Prognostic value of tumor-inltrating immune cells in primary colorectal cancer and metastases

Precise prognostic biomarkers are urgently needed to improve treatment and hence prognosis for patients with metastasized colorectal carcinoma. Tumor-inltrating CD3 and CD8 positive lymphocytes have been shown to robustly predict survival of colorectal cancer in stages I-III in retro- and prospective studies. Their clinical value in metastasized tumors or the role of other types of immune cells is not well known. We performed a detailed characterization of tumor inltrating immune cells of primary colorectal cancer and metastases of 55 patients. Immune inltrates were assessed visually and


Conclusions
A broad range of types of immune in ltrates predicts longer survival in metastasized colorectal cancer patients. Of all immune in ltrates, CD4 showed the highest prognostic precision. The combined assessment in primary colorectal cancer and metastases is valuable in most cases. Thus, automated digital analysis of immune in ltrates in colorectal metastases could improve the prognostic strati cation of patients.

Background
Colorectal cancer (CRC) is the third most common malignancy in the world and over one-half of these patients will develop liver metastases 1,2 . Currently, the prognosis and hence strati cation to treatment regimens is based solely on anatomical and histopathological criteria of the TNM classi cation 3 and tumor cell differentiation 4,5 . Although this approach has proven to be clinically useful in the past, its predictive accuracy is only moderate [6][7][8] . In order to improve the prognosis of metastasized colorectal cancer patients, there is a high clinical demand for more accurate prognostic markers to identify patients at high risk. A very promising biomarker is an immunoscore introduced by Galon et al. 6,9,10 . Here, the combined density of CD3 and CD8 positive lymphocytes at the invasive margin (IM) and tumor center (CT) of primary colorectal cancer and metastases is measured and a score calculated. Numerous studies have shown that patients with higher immunoscores are characterized by prolonged survival.
Furthermore, this immunoscore has successfully been validated in a large prospective multinational and multicenter study 11 and was described to exceed the prognostic power of the TNM classi cation system 8 . However, the strong evidence of its prognostic value is limited to UICC stages I-III [11][12][13][14][15][16][17][18][19][20][21][22] . Patients with metastasized colorectal carcinoma have rarely been investigated [23][24][25] , although they present a group of immense clinical importance due to their unfavorable prognosis. Furthermore, only CD3 and CD8 were included in the immunoscore, although tumor-in ltrating immune cells in general are thought to in uence tumor progression and metastasis. For example, preliminary data has shown that high densities of CD20 positive B cells and FOXP3 positive regulatory T cells may be associated with better prognosis in colorectal cancer 9,[26][27][28] . The role of CD4 positive T helper (TH) cells and macrophages, on the other hand, is still controversial, as is the prognostic relevance of key proteins of the checkpoint inhibition system PD1 and PD-L1 9, 29-40 .
In the present study, we performed a detailed characterization of a broad range of tumor in ltrating immune cells (CD3, CD4, CD8, CD20, CD45R0, CD68, FOXP3, PD1, and PD-L1) at the primary tumor site and in all available metastasectomy specimens of colorectal cancer patients. It was our aim to determine and con rm the suitability of immune in ltrates as prognostic biomarkers for patients with metastasized tumor stage. Additionally, we present a comparison of digital image analysis and conventional visual analysis.

Patients
A total of 55 patients with metastasized colorectal cancer were included in this study, which underwent surgical treatment of their primary cancer and metastases at the Department of General Visceral and Transplantation Surgery of the University Medical Center Mainz between 2000 and 2015. Clinical and pathological data were collected using pathology reports and medical charts. To obtain survival data, patients were followed up for a mean of four years [Min = 2 months; Max = 201 months]. Data of patients without the occurrence of death in the follow-up period were censored. 36 of the patients suffered from synchronous colorectal cancer and 19 patients from metachronous disease. The metastasectomy specimens were obtained from the liver (n = 84) or lung (n = 5). Of the 55 patients, 16 received preoperative systemic therapy or radiation treatment and 36 were treated with a postoperative systemic therapy. Table 1 depicts the patient's clinical characteristics. 1:250. For antigen retrieval, sections were heated in a steam cooker for either 20 min at pH 9 in EDTAbuffer (CD3, CD4, CD8, CD20, CD45R0, and CD68) or for 35 min at pH 6 in citrate-buffer (FOXP3, PD1 and PD-L1). The immunostained slides then were digitized at 40x magni cation using a NanoZoomer-Series Digital Slide Scanner (Hamamatsu Photonics, Hamamatsu, Japan). Analysis of the immune in ltrate was performed digitally and visually.

Analysis of in ltrates of immune cells and key proteins of the checkpoint inhibition system
The immune in ltrates (CD3, CD4, CD8, CD20, CD45R0, CD68 and FOXP3) and key proteins of checkpoint inhibition (PD1 and PD-L1) were quanti ed at the invasive margin and tumor center of primary colorectal cancer and at the invasive margin of all metastasectomy specimens. For the purpose of better readability, those will be subsumed as immune in ltrates in the remainder of this article. The tumor center of metastases was not considered due to extensive regressive changes and necrosis.
For visual analysis, immune in ltrates were assessed semiquantitatively based on the IC Score, de ned as the proportion of tumor area that is occupied by the positively stained cells 41 .
For digital analysis, a ready to use software based on the Halo platform from Indica Labs (Corrales, NM, USA) was used. On the one hand, the total number of stained cells was quanti ed in relation to the total area of the invasive margin or tumor center (cell-count/mm²). On the other hand, the area of the immune stain was assessed in relation to the total area of the invasive margin or tumor center (stainarea/mm²x10 5 ). For evaluation of the cell-count/mm², the CytoNuclear module v1.4 was applied and for the evaluation of the stain-area/mm²x10 5 the area quanti cation module v1.4 was applied. Firstly, the invasive margin and tumor center were manually selected for each whole slide. Then ten representative sections were used to de ne staining parameters in a training phase (e.g. minimum nuclear optical density, minimum staining optical density, nuclear and cellular size and roundness) for an optimal recognition of positively stained cells. All automated results were visually validated for accuracy and manually corrected for misclassi ed areas (e.g. artefacts due to air-bubbles, inking or tissue folding). The CD4, FOXP3, PD1 and PD-L1 stained sections showed a strong and non-speci c staining of hepatocytes. Therefore, a HALO tissue classi er was trained to recognize and exclude hepatocytes in whole slides of hepatic metastases for those immunostains. Supplement For digital analysis, ROC curve analysis was performed related to the patient's overall survival and the Youden`s index and its associated criterion was de ned as cut-off value 42 . Additionally, analysis was repeated using the median as cut-off value. Evaluated immune in ltrate values of all stains and cut-off values are depicted in Supplementary Table 1. Immune in ltrates with a value above the cut-off were given a Score 1 (high) and such below a Score 0 (low). CD3 and CD8 positive lymphocytes (immunoscore) as well as CD8 and CD20 positive lymphocytes (TB-cell score) were combined as de ned by Galon et al. and Mlecnik et al. 6,25 , to enhance comparability to those studies. Survival analysis was performed for primary carcinomas and metastases individually and combined. Since some patients had more than one metastasis, the least in ltrated, most in ltrated and last resected metastasis was considered in the survival analysis. In case of only one metastasis, the same metastasis was considered as least in ltrated, most in ltrated and last resected.

Metastasis = M CD4IM
Primary and metastasis combined = P CD4 IM + P CD4 CT + M CD4 IM Scores had a possible range from 0 to 6. Survival analysis was then performed for each score and again after dichotomization in patients with low grade and high grade immune in ltrates using the Youden`s index or the median as cut-off value.

Statistical Analysis
Quanti cation of immune in ltrates was presented in box plots indicating the minimum, rst quartile, median, third quartile and maximum. The Kaplan-Meier method was used for survival analysis. Survival curves were calculated for overall scores and for dichotomized groups (low grade and high grade). The signi cance of the differences between low grade and high grade immune in ltrate groups was compared using the logrank test chi². For the overall scores the logrank test chi² was used to determine the likelihood of a survival trend across the scores. Univariate and multivariate regression models based on Cox proportional hazards models were used to determine the hazard ratios (HR), which were visualized in forest plots. The clinicopathological parameters age, postoperative treatment, grading of differentiation, localization, number of metastases, preoperative treatment, pN, pT, biological sex and synchronicity of metastasis were analyzed in univariate analysis and those found to be statistically associated with survival included in the multivariate regression models. Since only 55 patients were included in the study and 29 of them died in follow-up, only 3 covariates should be included in the multivariate regression model. Therefore, not all immune in ltrates could be included at the same time.
Multivariate regression analysis was therefore performed for each type of immune in ltrate for primary colorectal cancer, metastases and the combined score individually and the type of metastasis with highest hazard ratio in univariate analysis was selected. The parameters regression coe cient b, its standard error SE, Wald statistic (b/SE) 2 , p-value P, Exp(b) and the 95% con dence interval for Exp(b) are provided. A p-value of less than 0.05 was considered as statistically signi cant. Statistical analysis was performed using MedCalc for Windows, version 18.11 (MedCalc Software, Ostend, Belgium).

Clinicopathological parameters
The prognostic signi cance of clinicopathological parameters (age, preoperative treatment, grade, localization, number of metastases, pTNM stage, biological sex and synchronicity of metastasis) was analyzed using univariate Cox proportional hazards models. Of all clinicopathological parameters, only high pT-stage (pT4 vs. pT1-3) and preoperative treatment were statistically associated with a poor outcome (see Table 2). Those were included in multivariate regression models. Primary tumor-in ltrating immune in ltrates as prognostic marker in metastasized CRCs The prognostic signi cance of the density of immune in ltrates in the invasive margin and tumor center of primary colorectal cancer was analyzed using univariate Cox proportional hazards models and Kaplan Meier Curves. Immune in ltrates were assessed digitally and visually and divided into a low grade and a high grade group based on predetermined cut-off values using ROC curve analysis (see Supplementary  Table 1). Scores were calculated as speci ed in the methods section.
Strikingly, survival analysis by the Kaplan Meier method revealed that a low density of all investigated immune in ltrate types (CD3-CD8, CD8-CD20, CD4, CD45R0, CD68, FOXP3, PD1 and PD-L1 positive cells) was statistically associated with a poor outcome (   Table 2). However, only three covariates were included in the regression model (each type of immune in ltrate, pT-stage and preoperative therapy), so comparison of the different types of immune in ltrates is not possible.

Metastatic-in ltrating immune in ltrates as prognostic marker in metastasized CRCs
Beyond evaluating the primary site of colorectal carcinomas, all metastasectomy specimens were investigated as well. For patients with more than one metastasis, the least and most in ltrated metastasis and the last resected metastasis were included in the survival analysis.
A low metastatic density of all investigated immune in ltrate types (CD3-CD8, CD8-CD20, CD4, CD45R0, CD68, FOXP3, PD1 and PD-L1 positive cells) was statistically associated with a poor outcome in digital analysis (Table 4; Fig. 2). Hazard ratios were consistently higher for digitally assessed stain area when compared to cell-count (see Supplementary Table 3). The highest hazard ratios were detected for CD4 immune in ltrates of the most in ltrated metastases (HR = 13.5) and CD3-CD8 immunoscore of the least in ltrated metastases (HR = 9.2). Except for PD-L1 in ltrates, all three types of metastases showed statistically signi cant hazard ratios. In visual analysis only CD3-CD8 immunoscore (last resected), CD8-CD20 TB cell score (last resected) and CD45R0 (all types), CD68 (all types), FOXP3 (least and most in ltrated), and PD1 (most in ltrated and last resected) -immune in ltrates were statistically associated with a poor outcome (see Supplementary  Table 3). Hazard ratios were consistently lower than those resulting from digital analysis. A detailed overview of all Kaplan Meier curves and hazard ratios is presented in Supplementary Figs. 4 and 5.
Multivariate Cox regression analysis suggests that all investigated types of metastatic immune in ltrates contribute to the prediction of survival in digital analysis, except for PD-L1 (see Supplementary Table 2).
In visual analysis, only CD3-CD8 immunoscore and CD45R0, CD68, FOXP3 and PD1 -immune in ltrates contributed to survival prediction. However, only three covariates were included in the regression model showed lower hazard ratios than evaluation of in ltrates in primary colorectal cancer in digital analysis.
As a next step, we investigated, whether a combined score of immune in ltrates of primary colorectal cancer and metastases shows increased prognostic power compared to the immune in ltrate score of each individually. Only for FOXP3 (most in ltrated: HR = 5.7), CD4 (least in ltrated: HR = 19.6) and PDL1 (least in ltrated: HR = 8.5) did the combined scores show higher hazard ratios (see Table 5). Multivariate Cox regression analysis suggests that all investigated types of combined immune in ltrates of primary colorectal cancer and metastases contributed to the prediction of survival in digital analysis.
In visual analysis, all but the CD8-CD20 TB-cell score contributed to the prediction of survival (see Supplementary Table 2). However, only three covariates were included in the regression model (each type of immune in ltrate, pT-stage and preoperative therapy), so comparison of the different types of immune in ltrates is not possible.

Summary of immune in ltrates with highest prognostic precision
The highest hazard ratios were calculated for the combined evaluation of immune in ltrates of primary colorectal cancer and metastases (quanti cation of stain area) for FOXP3 (HR = 5.7), CD4 (HR = 19.6) and PD-L1 (HR = 8.5) in digital analysis. For CD3-CD8 immunoscore (HR = 9.2; least in ltrated metastases), CD8-CD20 TB cell score (HR = 5.7, last resected metastases), CD45R0 (HR = 8.9, most in ltrated metastases), CD68 (HR = 5.7, most in ltrated metastases) and PD1 (HR = 6.0, most in ltrated metastases) evaluation of metastatic immune in ltrates alone showed highest hazard ratios for digital analysis. The evaluation of immune in ltrates in primary colorectal cancer alone led to consistently lower hazard ratios in digital analysis, as did evaluation using visual analysis.
We repeated all steps of the survival analysis with the median as cut-off value to divide immune in ltrates into low grade and high grade groups. This analysis con rmed ndings of ROC curve analysis driven cut-off values, although hazard ratios were predominantly smaller. Data is presented in Supplementary Table 5.

Discussion
All investigated primary tumor-in ltrating immune in ltrates are prognostic markers in metastasized CRC In order to improve prognosis of metastasized colorectal cancer patients, there is a high clinical demand for more accurate prognostic markers to identify patients at high risk. In the past years, prognostic relevance of tumor in ltrating lymphocytes was detected in an increasing number of solid tumors 9,28 .
Therefore, in this study, we performed a thorough characterization of a wide panel of tumor immune in ltrates of the primary site and its metastases in a patient collective with microsatellite stable metastasized colorectal carcinoma. Strikingly, a low density of all investigated immune in ltrate types (CD3, CD8, CD20, CD4, CD45R0, CD68, FOXP3, PD1 and PD-L1 positive cells) was statistically associated with a poor outcome in this collective.
The positive prognostic signi cance of high CD3, CD8 and CD45R0 cell in ltrates in the invasive margin and the tumor center of colorectal carcinomas has been demonstrated several times 11-22, 28, 43 . This led to a proposition of the so-called immunoscore, calculated from both CD3 and CD8 immune in ltrates to predict the clinical course of patients in early and late tumor stages (UICC I-III, but not IV) with high accuracy 6,10,11 . In the present study, a prognostic signi cance of a high CD3-CD8 immunoscore and a high CD45R0 immune in ltrate can now be con rmed in a group of patients with metastatic colorectal carcinoma. Similarly, Bindea et al. and Berntsson et al. showed a signi cantly prolonged overall survival for a high CD8-CD20 TB cell score for patients of early and late tumor stages 44,45 , as we con rmed in this study.
So far, studies have mainly concentrated on CD3, CD8, CD45 and CD20 lymphocytes. Far less is known about the prognostic signi cance of other immune cells. The prognostic signi cance of CD68 positive immune in ltrates has mainly been investigated for patients with non-metastasized colorectal carcinoma and several studies described a prolonged survival of patients with high CD68 intratumoral cell densities [33][34][35]37 . Only one study demonstrated this association in patients with advanced tumor stages (UICC stages III-IV) 36 , but only based on visual analysis. In this study, we were able to show a link between high CD68 positive cell in ltration of primary colorectal carcinoma and prolonged survival in both visual and digital analysis.
The prognostic signi cance of T helper cells (TH), however, is still largely unclear. Individual studies detected a positive prognostic signi cance of high TH1 in ltrates, a negative prognostic signi cance of high TH17 in ltrates and no signi cance of intratumoral TH2 cells, when primary colorectal cancer was analyzed 45,46 . In the present study, we did not subtype the TH cells, but used the well-established pan-TH Some studies have already detected superior survival of patients that show high FOXP3 positive immune in ltrates in primary colorectal cancer. However, predominantly early and consistently non-metastasized tumor stages were included in those studies. Here, we con rm this prognostic association in a collective of metastasized colorectal carcinomas.
The key proteins of the immune checkpoint system are currently receiving a lot of attention with regard to immunomodulation therapy in numerous malignancies. The prognostic value of immunohistochemical detection in colorectal carcinomas and their metastases, however, is yet to be determined. Only few studies investigated the possible role of PD1 and PD-L1 as prognostic biomarkers in tissue micro array based studies of microsatellite unstable 39 and stable 38, 40 colorectal carcinomas. All of these studies showed a favorable prognosis for the detection of high densities of PD1 and PD-L1 positive cells. Here, we con rmed this ability to predict a favorable prognosis on whole slides and in a metastasized disease stage.

Evaluation of metastatic tumor-in ltrating immune in ltrates increases prognostic power
An important strength of the present work is that not only primary colorectal cancer, but also all resected metastasectomy specimens were examined for immune in ltrates. Only few other groups have performed such a combined immunological characterization and data is only available for few immune in ltrate markers.
In case of the CD3-CD8 immunoscore, Mlecnik et al. 25 , Halama et al. 23 , Wang et al. 48 and Kwak et al. 24 analyzed the prognostic signi cance of the metastatic CD3-CD8 immunoscore, of which only the latter group also included the corresponding primary colorectal cancer in their evaluation. All groups reported that the immunoscore predicts prognosis following liver resection. Our data support this hypothesis, and we identi ed that the immunoscore of the least in ltrated metastases leads to the highest prognostic precision with the highest hazard ratio, compared to analysis of the primary tumor and other types of metastases. This nding is also supported by data of Mlecnik et al. 25 , who also described the least in ltrated metastasis as the strongest prognostic marker.
In case of the CD8-CD20 TB cell score, only Meshcheryakova et al. 49 and Halama et al. 23 analyzed in ltrates of metastases and reported prolonged survival for higher scores. Primary colorectal carcinomas were not investigated. Not only can we support their ndings, but since we evaluated both primary and metastatic in ltrates, we were able to identify the TB cell score of the last resected metastasis as the marker with highest prognostic precision, even compared to the evaluation of the primary colorectal cancer alone.
We found a signi cant association between high metastatic CD4 in ltrates and favorable survival. Only two groups have investigated the prognostic relevance of CD4 in colorectal cancer metastases 50-52 .
Katz et al. found a prognostic advantage of high metastatic CD4 positive immune in ltrates in one study, but prognostic disadvantages in another 51,52 . However, in contrast to our study, they analyzed only small tissue samples on a tissue micro array. Due to the expected heterogeneous spatial distribution of the CD4 positive cells (own unpublished data), an analysis of whole slides seems more preferable and might explain the contradictory ndings. Another group to investigate CD4 positive immune in ltrates was Tanis et al. that only performed a visual analysis 50 . In doing so, they found no association between a favorable prognosis and a high density of CD4 positive cells. However, this is not contradicting our data, as we too were not able to detect a prognostic relevance in visual analysis. Only by digital quanti cation a strong predictive value for survival was apparent. It is noteworthy that a combined evaluation of CD4 positive tumor in ltrates of primary cancer and the most in ltrated metastasis showed highest prognostic power of all other immune markers with a hazard ratio of 19.6.
In case of CD68, only Meshcheryakova et al. evaluated metastatic in ltrates, but did not detect a prognostic relevance 49 , that we were able to identify in our study. However, Meshcheryakova et al only evaluated the cell-count/mm² in one metastasis per patient that was selected randomly, ignoring all others. Using this method, only the most in ltrated metastases showed a prognostic signi cance in our data. Therefore, it seems probable that due to high temporal heterogeneity of distribution of the CD68 positive cells among the metastases (own unpublished data) this group missed the prognostic value. In performing digital quanti cation by stain area, all types of metastases showed a signi cant prolonged survival for high CD68 densities. This was also con rmed by visual analysis and emphasizes the advantage of using two separate methods of digital quanti cation.
High metastatic FOXP3 in ltrates showed a signi cant positive prognostic value in this study. Furthermore, a combined score of primary and metastatic in ltrates increased prognostic power even further. This is in line with a study performed by Nakagawa et al. that also demonstrated an inferior prognosis for patients with low in ltration of peritumoral regulatory T cells in resected metastases 53 . Katz et al. did not nd a prognostic value of FOXP3 positive immune in ltrates, but yet again did only analyze small tissue samples on a tissue micro array that might not su ciently represent the heterogeneity of spatial distribution (own unpublished data) 51 .
To our knowledge, metastatic in ltrates of PD1 and PD-L1 positive cells have not been investigated so far. In this study, we detected a signi cant prolonged survival in presence of high metastatic in ltrates and showed that evaluation of a combined score increases prognostic power even further in case of PD-L1.
Visual analysis of immune in ltrates carries some prognostic value for metastatic colorectal carcinoma The evaluation of the prognostic signi cance of different types of immune in ltrates in colorectal carcinoma is almost solely based on digital image analysis. Most noteworthy is that the immunoscore was recently validated in a prospective multinational and multicenter study without including a visual analysis 11 , although latter would be much easier to perform in clinical practice and is already common in breast cancer 54,55 . In the present study, we therefore additionally assessed the immune in ltrates visually. Remarkably, the visual quanti cation of the immune in ltrates in the invasive margin of primary colorectal cancer and metastases con rmed the prognostic value of all immunostains, except of PD1 and CD4. However, the hazard ratios were consistently inferior to those of the digital analysis. This is in line with ndings of Jakubowska et al. that visually assessed the in ammatory cell in ltrate using light microscopy with hematoxylin and eosin and found a signi cant shorter survival for patients with a weak in ltrate 56 . Both ndings support the prognostic value of visual analysis of immune in ltrates. However, a replacement of digital image analysis by visual assessment cannot be recommended at this time.

Limitations
The here presented study has some limitations. First, only 55 patients were enrolled reducing statistical power. Therefore, multivariate regression analysis could only be performed for each type of immune in ltrate for primary colorectal cancer, metastases and the combined score individually. However, we were able to con rm the well-known prognostic signi cance of the CD3-CD8 immunoscore in our patient collective, which has been described as robust in numerous studies [11][12][13][14][15][16][17][18][19][20][21][22]43 . Therefore, it seems plausible to assume that our case collection is suitable for investigating also other immune markers.
Secondly, the optimal cut-off value for dichotomizing patients in low-grade and high-grade in ltrate groups is not known. However, we analyzed our data using the two widely applied cut-off values Youden's index of ROC curve analysis and median and both showed substantially comparable ndings.
Thirdly, only cases with microsatellite-stable colorectal carcinomas were investigated.

Conclusion
In this study, we demonstrate the prognostic relevance of a broad range of tumor immune in ltrates in microsatellite stable primary colorectal carcinomas and their metastases. In addition to con rming the CD3-CD8 immunoscore as a robust prognostic marker in a metastasized disease stage, we identi ed a broad range of other less well investigated immune in ltrates as prognostically relevant (CD4, CD20, CD45R0, FOXP3, CD68, PD1 and PD-L1) in both primary tumors and metastases. Strikingly, evaluation of CD4 shows by far the highest prognostic power amongst all other; a combined score of primary colorectal cancer and metastases increases accuracy even further. Since visual analysis of immune in ltrates is easier to perform in clinical practice and is already common in breast cancer 54, 55 , we included a visual analysis in this study design. Notably, visual analysis did indeed con rm the prognostic value of most immune markers. However, its accuracy was consistently lower than digital image analysis.
Based on ndings in this study, tumor-in ltrating immune cells carry a high prognostic value in microsatellite stable and metastasized primary colorectal cancer and should be evaluated in prospective studies for clinical application. Automated digital analysis of immune in ltrates in colorectal metastases could improve the prognostic strati cation of patients even further. In addition to the well-established immunoscore, the marker CD4 seems especially promising.