CMTM3 links inflammation-related transcriptional dysregulation to immune and therapeutic stratification in gastric cancer

Highlights

• •CMTM3 reflects inflammation-metabolism dysregulation in gastric cancer.
• •CMTM3 links inflammatory signaling, metabolic pathways, and immune remodeling.
• •High CMTM3 expression identifies poor-prognosis tumors with immune suppression.
• •CMTM3 knockdown reduces tumor growth and macrophage recruitment in vitro.
• •CMTM3 may guide immunotherapy, chemotherapy, and targeted therapy selection.

Abstract

Gastric cancer (GC) shows strong biological heterogeneity and frequent disruption of inflammatory and metabolic programs, which affect tumor progression, immune escape, and treatment response. To identify biomarkers related to this inflammation-metabolism axis, we developed an integrative framework combining single-cell RNA sequencing, bulk transcriptomic cohorts, machine learning, immune and drug-response prediction, and in vitro validation. Inflammation-related gene modules in malignant cells were identified using hdWGCNA, and candidate prognostic genes were screened by CoxBoost and Random Survival Forest models. CMTM3 was prioritized as a candidate inflammation-associated gene and was further evaluated across independent GC cohorts. High CMTM3 expression was associated with poor survival, immune cell infiltration, increased immune checkpoint expression, and enrichment of multiple immunotherapy-related signatures, indicating an immune-infiltrated but functionally suppressed tumor state. Pathway analyses linked CMTM3 to inflammatory signaling and metabolic regulation, suggesting a role in coordinating tumor inflammatory and metabolic programs. Drug sensitivity prediction showed that low-CMTM3 tumors may be more responsive to selected chemotherapies and kinase inhibitors, whereas high-CMTM3 tumors may be more suitable for immune checkpoint-based treatment. Functional assays further showed that CMTM3 knockdown reduced GC cell proliferation and weakened macrophage recruitment. Overall, this study identifies CMTM3 as a prognostic and predictive biomarker in GC and provides a practical strategy for translating inflammation-metabolism-related molecular dysregulation into precision therapeutic stratification.

Graphical abstract

Introduction

Gastric cancer (GC) remains one of the most common malignant tumors worldwide and is a major cause of cancer-related death [1]. Despite progress in surgery, chemotherapy, targeted therapy, and immunotherapy, many patients are still diagnosed at advanced stages and have limited survival benefit from current treatment options [[2], [3], [4], [5]]. This clinical difficulty is partly caused by the strong heterogeneity of GC. Patients with similar clinicopathological features may show very different immune states, metabolic activity, drug sensitivity, and survival outcomes. Therefore, better biomarkers are needed to guide risk assessment and treatment selection.

Inflammation is closely involved in GC development and progression [6]. Chronic inflammatory stimulation can promote malignant transformation, support tumor growth, reshape immune cell infiltration, and contribute to therapy resistance. At the same time, cancer-related metabolic changes can affect both tumor cells and immune cells. Altered nutrient use, energy metabolism, cytokine release, and immune suppression often occur together rather than as separate events. This makes the inflammation-metabolism axis an important layer for understanding GC biology and for developing more precise therapeutic strategies. However, the molecular factors that connect inflammatory programs, metabolic dysregulation, immune remodeling, and treatment response in GC remain insufficiently defined [7].

Single-cell RNA sequencing provides a useful way to examine cell states within tumors, while bulk transcriptomic cohorts allow candidate markers to be tested in larger patient populations. When combined with machine learning methods, these data can help identify genes that are not only biologically relevant but also clinically useful. In this study, we focused on CMTM3, a member of the chemokine-like factor family, which has been implicated in immune regulation and tumor biology [8,9]. However, its role in GC, especially its relationship with inflammation-related transcriptional programs, metabolic pathway activity, immune response, and therapeutic stratification, has not been fully clarified.

To address this question, we built an integrative framework combining single-cell analysis, bulk transcriptomic validation, machine learning-based feature selection, immune and drug response prediction, pathway analysis, and experimental assays. We first identified inflammation-associated gene modules in malignant cells and then used CoxBoost and Random Survival Forest models to screen robust prognostic genes. CMTM3 was selected as a key candidate and was further evaluated across independent cohorts. We also examined whether CMTM3 expression could reflect immune activation, immune suppression, metabolic pathway changes, and sensitivity to immunotherapy, chemotherapy, or targeted agents. This study aimed to evaluate whether CMTM3 is associated with inflammatory transcriptional programs, immune microenvironmental features, prognosis, and exploratory therapy-response signatures in GC.

Materials and methods

Dataset collection and processing

We obtained single-cell data for GC from the GSE163558 dataset [10]. The single-cell RNA-seq dataset comprised 10 gastric cancer (GC) samples profiled using the 10x Genomics platform. After quality control, 44,563 cells were retained for downstream analysis. For bulk transcriptomic data (tissue samples), we used the TCGA database [11] and several GEO datasets (GSE15459, GSE26253, GSE26901, GSE62254, GSE84426, GSE84433, and GSE84437) to cover a large number of patients [[12], [13], [14], [15], [16]]. Inclusion criteria for cohort selection were as follows: complete overall survival information and histologically confirmed gastric adenocarcinoma. For RNA-seq cohorts, RNA-seq expression values were processed as log2(TPM + 1) values before downstream analysis. For microarray cohorts, normalized expression matrices were downloaded from the GEO database and log2 transformation was performed when required.

Single-cell and multi-model analysis

A set of inflammation-related genes was obtained from the HALLMARK database [17]. To keep high-quality cells, we removed cells with fewer than 200 genes, >5000 genes, or mitochondrial gene percentages above 20%. After normalization, Harmony was used to reduce batch effects among different samples. Normal cell populations were annotated based on canonical marker genes (Table S1). Cancer and non-cancer cells were then distinguished using inferCNV [18]. For malignant cells, we built a scale-free co-expression network using the hdWGCNA package [19]. The soft-thresholding power was chosen when the scale-free topology fit index reached 0.85. Inflammation-associated gene modules were then identified, among which the purple module showed the strongest association.

Feature gene selection using machine learning

To screen robust prognostic genes, we applied two machine learning methods: CoxBoost [20] and Random Survival Forest (RSF) [21]. The dataset was randomly divided into a training cohort and a testing cohort at a 7:3 ratio. This process was repeated 10 times to test the stability of the selected genes. CoxBoost model fitting was performed using the CoxBoost R package. The optimal boosting step was selected by cross-validation, and genes retained by the model were recorded in each repeated split. Random Survival Forest analysis was performed using the randomForestSRC R package for feature ranking. Genes repeatedly retained across the 10 training/testing splits were considered stable candidate markers. Genes consistently retained across the models were considered candidate prognostic markers.

Functional analysis and prediction

Drug response was estimated using the oncoPredict algorithm based on the GDSC2 drug sensitivity database [22]. These drug-response results were interpreted as computational estimates of predicted sensitivity, not as evidence of clinical treatment benefit. Immune infiltration was evaluated with several methods, including ESTIMATE [23], MCPcounter [24], Porpimol’s study via single-cell gene set enrichment analysis (ssGSEA) [25], and TIMER [26]. We also assessed the cancer-immunity cycle to explore anti-tumor immune activity [27]. Associations between CMTM3 and seven types of immunomodulators were further analyzed [28,29]. To estimate potential response to immunotherapy, nine published immune-related signatures were examined [30]. These included immune cytolytic activity (CYT) [31], interferon-gamma immune signature (IFNγIS) [32], expanded immune signature (AyersExpIS) [32], T cell-inflamed signature (GEP) [32], Roh immune score (RohIS) [33], Davoli immune signature (DavoliIS) [26], repressed immune resistance (RIR) [34], ImmuneScore [35], and network-based ICB immunotherapeutic signature (ICBnetIS) [30,36]. The BEST platform was also used to predict immunotherapy response. Pathway analysis of CMTM3 was performed using KEGG-based gene set enrichment analysis (GSEA) and the Metascape platform [37]. Genomic alterations were analyzed with GISTIC 2.0 [38] and maftools [39].

Cell culture and functional assays

The human gastric cancer cell line MKN-74 and the monocyte cell line THP-1 were purchased from Cellverse Co., Ltd. Cells were maintained in DMEM or RPMI-1640 medium containing 10% fetal bovine serum (FBS) and cultured at 37 °C in a humidified incubator with 5% CO2. Cell proliferation was measured using the Cell Counting Kit-8 (CCK-8) assay, and absorbance was read at 450 nm. DNA synthesis was further examined with the 5-ethynyl-2′-deoxyuridine (EdU) assay. To test macrophage recruitment, Transwell migration assays were performed with 8 µm pore-size polycarbonate membrane inserts. Gastric cancer cells were placed in the lower chamber to create a chemoattractant signal, while THP-1-derived macrophages were added to the upper chamber. Knockdown efficiency was confirmed by quantitative real-time PCR (qRT-PCR) using total RNA from cultured cells. In addition, levels of secreted CMTM3, IL-6, and IL-1β in the culture supernatant were measured with specific Enzyme-Linked Immunosorbent Assay (ELISA) kits according to the manufacturers’ instructions.

Statistical analysis

Statistical analyses were performed using R software. Differences between two groups were assessed using the Wilcoxon rank-sum test, and comparisons among more than two groups were assessed using the Kruskal-Wallis test. Survival curves were generated using the Kaplan-Meier method and compared with the log-rank test. Cox proportional-hazards models were used to estimate hazard ratios and 95% confidence intervals. Spearman correlation analysis was used to assess associations between CMTM3 expression and other markers.

Results

Identification of inflammation-associated gene modules at single-cell resolution

We mapped the major cell lineages in the TME based on canonical marker genes, identifying distinct populations of epithelial cells, immune cells, and stromal cells (Fig. 1A). We further examined the cellular distribution of CMTM3 in the single-cell dataset. CMTM3 expression was detected mainly in malignant cells. We further identified specific subtypes like B cells, macrophages, and T cells (Fig. 1B). Using the infercnv method, we distinguished cancer cells from non-cancer cells (Fig. 1C) and confirmed their distribution (Fig. 1D). We scored cancer cells for inflammation and used network analysis (hdWGCNA) to find related gene modules. The "purple" module was strongly linked to inflammation (Fig. 2A-C).

Fig. 1.

Mapping the cellular landscape of the gastric cancer microenvironment. (A) Visualization of major cell lineages in the tumor microenvironment (TME) using UMAP. (B) Detailed annotation of specific immune and stromal cell subtypes. (C) Genomic copy number variation (CNV) analysis used to distinguish malignant cells from non-malignant cells. (D) Final UMAP projection highlighting the distribution of identified cancer cells and TME components.

Fig. 2.

Screening for inflammation-driven gene networks in cancer cells. (A) Parameter selection for the hdWGCNA network analysis, showing the scale-free topology fit. (B) Identification of distinct gene co-expression modules within cancer cells. (C) Correlation analysis linking specific gene modules (colors) to the "Inflammation" signature, identifying the purple module as the most relevant.

CMTM3 is identified as a key inflammation-related prognostic biomarker

To find the most important gene in this module, we used a multi-model screening process. We started with Cox regression, followed by CoxBoost (Fig. 3A-B), and finally Random Survival Forest (RSF) (Fig. 3C). This repeated feature-prioritization process identified CMTM3 as a stable candidate prognostic gene. Patients with high CMTM3 expression exhibited significantly diminished overall survival across eight independent cohorts (Fig. 3E). Importantly, CMTM3 levels were significantly higher in tumor tissues than in normal tissues (Figure S1).

Fig. 3.

Multi-model screening identifies CMTM3 as a critical prognostic biomarker. (A) Univariate Cox regression screening of genes within the inflammation-related module. (B) Feature selection using the CoxBoost machine learning algorithm. (C) Final marker selection using the Random Survival Forest (RSF) algorithm, pinpointing CMTM3. (D) Correlation matrix showing relationships among nine established immunotherapy signatures. (E) Kaplan-Meier survival curves demonstrating the poor prognosis associated with high CMTM3 expression across eight independent cohorts.

In vitro assessment of CMTM3 knockdown in gastric cancer cells

To experimentally assess CMTM3 function, we knocked down CMTM3 in MKN-74 gastric cancer cells using siRNA. The PCR test showed that CMTM3 levels dropped significantly (Fig. 4A). CMTM3 knockdown reduced gastric cancer cell proliferation, as shown by lower OD values in the CCK-8 assay (Fig. 4B) and fewer dividing cells in the EdU assay (Fig. 4C-D). CMTM3 knockdown also reduced the migration of THP-1-derived macrophages toward gastric cancer cell-conditioned medium in the Transwell assay (Fig. 4E-F). Furthermore, ELISA tests showed that lowering CMTM3 reduced the protein levels of CMTM3, IL-6, and IL-1β (Figure S2A-C). These results suggest that CMTM3 knockdown may attenuate an inflammatory secretory phenotype in this cell-line model.

Fig. 4.

Experimental validation of CMTM3 biological function. (A) qPCR analysis confirming reduced CMTM3 mRNA levels in MKN-74 cells after siRNA knockdown. (B) Cell viability curves from CCK-8 assays showing reduced growth in CMTM3-knockdown cells. (C, E) Representative images and quantification of EdU staining, indicating decreased cell proliferation. (D, F) Transwell co-culture assays showing reduced migration of macrophages towards CMTM3-knockdown cancer cells.

Distinct genomic alteration profiles and chemotherapeutic therapy sensitivity stratified by CMTM3 expression

Genomic mutation analysis revealed that TTN was the most frequently mutated gene in the high-CMTM3 group (Fig. 5A), while TP53 mutations predominated in the low-CMTM3 group (Fig. 5B). We also analyzed other mutated genes (Figure S3) and gene pairs (Figure S4A-B). The oncoPredict analysis suggested that low-CMTM3 tumors may have lower predicted sensitivity scores for selected agents, including cisplatin and dasatinib (Fig. 6). These results should be interpreted as computational drug-sensitivity estimates rather than evidence of clinical benefit.

Fig. 5.

Distinct genomic mutation profiles associated with CMTM3 expression. (A) Top mutated genes (waterfall plot) in the high-CMTM3 expression group, highlighting TTN mutations. (B) Top mutated genes in the low-CMTM3 expression group, highlighting TP53 mutations.

Fig. 6.

Prediction of chemotherapy and targeted therapy sensitivity. Comparison of predicted drug sensitivity scores for standard chemotherapy agents and kinase inhibitors between high and low CMTM3 groups. Lower scores indicate higher sensitivity.

CMTM3 modulates the tumor-immune ecosystem and immunotherapy signature enrichment

Pathway analysis showed that CMTM3 is linked to immune responses (Fig. 7A-B). High CMTM3 expression was associated with increased numbers of immune cells (such as T cells, B cells, and NK cells) and higher microenvironment scores (Fig. 8A). It activated the cancer-immunity cycle (Fig. 8B). However, it was also linked to immune checkpoints like PD-1, CTLA-4, and PD-L1 (Fig. 8C). Finally, high CMTM3 levels were associated with higher scores across nine immunotherapy signatures (Fig. 9A) and accurately predicted response to immunotherapy across six cohorts (Fig. 9B).

Fig. 7.

Functional pathway enrichment analysis. (A) Enriched biological processes associated with CMTM3 identified via Metascape. (B) GSEA plots showing activation of KEGG pathways related to immune response and signaling in high-CMTM3 tumors.

Fig. 8.

CMTM3 shapes the tumor-immune ecosystem. (A) Heatmap showing positive correlations between CMTM3 expression and immune cell infiltration levels (estimated by TIMER, MCPcounter, etc.). (B) Activity levels of the seven steps of the cancer-immunity cycle in high vs. low CMTM3 groups. (C) Radar chart showing correlations between CMTM3 and key immune checkpoint molecules.

Fig. 9.

Predictive value of CMTM3 for immunotherapy response. (A) Box plots comparing levels of nine immunotherapy-related signatures (e.g., IFNγ, CYT) between high and low CMTM3 groups. (B) ROC curves demonstrating the accuracy of CMTM3 in predicting clinical response to immunotherapy across six independent datasets.

Discussion

This study suggests that CMTM3 is associated with inflammatory transcriptional programs, immune microenvironmental features, and prognosis in gastric cancer. It acts as a link between inflammation, the immune system, and treatment success. By using a multi-model approach, combining machine learning, single-cell data, and lab work, we prioritized CMTM3 as a candidate biomarker [40,41]. These findings raise the possibility that CMTM3 may reflect biologically relevant tumor-microenvironment interactions.

One key finding is that CMTM3 seems to amplify inflammation. High CMTM3 levels create an "immune-hot" environment. We saw more T cells and NK cells in these tumors [28]. Although immune infiltration can reflect anti-tumor immune activity, high CMTM3 expression was also associated with immune checkpoint molecules such as PD-L1, PD-1, and CTLA-4. This pattern suggests that CMTM3-high tumors may represent an immune-infiltrated but potentially immunosuppressed microenvironment [42]. This explains why high CMTM3 is linked to poor survival; the inflammation helps the tumor grow and hide [7].

Our lab tests support this. When we blocked CMTM3, cancer cells grew more slowly and stopped calling in macrophages. These results support an association between CMTM3 expression and proliferative and macrophage-migration phenotypes in the tested cell-line model, similar to findings in pancreatic cancer [8].

From a treatment-related perspective, the associations between CMTM3 expression and predicted therapy response should be interpreted cautiously. In the low-CMTM3 group, oncoPredict analysis suggested lower predicted sensitivity scores for several agents, including cisplatin and selected kinase inhibitors. The enrichment of TP53 mutations in low-CMTM3 tumors may partly reflect a distinct genomic background, and TP53 alterations are known to influence DNA-damage response and platinum-related treatment sensitivity [43]. However, our data do not establish a direct mechanistic interaction between CMTM3 expression, TP53 status, and cisplatin response. Similarly, the predicted association between low CMTM3 expression and kinase-inhibitor sensitivity may be related to altered signaling states, including pathways previously linked to CMTM3 and EGFR/Rab5 regulation [9], but this remains a hypothesis requiring experimental validation. For immunotherapy, high-CMTM3 tumors showed higher immune-infiltration features, immune-checkpoint expression, and immunotherapy-related signature scores [33]. Retrospective prediction models also suggested an association between CMTM3 expression and immune-checkpoint blockade response [32]. Nevertheless, these computational and retrospective findings do not prove that CMTM3 can guide treatment selection in clinical practice. Therefore, CMTM3 should currently be considered a hypothesis-generating marker for therapy-response studies rather than a ready-to-use precision oncology biomarker.

In short, CMTM3-driven inflammation seems to help the tumor at first, but also creates a weakness we can target with immunotherapy.

Conclusion

CMTM3 is associated with inflammatory transcriptional programs, macrophage-related immune features, poor prognosis, and exploratory therapy-response signatures in gastric cancer. Functional assays further suggest that CMTM3 knockdown can reduce gastric cancer cell proliferation and macrophage migration in the tested model. These findings support CMTM3 as a candidate biomarker that warrants further mechanistic, protein-level, and prospective clinical validation.

Limitation

Our study has several limitations. First, most analyses were based on retrospective public transcriptomic datasets, and differences in platform, preprocessing, clinical annotation, and cohort composition may affect generalizability. Second, bulk RNA-seq cannot fully distinguish tumor-cell-intrinsic CMTM3 expression from stromal or immune-cell contributions, although our single-cell analysis partially addresses this issue. Third, the immunotherapy-response and drug-sensitivity analyses are computational or retrospective estimates and should not be interpreted as evidence for clinical treatment selection. Fourth, the in vitro validation was performed in a limited gastric cancer cell-line model, and additional validation in other GC cell lines, in vivo models, clinical tissues, and protein-level datasets is required. Finally, prospective cohorts with standardized treatment and follow-up information are needed to determine whether CMTM3 provides independent prognostic or predictive value.

Funding

None.

CRediT authorship contribution statement

Changjian Li: Writing – original draft, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation. Qingxin Cai: Writing – review & editing, Writing – original draft. Yuan Tan: Writing – review & editing. Shifeng Yang: Supervision, Conceptualization. Feng Gao: Writing – review & editing, Validation, Supervision, Conceptualization.

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.