scRNA-seq data were sourced from the NCBI database (https://www.ncbi.nlm.nih.gov/) (GSE206785), and bulk RNA-seq data and clinical information for GC and control groups were obtained from TCGA, GSE84437, and GSE15459. Raw scRNA-seq data were filtered (300-10,000 genes per cell, count > 600, mitochondrial gene proportion ≤ 10%), normalized using LogNormalize function, and cell cycle effect removed with ScaleData function. Dimensionality reduction and batch effect correction were performed via RunPCA function and RunHarmony function; visualization was done using FindNeighbors function and RunUMAP function. These functions are all part of the Seurat package (version 5.2.1) in R.

Cell types were annotated based on single-cell atlas data of GC microenvironment (7, 15). Immune cell subtype proportions in GC vs. control were calculated using the prop.table function. Marker genes for different cell types were identified with the FindMarkers package (logFC ≥ 0.25, p < 0.05). ROUGE scores (Cell cluster purity) (16) for TAM subtypes were calculated using the CalculateRogue package, and Spearman correlation analysis was performed to explore relationships between cell subtypes. GO and KEGG enrichment analyses were performed on marker genes using the clusterProfiler package.

Pseudotime analysis of TAM subtypes was performed using Monocle 3 to construct differentiation trajectories. Cells were ordered using DDRTree dimensionality reduction, and Cytotrace scores were calculated to assess differentiation potential. The characteristic gene datasets of M1 and M2 type macrophages were used to quantitatively assess the M1/M2 functional bias of TAM subtypes using the AddModuleScore function from the Seurat package (17, 18). Genes with a Moran’s index > 0.1 were selected to identify differentially expressed genes across the three TAM developmental trajectories. The find_gene_modules function categorized genes into co-expression modules, and the relative average expression of module genes in different cell types was calculated.

Cell communication analysis of scRNA-seq was conducted using the CellChat package. The ligand-receptor database was used to predict signaling pathways, identify signaling molecules and receptors, and quantify interaction strength.

Molecular structure data for human CD74 (ID: P04233) and APP (ID: P05067) were obtained from UniProt (https://www.uniprot.org). The optimal spatial conformation of APP and CD74 was modeled using Autodock Vina, optimized with Maestro software, and visualized in PyMOL. Thermodynamic parameters were calculated using the PRODIGY (https://rascar.science.uu.nl/prodigy/) and PDBePISA (https://tess.elixir-europe.org/materials/pdbepisa) websites.

Module genes (Modules 1, 2, 4, 6) with high average expression related to the LA-TAM module were intersected with LA-TAM marker genes, forming a set of 473 key genes. This gene set was validated using TCGA, GSE84437, and GSE15459 datasets. A multi-model prognosis system with 101 machine learning algorithms was constructed using the ML.Dev.Prog.Sig function from the Mime1 package. Given the excellent performance of the StepCox[forward] + Ridge model in feature selection, dimensionality reduction, and handling high-dimensional data, and its comparable performance to other top-ranking models without significant differences, we have selected this model as the foundation for subsequent research; Kaplan-Meier survival analysis, and its C-index was evaluated. Model performance was validated using the cal_unicox_meta_ml_res function, and the cal_RS_pre.prog.sig function compared its efficacy with validated GC prognosis models (19–23).

We retrieved bulk RNAseq data for gastric cancer tissues and corresponding control tissues from the TCGA database and obtained gene expression data for normal gastric tissues from the Genotype Tissue Expression (GTEx) database. Using the “cor” function from the stats package in R, we conducted a Spearman correlation analysis to assess the expression correlation between APP and PD-1 across the data described above.

Tumor samples were collected from five GC patients in the First Affiliated Hospital of Anhui Medical University, all of whom had not undergone preoperative treatments (radiotherapy, chemotherapy or immunotherapy). Samples were stored at -80 °C, and diagnoses were confirmed by two experienced pathologists. Clinical data for all patients were collected, and informed consent was obtained. The study was ethically approved by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University (Approval No. PJ2025-01-61), adhering to the Helsinki Declaration.

THP-1 cells in the logarithmic phase of proliferation were treated with Aβ solutions of varying concentrations (1.2 μM, 2.5 μM, 5 μM, 10 μM, 20 μM, 40 μM) for 12h, 24h, and 36h (24–27). Cell Counting Kit-8 (CCK-8) (Servicebio, China, G4103) was added to the cells, and the cells were incubated at 37 °C for 15 minutes, followed by measuring the absorbance at 450 nm. Cell viability (%) was calculated using the formula: (OD_experimental/OD_control) × 100%.

β-amyloid (1–42) (Aβ) (MCE, USA, HY-P1363B) was dissolved in anhydrous DMSO to a concentration of 5 mM and then diluted with cold PBS at 4 °C. The mixture was incubated at 4 °C for 24 hours, followed by centrifugation at 14,000g for 10 minutes. The supernatant was collected and stored at -20 °C. In subsequent experiments, THP-1 cells were treated with oligomeric Aβ (10 uM, 24 hours) or PBS containing the same concentration of DMSO (24–27).

Protein expression in human GC and adjacent tissues was detected by Western blot (WB). Specifically, protein concentrations were quantified using the BCA protein assay kit. SDS-PAGE gel was used to separate proteins, followed by transfer to PVDF membranes. The membranes were incubated with primary antibodies at 4 °C for 12 hours, including APP protein expression (Affinity, China, AF6084; dilution 1:1000), PD-1 protein expression (CST, USA, 861635; dilution 1:1000), and β-actin (Proteintech, USA, 81115-1-RR; dilution 1:2000). Subsequently, incubate with the corresponding secondary antibody at room temperature for 1 hour. Finally, chemiluminescence detection was used to visualize the bands.

For immunohistochemistry (IHC), tissue sections were deparaffinized, rehydrated, and incubated with the first antibody against APP (Affinity, China, AF6084; dilution 1:100), followed by incubation with secondary antibody and hematoxylin staining for nuclear visualization.

All scRNA-seq and bulk RNA-seq data analyses and statistical tests were performed using R 4.4.3. One-way analysis of variance (ANOVA) was used for comparisons among multiple groups, and t-tests were used for comparisons between two groups. Statistical significance was defined as p-value less than 0.05.

Based on previous single-cell annotation studies of GC TME (15), manual cell type annotation was performed on single-cell transcriptome samples from the GSE206785 dataset (Figure 1A). The results showed significant declined epithelial cells and fibroblasts in GC tissues compared to normal tissues (Figures 1B, C). Plasma cells remarkably decreased, along with increased B cells, CD8⁺ T cells, and myeloid cells (Figures 1B, C). KEGG enrichment analysis of differentially expressed genes between GC and control groups revealed enrichment in immune-related pathways, including the PD-1/PD-L1 signaling pathway, chemokine signaling, cytokine-cytokine receptor interaction, and tumor necrosis factor (TNF) signaling (Figure 1D). These findings suggest that GC carcinogenesis is closely linked to immune activities, in which cells interact via signaling pathways such as NF-κB, NOD, TNF, and toll-like receptor(TLR).

Macrophages were analyzed from GC samples separately (Figure 2A). According to the previous study (7), three predominant subtypes consisted of GC TAM clusters (Figure 2B). LA-TAM exhibited a ROUGE score of approximately 0.6, significantly lower than the rest subtypes. It is also suggested that LA-TAM may stay in a transitional differentiation status, with higher dynamicity and plasticity (Figure 2C). A strong positive correlation (r > 0.9) between these three major TAM subtypes was indicated, with the highest correlation between Inflam-TAM and IFN-TAM. LA-TAMs clustered separately from Inflam-TAM and IFN-TAM (Figure 2D). LA-TAM enrichment in GC significantly decreased, with higher proportion of LA-TAM-like cells in control samples (Figure 2E). Functional enrichment analysis showed that Inflam-TAM and IFN-TAM were enriched in inflammatory, apoptotic, and immune-related pathways, while LA-TAM was abundant in carbohydrate, lipid, and amino acid metabolism pathways (Figure 2F). The data above proved the functional heterogeneity and unique metabolic characteristics of TAM subtypes in the GC TME. LA-TAM may play a role to drive tumor progression.

To investigate the differentiation process of TAM subtypes, pseudotime trajectory analysis was performed (Figure 3A). Inflam-TAM and IFN-TAM stay in early stages of differentiation with high Cytotrace scores (Figure 3B). It also indicated strong proliferation and differentiation potential of LA-TAM subtype. However, LA-TAM appeared at the terminal of differentiation, featured with metabolic reprogramming (Figure 2F). M1/M2 marker gene scoring showed that Inflam-TAM and IFN-TAM detents to exhibit a more M1-like functional phenotype, while LA-TAM was more M2-like (Figure 3C). Time-series analysis of characteristic gene clusters revealed a progressive increase in APOC1, APOE, C1QA, C1QB, C1QC, GPNMB, HLA-DPA1, and HLA-DPAB1 expression, indicating their potential involvement in immune response and lipid metabolism (28, 29). In contrast, FCN1 expression decreased over time, potentially reflecting a decline in immune activities (30). SELENOP and SLC40A1 expression climbed initially but followed by downregulation (Figure 3D). Functional enrichment analysis of pseudotime genes provided insights into the dynamic changes in TME. Inflammation and oxidative stress-related pathways like IL-17 and NF-κB enriched in early and mid-stages, while lipid metabolism, cell death, and immune response pathways such as cholesterol metabolism, lysosome function, and efferocytosis clustered in late stage (Figure 3E).

To clarify cell interactions in the TME, we conducted cell communication analysis on GC samples (Figure 4). It revealed that LA-TAM has significant autocrine regulatory capabilities and closely interacts with endothelial cells (Figure 4A). Visualization of three TAM subtype communication showed that four signaling pathways (MHC-I, MHC-II, CCL and MIF) exhibited autocrine regulation in LA-TAM (Figure 4B). APP and collagen signaling were significantly enriched and highly correlated with LA-TAM (Figure 4C). Ligand-receptor analysis revealed that APP on endothelial cells bound to CD74 of LA-TAMs (Figure 4D); molecular docking confirmed a strong interaction between APP and CD74 (Figure 4E). In quantitative analysis, binding free energy of the optimal conformation between APP and CD74 was -13 kcal/mol, with an extremely low dissociation constant. In the interface region, the two proteins can form 13 hydrogen bonds and 4 salt bridges (Table 1). It revealed a complex cell interaction network focusing on LA-TAM as potential therapeutic targets in TME.

We screened genes related to the developmental trajectory changes and visualized the module analysis for three TAM subtypes via WGCNA. The results showed high expression of Modules 3 and 5 in Inflam-TAM, Module 7 in IFN-TAM and Modules 1, 2, 4, 6 in LA-TAM (Figure 5A). We selected Modules 1, 2, 4 and 6 along with LA-TAM marker genes (Figure 5B). The intersection of LA-TAM marker genes with module genes yielded 473 key genes (Figure 5C). Validation was then performed from three independent cohorts (TCGA, GSE84437, and GSE15459); a multi-model prognosis analysis system from 101 machine learning algorithms (Figure 5D) proved “StepCox[forward]+Ridge” model with best predictive performance (C-index > 0.5) (Figure 5E). Univariate Cox meta-analysis confirmed the model’s stable and significant prognostic value (HR [95%CI] = 1.63 [1.34-1.97], p < 0.001) (Figure 5F), and Kaplan-Meier survival analysis demonstrated poorer overall survival of GC patients in the high-risk group across all datasets (Figure 5G). Compared to existing GC prognosis models, our model showed superior predictive efficacy (Figure 5H).

Immune checkpoint inhibitors constitute the integral treatments for GC. We found PD-1 and programmed death-ligand 1(PD-L1) upregulation within the tumor bulk among high-risk GC group (Figure 6A). Previous studies have shown that the extracellular matrix, particularly APP expressed by endothelial cells, significantly regulates TAMs (13, 31). APP expression significantly elevated in GC (p < 0.001) (Figure 6B), and higher APP expression correlated with poorer prognosis (Figure 6C). Single-cell atlas visualization showed that APP is primarily expressed in endothelial cells, epithelial cells and fibroblasts, with significant upregulation in endothelial cells within GC tissues (Figure 6D). WB and IHC analyses proved higher APP protein levels in tumor bulk compared to adjacent tissues (Figures 6F, G).

Spearman analysis showed a significant positive correlation between APP and PD-1 expression (R = 0.2, p < 0.001) (Figure 6E). In vitro, we identified the optimal conditions for Aβ stimulation of THP-1 (Supplementary Figure 1). Aβ stimulation of THP-1 significantly enhanced the expression of PD-1 (Figures 6H, I). Thus, we concluded that high APP level in GC endothelial cells may contribute to tumor immune suppression and evasion via the PD-1/PD-L1 pathway by promoting PD-1 expression in macrophages.