Ubiquitin-proteasome system reveals clinical subtypes with distinct molecular characteristics and prognoses in gastric cancer

Highlights

• •Ubiquitination-proteasome system modulates tumor microenvironment in gastric cancer.
• •Ubiquitination-proteasome system of gastric cancer is heterogeneity.
• •Double times unsupervised consensuses identify two distinct molecular subtypes.
• •Classification of different molecular subtypes helps risk stratification.
• •Our risk scoring system shows robustness of predicting effect of immunotherapy.

Abstract

Modulators in ubiquitin-proteasome system (UPS) has been implicated in regulating cancer-related genes, immune responses, and oncogenesis. However, the global UPS expression pattern and its role in gastric cancer (GC) pathology remain elusive. Herein, we integrated the modulators in UPS and dissected their associations with tumor microenvironment (TME), therapeutic response and prognosis in GC. Ten eligible GC cohorts (n = 2161) were collected in this comprehensive analysis. Unsupervised clustering based on expression profile of ubiquitination regulators was performed to identify distinct expression pattern. Then, pathway activation, and TME characteristics and prognosis were explored for patients in each pattern. Finally, a UPS scoring system in GC, termed UPSGC, is developed for individualized quantification of UPS expression pattern. Two prognosis-distinctive UPS expression patterns were identified and validated. Multiple interdependent characteristics were found in each pattern. Patients in the pattern with poor prognosis were found with activation of EMT, TNFα/NF-κB and IL6/JAK/STAT3 signaling, and more infiltration of immunosuppressive M2 macrophages and Th2 cells in TME. And another pattern was characterized by upregulation of angiogenesis, Notch and Wnt-β/catenin signaling, as well as enrichment of microvessels in TME. Based on the UPSGC system, two pattern-related clinical subtypes were identified. Finally, the UPSGC subtypes were validated as robust biomarker to predict patient's therapeutic responses and survival outcomes. In conclusion, this study proposes two previously unexplored UPS expression patterns in GC, in which patients have distinct survival outcomes and molecular characteristics. The findings provide new evidences to support the clinical relevance of ubiquitination with personalized therapy.

Introduction

Gastric cancer (GC) is the fifth most common malignance and the third leading cause of cancer death worldwide [1]. To date, there are still no efficient treatments and striking target medicines to alleviate and permanently cure GC, especially for those advanced diseases. One of the reasons responsible for above dilemma is the elusive molecular and clinical characteristics of GC which are complicated by its high heterogeneity in histology and etiology [2,3]. Therefore, it is urgently needed to uncover the complex genetic characteristics of GC and to characterize distinct clinical subtypes predictive of therapeutic responses and prognoses for patients, which could provide underlying foundation to improve the individualized strategies for clinical decision-making and follow-up schedule.

Ubiquitination is a widespread form of post-translational modifications found in nearly all eukaryotic organisms, in which the ubiquitin (a protein of ∼ 6-amino acid) is attached to the protein substrate [4]. The activity of ubiquitination modification is catalyzed by the ubiquitin-proteasome system (UPS), which maintains the dynamic reversibility between the ubiquitination and deubiquitination modifications of cellular protein [5]. The role of the UPS facilitates the regulation of proteasome stabilization and degradation and takes on most cellar activities, including autophagy, DNA damage repair, cell cycle, inflammatory signaling and etc. [6], [7], [8]. Over the past few decades, abundant studies have confirmed that the UPS plays a critical role in regulating cancer-related genes and the ingredients of tumor microenvironment [9,10]. For instance, the E3 ligase MDM2 in UPS can promote the degradation of tumor suppressor gene p53, which controls the regulation of cell cycle and DNA repair [11]. TRFA6, another E3 ligase, has been shown to play an important role in the regulation of NF-κB signaling pathway triggered by Helicobacter pylori infection [9]. The NF-κB signaling pathway participates in a broad spectrum of biological processes linking persistent infection and chronic inflammation to raise cancer risk [12]. Intriguingly, the UPS has also been proved to have tumor suppressive effects in multiple researches [13,14]. Given to the double-edged functions of the UPS in cancer, it is desperately necessary to comprehensively investigate the molecular landscape of the UPS and to uncover the underlying interaction and crosstalk among various ubiquitination modulators in UPS, thus helping to understand the complex effects of UPS on GC pathology. However, most of previous studies have confined to single gene or single signaling pathway to reveal the effects of ubiquitination modulators, owing to their limitations in inevitable technology or study design.

Herein, we hypothesized that the global effect of UPS on GC is characterized by numerous ubiquitination modulators in UPS that interact in a highly corporate manner. By integrating the expression pattern of 830 ubiquitination modulators in UPS, coupled with patient's clinicopathologic information in a meta-cohort of 1828 GC patients, we aimed to characterize distinct clinical subtypes predictive of therapeutic responses and survival outcomes in GC. We finally established and validated a UPS-associated risk scoring system for patient prognostication.

Materials and methods

Dataset source and preprocessing

Gene-expression data and full clinical annotation for GC were publically retrieved from Gene-Expression Omnibus (GEO, available at https://ncbi.nlm.nih.gov/geo/) database and the Cancer Genome Atlas (TCGA, available at https://portal.gdc.cancer.gov/) portal. For the TCGA dataset, the project “Stomach Adenocarcinoma (STAD)” were included in this study, with excluding those without available clinical information and expression profile. For the GEO datasets, we performed a systematic search in the GEO database by using the MeSH term “Stomach Neoplasms”, and selected the datasets derived from study type “Expression profiling by array” and organism “Homo sapiens”. To reduce the biological bias caused from different platforms, only arrays derived from platforms “Affymetrix” and “Illumina” were included, and those datasets with <50 samples or without available clinical information were excluded. In total, nine GC cohorts from GEO (GSE13861, GSE15459, GSE26253, GSE26899, GSE26901, GSE29272, GSE57303, GSE62254 and GSE84437) and one cohort from TCGA (TCGA-STAD) were gathered in this study for further analyses (Table S1). For microarray data from Affymetrix®, we downloaded the raw “CEL” files and adopted a robust multiarray averaging method with the affy package to perform background adjustment and quantile normalization. For microarray data from other platforms, the normalized matrix files were directly downloaded. Finally, nine GEO databases were integrated into a meta-cohort containing a total of 1828 GC patients. Batch effects across datasets were reduced by using a “Combat” algorithm implemented in R package sva [15]. For the TCGA-STAD dataset (n = 333), RNA sequencing data (Fragments Per Kilobase Million, FPKM) were obtained from the University of California Santa Cruz Xena (USCS Xena, available at https://xenabrowser.net/datapages/) [16]. Then FPKM values were transformed into Transcripts Per Kilobase Million (TPM) value for the final quantification of expression.

We also collected the IMvigor210 cohort by using the IMvigor210CoreBiologies R package. The IMvigor210 cohort (n = 298) was retrieved from a previous study observing efficacy of anti-PD-L1 blocker (i.e., atezolizumab) in patients with advanced urothelial cancer [17]. The complete expression data and the clinical outcome can be obtained from http://research-pub.gene.com/IMvigor210CoreBiologies. The raw count data were normalized by DEseq R package, and the count data were then transformed to TPM value as expression quantification for further analyses.

Modulators in UPS

A list of 1597 ubiquitination modulators in UPS were collected from the integrated annotations for Ubiquitin and Ubiquitin-like Conjugation Database 2.0 (iUUCD 2.0, available at http://iuucd.biocuckoo.org/), including 9 E1s, 43 E2s, 919 E3s, 126 DUBs, 387 UBDs and 113 ULDs [18]. Only 830 modulators with available expression in above meta-cohort were used to further explore the UPS expression pattern. Details of retrieved modulators were summarized in Table S2.

Identification of UPS expression pattern

The expression profiles of 830 ubiquitination modulators were extracted from the integrated meta-cohort (containing 1828 GC samples), and an unsupervised clustering algorithm was applied to identify the global expression pattern of UPS and to categorize samples for subsequent analyses. The optimal cluster number and corresponding robustness were evaluated by using a consensus clustering algorithm, which was implemented in the ConsensusClusterPlus package [19]. We set the parameters as 1000 iterations with 80% as resampling proportion.

Gene set variation analysis and functional annotation

To assess the variations of distinct patient subtypes in biological function, we performed gene set variation analysis (GSVA) to quantify the activation of pathways and related biological processes with a method of ssGSEA, by using a R package GSVA [20]. Two signature gene sets were downloaded from Molecular Signatures Database (MSigDB, available at http://www.gsea-msigdb.org/gsea/msigdb), including “c5.all.v7.5.1.symbols” and “h.all.v7.5.1.symbols”. An adjusted p value less than 0.05 was considered as statistical significance. The clusterprofiler package was used to perform the gene ontology (GO) enrichment analysis, including biological process (BP), cell components (CC) and molecular functions (MF), with the cutoff of false discovery rate (FDR) as 0.05 [21]. In addition, we also retrieved the gene set curated by Mariathasan et al. to perform GSVA enrichment analyses for quantifying the activation of epithelial-mesenchymal transition (EMT) and angiogenesis [17].

Estimation of tumor microenvironment infiltration

The overall abundances for immune and stromal infiltration in the tumor microenvironment (TME) were imputed by using the Estimate R package [22]. Further, the relative abundances of 40 specific TME-infiltrating cell types were estimated for each sample by using the xCell algorithm, based on transcriptomic profiles [23]. Wilcoxon test was employed to identify the abundance difference between distinct subtypes, and an adjusted p less than 0.05 was considered as statistical significance.

Construction of individual-level UPS scoring system and prognostication nomogram

To quantify the UPS expression pattern for individual patients, we applied a predefined analytic framework to generate an individual-level UPS scoring system in GC (UPS_GC) as followed.

• 1)Identify UPS expression pattern-related genomic subtypes. First, the limma method was used to identify differentially expressed genes (DEGs) between distinct UPS expression patterns [24]. The statistical threshold was set as an adjusted p less than 0.001 with the absolute value of log₂Fold-Change more than 1. Then, univariate Cox regression analyses were performed to screen out those DEGs with significant prognostic values under the cutoff of p set as 0.01. Finally, an unsupervised consensus clustering analysis was conducted by using the expression profiles of above prognostic pattern-related DEGs. And the UPS expression pattern-related genomic subtypes were identified from the optimal clustering results.
• 2)
  Construct UPS_GC scoring system. Based on the prognostic DEGs in above UPS expression pattern-related genomic subtypes, a UPS_GC scoring system was generated as followed:

  $$\text{UP}{\mathrm{S}}_{GC}=\sum \left(Ex{p}_i\times Bet{a}_i\right)$$

where Exp and Beta is the normalized expression value and corresponding Cox coefficient, respectively, for the prognostic DEGs.

Multivariate Cox regression analyses were then used to testify the value of UPS_GC scoring system as an independent prognostic predictor for GC patients. An associated nomogram was constructed as a prognostic model to predict the survival probability by using rms package. Calibration curve and receiver operating characteristic (ROC) curve were employed to validate the model efficacy on predicting 1-, 3- and 5-year overall survival probabilities.

Statistical analysis

Differences between two groups were tested using Wilcoxon test. The Kruskal-Wallis test was used to compare the differences across three or more groups [25]. By employing the “surv-cutpoint” function in survminer package, patients were categorized into high- and low-UPS_GC subtypes. The Kalan-Meier survival curves were used to show prognosis difference between distinct groups and log-rank tests were utilized to identify the statistical significance of differences. The forestplot package was used to visualize the hazard ratio and corresponding 95% confidence interval for survival-associated factors. The p value <0.05 was considered statistically significant. All data analyses were conducted in R (4.1.1 version) with related packages.

Results

Two distinct UPS expression patterns in GC

Nine GC datasets from GEO were collected to explore UPS expression patterns. We noticed significant batch effects across the expression profiles from different datasets (Fig. 1A). Thus, we employed a “Combat” algorithm to reduce the batch effects, and integrated nine datasets into a meta-cohort (n = 1828) for further analyses (Fig. 1B). Based on the expression profiles of 830 ubiquitination modulators in UPS, consensus clustering analysis identified two optimal clusters of patients in the meta-cohort (Fig. 1C, see Methods for details). Whereby, a total of 855 and 973 GC patients were classified into cluster_1 and cluster_2, respectively. The distinct UPS expression patterns for above two patient clusters were shown in Fig. 1D. The ubiquitination-associated GO enrichment analysis found different activities of ubiquitination-related processes between patients with distinct UPS expression patterns (Table S3). For instance, the activity of CUL2-ring ubiquitin ligase complex was upregulated in cluster_1, while CUL4B-ring ubiquitin ligase complex was activated in cluster_2. In cluster_2, protein K48-linked ubiquitination was enriched; however, the inversed process of protein K48-linked deubiquitination was activated in cluster_1 (Fig. 1E). These analyses indicated distinct ubiquitination characteristics exist in patients with distinct UPS expression patterns. We further evaluated the association of UPS expression patterns with patient's prognosis. The results revealed distinct survival outcomes between patients with distinct patterns; the cluster_1 had a significant survival advantage as compared against the cluster_2 (p <0.0001; Fig. 1F). These findings suggested that the UPS expression patterns might make underlying impacts on GC pathology and patient's prognosis through maintaining distinct ubiquitination landscapes in individual tumors.

Fig. 1.

Identification of UPS expression patterns in GC. (A-B) Principle component analysis of expression profiles from nine GEO datasets before (A) and after (B) removing batch effects. (C) The consensus matrix (optimal cluster number k = 2) from consensus clustering analysis of GC samples in the integrated meta-cohort by using ubiquitination modulators in UPS. (D) Heatmap showing the distinct expression patterns of 830 ubiquitination modulators in UPS for the two patient clusters in the meta-cohort. (E) Box plot showing differences in activity of ubiquitination-associated processes under distinct UPS expression patterns. (F) Kaplan-Meier curves comparing the differences in overall survival for GC patients with distinct UPS expression patterns in the meta-cohort. ****, p<0.0001.

Cancer hallmarks and TME infiltration characteristics under distinct UPS expression patterns in GC

To explore the biological functions involved in the effects of expression patterns of UPS on the GC tumorigenesis and prognosis, GSVA enrichment analysis was performed to investigate the activation of cancer hallmarks between the patterns of cluster_1 and cluster_2. Significant dysregulations of cancer hallmark pathways were revealed between the two patterns (Fig. 2A; Table S4). In detail, the activities of angiogenesis, Notch signaling and Wnt-β/catenin signaling were highly upregulated in cluster_1 pattern, while the cluster_2 pattern was characterized by the activation of epithelial-mesenchymal transition (EMT), TNFα/NF-κB signaling, and IL6/JAK/STAT3 signaling (Fig. 2A). These activations of cancer hallmarks might be responsible for, at least in part, the underlying molecular mechanisms as well as the malignant outcomes for GC with distinct expression UPS patterns.

Fig. 2.

Differences in cancer hallmark activity and TME infiltration between the two UPS patterns. (A) Bar plot showing the differences in hallmark activation between clusters_1 and cluster_2 patterns. (B) Box plot comparing immune score, stromal score, and tumor purity between the two patterns. (C) The bar plot showing the enrichment of immune and stromal cell types in each pattern. A positive (or negative) height of bar indicates enrichment of specific cell type in cluster_1 (or cluster_2). The color of the bar indicates the statistical significance for enrichment. **, p<0.01; ***, p<0.001.

TME infiltration has been demonstrated to play a critical role in GC pathology [26]. We next sought to analyze the TME infiltration characteristics in the tumor entities with distinct UPS expression patterns. We firstly estimated the overall score for immune and stromal fractions and tumor purity for each tumor. It was found that tumors in cluster_2 yielded significantly higher level of immune and stromal scores and had relative lower tumor purity, as compared with those in cluster_1 (Fig. 2B). We further compared the infiltration abundances for specific types of immune cells and stromal cells across tumors in distinct patterns (Fig. 2C; Table S5). The results showed that the cluster_1 had higher infiltration abundances of natural killer T (NKT) cells, CD8⁺ naive T-cells, type 1 T-helper (Th1) cells, microvascular (mv) endothelial cells, lymphatic (ly) endothelial cells and fibroblasts. Whereas, tumors in the cluster_2 were infiltrated by more M2 macrophages, type 2 T-helper (Th2) cells, CD4⁺ memory T cells and dendritic cells. This implied that the alterations of Th1/Th2 balance between tumors with distinct UPS expression patterns, and the increased infiltrations of immunosuppressive Th2 cells and M2 macrophages might facilitate immune tolerance in tumors of cluster_2. Consistent with abovementioned activated cancer hallmark of angiogenesis in cluster_1 (Fig. 2A), tumors in cluster_1 tended to harbor more microvessels (e.g., microvascular and lymphatic endothelial cells) in TME. The different survival outcomes for tumors with distinct UPS expression patterns might be associated with these distinct TME infiltration characteristics.

UPS_GC scoring system for individualized quantification of UPS expression patterns in GC

To quantify the individualized UPS expression patterns in GC, we further investigated the genomic signatures underlying each UPS pattern. A total of 113 pattern-related DEGs were identified across different patterns, including 68 and 45 significantly upregulated genes in cluster_1 and cluster_2, respectively (Fig. 3A; Table S6). As expected, we found these pattern-related DEGs efficiently distinguished patients into their original UPS expression patterns (Fig. 3B). This indicated that the identified set of DEGs contained genomic variations across distinct patterns, and could be taken as representative genomic signatures to further characterize the UPS expression patterns. Subsequent enrichment analyses revealed that the DEGs upregulated in cluster_1 pattern were associated with cellular protein transportation and localization, such as protein localization to endoplasmic reticulum (Fig. 3C). And beyond the protein localization, those DEGs upregulated in cluster_2 pattern were also involved in the process of immune response and immunocyte activities (Fig. 3D).

Fig. 3.

Identification of pattern-related DEGs in GC. (A) Volcano plot showing the fold change and statistical significance for DEGs upregulated in cluster_1 and cluster_2 patterns. (B) Heatmap showing the distinct expression profiles of the pattern-related DEGs across two UPS expression patterns. (C-D) Dot plots showing the GO enrichment analyses for the DEGs upregulated in cluster_1 (C) and cluster_2 (D) patterns, respectively. (E) Forest plot presenting hazard ratios and corresponding 95% confidence intervals for pattern-related prognostic DEGs. (F) The consensus matrix (optimal cluster number k = 2) from consensus clustering analysis of GC samples in the meta-cohort by using the prognostic pattern-related DEGs. (G) Heatmap showing the distinct expression profiles of the prognostic pattern-related DEGs between the two genomic subtypes, i.e., cluster_A and cluster_B. The classifications of UPS expression patterns and genomic subtypes are both used as sample (column) annotations for comparison. (H) Kaplan-Meier curves comparing the differences in overall survival for GC patients with distinct genomic subtypes in the meta-cohort.

Next, based on the above pattern-related DEGs, we further carried out univariate Cox analyses and identified a total of 28 pattern-related DEGs associated with prognosis, including 13 favorable prognostic DEGs and 15 unfavorable prognostic DEGs (Fig. 3E; Table S7). Based on these prognostic pattern-related DEGs, we employed similar unsupervised clustering algorithm to assign patients in the meta-cohort and identified two genomic subtypes, namely cluster_A and cluster_B (Fig. 3F). As a result, 856 patients were classified into the cluster_A genomic subtype which is related to cluster_1 pattern, while the remaining 972 patients were classified into the cluster_B genomic subtype which is associated with cluster_2 pattern (Fig. 3G). Interestingly, as shown in Fig. 3G, the genomic subtype cluster_A was characterized by the increased expression of the pattern-related DEGs with favorable effects on prognosis, while the genomic subtype cluster_B was observed with significant overexpression of those pattern-related DEGs with unfavorable effects on survival outcomes. In agreement with the UPS expression patterns, survival comparison analysis revealed that the cluster_1 pattern-associated genomic subtype cluster_A yielded better overall survival outcome than the cluster_2 pattern-associated genomic subtype cluster_B (Fig. 3H). These findings suggested that there did exist distinct UPS expression patterns associated with patient's prognosis in GC.

Finally, by integrating the prognostic pattern-related DEGs, we developed a UPS_GC scoring system for individualized quantification of UPS expression patterns (see Methods for details). By employing the scoring system, a UPS_GC score was calculated for each GC patient in the meta-cohort. We found that patients in cluster_2 pattern who yielded poor prognoses had significant higher level of UPS_GC score than those in cluster_1 pattern and with better survival outcomes (Fig. 4A). Similar results were also obtained between the two pattern-related genomic subtypes (Fig. 4B). To further evaluate the role of UPS_GC scoring system in risk stratification, we further dichotomized patients into high- and low-UPS_GC subtypes based on their UPS_GC scores, by using a method of maximally selected rank statistics [27]. Subsequent survival comparison revealed that the high-UPS_GC subtype had significant worse survival outcome than the low-UPS_GC subtype (Fig. 4C). We further performed multivariate Cox analyses to analyze the value of UPS_GC subtypes as an independent prognostic predictor. The results suggested that the high-UPS_GC subtype was a significant risk factor after adjusting clinicopathological factors, including age, sex and TNM stage (Fig. 4D). Then, we established a UPS_GC-associated prognostic nomogram model to predict 1-, 3- and 5-year overall survival probability for GC patients (Fig. 4E). The calibration curve validated well prediction performance of the model, when compared against actual survival probability (Fig. 4F). The area under the ROC curve was 0.766, 0.776 and 0.769 at 1-, 3- and 5-year overall survival probability, respectively (Fig. 4G). We further performed external validations for above results by using another independent TCGA-STAD cohort containing 333 GC patients. Similar findings were obtained that patients of low-UPS_GC subtype yielded better survival outcomes than those of high-UPS_GC subtype (Fig. 4H), and the high-UPS_GC subtype was an independent risk factor for prognosis (Fig. 4I). The developed prognostic nomogram could promisingly predict survival outcome and the area under ROC curve was 0.668, 0.671 and 0.729, respectively, at 1-, 3- and 5-year overall survival probability for patients in the TCGA-STAD cohort (Fig. 4J-K).

Fig. 4.

Generation of UPS_GC scoring system and prognostic nomogram in GC. (A) Box plot showing comparison of UPS_GC score between cluster_1 and cluster_2 patterns. (B) Box plot showing comparison of the level of UPS_GC score between cluster_A and cluster_B pattern-related genomic subtypes. (C) Kaplan-Meier curves comparing the differences in overall survival between GC patients of high- and low-UPS_GC subtypes in the meta-cohort. (D) Multivariate Cox regression analyses showing independent prognostic effects of UPS_GC subtypes on overall survival in the meta-cohort. (E) The UPS_GC-associated prognostic nomogram for prediction of 1-, 3- and 5-year survival probability for GC patients. (F) Calibration curve showing comparison between the nomogram-predicted survival probability and actual survival outcome for patients in the meta-cohort. (G) ROC curve showing efficacy of the nomogram to predict 1-, 3- and 5-year overall survival probability for patients in the meta-cohort. (H) Kaplan-Meier curves comparing the differences in overall survival between GC patients of high- and low-UPS_GC subtypes in the TCGA-STAD cohort. (I) Multivariate Cox regression analyses showing independent prognostic effects of UPS_GC subtypes on overall survival in the TCGA-STAD cohort. (J) Calibration curve showing comparison between the nomogram-predicted survival probability and actual survival outcome for patients in the TCGA-STAD cohort. (K) ROC curve showing the efficacy of the nomogram to predict 1-, 3- and 5-year overall survival probability for patients in the TCGA-STAD cohort.

Association of UPS_GC subtypes with canonical molecular characteristics and therapeutic response

We next analyzed the association of the UPS_GC score with clinical and molecular characteristics. In the meta-cohort, patients with metastatic tumor showed significantly higher UPS_GC score than those without metastasis (Fig. 5A). In previous TCGA-STAD project, GC patients were classified into four molecular subtypes, including genome stable (GS), microsatellite instability(MSI), EBV infection, and chromosomal instability (CIN) subtype [3]. In our analysis, the GS subtype had the highest UPS_GC score, followed by CIN, EBV and MSI (Fig. 5B). However, different from our UPS_GC subtypes, outcome data from TCGA-STAD project did not reveal survival differences across the four molecular subtypes. Another classification from Singapore proposed three subtypes in GC, including mesenchymal (including unstable and invasive), metabolic and proliferative subtypes [28]. We found that the metabolic subtype which was characterized by high sensitivity to 5-fluorouracil-based chemotherapy had lower UPS_GC score than other subtypes (Fig. 5C). Among four prognosis-distinctive subtypes proposed by Asian Cancer Research Group (ACRG) [2], the EMT subtype in which patients yielded worst prognosis had highest UPS_GC score than other three subtypes (Fig. 5D). As we mentioned above, the cluster_2 pattern was characterized by activation of EMT, and while the cluster_1 pattern was found with upregulated angiogenesis. We further used the gene signatures curated by Mariathasan et al. to quantify the activation of EMT and angiogenesis (see Methods for details). As expected, high EMT activity was found in the high-UPS_GC subtype (Fig. 5E), while active angiogenesis was observed in the low-UPS_GC subtype (Fig. 5F).

Fig. 5.

Association of UPS_GC subtypes with canonical molecular characteristics. (A) Comparison of the UPS_GC score between metastatic and non-metastatic patients in GEO meta-cohort. (B-D) Bar plot showing differences in UPS_GC score between distinct TCGA (B), Singapore (C) and ACRG (D) molecular subtypes. (E-F) Comparison of activation score for EMT (E) and angiogenesis (F) between low- and high-UPS_GC subtypes in GEO meta-cohort. The error bar embedded in the bar represents standard error.

Immunotherapy, such as immune checkpoint blockade of anti-PD-1 and anti-PD-L1, has increasingly been revolutionary treatment for malignances. We sought to examine the efficacy of UPS_GC score to predict patients’ responses to anti-PD-L1 immunotherapy in IMvigor210 cohort. It was found that patients in low-UPS_GC subtype yielded more survival benefits than those in high-UPS_GC subtype from immunotherapy (Fig. 6A). Increased UPS_GC score indicated less response for patients with anti-PD-L1 therapy (Fig. 6B), and the patients in low-UPS_GC subtype yielded higher response rate than those in high-UPS_GC subtype (Fig. 6C). We also assessed the variation of UPS_GC score among distinct immune phenotypes in IMvigor210 cohort. Relative higher UPS_GC score was observed in the “desert” and “excluded” immune phenotypes, in which the efficacy of immunotherapy was limited (Fig. 6D). It was demonstrated that patients with high expression of CD274 (also known as PD-L1) are more likely to benefit from immunotherapy. We found the CD274 expression was increased in the low-UPS_GC subtype (Fig. 6E). Moreover, tumor mutation burden (TMB) and tumor neoantigen burden (TNB) were also taken as underlying biomarkers for immunotherapy response. Similarly, the low-UPS_GC subtype had higher level of TMB and TNB than the high-UPS_GC subtype (Fig. 6F-G). Further, we found the combination of UPS_GC subtypes with TMB or TNB could increase the prognostication efficacy for patients with immunotherapy (Fig. 6H-I). These findings suggested that the UPS_GC subtypes could serve as a new biomarker to predict the response to immunotherapy.

Fig. 6.

The role of UPS_GC in anti-PD-L1 immunotherapy. (A) Kaplan-Meier curves comparing the differences in overall survival between patients of high- and low-UPS_GC subtypes after anti-PD-L1 therapy in the IMvigor210 cohort. (B) Differences in UPS_GC score among distinct anti-PD-L1 clinical response groups. SD, stable disease; PD, progressive disease; CR, complete response; PR, partial response. (C) Comparison of the response (CR/PR) rate between the high- and low-UPS_GC subtypes. (D) Differences in UPS_GC scores among distinct immune phenotypes in the IMvigor210 cohort. (E) Comparison of the expression of CD274 (PD-L1) between the high- and low-UPS_GC subtypes. (F-G) Differences in tumor mutational burden (F) and tumor neoantigen burden (G) between the high- and low-UPS_GC subtypes. (H-I) Kaplan-Meier curves showing survival comparison for patients receiving anti-PD-L1 immunotherapy stratified by combination of UPS_GC subtypes with tumor mutational burden (H) and tumor neoantigen burden (I).

Discussion

Increasing evidences have shown that UPS plays extensive roles in formation of TME, intracellular homeostasis and acquirement of malignant properties [29], [30], [31]. However, most of previous researches which are focused on the role of single ubiquitination modulator in prognostic efficacy or chemotherapy resistance provide incomplete understanding for this issue, since high heterogeneity exists not only across multiple ubiquitination types but also in the crosstalk among various modulators in UPS. In this study, we integrated known modulators in UPS and investigated the global variations in the landscape of UPS expression to explore underlying expression pattern of UPS in GC. Our findings could provide new evidences to support the association between the protein modification of ubiquitination and GC pathology. Two clinical subtypes identified in this study might provide new insights into developing strategies regarding personalized decision-making and follow-up schedule for GC patients in clinical setting.

Based on the global expression of ubiquitination modulators in UPS, we first identified two distinct UPS expression patterns in GC. Ubiquitination-associated functional enrichment analyses revealed the modification propensities were different across distinct patterns. The ubiquitination mechanism underlying the cluster_1 pattern tended to be the reverse of that within another pattern cluster_2. This seems to be linked to the variations in dynamic balance between the activities of ubiquitination and deubiquitination in these two patterns. Interestingly, we found significant association of such UPS expression patterns with patients’ prognosis, indicating the impact of UPS expression patterns on GC pathology. Our subsequent analyses found multiple interdependent characteristics for GC patients in the two prognosis-distinctive patterns. Patients in the cluster_2 pattern with poor prognosis were characterized by activation of EMT, TNFα/NF-κB and IL6/JAK/STAT3 signaling, as well as more infiltration of immunosuppressive M2 macrophages and Th2 cells in TME. Previous studies have demonstrated that the cytokine IL6 could induce acquisition of EMT in cancer cells via JAK2/STAT3 signaling pathway [32]. In addition, the TNFα/NF-κB signaling, as one of canonical proinflammatory pathways, has been verified to promote chronic inflammation and EMT during tumorigenesis [33,34]. It has also been demonstrated that the polarization of M2 macrophages has closed associations with EMT, as well as the invasion and migration of cancer cells [35]. In the cluster_1 pattern, patients were found with upregulation of angiogenesis, Notch and Wnt-β/catenin signaling. Previous works have reported that the process of angiogenesis could be enhanced by activated Notch and Wnt/β-catenin signaling pathway [31]. Notably, we found significant enrichment of microvessels (e.g. microvascular and lymphatic endothelial cells) in the TME of patients in cluster_1 pattern. Such specific TME-infiltrating characteristics might be explained by above activated angiogenesis in tumors of this pattern. So far it is hard to perform a quantitative comparison of different pathways (e.g. EMT and angiogenesis) in their contributions to GC development and survival outcome, thus it is elusive that the cluster_1 pattern which was characterized by angiogenesis had a survival advantage over the cluster_2 pattern in which EMT was activated. Actually, the difference in prognosis between different patterns was indicated by our developed UPS_GC scoring system, rather than above biological pathways (although they were found specifically in each pattern). Further, it remains an unsettled challenge for the elusive association between different pathways involved in cancer development and patient's survival outcome, and those new tools with incorporating multiple pathways, such the UPS_GC scoring system developed in our study, might provide underlying strategies to address above issue.

Further, in order to quantify the UPS expression pattern for individual GC patients, we developed a ubiquitination-proteasome scoring system in GC, named UPS_GC. Based on the UPS_GC score, we identified two clinical subtypes of GC, including low- and high-UPS_GC subtypes, which were associated with cluster_1 and cluster_2 patterns, respectively. Consistent with the findings of UPS expression patterns, we verified that patients in low-UPS_GC subtype yielded better survival outcomes than those in high-UPS_GC subtype. And high-UPS_GC subtype was proven to be an independent risk factor for overall survival in multivariate regression analyses. We finally constructed a UPS_GC associated prognostic nomogram to predict survival probability for GC patients. And prediction efficacy evaluation showed robust performance for our prognostication model. These analyses suggested that the UPS_GC clinical subtypes proposed in this study could serve as a promising tool for patient prognostication in GC.

Our data also revealed significant association of UPS_GC with previous molecular subtypes and therapeutic response. Metastasis is one of major reasons responsible for low therapeutic responses and poor clinical outcomes in GC [36]. We found patients with metastasis had higher UPS_GC score than those without metastatic tumors. We also analyzed the variations of UPS_GC score among multiple molecular subtypes proposed previously. The results revealed those molecular subtypes with poor therapeutic response and prognosis, such as genomic stable subtype, mesenchymal subtype, and EMT subtype, had relative higher UPS_GC score than other molecular subtypes. Notably, we have demonstrated that patients in the cluster_1 and cluster_2 patterns, related respectively to low- and high-UPS_GC subtypes, were characterized by upregulation of angiogenesis and EMT, respectively. In this analysis, we found consistent results that the low-UPS_GC subtype had high angiogenesis score while the high-UPS_GC subtype had elevated EMT score. We further evaluated the value of UPS_GC in predicting immunotherapy response and found the UPS_GC was associate with the efficacy of immunotherapy. Patients in low-UPS_GC subtype were more likely to benefit from anti-PD-L1 treatment. Our findings of elevated levels of CD274 (PD-L1) expression, TMB and TNB in low-UPS_GC subtype might be part explanations for above issue. The levels of PD-L1 expression, TMN and TNB have been taken as underlying biomarkers to predict benefits for immunotherapy in multiple malignances [37]. Our survival analyses revealed that the combination of UPS_GC subtypes with TNB or TNB could improve the risk stratification for patients received immunotherapy, indicating that the UPS_GC subtype could be a novel biomarker to help prognostication regarding clinical benefits from immunotherapy.

Conclusions

We propose two previously unexplored UPS expression patterns in GC, in which patients had distinct survival outcomes and molecular characteristics. Our developed UPS_GC scoring system could serve as a promising tool to predict prognosis and therapeutic response for GC patients. Our study provides new evidences to support the clinical relevance of ubiquitination with personalized therapy strategies.

Data availability statement

All datasets analyzed in this study are available in public databases. Details for dataset retrieval and utilized packages are presented in each subsection of “Materials and methods” section.

Funding

This work was supported by Hainan Provincial Natural Science Foundation of China [grant number 820QN390], National Natural Science Foundation of China [grant number 82203252], and National Natural Science Foundation of China [grant number 82103598].

Consent

Written informed consent was obtained from the patient for publication of this case report and accompanying images. A copy of the written consent is available for review by the Editor-in-Chief of this journal on request.

CRediT authorship contribution statement

Hao Wang: Conceptualization, Methodology, Writing – review & editing, Funding acquisition. Qinghua Lin: Formal analysis, Investigation, Writing – original draft, Visualization. Xiaosheng Wu: Validation, Funding acquisition, Conceptualization, Writing – review & editing. Baochun Wang: Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.