This study was a double-center retrospective clinical study registered in the Chinese Clinical Trial Registry (Approval No. ChiCTR1900024346). It was approved by the ethics committee of Shanghai Jiao Tong University Affiliated Sixth People’s Hospital (Approval No. 2019-KY-032K). The training cohort from Shanghai Jiao Tong University Affiliated Sixth People’s Hospital was used to construct the signatures, and then the external cohort from the Sixth Affiliated Hospital of Sun Yat-sen University to validate. All the patients underwent R0 radical resection and were pathologically diagnosed with stage I-III GC from January 2009 to December 2016. Written informed consent was waived because of the study design. The exclusion criteria were as follows: (1) stage IV GC; (2) non-adenocarcinoma; (3) no available preoperative serum tumor markers; (4) younger than 18 years old.

Clinical data was mainly provided by medical histories and the electronic medical record department. The pathological stage was accessed by the 8th AJCC criterion for GC. Preoperative serum tumor markers were examined within one week before R0 surgery (including subtotal gastrectomy or total gastrectomy). Gastric tumor sites include cardia, gastric body, and antrum carcinoma. Adjuvant therapy and follow-up were according to National Comprehensive Cancer Network (NCCN) guidelines. Disease-free survival (DFS) was defined as the time from surgery to tumor metastasis or recurrence. Overall survival (OS) was defined as the time from surgery to death.

Data processing and data analysis were performed in the R language (version 3.4.4). For prognostic signatures, we firstly performed a univariate analysis of all variables by Cox proportional hazards regression. Then, the significant covariates (score test p <0.05) were included in the multivariate Cox proportional hazards regression. Stepwise and backward selection processes were performed to achieve the final signatures according to the Akaike information criterion (AIC). Signature discrimination was examined by the concordance index (C-index) and corrected 1,000 times by bootstrapping. Furthermore, we classified the patients into three groups (high, intermediate, and low risk) based on the risk score. The Kaplan-Meier (K-M) curves were used to compare survival time among groups. The log-rank test was performed to calculate the significant differences. Nomograms were established to predict the survival rates. P values were two-sided, with statistically significant differences at p < 0.05.

From January 2009 to December 2016, a total of 995 patients were enrolled from two cohorts according to our inclusion and exclusion criteria. 585 patients with stage I-III GC from Shanghai Jiao Tong University Affiliated Sixth People’s Hospital were screened for the internal cohort to construct the signatures. The external cohort including 410 cases from the Sixth Affiliated Hospital of Sun Yat-sen University was used to validate the signatures ( Supplementary Figure 1 ).

Patients’ clinical information and follow-up data from the training cohort were shown in Supplementary Table 1 . The median age at the time of diagnosis was 63 years. 154 patients were diagnosed with stage I, 165 patients with stage II, and 266 cases with stage III. 557 patients had the pathological type of adenocarcinoma (95.2%), and 28 patients had the worse pathological type. LN metastasis occurred in 324 patients (55.4%). 103 patients had well and moderate tumor differentiation, while the others had poor differentiation. Lymphatic infiltration, vascular infiltration, and nerve infiltration occurred in 386, 141, and 461 patients, respectively. The median Ki67 was 55% (range: 5%-95%). The mean NLR and PLR were 2.92 and 158.12, respectively. The mean levels of CEA, CA19-9, and CA125 were 8.83 ng/ml, 17.47 U/ml, and 6.69 U/ml, respectively. The median follow-up time was 45 months. Recurrence occurred in 203 patients (34.7%). A total of 230 patients (39.3%) died during the follow-up.

To select the variables that were suitable for inclusion in our signatures, we performed univariate analyses. Our results suggested that 10 variables, including T stage, N stage, tumor CSA, differentiation, lymphatic infiltration, vascular infiltration, nerve infiltration, PLR, CEA, and CA19-9, were significantly associated with DFS and OS in GC patients ( Figures  1A, B ; p < 0.05). Pathological type affected DFS ( Figures  1A, B ; p < 0.05) but not OS, whereas NLR had the opposite effect. Therefore, there were 11 variables closely related to DFS or OS in GC.

Then, we used these variables as potential prognostic factors to construct prognostic signatures with the Cox proportional hazards model. AIC, stepwise, and backward analyses were performed as variable selection methods. Finally, 6 indicators were selected for the DFS signature in GC: T stage, N stage, vascular infiltration, nerve infiltration, CEA, and CA19-9. Our multivariate analysis results showed that they were all independent prognostic factors except nerve infiltration ( Figure 2A ; p < 0.05). The discriminative power, shown by the C index (concordance index), was 0.75, corrected with 1,000 permutations, for DFS in GC in the training cohort ( Supplementary Table 2 ). Our OS signature in GC also included 6 indicators, namely, T stage, N stage, lymphatic infiltration, vascular infiltration, CEA, and CA19-9. These variables were all independent factors for OS in GC ( Figure 2B ; p < 0.05). The C-index of our OS signature was 0.73, corrected with 1,000 permutations ( Supplementary Table 3 ). Overall, we constructed predictive prognostic signatures for DFS and OS in GC.

To evaluate the prognosis of patients more intuitively, we developed nomograms with Cox regression models. All the variables in the nomogram had a weighted score, and we could predict the 3-year or 5-year survival outcome by the sum of the scores ( Figure 3 ). To further verify the importance of these variables and calculate the risk score, we developed a nonparametric approach in our signature using random survival forest. Our results showed that the T stage had the largest influence on DFS and OS with a VIMP of positive value 0.6983 and 0.518, followed by vascular infiltration, N stage, CEA, and CA19-9 ( Supplementary Figure 2 ). The predictive accuracy of the signatures using time-dependent ROC analysis was relatively high. The AUC of the DFS signature based on the risk score was 0.808 at 3 years and was 0.799 at 5 years ( Figures 4A, C ), higher than the other single indicator. Similarly, the AUC of the OS signature was 0.785 at 3 years and was 0.793 at 5 years ( Figures 4B, D ). The calibration curves for GC based on our signatures showed an excellent correlation between predicted and observed outcomes for DFS and OS prediction at 3 years and 5 years ( Figure 5 ). All these results illustrated that the DFS and OS signatures had good accuracy and prediction ability.

We assessed patients according to the risk score achieved from the signatures using these different variables. Kaplan-Meier curves were applied to compare survival differences. In the DFS signature, compared with the low-risk group, the medium-risk group (HR 3.56; 95% CI, 2.30-5.49) and the high-risk group (HR 6.95; 95% CI, 4.95-9.75) had shorter survival times ( Figure 6A , p < 0.001). We found similar results for OS signature. Compared with the low-risk group, the intermediate- and high-risk groups had hazard ratios of 3.74 (95% CI, 2.65-5.28) and 8.18 (95% CI, 5.76-11.60), respectively ( Figure 6B , p < 0.001). Furthermore, the 5-year rate of the low-risk group was 84.0% for DFS and 83.0% for OS, much higher than the intermediate group (53.0% for DFS and 50.0% for OS) and high-risk group (35.0% for DFS and 24.0% for OS).

To examine whether the signatures were suitable for other centers, we collected data from an external cohort for validation. 410 patients with stage I-III GC were collected. Univariate analysis results showed that the variables included in the predictive DFS and OS signatures were all prognostic factors ( Supplementary Table 4 , p < 0.05) except CA19-9. The calibration curves suggested a perfect correlation between predicted and observed outcomes for DFS and OS prediction at 3 years and 5 years ( Figures 7A–D ). In addition, according to the scoring criteria, we divided the patients from the external cohort into three groups (high, intermediate, and low risk), and similar results were obtained. The DFS and OS curves in the validation cohort among the three groups were significantly different ( Supplementary Figures 3A, B , p < 0.001).

Based on the signatures, we developed a free website for predicting the prognosis of stage I-III GC (http://115.28.66.83/liuyuan/stad.php).