3.1 Transcriptomic profiling revealed three metabolic subtypes in GC: Chen et al. (29) extracted Metabolite-protein interactions from four data resources including Kyoto Encyclopedia of Genes and Genomes (KEGG) (38), Reactome (39), Human-GEM (40), and BRENDA (41) (Supplementary Table 2). According the above data, we obtained 2048 metabolic related genes with connectivity greater than 5 (Supplementary Table 3). Genes with greater connectivity indicate that this gene may play a significant role in metabolic pathways. After screening for prognostic related genes (p < 0.05), we obtained 162 metabolic genes (Supplementary Table 4). Based on the expression data of 162 metabolic genes, we used the consensus clustering algorithm to perform clustering analysis on the TCGA-STAD dataset (42). Our analysis showed that 3 is the optimal and stable clustering number (Fig. 1A). Among the three gastric cancer subtypes we obtained, cluster 1 (C1) has 127 patients, cluster 2 (C2) has 170 patients, and cluster 3 (C3) has 78 patients (Supplementary Table 5).
Interestingly, patients in C3 have worse overall survival (OS) than the other two subtypes (Fig. 1B). The prognosis of C1 and C2 is good, but there is no significant difference between them. Previous studies identified four subtypes of gastric cancer in the TCGA-STAD dataset (2). We found that C2 contains a higher proportion of MSI subtypes and EBV subtypes patients (Fig. 1C, Supplementary Table 6). There are four subtypes of patients with similar proportions in C1 and C3. C3 also contains a higher proportion of diffuse gastric cancer subtypes, which may explain the poor prognosis of C3 subtypes (Fig. 1C, Supplementary Table 6). Interestingly, we found that the C3 subtype actually has a very high proportion of patients under 51 years old, which means that the C3 subtype may have more congenital genetic reasons than the influence of acquired environment (Fig. 1C, Supplementary Table 6). There is no significant difference in T staging, N staging, M staging, AJCC staging, and tumor grading among the three subtypes (Fig. 1C, Supplementary Table 6), indicating that factors such as tumor staging did not play a role in the differential prognosis of the three subtypes.
3.2 Metabolic subtypes show distinct signaling pathway features: 10 tumor related pathways not only play an important role in the occurrence and development of tumors, but also play an important role in the metabolism reprogramming of tumors (43) (44) (45) (7) (46). C1 is only highly expressed in the NRF2 pathway, while C2 is enriched in the MYC, cell cycle, and P53 pathways (Fig. 2A). C3 has a high level of 6 pathways (Fig. 2A). The above results indicate that the C3 subtype has the highest tumor pathway activity. 50 hallmark feature scores were also calculated. Interestingly, C1 subtype exhibits low levels in almost all features, while both C2 and C3 have many enriched features (Fig. 2B). The epithelial-mesenchymal transition (EMT) affects the initiation, progression, metastasis, and treatment resistance of tumors (47) (48). We found that C3 is the most mesenchymal (Fig. 2C).
We found that among the 89 KEGG metabolic pathways, the C1 subtype is at the intermediate level in many pathways compared to C2 and C3 (Fig. 2D). The C2 subtype has significantly the highest metabolic activity, including energy metabolism, amino acid associated metabolism, and fatty acid metabolism (Fig. 2D). Hypoxia pathway is at a high level in C3, while glycolysis is at a high level in C2 (Fig. 2E-2F). The high levels of the glycolysis pathway in the C2 subtype may provide substrates for the elevated metabolic state observed in C2. Many studies have shown that hypoxia can promote tumor malignancy, which may explain the higher degree of malignancy in C3. Furthermore, we observed enrichment of immune-related pathways predominantly in the C3 subtype, whereas the C1 subtype exhibited an intermediate level of immune pathway enrichment (Fig. 2G).
3.3 Genomic features of three GC subtypes: We have already described that 10 oncogenic pathways play an important role in the metabolic reprogramming of tumors. Therefore, we calculated the proportion of samples with at least one gene mutation in the representative genes of each pathway in each subtype based on the representative genes of 10 oncogenic pathways provided in previous studies (Fig. 3A, Supplementary Table 7) (49). C2 has a higher mutation rate in five oncogenic pathways with significantly different mutation rate (PI3K, Hippo, WNT, NOTCH and RAS). For SCNA of 10 oncogenic pathways, there are significant differences among the three GC subtypes, especially amplification. The MYC and Hippo pathways only exhibit amplification in the C3 subtype, which indicates that the amplification of these two pathways may be a characteristic of the C3 subtype. Similarly, the amplification of the WNT pathway may be a characteristic of the C2 subtype. In addition, the amplification of PI3K, cell cycle, TGF- β and NOTCH pathways may be a common feature of C1 and C2 subtypes. Specifically, C2 and C3 subtypes also have a specific highly amplified pathway (P53). The NRF2 pathway only has subtype differences in somatic copy number deletion compared with other genomic changes, and is the highest deletion rate in C3 subtype. Unlike amplification, deletion of the MYC and NOTCH pathways is a feature of C1 and C3 subtypes. The significant differences in genomic changes among the three subtypes of 10 carcinogenic pathways may suggest that these genomic changes play an important role in the formation of metabolic subtypes in gastric cancer.
Genomic instability (GI) plays a key role in tumor initiation and progression (50). We found that tumor mutation burden (TMB) is significantly higher in C2 compared to C1 and C3 (Fig. 3B). We further compared the level of somatic copy number variation (SCNA) among three GC metabolic subtypes. As expected, C2 exhibits the highest level of SCNA and is consistent with the trend of TMB, with C2-C1-C3 ranking from high to low (Fig. 3C-3E). Next, we analyzed the top 20 differentially mutated genes among three subtypes (Fig. 3F). CDH1, encoding a classical cadherin of the cadherin superfamily, is more frequently mutated in C3 compared to C1 and C2 (Fig. 3F). This is consistent with the highest EMT score of C3.
Methylation of DNA inhibits gene transcription, thereby reducing the expression of the gene. We compared DNA methylation between each subtype and the remaining patients. We found that there seems to be no significant difference in DNA methylation genes between C1 subtype and the remaining patients (Supplementary Table 8). C2 and C3 both have numerous differentially methylated sites and differentially methylated genes (Supplementary Tables 9 and 10). These results suggest that DNA methylation may also play an important role in metabolic reprogramming of GC.
3.4 Construction of metabolism related prognostic risk model: As described earlier, we screened 162 prognostic related metabolic genes. These genes divide GC into three subtypes with different characteristics, so we wonder if it is possible to use these genes to establish a prognostic risk model. Because the low expression level of genes may reduce the credibility of the model, we selected out 124 genes with TPM expression values greater than 1. The existing sequencing methods are very diverse, and the magnitude of gene expression values obtained by different sequencing methods varies greatly. Therefore, we perform Z-score transformation on the expression values of TCGA-STAD. The Lasso-Cox algorithm was used to identify the most robust metabolic genes for prognosis risk prediction. Finally, risk models associated with metabolism were developed as fellow: -0.013258*AK1 + 0.031448*AKR1B1 + 0.021963*DUSP1 + 0.011265*ECH1-0.055403*FAAH-0.008266*GAD1 + 0.031188*NT5E + 0.001141*NTAN1 + 0.047563*SLCO2A1 + 0.021638*UPP1. Based on the optimal risk cutoff value calculated using the maxstat package, we found that low-risk patients had better OS than high-risk patients in TCGA-STAD (P = 1.8e-9, log-rank test; Fig. 4A). Similar results were also observed in GPL570 cohort, GSE26942 and GSE84437(GPL570 dataset: HR = 1.56, 95% CI = 1.21–2.01, p = 4.5e-4; GSE26942: HR = 2.02, 95% CI = 1.27–3.22, p = 2.6e-3; GSE84437: HR = 1.72, 95% CI = 1.29–2.29, p = 2.0e-4; Fig. 4B-4D).
We wonder the role of these metabolic risk model genes in GC. So, we exclude genes that have been studied and confirmed by previous researchers to promote the malignant characteristics of GC. Then, genes with significantly worse prognosis were selected in the high expression group (using median as cut-off value). Finally, we decided to investigate the role of ECH1 in GC. The expression of ECH1 was detected in a variety of GC cell lines and gastric mucosa epithelial cell line GSE-1. As shown in Fig. 4E, ECH1 has the highest expression in the AGS cell line. So, subsequent experiments will use the AGS cell line. We found that the expression of the ECH1 gene was significantly suppressed after siRNA transfection (Fig. 4F). The best efficiency of knockdown was observed with si-ECH1#1 and si-ECH1#2. We confirmed the effects of ECH1 on GC proliferation, which indicated that inhibiting ECH1 could significantly reduce the proliferation ability in vitro (Fig. 4G and 4H). Transwell assays showed that ECH1 knockdown inhibits the migration of AGS cells (Fig. 4I). 3.5 Metabolic subtyping of GC can distinguish chemotherapy response: Some studies show that tumor metabolism reprogramming plays an important role in chemotherapy resistance (14) (51). We used the pRRophic package to predict the sensitivity differences of five common chemotherapy drugs for GC across three metabolic subtypes. Excitingly, patients of C2 subtype have the highest sensitivity to all five types of chemotherapy drugs (Fig. 5A-5E). A study suggests that neutrophil extracellular traps (NETs) formed during chemotherapy are activated by TGF-β to confer therapeutic resistance (52). Analysis of the TCGA-STAD cohort showed that the C3 subtype had the highest NETs score, C2 had the lowest, and TGF-β pathway score is highest in the C3 subtype (Fig. 5F and 5G, Supplementary Table 11). The above results indicate that C3 is a chemotherapy resistant subtype. We used the XGBoost algorithm to predict metabolic subtypes in the GC chemotherapy cohort (GSE14209) based on the metabolic subtypes of the TCGA-STAD cohort (Supplementary Table 12). In this chemotherapy cohort for GC, 22 patients were sensitive to chemotherapy at the beginning of treatment and gradually became resistant to chemotherapy during the chemotherapy process. Interestingly, we found that 6 patients progressed from C1 and C2 subtypes to C3 subtypes during the process of acquiring chemotherapy resistance, which confirms that C3 subtype is a subtype of chemotherapy resistance (Supplementary Table 12).
3.6 Data processing and subtype prediction for single cell sequencing cohort: We selected some samples from the GSE183904 cohort for analysis. Detailed information about the samples can be found in Supplementary Table 13. After conducting standard quality control, we obtained gene expression data for 115070 cells from 26 samples. We performed unsupervised clustering to obtain 31 clusters and identified 9 major cell types based on cell type specific marker genes: epithelial cells, T cells and natural killer (NK) cells, B cells, fibroblasts, smooth muscle cells, endothelial cells, myeloid cells, mast cells, and undefined cells (Supplementary Table 14). In the subsequent analysis, we will perform re-clustering on these cells to identify cell groups with different functions. We found that epithelial cells have significantly higher metabolic activity compared to other cells, as well as fibroblasts and myeloid cells (Supplementary Fig. 1A). This indicates that metabolic reprogramming in GC is not dominated by epithelial cells, but also fibroblasts and myeloid cells play an important role. The glycolysis / gluconeogenesis and pentose phosphate pathway of myeloid cells showed high levels, which was related to the increase of glucose metabolism of macrophages in the tumor microenvironment (14).
We divided the TCGA-STAD dataset into two parts at 7:3, with the former being used as the training cohort and the latter as the validation cohort. The model trained in the training cohort was used to predict the metabolic subtypes of the validation cohort. Surprisingly, the receiver operating characteristic curve (ROC) value reached 0.9621. Based on this classification model, 26 samples from the single cell dataset were divided into three subtypes (Supplementary Table 13). We compared the KEGG pathway of three subtypes of bulk data and single cell data. As expected, there is a consistent trend in the metabolic pathway of KEGG between single cell samples and bulk samples, as well as the immune pathway of KEGG (Supplementary Fig. 1B and 1C). We calculated the proportion of each cell type in each single cell sample and found that there were significant differences in the proportion of 6 types of cells among the three subtypes (endothelial cells, enteroendocrine, gland mucus cells, myo-cancer associated fibroblasts, PMC − like and smooth muscle cells; Supplementary Fig. 1D-1I). Among them, four types of cells also showed a consistent trend in bulk (endothelial cells, myo-cancer associated fibroblasts, PMC − like and smooth muscle cells; Supplementary Fig. 1J-1M), and the trend of gland mucus cells was also quite similar (Supplementary Fig. 1N). The above results all demonstrated that our subtype prediction model also performed well on single cell datasets.
3.7 Epithelial cells exhibit heterogeneity: In order to distinguish malignant epithelium from all epithelial cells, we re-cluster all epithelial cells. Among them, two clusters highly express epithelial marker genes (CDH1, EPCAM and KRT19), but do not express normal epithelial marker genes of gastric mucosa (Fig. 6A). In addition, Tumor_2 highly express genes related to tumors (CEACAM5 and CEACAM6, Fig. 6A). Many studies used copy number variation to infer malignant epithelial cells, but we found that the variation of CNV between different epithelial cells is little, making it difficult to effectively distinguish malignant epithelial cells (Fig. 6D). This supports the research conclusion of TCGA that most GC patients have low levels of CNV, indicating that this method may not be suitable for gastric cancer. We conducted gene set enrichment analysis (GSEA) to compare the biological characteristics of malignant and non-malignant epithelial cells. Non-malignant epithelial cells were enriched in maturity onset diabetes of the young, nitrogen metabolism and linoleic acid metabolism (Fig. 6B, Supplementary Table 15). Multiple EMT related signaling pathways, glycolysis gluconeogenesis and multiple tumor-related signaling pathways are highly expressed in malignant epithelial cells (Fig. 6B, Supplementary Table 16). The above results also demonstrated the malignant epithelial cell identity of these two cell clusters.
According to the marker genes of normal gastric epithelial cells (Supplementary Table 17), non-malignant epithelial cells are divided into 8 categories, including: pit mucus cell (PMC), gland mucus cell (GMC), PMC-like, chief cell, proliferative cell (PC), goblet cell, enteroendocrine, and enterocytes. We found that the proportion of two clusters of malignant epithelial cells gradually increases along C1-C2-C3 (Fig. 6C). The increase of numbers in malignant epithelial cells may to some extent enhance the malignancy of tumors, making it difficult for drug therapy to target all cells. We also found that malignant epithelial cells in the three subtypes have different metabolic patterns, and the metabolic patterns of the two malignant epithelial cell clusters are similar (Fig. 6E and 6F). This also proved that the metabolic differences of malignant epithelial cells in the three GC subtypes are not caused by the different proportions of the two types of malignant epithelial cells, but rather an inherent difference in the metabolic subtypes of the three GC subtypes.
3.8 CD8 + T cells and NK cells aid in distinguishing patients who benefit more from immunotherapy: We re-clustered T cells and NK cells and identified 6 main cell groups, including three CD4 + T cell clusters, six CD8 + T cell clusters, two cycling T cell clusters, three regulatory T cell clusters, two NK T cell clusters, and three NK cell clusters (Supplementary Table 18). Previous studies have shown that progenitor exhausted CD8 + T cell subpopulations can respond to immune checkpoint therapy, but terminally exhausted CD8 + T cells cannot (53) (54). We scored individual cells for progenitor and terminally exhausted gene signatures of all CD8 + T cells and found that progenitor exhausted gene signatures were mainly enriched in CD8_1, CD8_2 and CD8_6, while terminally exhausted gene signatures are mainly enriched in CD8_3 (Fig. 7A-7C). We investigated the developmental trajectory of CD8 + T cells and found that CD8_2 and CD8_3 cells are located at both ends of the trajectory. Combining with the highest terminally exhausted signature in CD8_3 cells allowed us to roughly infer the evolutionary trajectory of CD8 + T cells, which also confirmed CD8_3 cells is a subgroup of terminally exhausted CD8 + T cells (Fig. 7D). We investigated the developmental trajectory of CD8 + T cells and found that CD8_2 and CD8_3 cells are located at both ends of the trajectory. Combining with the highest terminally exhausted signature in CD8_3 cells allowed us to roughly infer the evolutionary trajectory of CD8 + T cells, which also confirmed CD8_3 cells is a subgroup of terminally exhausted CD8 + T cells. TIGIT is only expressed on CD8_3 cells. This indicated that TIGIT may be a specific target for T cell exhausted in GC, but further research is needed to observe whether the treatment targeting TIGIT can achieve efficacy in GC patients (Supplementary Fig. 2A). In CD8_1, CD8_2 and CD8_6 cells with high progenitor exhausted signatures, we found that progenitor exhausted signatures were highly enriched in the C2 subtype, followed by C3, except for CD8_6 cells with only 207 cells (Fig. 7E and 7F). These analyses suggested that the C2 subtype may have more response to anti-PD-1/PD-L1 therapy.
To confirm the immunotherapy preference for the C2 subtype, we analyzed the dataset of anti-PD-L1 therapy for GC (PRJEB25780) and the immunotherapy cohort for urothelial carcinoma (IMvigor210). No significant differences were found among the three subtypes in the PRJEB25780 cohort (p = 0.262, Fig. 7G), which may be the reason for the small sample size in this dataset. However, in IMvigor210 with a sample size of 300 people, we observed a significantly higher proportion of CR/PR patients in the C2 subtype (Fig. 7H). We found that endothelial cells and inflammatory cancer associated fibroblasts 2 (iCAF_2) has a higher proportion in the progressive disease (PD) and stable disease (SD) groups for PRJEB25780, while monocyte cluster 2 (Mono_2) has a lower proportion in the PD/SD groups (Fig. 7I-7K). Endothelial cells and iCAF_2 also showed a lower proportion in the TCGA-STAD dataset in C2, a subgroup sensitive to anti-PD-1/PD-L1 therapy (Fig. 7L and 7M). And there is a similar trend in single cell datasets (GSE183904, Fig. 7N and 7O). Many clinical trials have shown that the combination of immune checkpoint inhibitors and anti-angiogenic drugs provides significant benefits to patients compared to monotherapy (55–58). Our analysis reveals that endothelial cells are significantly higher in the non-response group to immunotherapy. This finding leads us to consider that the combination of these two drugs may enhance the efficacy of immune checkpoint therapy for GC. KM analysis indicated that GC patients with high Mono_2 infiltration have a better prognosis, indicating that Mono_2 has an inhibitory effect on GC (Fig. 7P). The enrichment analysis of marker genes for two types of monocyte clusters indicated that the antigen processing and presentation pathway was enriched in Mono_2, which may be the reason why high Mono_2 infiltration enhances immunotherapy efficacy. (Fig. 7Q).
Previous studies have shown that HLA-E: KLRC1/KLRD1 mediates tumor cell escaping from NK cell recognition and phagocytosis (59) (60). Due to NK_3 only has 67 cells, we only compared NK_1 and NK_2 in three subtypes of GC. We here do not consider the impact of KLRC1 due to its extremely low expression level (Supplementary Fig. 2B). We found that KLRD1 exhibited the highest expression of the C3 subtype in two NK cell clusters (Supplementary Fig. 2C). For the ligand HLA-E of KLRD1, HLA-E is expressed highest in C1 and C3 among the two malignant epithelial categories, and lowest in C2 (Supplementary Fig. 2D). These findings suggested that immunotherapy targeting KLRD1 or HLA-E for GC has better response in C3 patients, followed by C1.
The study found that the membrane components of NK cells in the tumor microenvironment changed, and the content of sphingomyelin (SM) was significantly reduced. It was confirmed that the disorder of serine metabolism in the tumor microenvironment was the main cause of the decline of sphingomyelin. The using of inhibitors targeted sphingomyelinase can significantly increase the content of sphingomyelin in NK cell membrane in tumor, restore the formation of neurites, and improve the recognition and killing ability of NK cells to tumor cells. The combination of targeted sphingomyelinase and immune-checkpoint blocker can play a synergistic anti-cancer effect (61). Although the number of NK_3 cells is small, we found that the precursor of sphingomyelin, serine related metabolic pathway, is significantly enriched in NK_ 3 (Supplementary Fig. 2I). Sphingomyelin synthetase (SGMS1) is highly expressed in C1, while sphingomyelin lyase (SMPD1, SMPD2 and SMPD4) is highly expressed in C2 and C3, which indicated that C1 may have the highest membrane protrusions that recognize tumor cells, thus having the strongest ability to recognize and kill tumors (Supplementary Fig. 2E-2H). Only NK_3 cells exhibit activity in sphingolipid synthesis and degradation metabolism. Perhaps patients with high NK_3 cell composition may have a better response to combination therapy targeting sphingolipid enzymes and immune checkpoint inhibitors. However, in our dataset, the number of NK_3 cells is too small to draw such a conclusion.
3.9 Stromal cells exhibit heterogeneity: We obtained endothelial cells and smooth muscle cells during the first clustering. So, we only re-clustered the cancer associated fibroblasts (CAF) to obtain four main cell groups, including: iCAF_1, iCAF_2, myofibroblastic CAF (myCAF), and antigen-presenting CAF (apCAF) (Supplementary Fig. 2A). We found that patients with high infiltration of endothelial cells, smooth muscle cells, and myCAF in the TCGA-STAD cohort had shorter survival time (Supplementary Fig. 2B-2D).
Compared to other CAFs, myCAF is significantly enriched in ECM-receptor interaction and focal adhesion, indicating that myCAF may be related to the EMT characteristics of GC (Supplementary Fig. 2E). Interestingly, these two signaling pathways are C3-specific high expression pathways, suggesting that the highest level EMT of the C3 subtype may be dominated by myCAF (Supplementary Fig. 2F). The immune pathway analysis of KEGG indicated that the platelet activation pathway of endothelial cells is highly enriched in C3 (Supplementary Fig. 2G). Studies have shown that endothelial cells can promote tumor metastasis by activating platelets (62). This mechanism may also occur in GC and is strongest in C3.
3.10 SPP1 from tumor-associated macrophage plays a role in promoting cancer through CD44: We re-clustered the myeloid cells and identified four major cell populations, including three conventional dendritic cell (cDC) clusters, one plasmacytoid dendritic cell (pDC) cluster, two monocyte clusters, and five macrophage clusters (Supplementary Table 19). None of the five macrophage clusters showed specific classically activated (M1) and alternatively activated (M2) phenotypes, which was consistent with the research in tumor microenvironment of many other types of tumors (Fig. 8A and 8B, Supplementary Table 20) (14, 63). We studied the development trajectory of macrophages and monocyte, and found that two monocyte clusters were at the beginning, while Macro_1 and Macro_3 that highly expressed M2 signature, was at the end (Fig. 8C). By the angiogenic signatures (Supplementary Table 20), we evaluated the functional phenotype of each macrophage cluster. Consistent with previous studies (63), macrophage subpopulations with high expression of SPP1 exhibited higher angiogenesis scores (Fig. 8D and 8E). And SPP1 is highly expressed in Macro_3 (Fig. 8E). There is a strong correlation between SPP1 and Macro_3 in the TCGA-STAD cohort (Fig. 8F). We found that C2 and C3 had significantly higher angiogenesis scores in the Macro_3 with the highest angiogenesis score (Fig. 8D). Some studies found that after anti CSF1R treatment (an immunotherapy targeting macrophages), the composition of SPP1 + tumor-associated macrophage did not change significantly, while other macrophages decreased significantly (64). They also found that anti CSF1R resistant tumor-associated macrophage preferentially express genes involved in angiogenesis, which is consistent with the phenomenon that Macro_3 with the highest expression of SPP1 was also enriched in the angiogenesis signature. Based on the above results, C2 and C3 may have less benefit in anti CSF1R treatment.
Research has shown that cancer cells at the edge of the tumor are more capable of inducing macrophages to become SPP1 + macrophages compared to cancer cells at the center of the tumor (65). So, SPP1 + macrophages may be more distributed at the edge of the tumor. Many studies have identified the important role of SPP1 in tumors (66) (67). Therefore, we can reasonably infer that after cancer cells at the edge of the tumor induce macrophages to become SPP1 + macrophages, macrophages in turn secrete SPP1 protein to act on cancer cells and exert tumor promoting effects. We first compared the differences in cell-cell communication among all cells of the three subtypes and found that each subtype had its own prevalent signaling communication pathway (Supplementary Fig. 4A). Two pairs of SPP1-related ligand receptor pairs were found between macrophages and malignant epithelial cells (Supplementary Fig. 4B). But SPP1/PTGER4 is specifically present in macrophages and Tumor_1. We found that SPP1 is almost exclusively expressed in macrophages, indicating that the SPP1/CD44 interaction pair is macrophage specific (Supplementary Fig. 4C). SPP1 exhibits significant overexpression in macrophages of C1, and CD44 is also preferentially expressed in malignant epithelial cells of C1 (Supplementary Fig. 4D-4J). This indicated that SPP1/CD44 crosstalk is strongest in C1.
To confirm the tumor promoting effect of SPP1, the expression of SPP1 and CD44 was detected in various gastric cancer cell lines and normal gastric epithelial cell lines (Fig. 8G and 8H). In order to avoid the impact of endogenous SPP1 on cancer cells, we selected HGC-27 cell lines with extremely low expression of SPP1 and high expression of CD44 for subsequent experiments. Here, we use recombinant human SPP1 (rhSPP1) to simulate the effect of SPP1 secreted by macrophages. When CD44 is present, we found that SPP1 significantly promotes the proliferation and invasion of HGC-27 (Fig. 8J, 8L, 8N, 8P and 8Q). Then, after knocking down the CD44, the effect of SPP1 is significantly weakened, but there is still a partial promoting effect (Fig. 8K, 8M, 8O, 8P and 8R). We speculate that it may be related to the interaction of SPP1 with more than one receptor in gastric cancer cells. According to the analysis above, we can conclude that the SPP1 protein primarily secreted by macrophages promotes the proliferation and invasion of GC cells through interaction with CD44 on the surface of GC cells.