1. Department of Gastrointestinal Surgery, Zhongshan Hospital, Xiamen University, Xiamen, Fujian, China.
2. Institute of Gastrointestinal Oncology, School of Medicine, Xiamen University, Xiamen, Fujian, China.
3. Pancreatitis Center, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, Zhejiang, China.
4. State Key Laboratory Breeding Base of Marine Genetic Resources; Key Laboratory of Marine Genetic Resources, Third Institute of Oceanography, Ministry of Natural Resources, Xiamen, Fujian, China.
5. Laboratory of Xiamen Cancer Center, The First Affiliated Hospital of Xiamen University, Xiamen, Fujian, China.
6. Department of Pathology, School of Medicine, Jinan University, Guangzhou, Guangdong, China.
*These authors contributed equally to this work.
Rationale: The prognosis of gastric cancer (GC) patients is poor, and there is limited therapeutic efficacy due to genetic heterogeneity and difficulty in early-stage screening. Here, we developed and validated an individualized gene set-based prognostic signature for gastric cancer (GPSGC) and further explored survival-related regulatory mechanisms as well as therapeutic targets in GC.
Methods: By implementing machine learning, a prognostic model was established based on gastric cancer gene expression datasets from 1699 patients from five independent cohorts with reported full clinical annotations. Analysis of the tumor microenvironment, including stromal and immune subcomponents, cell types, panimmune gene sets, and immunomodulatory genes, was carried out in 834 GC patients from three independent cohorts to explore regulatory survival mechanisms and therapeutic targets related to the GPSGC. To prove the stability and reliability of the GPSGC model and therapeutic targets, multiplex fluorescent immunohistochemistry was conducted with tissue microarrays representing 186 GC patients. Based on multivariate Cox analysis, a nomogram that integrated the GPSGC and other clinical risk factors was constructed with two training cohorts and was verified by two validation cohorts.
Results: Through machine learning, we obtained an optimal risk assessment model, the GPSGC, which showed higher accuracy in predicting survival than individual prognostic factors. The impact of the GPSGC score on poor survival of GC patients was probably correlated with the remodeling of stromal components in the tumor microenvironment. Specifically, TGFβ and angiogenesis-related gene sets were significantly associated with the GPSGC risk score and poor outcome. Immunomodulatory gene analysis combined with experimental verification further revealed that TGFβ1 and VEGFB may be developed as potential therapeutic targets of GC patients with poor prognosis according to the GPSGC. Furthermore, we developed a nomogram based on the GPSGC and other clinical variables to predict the 3-year and 5-year overall survival for GC patients, which showed improved prognostic accuracy than clinical characteristics only.
Conclusion: As a tumor microenvironment-relevant gene set-based prognostic signature, the GPSGC model provides an effective approach to evaluate GC patient survival outcomes and may prolong overall survival by enabling the selection of individualized targeted therapy.
Keywords: gastric cancer, prognostic signature, machine learning, tumor microenvironment, targeted therapy