2 44 , suggesting a role of HER2 as a co-receptor transactivated by EGFR ligands. In summary, the EGF ligand and receptor families form a complex and incompletely understood network. The reason for the positive prognostic role of EGF is still unknown and leaves room for speculation: Dropping EGF serum levels after resection of CRC suggest the origin of elevated EGF levels within the tumor 46 . Low intratumoral MAPK/RAS/RAF signaling activity, suggesting low proliferative activity of the tumor cells and thus favorable prognosis, may lead to compensatory EGF excretion in tumor cells. Interestingly, EGF seems to be a prognostic factor in CRC independently of both activating RAS/RAF mutations and cetuximab treatment 47 , suggesting a role of EGF outside of the MAPK/RAS/RAF axis, possibly adding a second explanatory approach. Clearly, further research investigating the role of EGF in CRC is needed.
In this study, we compared the prognostic effects of serum proteins in some cohorts (TC and VC) and the prognostic effects of the respective mRNA levels in other cohorts. The correlation between mRNA and protein levels is highly dependend on the stability of both mRNA and protein. While in steady state mRNA the www.nature.com/scientificreports www.nature.com/scientificreports/ concordance between mRNA and protein abundance is 100%, the concordance between mRNA and protein variation is significantly lower 48 . Despite this, we found highly concordant prognostic effects between groups with protein and mRNA data; however, the comparison between mRNA and protein levels can still be considered a limitation of this study.
The here presented study includes a prospective training cohort and an independent prospective validation cohort which confirmed the main findings of the training cohort, thus representing high quality evidence. In addition, most findings could be validated in two more large cohorts, thus adding to the high level of evidence. The conclusions are clinically relevant, but the underlying molecular biology is still poorly understood. Therefore, molecular studies investigating the mechanistic principles as well as prospective clinical trials involving different centers and patient collectives are required prior to implementation of PlGF and EGF serum testing in clinical routine.
In summary, this study suggests preoperative serum PlGF and EGF levels as prognostic factors in non-metastatic CRC. The prognostic value of these cytokines in two independent prospective and another two independent retrospective cohorts supports their evaluation in larger multi-center studies on preoperative risk stratification and may ultimately enable clinicians to identify subgroups of patients who may benefit from adjuvant therapy while sparing others the side effects of potentially unnecessary systemic treatment.