An artificial intelligence method to assess the tumor microenvironment with treatment outcomes for gastric cancer patients after gastrectomy

The design of the entire study is shown in Fig. 1. A total of 4 independent cohorts, which included 936 GC samples, were enrolled in this study. One cohort (n 1 = 44) was from The Cancer Genome Atlas (TCGA) datasets (https://​portal.​gdc.​cancer.​gov/​) and 2 cohorts (GSE62254, n2 = 300, and GSE15459, n3 = 192) were from Gene Expression Omnibus (GEO) datasets (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​), and a cohort (n4 = 400) from Nanfang Hospital (Guangzhou, China), which were used to analyze the relationship between gene expression and immune microenvironment of tumors. We developed an SVM model to analyze the relationship between gene expression and CT images based on a TCGA cohort, aiming to learn the immune microenvironment from CT images. CT images were obtained from The Cancer Imaging Archive (TCIA, https://​www.​cancerimagingarc​hive.​net/​). One cohort that comprised 400 consecutive patients from Nanfang Hospital (Guangzhou, China) from October 2004 to September 2011 was used to develop the nomogram to predict survival and chemotherapy benefit for GC patients. All of these patients underwent partial or total radical gastrectomy. The inclusion criteria were histologically confirmed GC and patients who underwent standard unenhanced and contrast-enhanced abdominal CT < 30 days before surgical resection. Those who received preoperative chemotherapy, had severe surgical complications, did not have complete follow-up data over 3 years, or had other coexisting cancers were excluded. The entire study design is shown in Fig. 1.

The TNM staging was reclassified, and the postoperative chemotherapy regimen was taken based on the guidelines of the National Comprehensive Cancer Network (NCCN) [16]. In the Nanfang Hospital cohort, up to 223 patients were treated with adjuvant chemotherapy, including 114 (51.1%) patients treated with the XELOX (capecitabine-oxaliplatin, capecitabine 1000 mg/m²/days day1–day14, oxaliplatin 130 mg/m² once) regimen, 90 (40.4%) patients treated with the FOLFOX (fluorouracil-folinic acid-oxaliplatin, fluorouracil 400 mg/m² at day1 and 2400–3600 mg/m² for 46 h, folinic acid 400 mg/m² once, oxaliplatin 85 mg/m² once) regimen and 19 (8.5%) patients treated with 5-FU treatment alone or other regimens.

The follow-up duration was calculated from the day of surgery to the last follow-up. We defined the time to all-cause death as overall survival (OS) and defined the time to tumor recurrence at any site or death as disease-free survival (DFS). Ethical approval was obtained for this retrospective study, and informed consent was obtained from the patients.

To predict the prognosis of patients using gene information to assess immune infiltration, we included 44 cases from TCGA as the training cohort, 300 cases from GSE62254 and 192 cases from GSE15459 as 2 validation groups to establish and evaluate the role of the ImmunoScore (IS) in clinical prognosis.

To quantify the proportions of 22 types of infiltrating immunocytes of the TME, normalized RNA sequencing data were analyzed using the CIBERSORT algorithm [17]. The immunocyte type fractions of patients in the training cohort using the CIBERSORT algorithm could be downloaded directly from a comprehensive resource [18] (http://​timer.​cistrome.​org/​). For each sample, the sum of all estimates of the immunocyte type component equaled 100%. The least absolute shrinkage and selection operator (LASSO) method was used to screen the most effective prognostic markers among 22 immune cell subsets, and the optimal values of the penalty parameter λ were determined by 20-fold cross-validations.

The ImmunoScore (IS) was calculated by a Cox proportional hazards model with the LASSO results. The output was expressed as coefficients for estimating the immunocytes of the TME. The calculation string formula was as follows: where X_i represents the relative component of each immunocyte, β_i is the regression coefficient for the predictor and α is often a constant. We divided GC patients into 2 groups (High-IS and Low-IS) based on the median IS in each cohort.

Differential expression gene (DEG) analysis of the GSE15459 cohort was performed by applying the “limma” package at a corrected P < 0.05 and | logFC |≥ 1.5. The resulting data were used to generate volcano plots using R software (version 4.0.4). Data on the abundance of 22 types of immune infiltrates from the TIMER website were obtained. Using the “ggplot 2” package, we identified immunocytes with significant differences between the High-IS and Low-IS groups while analyzing the correlation between these cells and the IS using the “ggstatsplot” package.

To use CT information to obtain radiomics features associated with the IS of patients, we included 44 cases from TCGA as the training cohort and 400 cases from Nanfang Hospital as the validation cohort to establish a Radiomic ImmunoScore (RIS) to predict IS and estimate its connection with patient survival.

Using ITK-SNAP software (www.​itksnap.​org, version 3.8), two radiologists manually delineated the primary tumor on the CT images. The CT number, representing the absorptivity of tissue to X-rays, was set ranging from − 150 to − 50 Hounsfield units to exclude bone, muscle, blood vessels, and other intra-abdominal organs [19, 20]. The region of interest (ROI) was delineated by hand along the gastric wall at a distance of 1 mm from the stomach wall around the tumor lesion.

Based on extracted image features and ImmunoScore, the RIS model was developed by SVM, aiming to predict the ImmunoScore by CT images. The common image features included color, texture, shape and spatial relation. The software version we used was MATLAB 2018a, and the shape of the feature matrix extracted from a CT image was 1*9694. In the training cohort, the size of the original CT feature matrix was 44*9694, which made regression difficult. Therefore, the average weight of each feature was finally obtained after dimensional reduction using the Relief algorithm. The top 100 features were selected based on the feature weights. The AUC was regarded as the evaluation index. Six features were selected in the development of the Radiomic ImmunoScore (RIS) model when the AUC reached its peak (more details of this process are described in Additional file 1). The RIS of each cohort was also divided into High-RIS and Low-RIS groups according to the median.

Subsequently, the clinical significance of RIS was displayed intuitively by ROC, and the hazard ratios of each clinicopathological factor were analyzed between the High-RIS and Low-RIS groups.

We developed a comprehensive model, including RIS and other clinical features, and visualized it with a nomogram. To verify the accuracy and effectiveness of the nomogram, we randomly divided 400 patients from Nanfang Hospital into a training cohort and a validation cohort, with 200 cases in each group.

According to the influence degree of each factor on the outcome variable in the multifactor regression model, each value level of each variable was given a score, and the comprehensive nomogram obtained by adding was calculated by function to attain the prediction probability. Patients were then divided into High-nomogram and low-nomogram groups according to the median nomogram score.

Subsequently, the predictive ability of the nomogram for patients was demonstrated by a variety of evaluation methods, including ROC, survival curve, calibration plot decision curve analysis (DCA), net reclassification improvement (NRI) and risk factor association diagrams.

The association between the nomogram and adjuvant chemotherapy response was assessed in patients with stage II and III gastric cancer (n = 280, Nanfang cohort). All patients were divided into a chemotherapy cohort (n = 182) and a no chemotherapy cohort (n = 98). Both the total survival time and disease-free survival time were observed. Survival curves were generated with Kaplan–Meier analysis. Hazard ratios were generated with the Cox regression model.

All statistical tests were two-sided, and P < 0.050 was considered statistically significant. Statistical analysis was conducted with R software (version 4.0.4) and SPSS software (version 25.0). The packages for downloading data from TCGA and GEO are TCGAbiolinks and GEOquery. Correlations between the immune cell immune checkpoints and the Immunoscore were analyzed using Pearson’s correlation test. Survival curves were constructed by the Kaplan–Meier method and compared using the log rank test. The sensitivity and specificity of the survival prediction based on the 3 models were depicted by a time-dependent receiver operating characteristic (ROC) curve and a time-independent ROC curve, with quantification of the area under the ROC curve using the timeROC and pROC packages. The results of subgroup analysis were represented by a forest graph using the forestplot package, and the points of HR value in the graph were processed by log2 normalization. We performed Cox multivariate modeling using the rms package. After constructing the nomogram with the nomogramFormula package, we drew decision curves with the rmda package and risk factor correlation diagrams with the ggrisk package to measure the clinical value of the nomogram.

Up to 936 GC patients from 4 cohorts (TCGA, GSE62254, GSE15459 and Nanfang Hospital) were enrolled in this study. We collected and analyzed their clinical characteristics, which were summarized in Table 1: 44 patients in the TCGA cohort [38 (86.4%) were male, 6 (13.6%) were female and the mean age was 64.73 ± 9.33 years], 300 patients in the GSE62254 cohort [199 (66.3%) were male, 101 (33.7%) were female and the mean age was 61.94 ± 11.36 years], 192 patients in the GSE15459 cohort [123 (65.1%) were male, 69 (34.9%) were female and the mean age was 64.37 ± 13.24 years], 400 patients in the Nanfang Hospital cohort [276 (69.0%) were male, 124 (31.0%) were female and the mean age was 56.04 ± 10.87 years]. More information on tumors was analyzed and is listed in Table 1, including the depth of invasion, lymph node metastasis, metastasis, TNM stage, pathological type, venous invasion, lymphovascular invasion and perineural invasion.

The immunocyte composition of the TME was calculated with RNA sequencing data from the TCGA cohort (Fig. 2A),the GSE62254 cohort (Additional file 3: Figure S1) and the GSE15459 cohort (Additional file 4: Figure S2). As a training cohort, 6 types of immunocytes (plasma cells, CD8 + T cells, memory resting CD4 T cells, Tregs, M2 macrophages, and M0 macrophages) were selected by LASSO Cox regression analysis based on the TCGA cohort (Fig. 2B, C) and calculated the immunoScore by a Cox proportional hazards model (Additional file 2). The prognostic accuracy of the IS was assessed through time-dependent ROC analysis (2-year AUC = 0.798, and 4-year AUC = 0.873, Fig. 2D). The same calculations were performed for GSE62254 and GSE15459 as the validation cohorts (the results are shown in Additional file 2, Additional file 10: Figure S3 and Additional file 11: Figure S4). The survival curves of High-IS and Low-IS groups were generated with Kaplan–Meier analysis [TCGA cohort: p < 0.001, HR = 4.55 (95% CI 1.92–10.82), GSE62254 cohort: p = 0.004, HR = 1.60 (95% CI 1.17–2.21), GSE15459 cohort: p = 0.015, HR = 1.64 (95% CI 1.09–2.47), Fig. 2E–G].

Based on the gene expression data of the GSE62254 cohort, we obtained differentially expressed cells from the immune microenvironment (Fig. 2H). Notably, the infiltration of plasma cells, CD8 + T cells, activated CD4 + memory T cells, follicular helper T cells, activated NK cells and M1 macrophages significantly decreased with the higher immune risk score, while the infiltration of resting CD4 + memory T cells and M2 macrophages was significantly higher with the higher IS, which was likely related to the immunosuppressive reaction of the TME. Furthermore, we analyzed the relationship between IS and immune checkpoints, such as PD-1, PD-L1, and CTLA4 (more details are shown in Additional file 5: Table S1). In summary, IS has a strong relationship with immune cells and immune checkpoints.

As a training cohort from TCGA, radiomics signatures were compared with 44 cases based on the CT image data, which showed the best effect when 6 features were included (Fig. 3A, B). The ability of the RIS model to distinguish High-IS or Low-IS accurately and predict prognosis exhibited an AUC of 0.85 (Fig. 3D) and 0.81 (Fig. 3E), respectively. The ability of RIS to predict survival had an AUC of 0.72 (Fig. 3F) in the validation cohort from Nanfang Hospital. In the training cohort, overall survival curves were generated with Kaplan–Meier analysis, comparing High-RIS with Low-RIS [p < 0.001, HR = 4.55 (95% CI 1.91–10.82), Fig. 3G], and the result was similar in the validation cohort [p < 0.001, HR = 2.72 (95% CI 1.96–3.78), Fig. 3H]. Furthermore, in the forest plot of the Nanfang Hospital cohort, the significant Hazard Ratios were found for RIS in each subgroup, more specifically, in each age stage (< 50, 50–65, > 65), each differentiation stage (low, medium, high), and regardless of vascular invasion, nerve invasion and lymph node invasion (Fig. 3C).

A nomogram was developed by Cox regression based on the RIS and clinical features (Fig. 4A), and the Cox regression coefficient and corresponding score are shown in Additional file 6: Table S2. From the results of univariate survival analysis, RIS, tumor metastasis, lymph node positive detection rate, T stage and age were significant factors associated with OS and DFS (Additional file 7: Table S3, Additional file 8: Table S4). By calculating the total score and finding the corresponding prediction probability on the total point scale, the 1-year, 3-year and 5-year estimated survival probability could be obtained. Based on calibration plots for predicting 1-year, 3-year and 5-year survival (Fig. 4B–D), the predicted calibration curve was near the standard curve, or the 45-degree line, demonstrating that the derived nomogram had considerable prediction capability. Likewise, the decision curve analysis (DCA) depicted in Fig. 4E indicated that the nomogram had a higher net income than traditional TNM among a series of risk thresholds. The results of comparing the nomogram with the TNM staging system by Net Reclassification Improvement (NRI) are as follows: 1-year (NRI = 0.27, 95% CI = 0.07–0.44), 3-year (NRI = 0.03, 95% CI =  − 0.17–0.18), and 5-year (NRI = 0.05, 95% CI =  − 0.26–0.26), indicating that the nomogram had a relatively accurate prognosis. Simultaneously, the risk factor association diagram (Fig. 4F) showed the distribution difference of prognosis between High-nomogram and Low-nomogram groups distinguished by the nomogram. To predict the survival of GC patients, the performance of the nomogram was better than that of TNM staging (training cohort: nomogram AUC = 0.849 and TNM staging system AUC = 0.778, p = 0.007, Fig. 4G; validation cohort: nomogram AUC = 0.863 and TNM staging system AUC = 0.776, p = 0.001, Fig. 4J). The results of the time-dependent ROC curve analysis in the training cohort and validation cohort are shown in Fig. 4H, K. Comparing the High-nomogram and Low-nomogram groups, the survival curves of the training cohort and validation cohort are shown in Fig. 4I, L, which indicated that the prognosis of the Low-nomogram group was better than that of the High-nomogram group.

In addition, we analyzed differences in chemotherapy benefit among TNM stage II and III GC patients with different groups based on the nomogram. Chemotherapy dose leads to survival benefit for GC patients (OS: p = 0.023, and DFS: p = 0.005), but the degree of benefit varied among patients in different nomogram groups (High-nomogram group and Low-nomogram group) (Fig. 5). In the High-nomogram group, GC patients could significantly benefit with chemotherapy regarding OS and DFS (OS: p = 0.004, and DFS: p < 0.001), while in the Low-nomogram group, chemotherapy did not provide a survival benefit to patients compared with no chemotherapy (OS: p = 0.21, and DFS: p = 0.47). Furthermore, we analyzed the interactions between chemotherapy and death or recurrence in the High-nomogram and Low-nomogram groups with Cox proportional hazards regression, and High-nomogram patients benefited more from chemotherapy [OS: HR 0.41 (0.26–0.64), p < 0.001; DFS: HR 0.54 (0.34–0.84), p < 0.001] than Low-nomogram patients (Additional file 9: Table S5). In summary, High-nomogram GC patients benefit more from chemotherapy than Low-nomogram GC patients.