2.1 Single-cell RNA-seq

Five fresh GC samples were profiled via the BD Rhapsody system to obtain single-cell transcriptomic data, constituting our in-house cohort. Two public GC scRNA-seq datasets were obtained from the GEO database (GSE183904 and GSE231540). The outputs of three cohorts were subsequently imported into the Seurat (v5.0.1) R toolkit for quality control and downstream analysis. In detail, applied for normalisation to adjust for variations in transcript counts, and ‘FindVariableFeatures’ was used to identify highly variable genes that may be related to cellular heterogeneity. ‘ScaleData’ and ‘RunPCA’ were used for dimensionality reduction and clustering, while ‘IntegrateLayers’ and ‘HarmonyIntegration’ were utilised for sample integration and batch effect removal. ‘FindNeighbors’, ‘FindClusters’ and ‘RunUMAP’ were used for clustering and visualisation. ‘FindAllMarkers’ identified characteristic genes, and cell types were defined via the SingleR toolkit and known markers. All functions were run with default parameters, unless specified otherwise.

Data quality control and sample selection involve the following aspects: (1) Cell filtering: low-quality cells or doublets were excluded using thresholds of nFeature RNA > 300 and nCount RNA < 7500, which are commonly used thresholds to remove cells with insufficient gene expression coverage or excessive ambient RNA, respectively. Moreover, the DoubletFinder R package was utilised to identify doublets and the top 7.5% of cells on the basis of the doublet score were removed. (2) Gene filtering: Genes that were expressed in fewer than 100 cells were eliminated. (3) Sample filtering: After defining the cell types, samples that contained fewer than 200 epithelial cells were excluded. As a result, 44 samples were retained for downstream analysis.

In addition, two datasets, GSE167297 and GSE163558, both obtained from the GEO database, were employed to assess the differences in invasiveness and metastatic potential across various meta programs. Specifically, GSE167297 comprised single-cell data from five pairs of superficial and deep GC tissues, while GSE163558 encompassed single-cell data from three primary tumours and six metastatic sites.

2.2 Expression programs associated with ITH in GC

To capture GC heterogeneity, we employed non-negative matrix factorisation algorithm (NMF) on each tumour sample to generate MPs according to the tutorial²⁶ (https://github.com/tiroshlab/3ca/tree/main/ITH_hallmarks). First, NMF was run for each tumour with factors (K) ranging from 4 to 9, yielding 39 programs per sample. Each program was defined by the top 50 genes from the NMF coefficients, totalling 1716 programs. Second, robust NMF programs that appeared in more than one NMF rank within the sample and had similarity to an NMF program in a different sample were identified. Three criteria were used: (1) 70% gene overlap with different K programs in the same sample, (2) 20% overlap with any program in other samples and (3) no more than 30% overlap with other programs in the same tumour. This approach yielded a total of 210 robust NMF programs. Third, these robust NMF programs were clustered by Jaccard similarity, selecting programs with at least eight gene overlaps as potential cluster founders. If overlaps exceeded five cases, the program with the highest overlap was added, forming a cluster. The MP was initially defined by common genes and then completed for 50 genes based on NMF scores. This process was repeated, adding programs with at least 10 gene overlaps and updating the MP. Finally, seven GC MPs, each consisting of 50 genes, were emerged.

To assign each tumour cell to a certain MP, a cell score was calculated according to the tutorial.²⁶ Briefly, the scRNA expression matrix for each sample underwent normalisation to stabilise variance, and the ‘sigScores’ function from the scalop R package (v1.1.0) was used to score each cell, with at least 25 genes conserved in the MP. Cells were assigned to the MP with the highest score. An MP must be expressed in at least 5% of cells to ensure sample consistency, and cells not meeting these criteria are labelled ‘unresolved’. To quantitatively assess the expression of MPs within individual cells, MP scores were also calculated via the function ‘AddModuleScore’.

2.3 Signalling pathway enrichment analysis

To investigate the biological states or functional differences of MPs, Gene Set Variation Analysis (GSVA) was employed for pathway enrichment analysis of the scRNA-seq data. The enrichment analysis utilised signatures from Molecular Signatures Database (MsigDB), including Gene Ontology (C5-GO:BP), REACTOME (C2-CP:REACTOME) and Kyoto Encyclopedia of Genes and Genomes (C2-CP:KEGG) pathways. GSVA enrichment scores were calculated via the ‘gsvaParam’ and ‘gsva’ functions from the GSVA R package (v1.52.3). Signatures with adjusted p values < .05 were deemed significantly enriched.

2.4 Inferring MP regulators

To increase the signal-to-noise ratio in data and avoid biases from incorrect modelling or over-smoothing, the metacell algorithm was applied prior to gene regulatory network analysis.³⁵ Subsequently, pySCENIC (v0.12.1) was utilised to assess transcriptional factor (TF) activities in meta cells, following the protocol at https://github.com/aertslab/SCENICprotocol.

2.5 Spatial transcriptome

Spatial transcriptomics (ST) data used Visium 10× platform were obtained from the GEO database for the GSE251950 cohort (containing nine primary GC tissues). Downstream analysis of STs was performed according to the standard pipeline of Seurat (https://satijalab.org/seurat/articles/spatial_vignette). Spots were annotated using robust cell type decomposition (RCTD) with a reference subset derived from previous GC scRNA-seq data, which included 12 different cell types (excluding proliferating cell type).³⁶ The RCTD was run in ‘Multi mode’, outputting the estimated weights of each cell type within a spot. An alternative method was also used, integrating cell-type labels from ‘SCTransform’ processed scRNA-seq data into normalised ST data via the ‘FindTransferAnchors’ and ‘TransferData’ functions. The spots were annotated based on the highest prediction score. Tumour cell spots were selected for further analysis. For each spot, MP scores were calculated using Seurat's ‘AddModuleScore’ function, and the spot was assigned to the MP with the highest score.

2.6 Calculation of the tumour neighbourhood score

The tumour neighbourhood score was developed to investigate the cell composition around individual tumour cell spots. As shown in Figure 3E, the neighbourhood was defined as the 11-pixel area around the spot. The score for a cell type was calculated by summing the weights of that cell type within the neighbourhood.

To analyse the correlation between MPs and subcell types, subcell type scores were calculated with Seurat's ‘AddModuleScore’ function. The tumour neighbourhood score for a subcell type was found by summing scores around each tumour spot. The correlation between MP scores and subcell type neighbourhood scores was determined via Pearson's method.

2.7 Bulk RNA-seq

The TCGA-STAD dataset was retrieved from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov). Metastatic transcriptome data from 66 paired primary and metastatic tumours were obtained from GSE237876. Transcriptional data on the immunotherapy response of 45 metastatic GC patients treated with anti-PD-1 therapy (PRJEB 25780 dataset) were obtained from the Tumor Immune Dysfunction and Exclusion database (http://tide.dfci.harvard.edu/).

For deconvoluting bulk sequencing data to estimate the cellular composition, the BayesPrism algorithm was applied, using previous GC scRNA-seq data with 12 main cell types and 63 subcell types as references.³⁷ The analysis excluded mitochondrial and ribosomal genes, with a focus on protein-coding genes.

The epithelial expression matrix from BayesPrism was normalised with DESeq2 (v1.34.0) for further analysis, with MPs scores in each sample calculated by ssGSEA.

2.8 Pseudotime analysis

Monocle2 (v2.32.0) with the DDR-Tree method was utilised to analyse single-cell trajectories. The log-normalised data were used as the input, and DEGs or marker genes of clusters were applied as ordering gene sets. Trajectories were deduced post-dimension reduction and cell ordering. Additionally, the BEAM test was run to detect gene dynamics along branches.

2.9 Cell communication analysis

CellChat (v2.1.2) was used to analyse cell interactions in the TME. The communication probabilities were calculated using the trimean method via ‘computeCommunProb’, the cell–cell communications were filtered, and the pathway probabilities were computed with default parameters.

2.10 TME analysis

The cellular component abundances and microenvironment scores in the ST spot or bulk samples were deduced using the xCell pipeline (https://xcell.ucsf.edu/).

2.11 Survival analysis

To assess the impact of the abundance of each cell type on overall survival, we conducted Kaplan-Meier survival analysis and log-rank test statistics, utilising the survival (v3.5-8) and survminer (v0.4.9) R packages. Cell type fractions and MP scores for the TCGA-STAD cohort were determined using BayesPrism, as previously described. The ‘surv_cutpoint’ function from the survminer R package was utilised to determine the best cutpoint to divide each cell type into two groups.

2.12 Single-cell drug susceptibility assessment

The Beyondcell (v2.2.1) R package was employed to ascertain drug sensitivities from scRNA-seq data,³⁸ leveraging the package's integrated drug perturbation signature collection database.

2.13 In vitro experiments

The methods of in vitro experiments are detailed in the Supporting Information.

2.14 Statistics and reproducibility

Statistical analyses of the data were conducted as described in the methods or figure legends. All the results from representative experiments were collected at least three times independently.

3.1 MPs within the GC malignant compartment

To systematically reveal the tumour characteristics of GC tissue, we collected tumour tissue samples from five untreated GC patients and performed scRNA-seq analysis of these samples by BD Rhapsody. To reduce the impact of patient heterogeneity and sequencing platforms, we incorporated the in-house cohort with two public 10× Genomics scRNA-seq cohorts (GSE183904 and GSE231540). Following the stringent quality control for each dataset and double-cell removal, we ultimately obtained a total of 152 540 high-quality single-cell data from 44 GC samples (Figures 1A and S1 and Table S1). These cells were classified into 13 different cell types through dimensionality reduction and clustering (Figure 1B and Table S2).

Previous studies have reported significant heterogeneity in GC cells,^{23, 39-41} but their definition and function remain uncertain. We thus aimed to assess the diversity of the cellular state of malignant GC cells. Consistent with a previous study,²⁶ 20 555 tumour cells were extracted for analysis and exhibited significant ITH according to their sample origin (Figure 1C). By performing NMF clustering on the malignant cells of each sample, seven unique meta patterns (MP1-7), each containing a unique set of genes, were widely present in multiple samples reflecting common patterns of ITH (Figure 1D and Table S3). Based on gene sets of the MPs, we computed gene expression scores to define principal cellular states within malignant populations and quantify the proportional representation of each MP in individual patient samples. (Figure 1E,F).

To elucidate the biological function and define the state of each MP, we performed functional enrichment analysis on malignant cells in different MPs (Figure 2A). MP1 was characterised by expression of genes (e.g., CHGA and CHGB) involved in the endocrine hormone secretion and neurotransmitter secretion pathways, thus representing classic neuroendocrine cells. MP2 harboured the distinctive genes CEACAM5, GKN1 and GKN2, which are indicative of intestinal metaplasia cells. MP3 was characterised by the expression of genes associated with the regulation of chemokine-mediated signalling pathways and leukocyte chemotaxis (e.g., CCL5 and CXCR4), which might represent a subpopulation of epithelial cells involved in immune responses. MP4 had increased expression of genes associated with cell proliferation (e.g., TOP2A and MKI67). MP5 was enriched with ribosome related genes and characterised by active metabolic pathways, which probably reflected stress related cell state. MP6 displayed the hallmarks of mucous gland cells (e.g., MUC6, PGC and CXCL17) and was notably enriched in pathways related to intrinsic epithelial cell differentiation and the biosynthesis of fatty acids. Notably, MP7 exhibited a unique molecular signature enriched with histone-related genes and key transcription factors (e.g., early growth response 1 [EGR1], SOX4 and ATF3), along with significant up-regulation of tumour progression-associated pathways, particularly positive regulation of Wnt and Notch signalling pathways,^42-44 which were markedly enhanced in advanced GC (Figure 2B). Additionally, this subpopulation showed prominent activation of fibroblast differentiation and proliferation-related pathways, including positive regulation of smooth muscle cell proliferation and fibroblast growth factor receptor signalling. These molecular signatures collectively indicate that MP7 likely represents a distinct tumour cell state with active crosstalk with CAFs.

Furthermore, to clarify the specificity and robustness of the seven MPs in GC, we conducted Jaccard similarity analysis with the pan-cancer MPs. MPs in GC exhibited a significant level of congruence with pan-cancer MPs, especially MP2, MP4, MP5 and MP7, indicating the stability of these GC MPs (Figure S2A). The MPs were subsequently validated in four seminal studies on GC tumour cell subpopulations (Figure S2B and Table S4). Consistent with previous studies, MP1 was recognised as a classic endocrine subtype, MP2 as a pit cell subtype, MP4 as a proliferation subtype and MP6 as a typical neck cell subtype. Notably, MP7 was associated with stress meta program (MP5) in pan-cancer studies, sharing 14 common genes, including ATF3, EGR1, FOSB, GADD45B and NR4A1. In the context of GC, comparative analysis revealed that MP7 shared nine genes with the Proliferative Foveolar subpopulation identified by Cheng et al.,⁴¹ a non-malignant cell state characterised by overlapping markers such as CXCL3, HIST1H2AG, SOX4 and NR4A1. However, MP7 demonstrated partial molecular similarity to the C3 tumour cell subpopulation described by Zhang et al.,²³ sharing five key genes (KCNQ1OT1, GADD45B, FOSB, EGR1 and SOX4), an intermediate subgroup with poorly defined molecular features. Taken together, these results indicated that MP7 exhibited similarities to certain tumour cell subgroups, although the specific functions of these subgroups remain elusive.

Moreover, we utilised the R package Harmony to remove batch effects on malignant cells and re-clustered malignant cells through a graph-based clustering approach of the R package Seurat. We found that only the Cluster 2 and Cluster 4 subgroups in Seurat were exhibited a predominantly singular MP state (MP3 and MP6, respectively; Figure S2C–F). Collectively, these MPs were found to be both robust and adaptable, setting apart from previous single-cell clustering studies on GC and illuminating the heterogeneity of GC cells from a novel perspective.

The correlations and progression paths of these GC MPs were further analysed. The MP7 exhibited negatively correlation with MP5, suggesting that these states were mutually exclusive (Figure 2C, the left correlation matrix). For pseudotime analysis, MP1 was excluded due to its high enrichment of genes specific to endocrine cells, and we found that MP4 and MP5 were at the beginning of the trajectory path, whereas MP2/MP7 cells and MP3 cells were at the end of the trajectory path (Figure 2D–F). Furthermore, we analysed the trajectories of these meta programs using branched expression analysis modelling (BEAM analysis) and hierarchical clustering to identify genes enriched across states (Figure S3A). Finally, we showed the characteristic changes of MP2 (CLDN18), MP3 (VIM), MP6 (REG1A) and MP7 (SOX4) marker genes over pseudo-time (Figure S3B).

To gain a deeper understanding of the clinical implications of MP, we conducted an analysis of the MP scores within the TCGA-STAD dataset. Consistent with previous single-cell data, we observed that the MP7 score was significantly negatively correlated with the MP5 scores (Figure 2C, the right correlation matrix), while it was closely associated with tumour infiltration depth (Figure 2G) and poor prognosis in GC (Figures 2H and S3C). To further elucidate the correlation between MP7 and the invasion and metastasis of GC, we analysed the MP7 scores in GC tissues from various depths and found that the MP7 scores were greater in deep tumour tissues than that in superficial tissues (Figure S3D). Single-cell data analysis of GC metastasis and primary tumours revealed an increased MP7 score in the metastatic sites (Figure S3E). Due to the limited number of scRNA-seq samples, the correlation between MP7 and depth of invasion or metastatic foci did not reach statistical significance. To address this, we analysed bulk data from GC metastasis and primary tumour, which revealed significant up-regulation of MP7 in metastases (Figure 2I). Collectively, these findings confirm the significant correlation between MP7 and GC progression, metastasis and poor prognosis, indicating that it represents the most malignant state of tumour cells.

3.2 Association of MPs with TME

The interaction between tumour cells and other cellular constituents within the TME is pivotal in the reconfiguration of the TME and the genesis of tumour heterogeneity.⁴⁵ To further evaluate the relationship between the MPs and the TME, we correlated the summarised expression values of each MP with the percentages of TME cells inferred from the scRNA-seq data within individual patients (Figure 3A). Interestingly, MP2 was positively associated with B cells, whereas MP7 displayed a negative correlation with B cells. MP3 was predominantly positively associated with CD4 T cells, while MP4 was positively correlated with neutrophil infiltration. Furthermore, we analysed the correlation between MP scores and TME score in the TCGA-STAD data. The MP3 score was positively associated with immune scores, whereas the MP7 score was strongly correlated with the stromal score. (Figure 3B).

To gain a deeper understanding of the interaction between MPs and the TME at spatial resolution, we integrated our scRNA-seq data with a ST data from nine GC samples using 10x Genomics Visium technology. After defined the MP state of each malignant spot (see section Methods and Figure 3C), MP2 was found to be predominantly localised to the superficial layer of the mucosa, showing distinct spatial aggregation, whereas MP6 was situated near MP2. MP3, MP4, MP5 and MP7 presented in the submucosal layer of tumours and exhibited diffuse distribution, which facilitated their spatial engagement with diverse cell types in the TME. Furthermore, we have performed analyses focusing specifically on the tumour–stromal interface (Figure S4A) and found that MP2 was predominantly localised within the tumour region, while MP3 and MP7 show enrichment at the tumour–stromal interface (Figure S4B). Additionally, we examined the expression of marker genes that distinguish MP2 and MP6 by mIF. Tumour cells with high TFF1 expression (MP2 state) were clustered mainly at the top of the mucosal layer, while MUC6-positive cells (MP6 state) were predominantly aggregated around the neck of the gland lumen (Figure 3D).

To study the composition of TME cells associated with each MP, we defined a cell neighbourhood score by calculating the cell-type score surrounding the corresponding spot²⁷ (Figure 3E). In all nine ST slides, MP3 spots presented elevated immune cell neighbourhood scores, such as those for CD4 T cells, CD8 T cells, NK cells and macrophages (Figure 3F). Surprisingly, cytotoxic immune cells, especially CD8 T cells and NK cells, were mainly aggregated around MP4 cells, confirming that the MP3 and MP4 spots may be involved in different tumour immune responses. Moreover, MP7 spots were surrounded primarily by accumulated fibroblasts, indicating their key role in matrix remodelling. Concurrently, there was a significant decrease in the enrichment of immune cells, including B cells, CD4 T cells, CD8 T cells and NK cells, which suggested that fibroblasts surrounding MP7 might create an immune-deprived microenvironment around it (Figure 3F). Additionally, we analysed tumour neighbourhood scores at the tumour–stromal interface versus whole-tumour analysis. As shown in the Figure S4C, the interface scores closely matched the whole-tumour results: MP2 maintained epithelial cells and plasma cells; MP3 enriched CD4 T cells, endothelial cells and macrophages; MP4 exhibiting elevated CD8 T cells, mast cells, monocytes and NK cells; and MP7 accumulated fibroblast. These results confirm the consistency of the whole-tumour findings, strengthening the reliability of our conclusions.

Based on these observations above, we subsequently investigated the potential roles of MPs and TME components in immunotherapy. By analysing the groups responsive and non-responsive to PD-1 immunotherapy in the PRJEB25780 dataset, we found that fibroblasts were significantly enriched in the group non-responsive to PD-1 treatment (Figure S4D). MP4, which has the closest spatial distance to immune killer cells such as CD8 T cells and NK cells, significantly increased in the responsive group, whereas there was no significant difference in the scores of MP3 and MP7 between the two groups (Figure 3G). Together, these findings suggest that MP3 is closely associated with tumour immune infiltration, MP4 more effectively reflects the cytotoxic capacity of immune cells compared with MP3, and MP7 is characteristically related to fibroblast infiltration.

3.3 MP3 and MP4 were involved distinct functional states of CD8 T cells

In ST, significant enrichment of T cells and NK cells was observed surrounding both MP3 and MP4 (Figure 3F). However, MP3 scores were positively correlated with immune scores, while MP4 scores demonstrated a negative correlation, suggesting distinct immune response mechanisms mediated by these cell populations (Figure 4A). To clarify the distinct immune activities as characteristic responses of the two states of tumour cells, we performed further dimensionality reduction and clustering on NK and T cells, revealing seven subtypes of CD4 T cells, six subtypes of CD8 T cells and four subtypes of NK cells (Figure 4B). Furthermore, all subpopulation states were evaluated via TCellSI⁴⁶ (Figure 4C), and the correlations between the MP scores and the proportions of immune cell subgroups in the single-cell data were compared (Figure 4D). MP3 exhibited a strong correlation with diverse CD8 T cell populations, particularly the cytotoxic subgroups, such as CD8-GZMK, CD8-HAVCR2 and CD8-ZNF683. In the spatial transcriptome samples, by evaluating the expression of CD8 T cell signatures in MP3 or MP4 neighbourhoods, we observed a significant positive correlation between MP3 and CD8-GZMK or CD8-ZNF683, which was consistent with previous results (Figure 4E, upper). However, the increase in the tumour neighbourhood CD8-HAVR2 score was not related to the MP3 score but rather to the MP4 score (Figure 4E, bottom), indicating that MP3 and MP4 seem to be associated with different functional states of CD8 T cells. Moreover, within the GC immunotherapy dataset, we observed significant enrichment of exhausted T cells (CD8-HAVCR2) and MP4 scores in the anti-PD1 treatment response cohort, whereas MP3 showed no significant difference between responders and non-responders to immunotherapy (Figures 4F and 3G). Furthermore, MP4 and CD8-HAVCR2 displayed the highest correlation in the immune-responsive group, outperforming their correlation in the non-responsive group, as well as MP3 and CD8-HAVCR2 in both groups (Figure 4G). These findings indicate that although MP3 shows a significant correlation with immune cell infiltration in GC, MP4 appears to be more strongly associated with immune elimination function of CD8 T cells.

Additionally, to highlight the differences between the MP3 and pan-cancer immune programs, we analysed their spatial expression patterns. MP3 predominantly displayed features characteristic of the tumour epithelium (Figure S4A,B). This contrasted to the immune score and CD8 T cell signature, indicating that MP3 tends to be more prevalent at the tumour periphery. Using mIF in GC tissue (Figure 4H), we observed frequent co-localisation of CD8 T cells with MP3 tumour cells (dashed regions). Moreover, TIP analysis of MP3-dominant tumours revealed CD8 T cell infiltration in tumours with high RAC2 expression (Figure S5C). Flow cytometry of isolated epithelial cells further confirmed MP3 presence in GC tissue, with characteristic markers (RAC2, CXCR4 and IL-7R) detected in a subset of epithelial cells (Figure S5D). Together, these findings indicate that MP3 defines an epithelial subpopulation engaged in immune interactions, ruling out technical artifacts or immune cell contamination.

3.4 MP7 spots harboured an immune lockdown microenvironment and associated with myCAFs

The aforementioned findings indicated that fibroblasts predominantly accumulated around MP7, while NK and T cells were negatively correlated with MP7 (Figure 3F). We further analysed both the stromal score and the immune score around MP7 and found that the stromal score around MP7 significantly increased; conversely, the immune score significantly decreased, indicating the formation of an immuno-deprived microenvironment around MP7 (Figure 5A). Additionally, we characterised the expression of marker genes that distinguish various MPs and TIPs. Notably, EGR1-high tumours (MP7-dominant) exhibited strong enrichment of the CAF marker FAP and a striking absence of CD8 T cell infiltration. In contrast, RAC2-high tumours (MP3-dominant) showed lacked FAP expression and maintained abundant CD8 T cell infiltration (Figure S5D). We next presented a ST sample enriched with MP7 and observed a significant decrease in immune scores; in particularly, CD8 T cells were present only at the tumour margins, indicating a significant disruption in the infiltration process, which may be caused by the significant enrichment of fibroblasts (Figure 5B).

To strengthen the evidence that CAFs promoted the formation of an immunologically isolated microenvironment around MP7, we proceeded to perform further dimensionality reduction and clustering on fibroblast subsets. We identified a group of Schwann cells and eight clusters of fibroblasts, including inflammatory CAFs (iCAFs, Fib_C01), metabolic CAFs (meCAFs, Fib_C02), myofibroblasts (myCAFs, Fib_C04), antigen-presenting CAFs (apCAFs, Fib_C05), tumour-like CAFs (Fib_C06), proliferative CAFs (Fib_C09), pericyte (Fib_C03) and smooth muscle cells (SMC, Fib_C08) (Figure 5C,D). By analysing the relationship between MP scores and CAF subgroup proportions, we found that MP7 scores escalated with increasing proportions of iCAFs; in contrast, MP4 scores exhibited a markedly negative correlation with iCAFs (Figure S6A). Spatial distribution analysis revealed that iCAFs predominantly aggregated around MP2 and MP5, whereas the scores for iCAFs around MP7 notably decreased (Figure S6B,C). The discrepancy between ST and single-cell transcriptomic data may arise from limitations in sample size or analytical methodologies, but it also likely reflects intrinsic, tumour state-dependent preferences in CAF spatial distribution within the TME. Additionally, apCAFs were more dominant around MP3, potentially participating in tumour immune responses (Figures 5E and S6B). Notably, myCAFs were predominantly enriched around MP7 (Figures 5E and S6B), and the ST image revealed that myCAFs were primarily situated between the tumour and CD8 T cells, constituting the main subtype of CAFs that undergo immune exclusion (Figure 5F).

To elucidate the contribution of different CAF subgroups to tumour progression, the prognostic implications of each fibroblast subgroup within the TCGA-STAD dataset was examined. Elevated levels of iCAFs were associated with a favourable prognosis, while a higher prevalence of myCAFs was linked to a poor outcome (Figures 5G and S6D,E). Additionally, in the GC immunotherapy cohort, there was a marked increase in the prevalence of myCAFs among patients who did not respond to PD-1 treatment, suggesting that myCAFs might contribute to the resistance to immunotherapy through the promotion of immune exclusion (Figure 5H). Consequently, myCAFs, the predominant CAF subtype surrounding MP7, likely contribute to immune evasion mechanisms, particularly in the microenvironment adjacent to MP7-type GC cells.

3.5 MP7 cells recruited fibroblasts and induced them to myCAFs via GDF15

In the mIF images of GC, myCAFs (Figure 6A, arrows, FN1⁺ cells or FAP⁺ cells) were confirmed to be significantly enriched around the MP7 subpopulation (Figure 6A, dashed circle, MUC1⁺ EGR1⁺ cells). Further exploration of the potential molecular mechanisms involved in the aggregation of myCAFs around MP7 was conducted using CellChat. We observed that GDF15 derived from MP7 significantly interacted with the TGFBR2 receptor on the surface of CAF subpopulations, including iCAFs (Fib_C01), myCAFs (Fib_C04) and apCAFs (Fib_C04), exhibiting distinct MP subpopulation specificity (Figure 6B,C). To determine whether GDF15 is involved in the aggregation of myCAFs around MP7, we first used the supernatant of GC cells or exogenous GDF15 to induce chemotaxis in GC fibroblasts. Transwell assay revealed that both the supernatants from GC cell lines (NUGC3 and HGC27) and exogenous GDF15 could recruit CAFs and normal fibroblasts (NFs) (Figure S7A). Furthermore, we employed siRNA to knockdown of GDF15 expression in GC cells (Figure S7B) and assessed the chemotactic effect of the tumour cell supernatant on CAFs. We observed that the supernatant from NUGC3 and HGC27 cells with GDF15 knockdown exhibited a significant reduction in their ability to induce CAF migration (Figure 6D). These findings confirm that the induction of CAFs by GC cells is dependent on GDF15. Besides, GDF15 significantly promoted the expression of DCN (decorin), EGFR (epidermal growth factor receptor), FN1 (fibronectin 1) and PLXNB2 (plexin B2) in both NFs and CAFs, indicating that GDF15 could induce the switch to myCAFs (Figure 6E,F).

Subsequently, we investigated the primary receptor mediating GDF15-induced CAF chemotaxis. Expression analysis revealed that of TGFBR2, identified by CellChat as the primary GDF15 receptor, was significant up-regulated in both GC tissues and CAFs, while GFRAL, another known GDF15 receptor, was nearly undetectable in GC tissues and CAFs (Figure S7C,D). These findings established TGFBR2 as the predominant receptor for GDF15 in GC. To further elucidate the role of TGFBR2 in GDF15-mediated chemotaxis, we knocked down TGFBR2 in CAFs. While TGFBR2 depletion did not significantly altered the intrinsic migratory ability of CAFs, it markedly attenuated GDF15-induced chemotaxis (Figure S7E,F). Consistent with this observation, the chemotactic effect of NUGC3 and HGC27 cell supernatants on CAFs was significantly reduced following TGFBR2 knockdown (Figure 6G). Thus, these findings suggest that MP7 induces the aggregation of NFs and CAFs and then promotes their transition to myCAFs through the GDF15–TGFBR2 ligand–receptor interaction.

3.6 MyCAFs-derived RSPO3 facilitated the formation of the MP7 state via EGR1

To determine the potential involvement of myCAFs in the development of GC cells, pathway enrichment analysis was conducted in myCAFs. ECM receptor interaction, Wnt signalling pathway and TGFβ signalling pathway were significantly enriched in myCAFs (Figure S7G,H). Notably, RSPO3 (R-spondin-3), a potent activator of the canonical Wnt pathway, was markedly increased in myCAFs induced by GDF15 (Figure 6E), indicating its potential as a key regulator in myCAFs-mediated modulation of GC cells. Consequently, the impact of exogenous RSPO3 on GC cells was assessed. The results demonstrated that RSPO3 markedly enhanced the expression of MP7 markers in the GC cell lines NUGC3 and HGC27 (Figure 7A,B).

Furthermore, to elucidate the key TF of MP7 transformation, we performed an enrichment analysis of TF activity in MP7 cells and identified EGR1 as a highly active TF, as well as a marker gene of MP7, potentially playing a crucial role in the transformation or maintenance of the MP7 cell state (Figure 7C). We then silenced EGR1 in the GC cell lines NUGC3 and HGC27 via siRNA and observed a significant decrease in the expression of MP7 marker genes, such as ATF3, HES1, NR4A1 and SOX4 (Figure 7D,E). Moreover, EGR1 knockdown significantly inhibited the migration ability of GC cells (Figure 7F). Consistent with these findings, treatment with EGR1 inhibitor (EGR1-IN) also resulted in a markedly reduction of GC cell migration (Figure 7G). In GC patient-derived organoids (PDOs), EGR1 knockdown and pharmacological inhibition of EGR1 were also led to a significant reduction in the size of PDOs, along with a decrease in the expression of marker genes for MP7 (Figure 7H,I), indicating the crucial role of EGR1 in the formation and maintenance of MP7. Additionally, the chemotactic effect of GC cells pretreated with RSPO3 (MP7) on CAFs was assessed. The results showed that the migratory capacity of CAFs and NFs induced by RSPO3-pretreated HGC27 and NUGC3 cells was enhanced (Figure 7J). The above results suggest that myCAFs-derived RSPO3 facilitated the formation of MP7 state via EGR1.

3.7 MP7 served as a potential therapeutic target in GC treatment

MP7 tumour cells, characterised by their high malignancy and propensity to induce myCAFs transformation, represented promising therapeutic targets for GC. While EGR1-IN was initially considered as a candidate compound for targeting MP7, in vitro experiments revealed no significant cytotoxic effect on GC cells (Figure S8A). To find more effective drugs targeting MP7, we used BeyondCell³⁸ to analyse drug sensitivity profiles. Tumour cells were categorised into three different BC clusters based on drug sensitivity (Figure S8B). Notably, MP3 displayed unique drug sensitivity, while MP2 and MP7 exhibited similar sensitivity (Figure S8C). For the MP7 cell subpopulation, taselisib, a potent and selective inhibitor of phosphoinositide 3-kinase (PI3K),^{47, 48} and lapatinib, a dual tyrosine kinase inhibitor that inhibits both EGFR and HER2,^{49, 50} were predicted as highly sensitive drugs (Figure 8A,B).

To elucidate the relationship between MP7 transformation and PI3K/AKT signalling pathway activation, we first attenuated the MP7 state by depleting EGR1 in GC cells. EGR1 down-regulation significantly reduced the phosphorylation of PI3K, AKT, ERK and MAPK, indicating that EGR1 regulates MP7 transformation through activation of the PI3K/AKT signalling pathway (Figure S8D).

We further assessed the cytotoxic effect of taselisib and lapatinib on MP7 using two conventional drugs of GC, cisplatin and 5-Fu, as controls. In NUGC3 cells, the IC50 values for lapatinib and taselisib were lower than that of 5-Fu, while in the HGC27 cells, both drugs exhibited lower IC50 values compared with 5-Fu and cisplatin, demonstrating their potent cytotoxic effects against GC cells (Figure 8C).

To explore whether these drugs impair MP7-mediated CAF recruitment, NUGC3 and HGC27 GC cell lines were pretreated with IC50 concentrations of lapatinib and taselisib for 48 h. CAF chemotaxis assays revealed a significant reduction in CAF recruitment by drug-pretreated tumour cells compared with the DMSO-pretreated control group (Figure 8D). Meanwhile, taselisib-pretreated GC cells exhibited a more pronounced attenuation of CAF chemotactic induction than lapatinib-pretreated cells, likely due to the stronger targeting effect of taselisib on MP7 cells (Figure 8D).

Additionally, both drugs were tested in GC PDOs. In PDOs derived from two patients, both lapatinib and taselisib demonstrated higher sensitivity compared with 5-Fu and cisplatin, suggesting their potential therapeutic value in GC treatment (Figure 8E). Collectively, these findings highlight MP7 as a potential therapeutic target in GC and suggest that taselisib, with its high specificity and efficacy, may provide a promising strategy for MP7-targeted GC therapy.