Tumor Immune Microenvironment and Chemosensitivity Signature for Predicting Response to Chemotherapy in Gastric Cancer

As one of the most common cancers, gastric cancer is the third leading cause of cancer-related deaths worldwide (1). Surgical resection is the primary treatment for resectable stage II and III gastric cancer, although the rate of relapse remains high. For the roughly 40% of patients who relapse within 2 years after surgery, adjuvant chemotherapy is the first-line postoperative treatment (2–4). Although patients with stage II and III gastric cancer who received chemotherapy have improved survival (2, 4, 5), in the 5-year follow-up of the CLASSIC trial, the median overall survival (OS) in the adjuvant chemotherapy group was 78% compared with 69% in the surgery-only group, with a moderate absolute benefit from adjuvant chemotherapy of 9% (2). In the Adjuvant Chemotherapy Trial of S-1 for Gastric Cancer, the 5-year OS rate in patients with stage IIIB disease was 50.2% in the adjuvant chemotherapy group and 44.1% in the surgery-only group, suggesting that not all patients with stage II to III gastric cancer benefit from adjuvant chemotherapy (6). The tumor–node–metastasis (TNM) staging system is used to predict prognosis and stratify risk for treatment decisions (4). However, variations in clinical outcomes have been shown among patients who receive similar treatment for cancers of the same stage, suggesting weakness in the current prognostic model (3, 4). Improved stratification of gastric cancer is needed to more accurately predict patient prognosis and chemotherapy response.

Immune classification of cancers has provided prognostic and predictive values for chemotherapies and immunotherapies (such as immune-checkpoint inhibitors; refs. 3, 7–9). Gastric cancer could be a paradigmatic tumor for immune classifications (3, 10, 11). Patients whose tumors were highly infiltrated by T lymphocytes, particularly cytotoxic CD8⁺ T lymphocytes and memory T cells, had a longer disease-free survival (DFS) and OS, and patients whose tumors were highly infiltrated with neutrophils had a poor prognosis (3, 10, 12). On the basis of the quantification of lymphocyte and myeloid cell populations in the invasive margin (IM) and the center of primary tumors (CT), an immunoscore may be associated with survival and chemotherapy response (3, 10). Both malignant cells and stromal cells orchestrate tumor-associated inflammation, tumor progression, metastasis, and treatment response (13, 14). We hypothesize that integrating immune and stromal cell features, cancer cell features, and angiogenesis in tumor may improve prediction of outcome for patients with gastric cancer.

Support vector machine (SVM) was introduced by Vapnik (15) in 1999 for data classification and function approximation. SVM has been applied to problems in several cancer types, such as nasopharyngeal carcinoma, hepatocellular carcinoma, and non–small cell lung cancer (13, 16–19). As a supervised machine learning method, SVM classifies data points by maximizing the margin between classes (20). SVMs can be used to select a subset of discriminating biomarkers and patient or disease attributes to build reliable cancer classifiers (17, 20, 21). In this study, we developed an SVM-based prognostic classifier for predicting survival and benefit from adjuvant chemotherapy for patients with stage II and III gastric cancer who already had surgery. We also compared the prognostic and predictive efficacies of the SVM-based prognostic classifier to predictions based on single features and clinicopathologic risk factors.

We enrolled three independent panels of 668 patients with stage II and III gastric cancer in this study. For the training cohort, data were obtained from 223 patients with incident, primary, biopsy-confirmed gastric cancer diagnosed between January 2005 and December 2007 at Nanfang Hospital of Southern Medical University (Guangzhou, China). Inclusion criteria were availability of hematoxylin and eosin slides with invasive tumor components, availability of follow-up data and clinicopathologic characteristics, no history of cancer treatment, and appropriate patient informed consent. We excluded patients if they had no formalin-fixed, paraffin-embedded (FFPE) tumor sample (CT and IM) from initial diagnosis, or they received previous treatment with any anticancer therapy. We also included 218 patients in the internal validation cohort, with the same criteria as above, diagnosed between January 2008 and December 2009. Moreover, we included an external validation cohort (227 patients diagnosed between January 2005 and July 2008 at the First Affiliated Hospital of Sun Yat-sen University). All the patients underwent D2 gastrectomy. Two independent pathologists reassessed all these samples. Clinical baseline information was collected for each patient from the two hospitals' clinical databases. The clinical sources of the 668 patients with stage II or III gastric cancer are listed in Supplementary Table S1. A total of 641 (96.0%) patients were Han Chinese, and 27 (4.0%) patients belonged to one of the Chinese minority ethnic groups (the Zhuang, She, Li, and others). We restaged all patients according to the eighth version of the American Joint Committee on Cancer/International Union Against Cancer criteria. The two clinical databases were prospectively maintained with regularly scheduled follow-up. Patients were postoperatively followed up with abdomen CT scans every 6 to 12 months for the first 2 years and then annually thereafter, according to the follow-up protocol of our institution. The endpoints of this study were DFS and OS. The DFS was defined as the time to recurrence at any site, or all-cause death, whichever came first. OS was defined as the time to death from any cause. In total, there were 92 (41.3%), 124 (56.9%), and 112 (49.3%) patients who received adjuvant fluorouracil–based chemotherapy for 6 months in these three cohorts. Of the 328 patients treated with postoperative chemotherapy, 131 (39.9%) patients received the XELOX (capecitabine–oxaliplatin) regimen, and 197 (60.1%) patients received the FOLFOX (fluorouracil–folinic acid–oxaliplatin) regimen (Supplementary Table S2). None of the patients received radiotherapy. All samples were anonymously coded in accordance with the local ethical guidelines (as stipulated by the Declaration of Helsinki). Written informed consent was obtained from the patients, and this study was approved by the Review Boards of Nanfang Hospital of Southern Medical University and the First Affiliated Hospital of Sun Yat-sen University.

On the basis of previous study findings (3, 9, 10, 12, 13, 22–25), we selected 17 molecular markers that were involved in gastric cancer development and/or metastasis, which included nine immune cell biomarkers [CD3 (pan T cells), CD8 (cytotoxic T cells), CD20 (B cells), CD45RO (memory T cells), CD45RA (native T cells), CD57 (natural killer cells), CD68 (macrophages), CD163 (M2 macrophages), and CD66b (neutrophils)], two fibroblast markers [α-smooth muscle actin (SMA), and periostin (POSTN)], a microvascular marker (CD34), and five chemotherapy response–associated tumor-cell biomarkers [cyclooxygenase 2 (COX2), excision repair cross-complementing gene 1 (ERCC1), Ki-67, p21^WAF1, and thymidylate synthase (TS)]. FFPE samples were cut into 4-μm sections, which were then processed for IHC as described previously (3, 10, 13). Following incubation with an antibody against human CD3 (NeoMarker, clone SP7), CD8 (NeoMarker, clone SP16), CD45RO (Invitrogen, clone UCHL1), CD45RA (Zhong-shan Goldenbridge), CD20 (Invitrogen, clone L26), CD57 (NeoMarker, clone NK1), CD66b (BD Pharmingen), CD68 (Dako, clone PG-M1), CD163 (Zhong-shan Goldenbridge), CD34 (Abcam, ab81289), a-SMA (Abcam,ab5694), POSTN (Abcam,ab92460), COX2 (Abcam), ERCC1 (Thermo Fisher Scientific, clone 8F1), Ki-67 (Zhong-shan Goldenbridge, clone SP6), p21WAF1 (Santa Cruz Biotechnology), and thymidylate synthase (Millipore, clone TS106), the sections were stained in an EnVision System (Dako). The antibody dilutions and antigen retrieval are shown in Supplementary Table S3. Every staining run contained a slide treated with PBS buffer in place of the primary antibody as a negative control. Every staining run contained a slide of positive control. Prior to staining, sections were blocked with endogenous peroxidase (prepared in 1% H₂O₂/methanol solution) for 10 minutes and then microwaved for 30 minutes in 10 mmol/L citrate buffer, pH 6.0. The sections were blocked using 10% normal rabbit serum for 30 minutes. Furthermore, all slides were stained with the same concentrations of primary antibody for each antibody and incubated with monoclonal primary antibody overnight at 4°C, followed by incubation with an amplification system with a labeled polymer/HRP (EnVision, DakoCytomation) at 37°C for 30 minutes. The sections were developed with 0.05% 3, 3′-diaminobenzidine tetrahydrochloride (DAB) and counterstained with modified Harris hematoxylin. And all slides were stained with DAB dyeing for the same time for each antibody (Supplementary Table S3).

For the evaluation of IHC staining results, analysis was performed by two independent gastroenterology pathologists who were blinded to the patients' clinical and survival data. At low power (100×), the tissue sections were screened using an inverted research microscope (model DM IRB; Leica Germany), and the five most representative fields were selected. Thereafter, to evaluate the density of stained immune cells, CT and IM were measured at 200× magnification. The nucleated stained cells in each area were quantified and expressed as the number of cells per field. The microvessel density (MVD) in gastric cancer tumor tissues was evaluated by staining for CD34. Any discrete cluster or single-cell stained for CD34 was counted as one microvessel (10, 12, 26). Five representative fields were quantified, and the average number of microvessels per field (200× magnification) was presented as the MVD. Intensity of fibroblast cell staining (α-SMA and POSTN) in tumor specimens was graded as 0 (negative staining), 1 (weak staining), 2 (moderate staining), and 3 (strong staining). Staining extent was graded as 0 (0%–4%), 1 (5%–24%), 2 (25%–49%), 3 (50%–74%), and 4 (>75%; ref. 27). Values of the intensity and the extent were multiplied as an immunoreactive score. A global assessment of the entire tumor was made without selection for the IM or CT regions. For the four tumor-cell biomarkers (COX2, p21^WAF1, ERCC1, and TS), each slide was assigned a score at 200× magnification as established previously (13): the average of the score of tumor cells staining multiplied by the score of staining intensity. Tumor cell staining was assigned a score using a semiquantitative five-category grading system: 0, none of tumor cells were stained; 1, 1% to 10% of tumor cells stained; 2, 11% to 25% of tumor cells stained; 3, 26% to 50% of the tumor cells stained; 4, 51% to 75% of the tumor cells stained; and 5, more than 76% of the tumor cells stained. Staining intensity was assigned a score using a semiquantitative four-category grading system: 0, no staining; 1, weak staining; 2, moderate staining; and 3, strong staining. Values of the extent and the intensity were multiplied as an immunoreactivity score. The Ki-67 scoring system was based on the percentage of tumor cells stained (13). Two pathologists independently scored all samples blindly with regard to clinical characteristics and prognosis. A third pathologist was consulted when different opinions arose between the two primary pathologists. If the third pathologist agreed with one, then that value was selected. If the conclusion of the third pathologist was different, then the three pathologists worked collaboratively to reach a common decision. The optimum cut-off score for each feature was selected on the basis of the association with the patients' OS by using X-tile software version 3.6.1 (Yale University School of Medicine, New Haven, CT; refs. 3, 13, 28).

SVM, a method for building a classifier, aims to create a decision boundary between two classes that enables prediction of labels from one or more feature vectors (29). This decision boundary, known as the hyperplane, is orientated in such a way that it is as far as possible from the closest data points from each of the classes. These closest points are called support vectors. In this study, we addressed the prognosis prediction problem of gastric cancer as a classification problem where the input was a vector that we call a “pattern” of n components which were called “features.” The features consisted of IHC markers. Each pattern corresponds to a patient. We limited ourselves to a two-class classification problem (i.e., whether a patient died within 5 years or not). The recursive feature elimination method was applied for features selection and ranking in the training data set (17). The pruning method was applied to exclude useless features. We evaluated how well an individual feature contributed to the prognosis predication (e.g., ≥5 years vs. <5 years), and then all candidate features were ranked based on their contributions. To evaluate the contribution of one feature, we excluded the feature to develop a reduced feature set. The contribution of each feature could be measured on the basis of the performance of the SVM trained with this reduced feature set. After ranking, features with the least contribution to the prognosis predication could be excluded. The radial basis function kernel was adopted, because our classification problem was nonlinear (17, 19). After the feature subset with the best performance was selected, we predicted the labels of tested samples and kept a record of the actual performance of the trained SVM models. To investigate the possibility of identifying different prognostic subgroups of patients with stage II and III gastric cancer based on these immunomarkers using SVM, we performed a set of experiments in the training cohort of 223 patients with stage II and III disease; the developed SVM classifier was further validated in two independent validation cohorts of 445 patients. The SVM data processing methods were conducted as described previously (13, 17–19). The programs were coded using R software.

We compared two groups using the t test for continuous variables and χ² test for categorical variables. Survival curves were generated using the Kaplan–Meier method and compared using the log-rank test. Univariate and multivariate analyses were performed using the Cox proportional hazards model. Variables that achieved statistical significance at P < 0.05 were entered into the multivariate Cox regression analyses. Interactions between the signature status and treatment were detected using the Cox model as well. All the statistical tests were done with the R software (version 3.4.0) and SPSS software (version 19.0). A two-sided P < 0.05 was considered significant.

Supplementary Table S1 lists the clinicopathologic characteristics of the patients in the training (n = 223), and internal (n = 218) and external (n = 227) validation cohorts. Of the 668 patients included in the study, 450 (67.4%) were men, and the median (interquartile range) age of all patients was 57.0 (49.0–65.0) years. The number of patients with stage II or stage III gastric cancer who received adjuvant chemotherapy was 91 (40.8%) in the training cohort, 128 (56.4%) in the internal validation cohort, and 112 (49.3%) in the external validation cohort.

We first analyzed the IHC staining of 223 patients in the training cohort, and selected the optimum cut-off score for all 26 IHC features by X-tile plots. Supplementary Table S4 shows the results of the univariate analysis between each of the 26 features and survival in the training cohort. On the basis of the SVM analysis of the training data set, the SVM signature selected and integrated seven features, including CD3 IM, CD8 IM, CD45RO CT, CD66b IM, CD34, POSTN, and COX2 (Supplementary Fig. S1), as pivotal factors.

In the training cohort of the 223 evaluated patients, we defined 136 patients as low SVM and 87 patients as high SVM, according to the SVM signature. The 5-year OS and DFS rates were 10.3% and 7.4%, respectively, for the low-SVM patients, and 72.4% and 62.1%, respectively, for the high-SVM patients [HR 0.123 (0.076–0.199) and 0.153 (0.099–0.235), respectively; both P < 0.0001; Fig. 1A]. We then performed the same analyses in the internal validation cohort. The 5-year OS and DFS rates were 24.8% and 19.4%, respectively, for the low-SVM patients, and 89.9% and 87.6%, respectively, for the high-SVM patients [HR 0.158 (0.096–0.260) and 0.161 (0.101–0.258), respectively; both P < 0.0001; Fig. 1B]. To confirm that the SVM signature had prognostic value in different populations, we applied it to the external validation cohort and demonstrated the similar results (Fig. 1C). We also assessed the distribution of the SVM signature, recurrence, and survival status as well as the expression of the seven features in the training, internal, and external validation cohorts (Supplementary Figs. S2–S4). In addition, the SVM signature remained a significant predictor of DFS and OS after stratification by clinicopathologic factors (Supplementary Figs. S5–S8).

In univariate analysis, low-level expression of CD3 IM, CD8 IM, and CD45RO CT, high-level expression of CD66b IM, CD34, POSTN, and COX2 in the SVM signature, and low-SVM patients were associated with significantly poorer DFS and OS (Supplementary Tables S4–S6). Multivariate Cox regression analysis after adjustment for clinicopathologic variables and TNM stage revealed that the SVM signature remained an independent predictor for OS and DFS in the training, internal, and external validation cohorts (Table 1).

To further determine whether the SVM signature could stratify patients by TNM stage, we evaluated the prognostic value of the SVM signature in patients with stage II and III disease, respectively (Supplementary Fig. S9). Compared with patients with low-SVM, high-SVM patients with stage II or III disease had a significantly longer DFS and OS in both the internal and external cohorts.

Moreover, in the ROC curves for traditional clinicopathologic factors, including tumor size, differentiation status, amounts of CEA and CA19-9, depth of invasion (T stage), lymph node metastasis (N stage), and TNM stage, as well as each single immunomarker feature and the SVM signature, the point with the maximum AUC was illustrated for each factor. The SVM signature exhibited a higher prognostic accuracy for 3- and 5-year DFS and OS than TNM stage, any clinicopathologic risk factor, and single feature alone (Fig. 2; Supplementary Figs. S10–S12). In the three cohorts, the AUCs of the SVM signature for 5-year DFS and OS [training cohort: 0.818 (0.755–0.881), 0.827 (0.766–0.888); internal validation cohort: 0.831 (0.773–0.889), 0.815 (0.755–0.874); and external validation cohort: 0.815 (0.755–0.875), 0.817 (0.758–0.877)] were significantly greater than the AUCs for all the other prognostic factors considered. AUCs for TNM stage were: training cohort: 0.659 (0.577–0.740), 0.640 (0.561–0.718); internal validation cohort: 0.665 (0.592–0.739), 0.684 (0.613–0.755); and external validation cohort: 0.628 (0.553–0.703), 0.642 (0.568–0.716; Fig. 2). Similar results were also observed for 3-year DFS and OS in all the three cohorts (Supplementary Figs. S11 and S12).

To evaluate whether high- or low-SVM patients might benefit from adjuvant chemotherapy, we investigated the association between SVM status and survival among patients who either did or did not receive adjuvant chemotherapy (Supplementary Table S7). For patients who either did or did not receive chemotherapy, the SVM signature was significantly associated with DFS and OS in stage II or III patients (Supplementary Fig. S13). However, the SVM signature had a stronger association with the prognosis of patients who received chemotherapy than those who did not receive chemotherapy (Supplementary Fig. S13). Therefore, we performed a subset analysis according to the SVM signature. A test for an interaction between SVM signature and chemotherapy revealed that, in either stage II or III disease, the benefit observed in high-SVM patients [stage II: DFS, HR 0.303 (0.124–0.738), 0.009; OS, HR 0.238 (0.077–0.739), 0.013; stage III: DFS, HR 0.467 (0.270–0.809), 0.007; OS, HR 0.504 (0.282–0.900), 0.020; and combined stage II and III: DFS, HR 0.406 (0.255–0.647), <0.001; OS, HR 0.411 (0.247–0.686), 0.001; all P_interaction < 0.05; Table 2] was superior to that observed in low-SVM patients. The corresponding Kaplan–Meier survival curves for patients with stage II or stage III gastric cancer, which compared low SVM with high SVM by treatment, are shown in Fig. 3. The results from the subset analysis using the SVM signature revealed that chemotherapy was significantly associated with increased DFS and OS in the high-SVM group (stage II: P = 0.009 and P = 0.013; stage III: P = 0.007 and P = 0.020; combined stages II and III: P < 0.0001 and P = 0.0001, respectively), but had no significant effect in the low-SVM group.

Accurate prognostic assessment is vital for the formation of appropriate treatment decisions. Cancer classification based on molecular biomarkers has improved prognosis estimates for some malignancies (18, 30, 31). Because gastric cancer is a clinically heterogeneous disease, with various outcomes even among patients with the same clinical stage, we sought to improve the prediction of prognosis. We developed a seven-feature SVM signature to categorize patients with gastric cancer into high-risk and low-risk groups with large differences in DFS and OS.

In clinical practice, adjuvant chemotherapy is the standard of post-surgery care for patients with stage III gastric cancer; however, means to identify low-risk patients with stage III gastric cancer who may not need adjuvant chemotherapy are lacking (4, 10, 32–34). Because the results of several trials, including CLASSIC, adjuvant chemotherapy for stage II gastric cancer has become standard (2, 35). However, clinical or laboratory method other than TNM staging system that could identify stage II or III patients likely to benefit from adjuvant chemotherapy have been lacking (4). Stratification of patients into stage II and III gastric cancer according to their risk of disease recurrence and progression allows identification of those who should derive the most benefit from adjuvant chemotherapy and limits overtreatment in patients for whom the toxicities of adjuvant chemotherapy outweigh the benefits. Assignment of treatment partly based on tumor molecular characteristics is a promising approach. Research suggests that tumor-infiltrated immune cells were related to chemotherapy response in a variety of tumors (3, 36–38). In addition, stromal POSTN and tumor-cell COX-2 expression might also predict tumor response to chemotherapy and/or chemoradiotherapy (13, 39–41). In our current research, we integrated multiple immune and stromal cell features, cancer cell features, and angiogenesis in tumor to build a model to improve the overall prediction of outcome for patients with gastric cancer. First, our SVM signature was a predictor of DFS and OS after stratification by clinicopathologic factors; second, multivariate Cox regression analysis revealed that the SVM signature was an independent predictor for OS and DFS in the cohorts; third, the SVM signature could stratify patients by TNM stage. We evaluated the prognostic value for patients with stage II and III gastric cancer. Our model exhibited a higher prognostic accuracy than TNM stage, any clinicopathologic risk factor, or single feature alone. Therefore, our seven-feature–based SVM signature for patients with stage II and III gastric cancer is both a prognostic and predictive method, and such stratification might lead to more individualized treatment for gastric cancer patients. Our results are also supported by studies in colon cancer, gastric cancer, and other kinds of cancers (13, 24, 39, 40, 42–46).

There were some limitations for this study. First, our data were obtained in one endemic area in China; the pathologic subtype and distribution of clinical characteristics might be different in other areas. Second, the use of adjuvant chemotherapy was not within a randomized comparison, and the decision to treat or not to treat patients after surgery was made by the patients and/or clinicians. Third, the biological mechanisms by which candidate markers that have been included in the signature, such as COX2 and POSTN, contribute to progression and chemoresistance of gastric cancer remain unclear. Further investigations into their functions might provide clues for targets and treatment strategies. Moreover, the signature requires further validation in prospective studies and multicenter clinical trials.

Our findings show that the signature can classify patients with stage II and III gastric cancer into groups with different survival. Moreover, the signature might be a useful predictive tool for identifying stage II and III patients who would benefit from adjuvant chemotherapy.

No potential conflicts of interest were disclosed.

Conception and design: Y. Jiang, J. Xie, S.-R. Cai, T. Li, G. Li

Development of methodology: Y. Jiang, J. Xie, Z. Han, L. Huang, L.-Y. Zhao, T. Li, G. Li

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Y. Jiang, J. Xie, W. Huang, H. Chen, S. Xi, Z. Han, T. Lin, L.-Y. Zhao, Y.-F. Hu, J. Yu, S.-R. Cai, T. Li, G. Li

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y. Jiang, J. Xie, W. Huang, H. Chen, S. Xi, Z. Han, L. Huang, L.-Y. Zhao, Y.-F. Hu, J. Yu, S.-R. Cai, T. Li, G. Li

Writing, review, and/or revision of the manuscript: Y. Jiang, J. Xie, W. Huang, H. Chen, S. Xi, Z. Han, L. Huang, T. Lin, L.-Y. Zhao, Y.-F. Hu, J. Yu, S.-R. Cai, T. Li, G. Li

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Y. Jiang, J. Xie, W. Huang, H. Chen, S. Xi, Z. Han, T. Lin, L.-Y. Zhao, Y.-F. Hu, J. Yu, S.-R. Cai, T. Li, G. Li

Study supervision: Y. Jiang, S.-R. Cai, T. Li, G. Li

This work was supported by grants from the National Natural Science Foundation of China (81872013, 81672446, and 81600510), the National Key Research and Development Program of China (2017YFC0108300), the Key Clinical Specialty Discipline Construction Program (to Department of General Surgery, Nanfang Hospital), the Outstanding Youths Development Scheme of Nanfang Hospital, Southern Medical University (2018J007), and Director's Foundation of Nanfang Hospital (2016B010).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.