Analyses were performed in R (≥4.3) using Seurat (v4) and custom code. Quality control retained cells with 200–6,000 genes, and mitochondrial RNA < 10%; embryonic-biased cells were excluded. Data were normalized via SCTransform, integrated across batches with Harmony or Seurat’s integration workflow, and visualized by UMAP or t-SNE (20, 21).

Amino-acid–related gene sets (KEGG/GO) and data-driven features were decomposed by NMF to derive metabolic programs (MPs). Optimal k was chosen by silhouette, consensus stability, and reconstruction error. Program scores were computed with AddModuleScore/AUCell/GSVA (ssGSEA) and compared across lineages/states.

Communication was inferred using CellChat or CellPhoneDB with curated ligand–receptor databases. Pairs required ≥10% cell positivity and sufficient mean expression. Global networks and pathway-level views were generated and decomposed into incoming/outgoing heatmaps. Graph-theory centralities identified hubs. Myeloid subclusters were profiled separately (22).

Spatial transcriptomic data were processed using the Seurat (v5.0) and RCTD (v1.2.0) packages in R. Quality control at the spot level was performed by filtering out spots with low total gene counts or an excessively high proportion of mitochondrial reads. Data normalization was carried out using either the SCTransform method or the NormalizeData function, depending on the downstream analysis requirements.

To infer cellular composition within each spatial spot, we performed deconvolution using RNA-seq for Cell Type Decomposition (RCTD, v1.2.0), which leverages a matched single-cell RNA sequencing (scRNA-seq) reference to map cell-type proportions onto spatial coordinates. Spatial expression patterns of LAP3, pathway activity scores, and other genes of interest were visualized using SpatialFeaturePlot. Furthermore, spatial co-localization between specific cell types or gene programs was quantitatively assessed using established metrics (23, 24).

At single-cell resolution, expression and positivity of immune checkpoints (PDCD1, CD274, CTLA4, LAG3, TIGIT), antigen presentation molecules (HLA-I/II, B2M), and chemokines/receptors were quantified. In bulk cohorts, immune infiltration (estimated by ssGSEA, xCell, or EPIC) was correlated with LAP3 expression using Spearman’s rank correlation with FDR correction.

Kaplan–Meier curves and log-rank tests compared groups. Cox proportional hazards modeled associations with PH tested via Schoenfeld residuals. Time-dependent AUCs used timeROC. Unless stated, tests were two-sided with FDR<0.05. Random seeds were fixed; parameter lists and sessionInfo are provided in Supplementary Materials (25–29).

Cell lines and culture: A549 and PC9 were cultured in DMEM or RPMI-1640 (Gibco) with 10% FBS and 1% penicillin/streptomycin at 37 °C, 5% CO₂; logarithmic-phase cells were used.

A549 and PC9 cells were grouped as:①OE-NC, OE-LAP3. It onto the bottom of a 6-well plate and mark scratch lines to create wounds. Place the plate in a 37 °C, 5% CO2 cell culture incubator. At specified time points such as 0 hours and 48 hours, remove the cells from the plate and observe the width of the scratch at the same position under a microscope, taking pictures. Finally, use ImageJ software to analyze the distance of cell migration and the area of the scratch.

Cells (500–1,000/well) were plated in 6-well plates and cultured for 10–14 days. Colonies were fixed (4% PFA, 15 min), stained (0.5% crystal violet, 20 min), imaged, and counted (>1 mm diameter or >50 cells). For dye quantification, stain was eluted with 10% acetic acid and read at 590 nm.

The single-cell landscape reveals that tumor-associated epithelial and myeloid compartments (chiefly macrophages/monocytes) dominate the atlas, followed by endothelial and smooth-muscle–like cells, whereas lymphoid lineages (T, NK, B) are comparatively diffuse, jointly outlining the tumor’s cellular ecology (Figure 1A). In parallel, amino-acid–metabolism scores vary by lineage, being consistently higher in epithelial and myeloid cells and only modestly elevated in lymphoid subsets, indicating lineage-biased metabolic activity (Figure 1B). At the sample level, tumors display an enrichment of epithelial/myeloid fractions and a relative contraction of selected effector lymphocytes compared with normals, thereby shifting the system toward an “epithelium–myeloid axis” that mirrors the metabolic upregulation (Figure 1C). Dot plots reveal that LAP3 is preferentially expressed in dendritic cells, macrophages, and other monocycle lineages under normal conditions, but its expression is markedly reduced in tumor tissues (Figures 1D, E).

Building on the observed cellular compositional and transcriptional changes, we next interrogated the intercellular communication network with a specific focus on amino acid–related immunometabolic programs. Initial validation of cell identities was confirmed through cluster-marker dot plots, which reinforced stable lineage annotations and established a reliable foundation for downstream signaling inference (Figure 2A). At the global network level, epithelial cells, macrophages, and endothelial cells emerged as central communication hubs, characterized by dense connectivity and high interaction weights. Notably, the epithelium–macrophage, epithelium–endothelium, and monocyte–macrophage axes exhibited particularly strong signaling activity, whereas lymphoid-derived circuits were comparatively sparse (Figure 2B). Dissecting the ligand–receptor interactions revealed that macrophages serve as key signaling coordinators, engaging in enriched exchanges via CCL/CXCL chemokine families, integrin- and adhesion-mediated pathways, and immune receptor networks—all of which are closely associated with metabolic stress responses and immune modulation (Figure 2C). This is further corroborated by lineage-level centrality heatmaps, which highlight robust bidirectional communication involving macrophages; epithelial cells act as primary signal emitters toward stromal and immune compartments, while endothelial cells function as critical relays that integrate and propagate signals across multiple cellular circuits (Figure 2D).

At the molecular level, enzymes and transporters involved in amino acid metabolism form tightly co-expressed modules, with pronounced enrichment in both epithelial and myeloid populations. LAP3 is consistently embedded within these functional modules, reinforcing its association with core immunometabolic processes (Figure 2E). Finally, gene ontology (GO) enrichment analysis delineated distinct functional programs across lineages: epithelial cells are enriched for amino acid and protein processing; myeloid cells for antigen presentation and inflammatory activation; endothelial cells for angiogenesis and extracellular matrix (ECM) remodeling; and T/NK/DC populations for immune activation and antigen presentation (Figure 2F). Collectively, these findings delineate an integrated epithelial–immune communication circuit—orchestrated in part through endothelial intermediaries—in which metabolic activity and intercellular signaling are tightly coupled. LAP3 resides at the intersection of this network, positioning it as a potential regulator and biomarker of this immunometabolic axis.

To dissect directionality, incoming vs. outgoing communication heat maps delineate pathway sources and sinks. Matrix/adhesion programs (COLLAGEN, LAMININ, FN1, THBS/integrins) are predominantly emitted by epithelial/endothelial cells and received by macrophages, monocytes, and smooth-muscle–like cells. In contrast, inflammatory/immune-modulatory routes (MIF, GALECTIN, CD45/CD99, CXCL, RESISTIN, CLEC, GRN, COMPLEMENT, MHC-II) display a myeloid “high-in/high-out” signature, underscoring their hub role. Endothelium shows strong incoming ICAM/PECAM signals, consistent with a relay function. Lymphoid bands are weaker or more selective, with notably low T-cell output, matching the constrained lymphoid communication in this cohort. These patterns align with Figures 1, 2: metabolically active epithelial/myeloid cells also maintain tightly coupled interaction networks, reinforcing metabolism–immunity coupling (Figure 3).

To move from genes to programs, we decomposed transcriptional co-variation into discrete amino-acid metabolic programs. The similarity matrix exhibits a clear block-diagonal architecture, separating MP1–MP8 from a diffuse background, thereby indicating cohesive yet distinct programs (Figure 4A). A representative gene dot plot shows that these programs comprise enzymes and transporters; notably, LAP3 clusters with PLAC8 within the same feature set and is more highly expressed, supporting its role as a core marker of this axis (Figure 4B). When scores are projected onto the single-cell embedding, program topographies emerge—some concentrated in tumor epithelium and others in myeloid cells—suggesting a division of metabolic labor across lineages (Figure 4C). Spatial transcriptomics then anchors these signals: deconvolution maps juxtaposed belts of tumor cells and macrophages, while LAP3 peaks in tumor-rich regions and at interfaces abutting myeloid zones, consistent with a metabolic–immune intersection (Figure 4D). In sum, these analyses translate reprogramming into reusable metabolic programs and spatially align them with epithelial–myeloid interactions, again highlighting LAP3 as a hub.

Given the myeloid centrality, we resolved macrophage heterogeneity along the amino-acid axis. The gene–cell heat map defines modular enzyme/transport patterns and delineates NMF-derived subclusters (Figure 5A). In low-dimensional space, these subclusters separate cleanly; the MAC-LAP3-C4 (LAP3-high) group occupies a distinct territory relative to ME1-, MGAT1-, and PPM1B- states, indicating metabolic division of labor within macrophages (Figure 5B). Communication graphs then highlight hub-like behavior for LAP3-high cells—with high edge counts and weights to the pan-macrophage pool and to other macrophage subclusters—forming a tightly coupled myeloid subnet (Figure 5C). Marker plots corroborate this: LAP3 and PLAC8 are co-upregulated in MAC-LAP3-C4 alongside ribosomal/protein-processing features (Figure 5D). Finally, incoming/outgoing patterns show a myeloid high-in/high-out signature, most pronounced in MAC-LAP3-C4 across GALECTIN, RESISTIN, ANNEXIN, MIF, GRN, TNF routes, with notable CSF/chemokine outputs that position this cluster as a network organizer of stromal–immune interactions (Figure 5E).

We next assessed model robustness and clinical relevance for LAP3. Cross-validated Least Absolute Shrinkage and Selection Operator (LASSO) identified a stable penalty window, with λ_min and λ_1SE bracketing a sparse yet discriminative solution (Figure 6A); coefficient paths illustrate early entry of metabolism-related features as regularization relaxes (Figure 6B). In parallel, random-forest error plateaued rapidly with tree number (Figure 6C). Repeated runs yielded stable variable-importance trajectories (Figure 6D) and a consistent ranking across resamples (Figure 6E). At the expression level, LAP3 is upregulated in tumors vs. normals (Wilcoxon P = 2.592×10^-5; Figure 6F). Moreover, combined stratification by LAP3 and MKI67 produced stepwise overall-survival separation—double positive worst, double negative best, intermediates in between (log-rank P = 0.004; Figure 6G). Collectively, feature selection, stability analyses, differential expression, and prognosis all converge to implicate LAP3 as a clinically meaningful metabolic node.

Our analysis reveals that LAP3 is indeed dysregulated in multiple cancer types beyond lung cancer. Specifically, LAP3 expression is significantly upregulated in breast invasive carcinoma (BRCA), esophageal carcinoma (ESCA), glioblastoma multiforme (GBM), head and neck squamous cell carcinoma (HNSC), prostate adenocarcinoma (PRAD), and stomach adenocarcinoma (STAD). Conversely, LAP3 expression is markedly downregulated in cholangiocarcinoma (CHOL), kidney renal papillary cell carcinoma (KIRP), liver hepatocellular carcinoma (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), and pheochromocytoma and paraganglioma (PCPG) (Supplementary Figure 1A). Furthermore, we also discovered that the mutation of LAP3 is a key factor contributing to the progression of lung cancer (Supplementary Figure 1C).

In this study, we utilized the TCGA-LUAD (Lung Adenocarcinoma) cohort to develop and evaluate multiple machine learning models. Figure 7A presents a comparative analysis of model fitting accuracy through reverse cumulative distribution plots and residual boxplots. The results highlight that the Support Vector Machine (SVM) model exhibits the lowest Root Mean Square Error (RMSE) with residuals tightly clustered around zero, indicating superior predictive performance compared to other models such as PLS, Elastic net, and Random Forest. These findings suggest that SVM effectively captures the complex immune-related signals within the dataset. Figure 7B evaluates the classification performance of each model using Receiver Operating Characteristic (ROC) curves. All models demonstrate excellent discrimination capabilities (AUC > 0.95). Notably, the SVM model achieves the highest AUC value (0.997), significantly outperforming alternatives like Random Forest (AUC = 0.992), PLS (AUC = 0.992), and Gradient Boosting Machines (AUC = 0.987). Other models, including Logistic Regression (AUC = 0.955), K-Nearest Neighbors (AUC = 0.958), and Naive Bayes (AUC = 0.992), also show strong discriminatory power, whereas stepLDA (AUC = 0.583) performs notably worse, indicating its limitations in handling non-linear data structures. We also using an independent dataset from TCGA lung adenocarcinoma (LUAD) cohorts. Instead of relying solely on the multi-gene immune signature, we evaluated the diagnostic performance of LAP3 expression alone as a binary classifier to distinguish tumor from normal tissue. This analysis yielded an AUC of 0.886, demonstrating robust discriminatory power without overfitting (new Supplementary Figure 1B).But, we acknowledge certain limitations in this predictive analysis. The original high-performance models were derived from a single institutional cohort with relatively limited sample size, which increases the risk of overfitting despite cross-validation.

Having established modeling robustness, we delineated LAP3’s immune context. Cross-cohort correlation heat maps show reproducible associations: LAP3 tracks positively with myeloid and epithelial programs and relates to effector lymphocyte and antigen-presentation modules, with coherent red/blue patterns and frequent significance marks (Figure 8A). When stratifying tumors by LAP3 levels, we observe coordinated increases in immune-inhibitory checkpoints, chemokines/receptors, stimulatory molecules, and HLA genes in LAP3-high tumors, suggesting that metabolic activation coexists with adaptive immune suppression (Figure 8B). A checkpoint compendium integrating mRNA, expression–methylation coupling, and copy-number alterations maps both inhibitory and stimulatory axes; LAP3-associated upregulation frequently coincides with local hypomethylation or focal gains, indicating potential genomic/epigenetic co-regulation (Figure 8C). Single-cell visualization places LAP3 predominantly in tumor epithelium and myeloid cells, forming gradients at their interfaces—precisely where communication hubs and metabolic programs converge (Figure 8D). Overall, the LAP3-high state aligns with “chemokine activation + checkpoint upregulation + enhanced antigen presentation,” offering actionable cues for immunotherapy stratification and metabolic-immune combinations.

To functionally validate LAP3 as a key regulated gene, we established stable LAP3-overexpressing A549 and PC9 NSCLC cell lines. qRT-PCR and immunoblotting confirmed significant upregulation of LAP3 at both mRNA and protein levels compared to vector controls (P < 0.01; Figures 9A–C). Using these models, we assessed the phenotypic consequences of LAP3 elevation. CCK-8 proliferation assays revealed significantly reduced cell growth over 24–96 hours in both cell lines (Figures 10A, B), and colony formation assays demonstrated markedly decreased clonogenic potential (A549: P < 0.05; PC9: P < 0.01; Figures 10C, D). Furthermore, LAP3 overexpression impaired cell motility and invasiveness: wound-healing assays showed delayed closure at 48 hours (A549: P < 0.01; PC9: P < 0.05; Figures 11E, F), and Transwell assays confirmed fewer migrated and invaded cells (P < 0.01; Figures 11A–D). Critically, these in vitro effects translated to in vivo tumor suppression. In a subcutaneous xenograft mouse model, tumors derived from LAP3-overexpressing cells exhibited significantly reduced volume compared to controls (P < 0.01; Figures 12A, B). Collectively, these orthogonal functional experiments demonstrate that LAP3—initially identified through integrated single-cell, spatial, and computational analyses—not only serves as a biomarker but also exerts tumor-suppressive activity by dampening proliferation, migration, invasion, and in vivo tumor growth, thereby reinforcing its central role within the proposed metabolic-immune regulatory framework.