Identification of cancer-associated fibrolast subtypes and distinctive role of MFAP5 in CT-detected extramural venous invasion in gastric cancer.

Introduction
Gastric cancer (GC) is a heterogeneous disease, and treatment outcomes vary widely even among patients with the same stage [1]. In clinical practice, over 90% of patients diagnosed with GC present with advanced-stage disease, and the rate of radical resection for these patients is only around 50% [2]. Extramural venous invasion (EMVI) is characterized by the infiltration of tumor cells through the gastric wall and into the lumen of extramural vessels, which can be visualized on computed tomography (CT) images [3]. Compared with tumor invasion depth, lymph node metastasis and distant metastasis, EMVI has been confirmed to be an independent risk factor for predicting patient overall survival [4,5]. The high-resolution images and multiplanar reconstruction technology currently available via contrast-enhanced multiple-row detector computed tomography (ceMDCT) can be used to assess the EMVI status of GCs before surgery [6]. In the previous work, we found that emvi score was related to microsatellite instability, tumor mutational burden and immune escape status of GC, and could effectively predict the prognosis of GC patients. CT imaging can provide additional comprehensive information about tumor heterogeneity, and radiogenomics has great potential in predicting the prognosis of GC and guiding treatment [7,8].
The occurrence and progression of GC are affected by many factors, one of the most important of which is the tumor microenvironment (TME) [9]. A large number of CAFs in tumor tissues are core components of the TME [10]. CAFs not only promote tumor occurrence, proliferation, invasion, metastasis and drug resistance but also participate in angiogenesis, lymphangiogenesis, extracellular matrix remodeling, microenvironment reconstruction and other cancer-inducing events [11]. CAFs undergo glycolysis and secrete large amounts of lactic acid and hydrogen ions, forming an acidic microenvironment and inhibiting immune cell activity [12]. Increased glycolysis leads to increased lactic acid secretion [13]. In addition to being a metabolic product, lactic acid can also cause lactylation of histone lysine residues, thus regulating gene transcription [14,15]. In addition, CAFs can induce epithelial cells to undergo epithelial-to-mesenchymal transition (EMT), which promotes tumor invasion and migration [16,17]. Because CAFs are highly heterogeneous and there is a lack of comprehensive understanding of the functions of different types of CAFs, it is still impossible to develop individualized treatment plans based on the features of CAFs in patients' tumors [18].

Materials and methods

Patient selection and collection of tissue samples
This retrospective analysis was supported by the institutional review board (Approval number: 2019PHB171–01), and the requirement for informed consent was waived. A total of thirteen pathologically proven T4aN+M0 GC patients who underwent preoperative contrast-enhanced multidetector CT followed by standard D2 gastrectomy and adjuvant chemotherapies were included. Moreover, the frozen tumor samples were stored in the institute biobank. The clinical information of the 13 patients with EMVI is shown in supplementary Table 1. The flowchart of the study is shown in Fig. 1A.

CT-detected EMVI scoring and mRNA sequencing
The CT-detected EMVI of GC was scored from 0–4. The following scoring criteria were used: 0, tumor outline not nodular without adjacent vessel; 1, tumor outline irregular or nodular without adjacent vessel; 2, stranding in the vicinity of the normal caliber of the extramural vessel without tumor attenuation within the vessel lumen; 3, tumor attenuation within the extramural vessel, where the caliber of the vessel is slightly expanded; and 4, irregular vessel contour or nodular expansion of the extramural vessel by tumor attenuation. Of these 13 included patients, 2, 5, 3, and 3 patients had EMVI scores of 1, 2, 3, and 4, respectively. There was no score of 0 in this cohort because all patients had stage T4a disease, which indicated that the masses all had a nodular appearance.
After the quality control test, frozen tumor tissue samples from these 13 patients were subjected to whole-genome sequencing using the Illumina HiSeq 4000 system. We used the R package to calculate the Spearman correlation coefficient between gene expression and the EMVI score. The threshold of significance for correlations was a P value < 0.001. A heatmap illustrating EMVI-related genes and clinical data was generated. The sequencing data were uploaded to the GEO database (GSE182831) [19].

Data retrieval and processing
Using the TCGA database (https://portal.gdc.cancer.gov/, April 2022), we obtained the original mRNA matrix and clinical data of patients with GC. The original mRNA matrix data were processed to remove duplicate samples. We downloaded the GSE15459 dataset from the GEO database (https://www.ncbi.nlm.nih.gov/geo/, April 2022) and obtained the mRNA matrix.
With respect to the GEO database (https://www.ncbi.nlm.nih.gov/geo/, the GSE183904 single-cell dataset was downloaded in May 2022 [20]. These data included data on samples from 40 GC patients, including 29 tumor samples (26 in situ tumors, 3 peritoneal metastasis tumors) and 11 adjacent healthy control samples (10 in situ controls, 1 peritoneal metastasis control). We selected all the samples for subsequent integrated analysis and research. Quality control was performed on the data to remove low-quality cells or cell debris, as well as possible double cells. We selected cells with a gene number of 500–6000 and a mitochondrial gene percentage of less than 20% for subsequent analysis. Overall, we obtained a total of 158,500 high-quality single-cell expression data points.

Integrated analysis of single-cell sequencing data and cell type annotation
After data preprocessing and quality control, we obtained high-quality expression data for 40 samples. To avoid the influence of sample batch effects, we used the canonical correlation analysis (CCA) algorithm to analyze multiple samples. For unbiased integration, the SplitObject function was used to divide the matrix according to sample groups, the NormalizeData function was used to standardize the samples, the influence of sequencing depth was removed from the original expression counts, and the data were converted into standardized and comparable data. The FindVariableFeatures function was subsequently used to define the highly variable genes of each sample; 2000 highly variable genes were selected by default. The FindIntegrationAnchors function was subsequently used to find similarity anchors between pairs of data, and the IntegrateData function was used to integrate multiple sample groups. After the sample integration step, we performed dimensionality reduction clustering and cell type annotation on the data.

Survival analysis
Survival analysis was completed through the online website GEPIA (http://gepia.cancer-pku.cn/), and the survival data of GC patients in the TCGA were selected for analysis [21]. Patients were sorted according to gene expression levels and divided into high- and low-expression groups using upper and lower quartiles, and overall survival and progression-free survival were analyzed. The logrank p value was calculated, and a logrank p value < 0.05 was considered to indicate a significant difference in survival.

Extraction of gastric fibroblasts by the collagenase method
Surgically resected normal gastric tissue and GC tissue adjacent to the cancer were collected. After the specimens were removed, they were immediately placed in DMEM supplemented with penicillin–streptomycin antibiotics. Next, 40 mg of type II collagenase (Thermo Fisher, USA) was weighed on an electronic balance, and 10 ml of DMEM was added to the sample, which was pipetted and mixed to fully dissolve the collagenase to yield a 0.4% type II collagenase solution; the solution was filtered with a 0.22 µm filter and stored in a refrigerator at 4 °C. The paracancerous tissues and GC tissues were transferred to a purification workbench under sterile low-temperature conditions and rinsed with PBS containing two antibiotics (penicillin 200 U/ml, streptomycin 200 µg/ml) and amphotericin (1 µg/ml). Ophthalmic scissors were used to cut off fat and other tissues, and the tissues were cut into 1–2 mm3 pieces and placed in a 0.1% type II collagenase solution approximately 5 times the volume of the tissue. The sample was shaken and mixed at 37 °C. Afterward, the sample was incubated at 37 °C and 5 °C in a % CO2 incubator for 4 hours. The cells were filtered through a 40 µm cell sieve to remove tissue debris, the filtrate was centrifuged at 1500 rpm for 5 minutes. Cells were transferred to a T25 culture flask and cultured in a 37 °C 5% CO2 incubator. After 24 hours, the nonadherent cells were removed. The cells were used for cytological experiments at passages 3–6.

Transfection
CAFs were cultured in 24-well plates and transfected with the constructed interfering lentivirus (Genechem, China) or the negative control(NC) lentivirus. After 24 hours, the medium was replaced, and the culture was continued. The interfering sequences used for the lentiviruses are shown in supplementary Table 2. The cells were observed under a fluorescence microscope after 72 hours. Visible fluorescence indicates successful lentiviral transfection. The lentivirus was resistant to puromycin. One week after transfection, the culture medium was replaced with complete medium containing puromycin to select cell lines that stably expressed the lentivirus. During the culture process, after the cells had grown to confluence in a 24-well plate, they were gradually passaged to the 12-well plate and the 6-well plate.

Coculture of CAFs and GC cells
The concentrations of GC cells (MKN-45 and HGC-27) and CAFs were adjusted to 1 × 106/ml in complete DMEM. A six-well plate cell culture chamber (pore size 0.4 µm; LABSELECT, China) was used to coculture CAFs and GC cells (Fig. 8A). Then, 200 µl of fibroblast suspension was added to the chamber, and 800 µl of GC cell suspension was added to the six-well plate.

Immunofluorescence assay
The cells were cultured in a 24-well plate. When the cells reached an appropriate density, the culture medium was aspirated, the cells were fixed with 1% paraformaldehyde for 15 minutes, washed twice with PBS, incubated with 3% bovine serum albumin for 30 minutes, and washed twice with PBS. Specific primary antibodies were added, and the samples were incubated overnight at 4 °C. The sections were incubated with secondary antibodies conjugated to Alexa Fluor 488 and 594 (1:200; Abcam, USA) for 30 min and washed twice with PBS. The sections were incubated with DAPI (Solarbio Life Science) for 10 min for nuclear staining and then washed twice with PBS. Forceps were used to remove the slide from the 24-well plate, after which the cells were mounted side-down on the glass slide. Glycerol was added to prevent fluorescence quenching, and a confocal microscope was used to observe and photograph the cells.

Nude mouse GC EMVI model
All animal experiments were performed according to protocols proved by Institutional Animal care and use Committee of the Peking University People's Hospital. Nine BALB/c nude mice aged 5–7 weeks were purchased from Weitong Lihua Experimental Animal Technology Co., Ltd. CAFs and MKN-45 cells in the logarithmic growth phase were digested and resuspended in PBS, the concentration was adjusted to 1 × 108/ml, and CAFs and MKN-45 cells were mixed at a ratio of 1:1. After all the mice were anesthetized with a 2% isoflurane mixture, their limbs were fixed with tape. An incision approximately 5–10 mm in length was made on the left side of the midline, and the height of the incision was not higher than the lower edge of the stomach to avoid damaging the liver during dragging. The muscles and peritoneum were separated layer by layer, tweezers were used to gently pull out the gastric wall, and an insulin needle was used to slowly inject approximately 0.1 ml of cell suspension into the front wall of the stomach. Success was considered the appearance of translucent skin on the gastric wall. After 28 days of feeding, the experiments were performed on a 9.4T small animal MRI scanner (Bruker Mice were scanned using a 40 mm volume transmit/receive coil in BioSpin, Germany). After inducing anesthesia using 4% isoflurane, the mouse was placed in the coil, and anesthesia was maintained with 1%-1.5% isoflurane through an animal nasal mask. Body temperature was maintained at 37 ± 0.5 °C using a thermostatically controlled waterbed, and the respiratory rate was monitored via an MR-compatible remote monitoring system (Model 1030 Monitoring Gating System, SA Instruments, USA). T2-weighted multiecho spin‒echo sequences were triggered by breathing.

Statistics
Data were analyzed and visualized using the GraphPad Prism 8.0 software. The Student's t-test was used to compare means between two groups, and one-way ANOVA was conducted to determine the significance of differences among multiple groups (>2). P < 0.05 was considered statistically significant.

Results

Correlation between the degree of CAF infiltration and the ceMDCT EMVI score in patients with GC
Based on the expression of PDGFRA, PDGFRB, ACTA2, THY1, PDPN, COL1A1 and FAP, we divided all the samples in the EMVI, TCGA and GSE15459 datasets into high- and low-CAF groups (Fig. 1B-D) [[22], [23], [24]]. The results showed that the ESTIMATE score, stromal score and immune score were significantly higher in the high-CAF subgroup than in the low-CAF subgroup. We also calculated the ceMDCT EMVI score for the patient samples in the EMVI dataset. We found that the patient samples with an EMVI score of 4 were all located in the high-CAF subgroup, while the patients and samples with an EMVI score of 1 were all from the low-CAF subgroup (Fig. 1B). PCA clustering verified the effectiveness of CAF grouping through data dimensionality reduction. The results showed that the similarity within the group was high, while the similarity between the groups was low (Fig. 1E-G).
We also performed MCP-counter and xCell score analyses on the high- and low CAF groups, and the results showed that the scores of the vast majority of the high-CAF groups were significantly higher than those of the low CAF group (Supplementary Fig. 1A-C). The difference in the xCell score between the EMVI dataset and the control dataset was not statistically significant, possibly due to the small sample size. We analyzed the expression levels of iCAF and myCAF markers, and the results showed that their expression levels were higher in the high-CAF subgroup (Supplementary Fig. 1D-F). We also calculated the fold changes in immune-related genes in the high- and low CAF groups (Supplementary Fig. 1G-I). The EMT pathway was significantly enriched in the high-CAF group in the three datasets (Supplementary Fig. 2A). To further study the relationship between CAFs and immune cells, we also analyzed the MCP-counter and xCell scores of CAFs (Supplementary Fig. 2B).

Single-cell sequencing clustering and cell type annotation
Through dimensionality reduction clustering and cell type annotation of the data, a total of 151,874 high-quality single-cell expression data points were ultimately obtained. Using principal component analysis, all cells were assigned to different clusters. We used the UMAP method to integrate fibroblast cluster to construct a landscape map (Fig. 2A) and used the featureplot function to divide fibroblasts into 8 subtypes [25]. Fibroblast came from different tumor samples and exhibited similar gene expression patterns (Fig. 2B), indicating the effectiveness of fibroblast clustering. We also divided the cells according to their different origins into normal gastric tissue origin and tumor tissue origin (Fig. 2C) and into GC primary tumor origin and peritoneal metastasis origin (Fig. 2D). We selected specific markers that were more highly expressed in the fibroblast subtypes and visualized them (Fig. 2E). The bubble chart shows the expression of the remaining eight markers in the eight fibroblast subtypes (Fig. 2F).

The relationship between fibroblast subtypes and GC invasion and metastasis
To explore the relationship between fibroblast subtypes and GC infiltration and invasion, we analyzed the relationships between 8 fibroblast subtypes and normal gastric tissue, GC primary lesions and peritoneal metastases and constructed landscape diagrams and percentage histograms (Fig. 3A and B). The CCL11+ Fibro amount was the highest in normal gastric tissue, the DIO2+ Fibro amount was the highest in primary GC lesions, and there was no significant difference in the amount of each fibroblast subtype in peritoneal metastases (Fig. 3B). In normal gastric tissue, GC primary lesions, and peritoneal metastases, the expression of the CCL11+ and PCSK6+ Fibro gradually decreased, while the amount of the MFAP5+ CAFs gradually increased; moreover, there was no obvious trend for the other subtypes (Fig. 3C). These findings indicate that these three subtypes may play important roles in the occurrence, development and invasion of GC. A statistical analysis (Fig. 3D) revealed that there was no significant difference in the amount of the CCL11+ Fibro (p=0.16) or the PCSK6+ Fibro (p=0.15) between normal gastric tissue and peritoneal metastases. The reason may be that these two subtypes are present at lower levels in peritoneal metastases (Fig. 3A). The expression of MFAP5+ CAFs gradually increases in normal tissues, primary GC lesions and peritoneal metastases.

Relationships between fibroblast subtypes and signaling pathways
We also performed metabolic pathway gene set analysis on the fibroblast subtypes (Fig. 4A). The results showed that the CXCL2+ and IGFN1+ Fibro were mostly negatively correlated with metabolic pathways, while the other six subtypes were mostly positively correlated with metabolic pathways. Among them, the glycolysis/gluconeogenesis pathway was closely related to the MFAP5+, ANXA3+ and CH13L1+ Fibro. The relationship between fibroblast subtypes and hallmark pathways was shown in Fig. 4B. The CXCL2+ Fibro was positively correlated with oncogenic pathways such as P53, KRAS, and TNFA pathways. To further study the mechanism by which MFAP5+ CAFs promote GC invasion, we also conducted signaling pathway analysis of MFAP5+ CAFs separately. The MFAP5+ CAFs intersected with other subtypes and was enriched in metabolism-related hallmark pathways such as oxidative phosphorylation, EMT, glycolysis and hypoxic pathway (Fig. 4C). GSEA showed that in addition to the above pathways, MFAP5+ CAFs are also closely related to angiogenesis. Therefore, we infer that this subtype is closely related to lactylation and EMT processes.

Immune cell infiltration in the MFAP5+ CAFs
Fig. 5A shows the percentages of immune cells in the eight fibroblast subtypes. The results showed that the immune cell amount in peritoneal metastases enriched in MFAP5+ CAFs was significantly higher than that in primary GC lesions, while the immune cell amount in peritoneal metastases enriched in DIO2+ and PCSK6+ Fibro was significantly higher. The amount of immune cells was found to be significantly higher in peritoneal metastases compared to that in primary GC lesions. The bubble chart shows the expression of fibroblast markers in the different immune cell high and low groups. MFAP5 expression was increased only in the Fibroblast-FAP high-expression group, indicating that MFAP5 was a highly specific fibroblast marker (Fig. 5B). The genes with increased expression in the Fibroblast-FAP high-expression group also included IGFBP6, FAP, VEGFA and CD276. Moreover, Fibroblast-FAP, macrophages and endothelial immune cells were significantly enriched in the MFAP5+ CAFs high-expression group (Fig. 5C). Fig. 5D showed that the levels of MFAP5+ CAFs were significantly positively correlated with the levels of Fibroblast-FAP and endothelial cells.

Analysis of the correlation between the MFAP5+ CAFs amount and survival in GC patients
We selected the top ten highly expressed markers in the MFAP5+ CAFs and constructed a violin plot (Fig. 6A). We also analyzed the relationships between the expression of these ten markers and overall survival and disease-free survival rates in GC patients (Fig. 6B and C). The results showed that generally, the higher the expression of most of these markers was, the worse the prognosis was. MFAP5 was closely related to overall survival but not significantly related to disease-free survival. IGFBP6, FBLN2, PCOLCE2, C1QTNF3, and SEMA3C were closely related to both overall survival and disease-free survival.

MFAP5 regulates lactylation in CAFs through LDHA
Immunohistochemistry revealed that the expression of MFAP5, L-lactyl and N-cad in samples with a ceMDCT EMVI score of 3, 4 was significantly higher than that in samples with a ceMDCT EMVI score of 1–2, while E-cad expression was significantly lower in samples with a ceMDCT EMVI score of 3–4 than in samples with a ceMDCT EMVI score of 1–2 (Fig. 7A). Fibroblasts from normal gastric tissue and GC tissue were extracted by the collagenase digestion method. FAP is considered to be a protein specifically expressed by CAFs, and SMA is a marker of activated CAFs. Immunofluorescence staining revealed that the expression levels of FAP and SMA in fibroblasts extracted from GC tissue were significantly higher than those in normal gastric tissue (Fig. 7B). These findings showed that fibroblasts extracted from GC tissue more closely resemble CAFs. MFAP5 was knocked down in GC CAFs, and the WB results showed that all three lentiviruses could effectively knock down MFAP5 expression, among which sh-3742 had the highest knockdown efficiency (Fig. 7C). L-lactyl was a pan-L-lactyl-lysine antibody for measuring the level of lactylation modification. Fig. 7D and E show that the expression level of L-lactyl in the MFAP5-KD group was significantly reduced, suggesting that MFAP5 can regulate the overall lactylation modification level in GC CAFs. Confocal microscopy revealed that MFAP5 is located mainly in the cytoplasm, while L-lactyl is located mainly in the nucleus.
Currently, a large number of studies have focused on the lactylation modification of histones, and the nucleus contains a large number of histones. Therefore, we speculate that the lactylation modification of GC CAFs mainly occurs through histones. We selected multiple H3 histone lactylation modification site antibodies (H3K9, H3K18, H2K56) for WB detection and found that the lactylation modification levels at these sites were significantly reduced in the MFAP5-KD group (Fig. 7F). LDHA serves as the key enzyme that catalyzes the conversion of pyruvate to lactic acid, which is the precursor substance for lactic acid modification. We also studied the relationship between LDHA and MFAP5 in CAFs, and the results showed that after knocking down MFAP5 in CAFs, the expression level of LDHA also decreased significantly (Fig. 7G). After overexpressing LDHA in CAFs, the expression of L-lactyl also increased significantly (Fig. 7H). Co-IP experiments also showed that MFAP5 and LDHA can directly bind to GC CAFs (Fig. 7I). A colony formation experiment showed that after knocking down MFAP5 in CAFs, the colony-forming ability of CAFs was significantly reduced (Fig. 7J). In the tube formation experiment, GC CAF-conditioned medium was added to a 24-well plate for coculture with HUVECs for 24 hours. The results showed that the tube formation ability of HUVECs in the MFAP5-KD group was significantly reduced (Fig. 7J).

MFAP5+ CAF subtype promotes GC development and invasion
We used Transwell chambers to coculture three groups of CAFs (NC group, MFAP5-KD group and MFAP5-KD+LDHA-OE group) with GC cell lines (HGC-27 and MKN-45) (Fig. 8A). The cocultured HGC-27 and MKN-45 cells were subjected to protein extraction. The WB results showed that after coculture with CAFs from the MFAP5-KD group, the expression of N-cad and vimentin in HGC-27 and MKN-45 cells was significantly reduced, while that in the E-cell population was increased (Fig. 8B and C). After LDHA was overexpressed in CAFs, the expression levels of N-cad, E-cad and vimentin in the GC cell lines were not significantly different from those in the NC group. The scratch test and Transwell test results showed that the metastatic and invasive abilities of GC cell lines cocultured with CAFs in the MFAP5-KD group were significantly weakened, while the metastatic and invasive abilities of the GC cell lines cocultured with CAFs in the NC group and CAFs in the MFAP5-KD + LDHA-OE group were not significantly different (Fig. 8D and E).
We also constructed an EMVI model in nude mice. Three groups of CAFs were mixed with MKN45 cells at a ratio of 1:1 and injected into the anterior gastric wall of nude mice. The nude mice were scanned in a 9.4T small animal MRI system 28 days later (Fig. 8F). The MRI images in Fig. 8G show that in the nude mouse EMVI model, both the NC group and the MFAP5-KD + LDHA-OE group had obvious tumor invasion outside the gastric wall according to the MRI system, and obvious tumor invasion was observed in the HE-stained sections. Immunohistochemical staining of tumor tissue revealed high expression of CD34 and SMA in some of the tumor tissues. However, in the nude mouse EMVI model of CAFs in the MFAP5-KD group, only local thickening of the gastric wall was observed via MRI, and no tumor tissue was observed. Only local tumor cell clumps were observed via HE staining. There was no obvious high expression of CD34 or SMA according to immunohistochemical staining.

Discussion
EMVI is an independent risk factor for predicting the overall poor prognosis of patients with GC [26]. Traditionally, EMVI is detected in postoperative pathology specimens and therefore cannot be used to predict treatment efficacy [27]. If the EMVI status of GC patients can be assessed through preoperative imaging, individualized treatment plans can be formulated accordingly to improve the prognosis of GC patients. In recent years, with the development of CT imaging, ceMDCT has been shown to reveal the relationships between the location, shape, and size of GCs and surrounding structures, blood supply vessels and lymph node metastasis [28,29].
EMVI-related GC radiogenomic analysis revealed that the higher the ceMDCT EMVI score was, the more obvious the CAF infiltration was. By downloading and sorting the TCGA dataset and GEO dataset, we divided all the dataset samples into high and low CAF groups in GC. The TME score and immune cell infiltration increased in the high-CAF subgroup, and GSVA revealed enrichment of genes associated with the EMT signaling pathway. These findings show that increased infiltration of CAFs accelerates the EMT process in GC and eventually leads to extramural venous invasion, ultimately leading to a poor prognosis. The results of our analysis are basically consistent with the results of most current studies. Liu et al. reported that CAF-derived IL32 specifically binds to integrin β3 through the RGD motif, thereby activating intracellular downstream p38 MAPK signaling in breast cancer cells. This signal increases the expression of EMT markers and promotes tumor cell invasion [30]. Moreover, CAFs can coevolve into an activated state through paracrine and autocrine communication during malignant tumor progression, thereby creating a dynamic signaling circuit that helps reshape the extracellular matrix and regulate the TME to promote tumor growth and progression [31]. Invasion and metastasis are characteristics of malignant tumors, and CAFs are extensively involved in these processes [32]. The EMT process is the process by which tumor cells acquire a migratory and invasive phenotype; in this process, cancer cells lose epithelial markers such as E-cadherin and acquire mesenchymal markers such as N-cadherin [33,34]. The immunohistochemistry results of this article showed that N-cadherin is relatively highly expressed in GC tissue with high EMVI scores, while E-cadherin expression is low. These findings showed that patients with high EMVI scores had enhanced EMT.
CAFs are the most abundant cell type in the TME and are the center of cross-communication between various cells in the tumor stroma [35]. CAFs have multiple sources and are highly heterogeneous, and the roles of different types of CAFs in tumor progression are multifaceted [36,37]. At present, CAFs can be divided into three main categories: myofibroblastic CAFs, immunoregulatory and inflammatory CAFs and antigen-presenting CAFs [38]. Single-cell sequencing is a technique that interprets information such as the gene structure and expression of a single cell and reflects the heterogeneity between cells [39]. Therefore, we used single-cell sequencing to identify fibroblast subtypes and screen for specific markers. The degree of infiltration of the MFAP5+ CAFs gradually increased from normal gastric tissue to GC primary lesions to peritoneal metastases. Therefore, it can be speculated that the MFAP5+ CAFs plays an important role in the invasion and metastasis of GC. Several studies have used single-cell sequencing and spatial transcriptomics to determine that MFAP5+ CAFs interact with myeloid cells, such as C1QC+ macrophages, through tumor-promoting signaling pathways, such as MIF/CD74 and IL34/CSF1R, to reshape the malignant behavior of colorectal cancer cells [40]. Shi et al. reported that MFAP5 secreted by CAFs promotes the proliferation and metastasis of bladder cancer cells by activating the NOTCH2/HEY1 signaling pathway [41]. These study also provide sufficient theoretical support for our research.
This work found that MFAP5+ CAFs are closely related to lactylation-related pathways (oxidative phosphorylation, glycolysis, hypoxia, etc.). CAFs primarily use glycolysis as an energy source, resulting in increased lactate secretion [42,43]. Lactic acid was previously thought to be a metabolic waste product, but in recent years, an increasing number of studies have proven that lactic acid can promote tumor progression [44]. Research shows that the acidic environment formed by high concentrations of lactic acid is very important for tumor cell metastasis, angiogenesis, and treatment resistance [45,46]. In addition to serving as a fuel substrate to provide the energy needed by cells, lactic acid can also be used as a precursor substance to cause lactylation modification of histone lysine residues, thus regulating gene transcription [47,48]. Many studies have shown that lactic acid chemistry is closely related to hypoxia. A study by Pascual revealed that during the maturation of red blood cells, a metabolic switch mediated by hypoxia-inducible factor is responsible for activating the ubiquitin‒proteasome system (UPS). Because lactate in the red blood cells of lupus patients regulates lysine lactylation in the UPS, USP activation mediated by the regulatory metabolic switch prevents mitochondria in red blood cells from being cleared through autophagy [49]. After we knocked down MFAP5 in GC CAFs, the overall level of lactylation was significantly reduced, and the levels of lactylation at multiple H3 histone modification sites were also significantly inhibited. These findings indicate that the MFAP5+ CAFs isoform plays a key role in lactylation-related modifications in GC.
Through radiogenomics, transcriptomics, single-cell sequencing, basic experiments and other methods, this study revealed that MFAP5+ CAFs can promote EMT in GC through lactylation, which ultimately leads to the invasion and progression of GC and the development of EMVI. The innovation of this article is the close combination of clinical imaging with gene transcriptomics, which effectively converts the imaging feature of EMVI of GC into information on the infiltration of CAFs in GC based on the degree of marker enrichment. Through single-cell sequencing, the impact of CAFs heterogeneity on the analysis was eliminated, a type of MFAP5+ CAFs subtype was ultimately identified, and its function in GC was elucidated. Although our article has many advantages, it has several limitations. For example, recent studies have shown that lactic acid mainly induces lactic acidification of histones in macrophages. However, this article analyzed only CAFs and did not conduct in-depth research on the interaction between fibroblasts and macrophages.

Conclusion
This study explores the relationship between radiographic features of GC and the tumor microenvironment, and preliminarily elucidates its potential mechanisms. This innovative work may provide important new insights into the diagnosis and treatment of GC.

Funding
National Natural Science Foundation of China, Grant/Award Number: 81901819; Beijing Natural Science Foundation, Grant/Award Number: 7232194; Peking University People's Hospital Research and Development Funds, Grant/Award Number: RDX2019–01, RS2021–08 and RDJP2022–13.

CRediT authorship contribution statement
Bo Gao: Writing – review & editing, Writing – original draft, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Xinyi Gou: Methodology, Investigation, Formal analysis, Data curation. Caizhen Feng: Investigation. Yinli Zhang: Investigation. Huining Gu: Methodology, Investigation. Fan Chai: Investigation. Yi Wang: Investigation. Yingjiang Ye: Resources. Nan Hong: Investigation. Guohua Hu: Software. Boshi Sun: Writing – review & editing, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation. Jin Cheng: Writing – review & editing, Visualization, Validation, Supervision, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Hao Yang: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Project administration, Methodology, Investigation, Formal analysis, Data curation, 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.