Mapping the Tissue Architecture and Endothelial‐PEC Network in HAML Using Single‐Cell and Spatial Transcriptomics Authors: Xinyang Liu, Yang Guo, Xuehong Tie, Yu Liu, Fuyu Li, Jingkai Liu, Xiaohang Sun, Jiajun Zhao, Feng Geng, Yingjian Liang Published: 2026 ABSTRACT ABSTRACTHepatic angiomyolipoma (HAML) is a rare variant of perivascular epithelioid cell neoplasms (PEComas). Unlike its more common renal angiomyolipoma (RAML), the unique spatial microenvironment of HAML and its specific progression mechanisms remain poorly characterized. Here, we integrated single‐nucleus RNA‐sequencing and spatial transcriptomics from three sporadic HAML cases. We mapped functionally heterogeneous PEComas subpopulations and their localized tissue niches. Transcriptomic alignment and spatial colocalization demonstrated a high similarity between endothelial cells and a subset of PEComas, supporting a potential transition trajectory. Furthermore, spatial intercellular communication analysis identified an endothelial‐derived PDGFD‐PDGFRB axis. We propose that this paracrine signaling engages downstream Ras cascades to sustain localized autophagic flux in adjacent PEComas, potentially acting as a compensatory survival mechanism parallel to intrinsic TSC‐mTOR defects. While definitive in vivo functional fate‐mapping and perturbation studies are requisite, this multi‐omics atlas provides a novel spatial framework for HAML pathogenesis and highlights the endothelial‐PEComa signaling network as a potential therapeutic vulnerability.Keywords: autophagy, hepatic angiomyolipoma, PEC, snRNA‐seq, spatial transcriptomics Abbreviations copy number variations differentially expressed genes endothelial‐mesenchymal transition Gene Set Variation Analysis hepatic angiomyolipoma immunofluorescence immunohistochemistry Kyoto Encyclopedia of Genes and Genomes lymphangioleiomyomatosis Multimodal intersection analysis perivascular epithelioid cell renal angiomyolipoma robust cell type decomposition single‐nucleus RNA‐sequencing spatial transcriptomics tumor microenvironment tuberous sclerosis complex unique molecular identifier vascular smooth muscle cells Introduction Perivascular epithelioid cell neoplasms (PEComas) are characterized by the co‐expression of melanocytic and smooth muscle markers and are frequently associated with TSC1/2 mutations and mTORC1 hyperactivation [1]. While angiomyolipoma most commonly arises in the kidney, hepatic angiomyolipoma represents a rare and clinically distinct entity [2]. Unlike RAML, which is strongly linked to tuberous sclerosis complex (TSC) genetic syndromes, HAML occurs primarily sporadically within the unique metabolic niche of the liver [3, 4]. The unpredictable clinical behavior of HAML—ranging from benign courses to aggressive metastasis—and the variable clinical responses to mTOR inhibition highlight the need to better understand additional regulatory mechanisms in HAML [3]. The specific microenvironmental cues and non‐classical regulatory pathways driving the pathogenesis of HAML, compared to its well‐characterized renal counterpart, remain largely unexplored. A critical limitation in the current understanding of PEComas is the over‐reliance on renal models, which are predominantly driven by canonical TSC‐mTOR hyperactivation. The rationale for investigating HAML is that the liver's distinct metabolic and immunological landscape may impose specific selective pressures, thereby shaping a unique evolutionary trajectory. We hypothesized that this hepatic microenvironmental pressure selects for parallel, non‐classical survival networks that drive its unpredictable clinical behavior—mechanisms that might differ fundamentally from those in RAML. By comparing HAML with its well‐characterized renal counterpart, we aim to uncover organ‐specific vulnerabilities that could explain the heterogeneous responses to mTOR inhibition observed in clinical practice [5, 6]. Therefore, we integrated single‐nucleus RNA‐sequencing (snRNA‐seq) and spatial transcriptomics (ST) to construct a high‐resolution, spatial multi‐omics atlas of HAML. By precisely mapping functionally distinct PEComas subpopulations, we delineate the intra‐tumoral heterogeneity and the spatial architecture of this rare tumor. Based on transcriptomic alignments and spatial trajectory inferences, we computationally formulated a hypothesis indicating a lineage transdifferentiation continuum between endothelial cells and specific PEComas subsets. Furthermore, spatial interactome modeling identified an in silico‐predicted paracrine module—the endothelial‐derived PDGFD‐PDGFRB axis. We propose that this local signaling interacts synergistically with intrinsic TSC‐mTOR defects to activate Ras pathways and maintain localized autophagy. While we acknowledge that definitive in vivo functional perturbations and larger clinical cohorts are required to unequivocally validate these causations, our findings provide a foundational spatial framework. This study delineates a hypothetical, compensatory microenvironmental circuit that substantially broadens our pathophysiological understanding of HAML beyond canonical genetic alterations. Materials and Methods Single Nucleus RNA‐Seq Data Processing Raw sequencing core data (BCL files) were converted into FASTQ format using the mkfastq pipeline from the Cell Ranger toolkit (10× Genomics). Given the single‐nucleus nature of our assay, FASTQ reads were subsequently aligned to the human reference genome (GRCh38) utilizing the STAR aligner incorporated within Cell Ranger, specifically configured to include intronic reads (pre‐mRNA mapping) to capture unspliced nuclear transcripts. A custom gene‐barcode matrix containing Unique molecular identifier (UMI) counts per nucleus was then generated. Downstream analyses were performed using the Seurat R package. Initial quality control was conducted to filter out low‐quality nuclei and technical artifacts. Specifically, we retained nuclei that expressed between 200 and 6000 genes, while genes detected in fewer than three nuclei were discarded. To exclude damaged or dying nuclei contaminated with cytoplasmic debris, we removed nuclei exhibiting a mitochondrial transcript proportion greater than 10%, calculated via the Percentage FeatureSet function. After QC filtering, the gene expression matrix was normalized using the LogNormalize method. Highly variable genes were identified globally using a variance‐stabilizing transformation method. To correct for potential batch effects across different datasets, multiple experimental runs were integrated applying Seurat's FindIntegrationAnchors and IntegrateData functions, thereby projecting the dataset into a shared reduced‐dimensional space for downstream clustering analysis [7]. Spatial Transcriptomic Analysis Raw sequencing reads generated from the ST libraries were aligned to the human reference genome (GRCh38/hg38) utilizing the Space Ranger pipeline (10× Genomics). The pipeline generated spot‐level spatial gene expression matrices, which were subsequently imported into the Seurat R package for downstream quality control, normalization, and clustering. Following data normalization and identification of highly variable genes, principal component analysis was conducted for initial linear dimensionality reduction. Subsequently, graph‐based clustering and non‐linear dimensional reduction visualizations were performed using t‐SNE and UMAP algorithms. Because individual spatial spots often encompass multiple cells, we performed spatial deconvolution to resolve the microenvironmental cellular composition. Specifically, we integrated our snRNA‐seq dataset with the ST data applying robust cell type decomposition (RCTD) and Multimodal intersection analysis (MIA) algorithms. This dual integrative approach enabled the precise prediction and mapping of specific cell clusters (such as endothelial cells and PEC subpopulations) onto their original spatial tissue architecture [8]. Clustering and Marker Gene Identification To account for potential batch effects across diverse samples, the Harmony algorithm was applied to the PCA space, generating a batch‐corrected joint embedding. Utilizing these integrated coordinates, a shared nearest neighbor graph was constructed via the Seurat FindNeighbors function. Subsequently, cell clustering was performed utilizing the graph‐based Louvain clustering algorithm implemented in the FindClusters function. The resulting cellular clusters were visualized in a two‐dimensional space employing t‐SNE and UMAP non‐linear dimensionality reduction algorithms. For high‐resolution subclustering of specific lineages (such as PEComas sub‐populations), the identical analytical pipeline—comprising highly variable gene selection, dimensionality reduction, and clustering—was iteratively applied to the subsetted data [9]. To identify cluster‐specific marker genes, differential expression analysis was performed using the FindAllMarkers function, deploying the non‐parametric Wilcoxon rank‐sum test. The resulting p‐values were adjusted for multiple testing utilizing the Bonferroni correction. Finally, biological cell types and states were comprehensively annotated by integrating automated cellular recognition via the SingleR package with meticulous manual curation guided by classical cell‐type marker genes established in canonical literature [10, 11, 12, 13]. Cell and Cluster Similarity Analysis To perform robust cluster‐level comparisons across individual patients, pseudobulk expression profiles were generated. Specifically, the log‐normalized expression values within each predefined cell cluster were averaged for each patient to represent their sample‐specific expression signatures. To mitigate the confounding influence of extreme biological or technical anomalies, a rigorous outlier removal step was implemented. Samples (or specific intra‐cluster pseudobulk profiles) whose first three principal component scores deviated beyond the mean ±3 standard deviations of their respective clusters were identified as outliers and subsequently excluded from the analysis. Furthermore, to refine the feature space for downstream comparative analyses, low‐expression and low‐variance genes were filtered out; only genes demonstrating a mean expression level > 0.1 and an expression variance > 0.05 across all aggregated cluster vectors were retained [14]. Detailed information can be found in Data S1. Results A Single‐Nucleus Transcriptomic and Spatial Landscape of Paired HAML and Adjacent Normal Tissue To elucidate the microenvironmental landscape of HAML, we generated paired snRNA‐seq profiles from a treatment‐naïve HAML tumor and its matched adjacent normal liver tissue. Concurrently, ST data were acquired from three HAML sections encompassing two tumor‐normal interfaces and one core tumor region (Figure 1A, Table S1). This clinical characterization supports the sporadic presentation of the cases in our cohort. Up to the latest clinical follow‐up, all patients remained disease‐free post‐resection; however, long‐term prognostic outcomes remain to be continuously monitored. Following stringent quality control, a total of 18,214 high‐quality single nuclei (11,129 from HAML and 7085 from adjacent normal tissue) were captured, with a median of 2880 UMIs and 1732 genes per nucleus (Figure S1A). Unsupervised clustering visualized via t‐SNE grouped these nuclei into 15 distinct subpopulations (C0–C14) (Figure 1B). Based on the expression patterns of established lineage markers, we annotated six major cell types [15, 16, 17]: hepatocytes (n = 5,578; C1, C4, C6, C10, C12, C14; characterized by CUX2 and PCSK6); PEComas (n = 10,550; C0, C2, C3, C5, C8; driven by MITF, PMEL, and MLANA); endothelial cells (n = 1,072; C7; expressing FLT1 and VWF); fibroblasts (n = 610; C9; expressing DCN and COL3A1); vascular smooth muscle cells (VSMCs, n = 95; C13; expressing MYH11 and ACTA2); and a minor population of epithelial cells (n = 309; C11; expressing KRT7) (Figure 1B–D, and Figure S1B,C). Single‐nucleus transcriptomic profiling of hepatic angiomyolipoma. Experimental workflow and data preprocessing steps. (B) t‐distributed stochastic neighbor embedding (t‐SNE) plot of the six identified main cell types and 15 subclusters in HAML and normal tissue. (C) t‐SNE plot showing the distribution of 6 cell types in normal and tumor tissues. (D) Heatmap showing the average expression of top 3 differentially expressed genes (DEGs) among six cell types. The color scale indicates normalized gene expression levels among the clusters. (E) Relative proportion of cells originating from tumor and adjacent normal tissues across the six major cell populations. (F) Heatmap showing Spearman correlation coefficients between transcriptional profiles of six cell types. |Rs| > 0.3 was considered a significant correlation. Cellular composition analysis revealed a profound reshaping of the HAML tumor microenvironment (TME). While hepatocytes natively constituted the vast majority of the adjacent normal tissue, the tumor compartment was robustly dominated by PEComas, accompanied by regional expansions of endothelial cells and fibroblasts (Figure 1E). To further infer the intercellular relationship within this highly remodeled TME, we assessed pairwise Spearman correlations based on cellular co‐infiltration patterns. We observed a striking positive correlation specifically between PEComas and endothelial cells (|Rs| = 0.834) (Figure 1F), highlighting a tightly coordinated spatial colocalization that warrants further exploration to determine their potential developmental or functional crosstalk. Spatial Transcriptomics Reveals the Intratumoral Architecture and Microenvironmental Niches of HAML To construct an unbiased spatial architecture of this extremely rare tumor entity, we mounted paraffin‐embedded sections from three independent HAML specimens (HAML#1–3) onto spatially barcoded microarrays for ST analysis (Table S1). To capture the tumor heterogeneity, HAML#1 and HAML#2 sections were deliberately selected to encompass the tumor‐normal interfaces (encompassing both distinct compartments, designated as T and N), whereas HAML#3 was representative of a dense tumor core (Figure 2A). Guided by H&E morphological annotations, the ST platform successfully captured robust transcriptomic signals. Depending on local tissue histology, each 55 μm‐diameter spot encapsulated approximately 1 to 9 cells. Predictably, lipid‐rich vacuolated domains within the tumor yielded lower cell contents compared to spatially dense areas like normal hepatocyte cords (Figure S2A). Sequencing depth was highly faithful, yielding median UMI counts of 22,657 (HAML#1), 16,668 (HAML#2), and 26,655 (HAML#3) per spot (Figure S2B). Spatial deconvolution and microenvironmental mapping of human HAML tissue. (A) Spatial transcriptomic profiling of three HAML specimens (HAML#1–3). For each specimen, panels from left to right show: Representative H&E staining with morphological annotations (T: Tumor, N: Adjacent normal tissue; prominent morphological boundaries are outlined with dotted lines); spatial domain distributions defined by unsupervised clustering; and spatial mapping of the predominant predicted cell types. (B) Stacked bar plots illustrating the relative composition of deconvoluted cell types within each identified spatial cluster across the three sections. Areas are magnified on the right for higher‐resolution visualization. Scale bars, 100 μm. (C) Multimodal Intersection Analysis (MIA) heatmaps evaluating the spatial colocalization between scRNA‐seq‐identified cell lineages (y‐axis) and ST‐defined tumor spatial clusters (x‐axis). Red indicates spatial enrichment (significantly high overlap), whereas blue indicates depletion. (D) Representative immunohistochemical (IHC) staining of the PEComas marker PMEL and the endothelial marker VWF in specific HAML tissue spots matching the ST coordinates. Boxed areas are magnified on the right for higher‐resolution visualization. Scale bars, 100 μm. Unsupervised spatial clustering segregated the spots into 13 distinct spatial domains (C0–C12) for HAML#1, and 16 regions (C0–C15) for both HAML#2 and HAML#3 (Figure 2A), with each cluster characterized by specific spatial transcriptomic signatures (Tables S2–, S4). Hierarchical clustering confirmed that spatial clusters anatomically corresponding to tumor or normal tissues naturally segregated based on their overarching expression profiles (Figure S2C). Given the multicellular nature of ST spots, we employed robust cell type decomposition (RCTD) to integrate our snRNA‐seq reference for spatial deconvolution. This analysis demarcated clear compartmentalization: hepatocytes exclusively dominated the adjacent normal regions, regular vascular tracks comprised predominantly VSMCs and fibroblasts, while PEComas robustly aggregated within the tumor tissue (Figure 2A,B). Notably, mapping co‐infiltration patterns highlighted specific intratumoral niches where PEComas and endothelial cells frequently co‐localized. Multimodal intersection analysis (MIA) statistically reinforced this intimate spatial overlap across the tumor regions (Figure 2C). While this strict spatial adjacency strongly implies a shared microenvironmental niche or potential crosstalk, it is essential to acknowledge that this spatial co‐localization relies on correlative profiles and cannot definitively establish a lineage transition, thereby warranting future functional fate‐mapping validations. To morphologically validate these spatially defined transcriptomic compartments, tissue sections were assessed by immunohistochemistry (IHC) (Figure 2D). The canonical markers PMEL (for PEComas) and VWF (for endothelial cells) exhibited robust protein‐level enrichment that precisely mirrored the expression patterns mapped by ST in the corresponding tumor zones (Figure S2D), thereby firmly anchoring our sequencing findings to the pathological reality. Transcriptomic Profiling Delineates the Heterogeneity and Organ‐Specific Divergence of PEComas As the cardinal neoplastic cell type in PEComas, the typical PEComas compartment was deeply interrogated to identify its precise transcriptomic signatures. Beyond the canonical angiomyolipoma diagnostic markers (PMEL, MLANA) (Table S5), our differential expression analysis delineated a robust set of novel PEComas‐selective markers, notably including SCGN, CPA6, F13A1, and ROR1 (Figure 3A and Figure S3A). The spatial expression mapping derived from our ST data corroborated their tumor‐specific distribution. These computational findings were orthogonally validated via multiplex immunofluorescence (IF) staining, which demonstrated stringent spatial co‐localization of the newly identified markers SCGN and F13A1 with the canonical marker MLANA explicitly within the tumor core (Figure 3B). Spatiotemporal molecular architectures and lineage dynamics of PEComa subpopulations. t‐SNE feature plots visualizing the log2‐transformed expression distribution of newly identified representative PEComas marker genes. (B) Representative multiplex immunofluorescence (mIF) images validating the spatial co‐localization (indicated by white arrows) of the canonical PEComas marker MLANA (pink) with the novel markers SCGN (green, top panel) and F13A1 (green, bottom panel) within the tumor region. Nuclei were counterstained with DAPI (blue). Boxed areas are magnified on the right. Scale bars, 100 μm. (C) Hierarchical clustering heatmap demonstrating the transcriptional distinctiveness based on differentially expressed genes (DEGs) between PEComas and adjacent non‐neoplastic mesenchymal lineages (VSMCs and fibroblasts). (D) Violin plots comparing the expression levels of nine designated marker genes across fibroblasts, PEComas, and VSMCs. (E) Monocle 2 pseudotime trajectory analysis revealing the dynamic developmental continuum and distinct branching fates (fates 1 and 2) associated with specific PEComa subclusters. (F) Dot plot illustrating the functional pathway enrichment across distinct PEComa subclusters. Circle size indicates the gene count, while color intensity represents the statistical significance (p‐value). (G) Multimodal Intersection Analysis (MIA) heatmaps displaying the spatial colocalization strength (MIA score) between snRNA‐seq‐identified PEComa subclusters and ST‐defined microenvironmental regions across the three HAML specimens. To precisely isolate PEComas from adjacent non‐neoplastic liver stroma, we employed rigorous binomial probability algorithms to establish a definitive PEComas‐specific expression matrix (Figure 3C). PEComas exhibited heterogeneous retention of mesenchymal signature genes (Figure 3D and Figure S3B). Furthermore, gene ontology analysis revealed that this overarching PEComas signature was heavily driven by programs dictating cellular plasticity and microenvironmental sensing, particularly ECM‐receptor interactions, focal adhesion, and MAPK/AMPK metabolic signaling hubs (Figure S3C), highlighting the well‐documented hybrid epithelioid‐mesenchymal phenotypic plasticity inherent to PEComas biology. Given the recognized clinico‐biological heterogeneity of PEComas, determining the substructure of the PEComas compartment is critical. Unsupervised sub‐clustering delineated multiple transcriptomic subpopulations. Pseudotime trajectory inference and functional enrichment analyses revealed that intra‐tumoral heterogeneity in HAML is not random, but rather represents biologically distinct states executed by specific PEC subpopulations (Figure 3E). Subcluster C1 maps to the root of the developmental trajectory, retaining prominent EndMT‐like transcriptional features, suggesting that it may represent a relatively undifferentiated intermediate cellular state. As the trajectory branches, the cellular functions strictly diverge. Subsets C2/C3 are strongly enriched for intracellular catabolism programs—specifically autophagy, lysosomal activity, and response to hypoxia. This transcriptomic profile defines an ‘adaptive metabolic state’ that likely facilitates tumor cell survival within the necrotic or nutrient‐deprived tumor core. Conversely, Subcluster C4 exhibits a specialized ‘motile phenotype’, characterized by the robust upregulation of focal adhesion, ECM‐receptor interaction, and actin cytoskeleton remodeling networks, which are features associated with cellular motility and matrix interaction (Figure 3F). Spatially mapping these subclusters via MIA confirmed that PEComas functional heterogeneity is closely tied to distinct localized niches, indicating that tumor progression in HAML is driven by a cooperative division of labor among these specific subpopulations (Figure 3G). HAML is an exceedingly rare variant compared to the classic RAML, identifying the organ‐specific divergence is clinically imperative. To contextualize our findings within the broader PEComa spectrum, we comparatively integrated our dataset with published RAML and lymphangioleiomyomatosis (LAM) single‐cell cohorts [18, 19]. Deconvolution analysis revealed a divergent distribution of our defined PEComas subclusters: while C2, C3, and C8 mapped predominantly to pulmonary LAM, subclusters C0, C5, and C8 aligned closely with the RAML microenvironment (Figure S3D). Strikingly, comparative phenotypic profiling unmasked profound organ‐specific tissue architectures [20] (Figure S3E). While canonical PEComas pathogenesis is dogmatically driven by the TSC‐mTOR axis, our comparative expression matrix (Figure S3F) suggests that HAML tumors exhibit unique compensatory dependencies. Specifically, alongside the expected mTOR pathway engagement, HAML uniquely enriched for PDGF signaling, extensive vascular angiogenic programs, and specific immunosuppressive signatures compared to RAML (Figure S3G,F). These divergent evolutionary trajectories not only underscore the rational necessity of studying this rare hepatic subtype but also suggest that HAML may integrate organ‐specific metabolic signaling parallel to the canonical TSC‐mTOR axis to support tumor progression. Transcriptomic Continuum and End‐MT‐Like Plasticity Reveal a Putative Endothelial Ancestry of PEComas Given the technical hurdles in isolating primary PEComas and maintaining their phenotypic fidelity in vitro, we sought to delineate their developmental hierarchy utilizing a computational approach. By matching our bulk and single‐cell PEComas signatures against 36 classical human cell‐type modules (CellMarker database), we identified that endothelial cell signatures shared the most profound transcriptomic overlap with PEComas (Fisher's exact test, p = 1.39 × 10−7) (Figure 4A; Table S6). Accordingly, representative endothelial markers—including CD36, EBF1, FILIP1, and LAMB1—were significantly enriched within the HAML tumor compartment compared to adjacent normal tissues (Figure 4B, Figure S4A). Transcriptomic state transition from endothelial cells to PEComas. Bar plot indicating the statistically significant scores (−log10(p‐value)) of transcriptomic similarity between PEComa and reference cell types. (B) Violin plots showing the normalized expression levels of 5 signature genes in endothelial cells from tumor and normal tissues. (C) Representative multiplex immunofluorescence images of HAML tissue. Magnified views display individual transitional cells co‐expressing the PEComa markers MLANA and PMEL, along with the endothelial markers CD36 and EBF1. (D) Monocle 2‐inferred developmental trajectory detailing the continuum of PEComas and endothelial cells (top). Heatmap characterizing coordinate changes in mesenchymal marker gene expression progressing through pseudotime (bottom). (E) Heatmap showing the expression levels of transcription factors across each subpopulation. (F) t‐SNE plot showing the expression level of PPARG (Left). Corresponding spatial expression pattern of PPARG (Right). (G) Bubble plot showing the enriched KEGG pathways in PEComa and endothelial cells. (H) Western blot analysis evaluating the protein expression of mesenchymal (ACTA2) and endothelial (CD36, VE‐cadherin) markers. Importantly, in situ mIF confirmed the spatial co‐localization of canonical PEComas biomarkers (MLANA and PMEL) with endothelial markers (CD36 and EBF1), suggesting a transitional cellular state (Figure 4C). To further interrogate this biological continuum, pseudotime trajectory mapping via Monocle 2 identified a continuous transcriptional trajectory from an endothelial‐associated state toward a PEC‐associated state. Along this in silico trajectory, critical transcription factors (EBF1, MITF, and PBX1) and actin cytoskeleton‐associated regulators (VCAM1, VCL, DDR2, and FYB) exhibited temporal upregulation, aligning with cell‐state reorganization (Figure 4D). To decode the transcriptional forces driving this transition, pySCENIC regulatory network analysis was employed, highlighting STAT2, TCF12, and FOXO3 as core regulators exhibiting robust activity within the PEComas compartment (Figure 4E). Prior evidence establishes FOXO3 as a critical orchestrator of adipogenesis through the AMPK/PPARG cascade [1]. Mirroring this mechanism, we observed significant coordinate enrichment of the AMPK signaling pathway alongside augmented PPARG expression specifically within tumor niches (Figure 4F,G and Figure S4B). The acquisition of a mesenchymal‐like phenotype by endothelial cells typically proceeds via the endothelial‐mesenchymal transition (EndMT) program [21]. Consistent with EndMT, western blot analysis of HAML tissues demonstrated a profound downregulation of endothelial adhesion molecules (VE‐cadherin and CD31) concurrent with a robust elevation of the mesenchymal marker ACTA2, relative to normal hepatic controls (Figure 4H). While in vivo lineage tracing is warranted to establish direct causality, our integrated computational and phenotypic findings are compatible with an EndMT‐like cellular state. These correlative data support a compelling hypothetical model wherein local endothelial cells may serve as a potential contributor to the PEC population within the HAML microenvironment. Spatial Architecture and Transcriptomic Heterogeneity of HAML To explore the spatial genomic drivers within distinct tumor clusters, large‐scale CNVs were inferred based on ST data. By utilizing normal spots within the tissue as a reference background, we evaluated chromosomal amplifications and deletions across different spatial regions. Consequently, this approach successfully identified distinct tumor regions in HAML characterized by highly aberrant CNV profiles (Figure 5A). Functional enrichment analysis of these spatially resolved tumor regions demonstrated a robust hyperactivation of autophagy‐centric pathways—specifically phagosome, lysosome, and Endocytosis—across both L and T sections (Figure 5B, Figure S5A). Transcriptomic characterization of the HAML neoplastic compartment. Heatmaps illustrating the inferred copy number variation (CNV) landscape across HAML#1–3 samples. (B) Dot plot illustrating KEGG pathway enrichment across distinct spatial clusters in HAML#1. Dot size corresponds to the gene count in each pathway, and color gradient indicates statistical significance (p‐value). (C) Venn diagram showing the intersection of differentially expressed genes (DEGs) shared among HAML#1–3 spatial transcriptomics (ST) datasets and the PEComa scRNA‐seq dataset. (D) KEGG pathway enrichment of the 248 shared DEGs identified in Figure 4C. (E) Immunoblot analysis of autophagy‐related proteins (LC3B and Beclin‐1) in matched normal (N) and tumor (T) tissues. (F) Spatial mapping delineating the interacting boundaries between tumor and adjacent microenvironmental clusters in HAML#1 and HAML#2. (G) Bubble heatmap illustrating the communication probability of ligand‐receptor (L‐R) pairs across spatial interfaces. Dot size indicates statistical significance derived from permutation testing, while color intensity represents the communication probability. By intersecting upregulated gene sets from the single‐cell PEComas populations with spatially defined tumor regions (HAML#1–3), we mapped a core pathogenic signature comprising 248 co‐upregulated genes (Figure 5C). KEGG evaluation of this core signature confirmed prominent enrichment in the AMPK/MAPK signaling cascades and actin cytoskeleton regulation (Figure 5D). Given the established interplay between AMPK signaling and cellular catabolism, we experimentally validated this in silico finding via Western blot. Notably, the pivotal autophagy mediators LC3B (demonstrating canonical lipidation from LC3B‐I to LC3B‐II) and Beclin‐1 were markedly overexpressed in HAML tissues compared to normal liver, corroborating the functional reliance of this tumor on autophagic flux (Figure 5E). Furthermore, tumor‐stroma crosstalk represents a critical dynamic in shaping the local microenvironment. Spatially restricted analysis along the tumor‐normal boundary (Figure 5F) unveiled precise reciprocal cell–cell communication loops. Interestingly, tumor‐derived signals strongly targeted adjacent normal areas, potentially dictating local angiogenesis and stromal remodeling. Conversely, the adjacent uninvolved regions reciprocated by enforcing tumor autophagic activity via pathways such as MIF‐CD74‐CXCR4, ANGPTL4‐SDC3, and C3‐ITGAM‐ITGB2 (Figure 5G, Figure S5B). Together, these spatially anchored signaling hubs highlight a local symbiotic ecosystem, presenting specific ligand‐receptor interactions as potential vulnerabilities to disrupt HAML progression clinically. Cell–Cell Communication Networks Implicate a PDGFD‐PDGFRB‐Ras Axis in HAML Pathogenesis To decode the functional interactome within the HAML microenvironment, we executed CellChat to infer spatial crosstalk among the co‐localized cell populations (Figure 6A). Consistent with their spatial juxtaposition, we identified robust paracrine interactions between the endothelial cluster (C7) and distinct PEComas subpopulations (C0, C3, and C5), reflecting pronounced intra‐tumoral heterogeneity and potential niche‐driven differentiation (Table S7). Network centrality analysis underscored that specific PEComas clusters (C0 and C3) alongside endothelial (C7) and epithelial cells (C11) served as dominant signaling sources aiming at the C0 and C5 PEComa sinks (Figure 6B). Notably, the bulk of these interactions operated via paracrine mechanisms, though the major PEComa cluster C0 also exhibited potent autocrine signaling (Figure 6C). Inference of spatial crosstalk among colocalized cell populations using CellChat. Circle plot illustrating the number of significant ligand‐receptor (L‐R) interactions among distinct cell clusters (C0–C13). Edge width is proportional to the number of interacting L‐R pairs. (B) Hierarchical plot depicting the inferred intercellular signaling network between source and target cell clusters. (C) Heatmap illustrating the relative importance of each cell cluster within the PDGF signaling network, based on four computed network centrality measures. (D) Bar plot showing the relative contribution of individual L‐R pairs to the overall PDGF signaling network. (E) Immunoblot analysis of key PDGF signaling pathway proteins and their downstream effectors in matched normal (N) and tumor (T) HAML tissue. (F) Feature plots (t‐SNE) visualizing the spatial expression patterns (log2 transformed) of key PDGF pathway‐related genes. (G) Alluvial plots delineating the inferred outgoing (left) and incoming (right) communication patterns, mapping the correspondence among cell groups, latent patterns, and signaling pathways. Among the identified pathway architectures, PDGF signaling prominently anchored this multicellular crosstalk, primarily driven by the PDGFD ligand and its cognate PDGFRB receptor (Figure 6D). Integrating these data with the KEGG pathway knowledgebase, we delineated that PDGFD‐PDGFRB engagement canonically triggers the downstream Ras signaling cascade. We mapped this axis to our snRNA‐seq dataset, revealing coordinate upregulation of critical Ras pathway effectors—including PDGFRB, GRB2, SOS1, MRAS, RGL1, RALA, and PLD1—specifically within the tumor compartment (Figure 6F). Crucially, to provide robust protein‐level validation of this computational prediction, Western blot analysis confirmed the marked overexpression of PDGFRB and these core Ras/autophagy mediators in HAML tissues relative to normal liver (Figure 6E). While these protein‐level findings provide additional evidence that components of the predicted PDGFD–PDGFRB–Ras/autophagy network are enriched in HAML tissues, direct functional perturbation will be required to establish pathway activation and its causal regulation of autophagic flux. Furthermore, integrating our findings with the classical PEComa paradigm provides novel mechanistic insights. While conventional PEComa genesis is driven by the TSC1/2‐mTORC1 pathway, our identification of a PDGFD‐Ras associated signaling module suggests a parallel or compensatory oncogenic network. Given the established crosstalk between Ras signaling and autophagy modulation, this paracrine‐driven Ras activation may synergize with intrinsic TSC‐mTOR defects to fuel the profound plasticity and metabolic adaptation seen in the hepatic variant. Finally, to unravel how global communication networks are coordinated, we performed pattern recognition analysis. This uncovered distinct modes of multi‐cluster integration: for instance, incoming signals to the C0 and C5 PEComas were dominantly governed by Pattern 1 (PDGF, BMP, and ANGPTL pathways), whereas endothelial C7 reception was characterized by Pattern 2 (VEGF signaling) (Figure 6G). Collectively, these transcriptomic and protein‐level topological analyses illuminate a highly coordinated microenvironmental signaling web, offering a hypothetical yet actionable mechanistic framework for distinct PEComa subpopulation behaviors in HAML. Discussion Although angiomyolipoma predominantly arises in the kidney and is frequently associated with the TSC [4], the hepatic form is an extremely rare entity with distinct clinical and biological features. While RAML has been extensively studied within the classic TSC1/2‐mTOR activation paradigm, the distinct cellular landscape of HAML—dictated by the unique hepatic niche—remains largely unexplored [3]. Furthermore, the limited clinical efficacy of mTOR inhibitors (sirolimus) in a subset of patients suggests that alternative or compensatory mechanisms drive HAML pathogenesis [4]. To address this gap, we utilized spatially resolved snRNA‐seq integrated with ST to systematically delineate the TME of HAML. By mapping the cellular architecture, we aim to uncover how liver‐specific microenvironmental cues dictate the behavior and heterogeneity of PEComas, the hallmark neoplastic cells of PEComas. The marked biological heterogeneity of PEComas is reflected in the morphological diversity of PEComas, which can exhibit epithelioid, spindle‐shaped, or vacuolated phenotypes expressing melanocytic and smooth muscle markers [22, 23]. In this study, our transcriptomic analysis identified multiple distinct PEComa subpopulations within HAML. These subgroups likely reflect divergent functional states, such as varied metabolic activity, proliferative capacity, or progressive stages of cellular differentiation within the tumor core versus the invading edge. Interestingly, when cross‐referencing our data with publicly available AML datasets from other organs, we observed variations in the proportional distribution of these PEComa subgroups. This suggests that while PEComas share core characteristic markers across diverse organs, the hepatic stromal niche exerts specific selective pressures that shape the discrete intratumoral sub‐clonal architecture of HAML. The histogenesis of PEComas has long been a subject of debate, with hypotheses suggesting neural crest, smooth muscle, or pericytic origins [24, 25, 26]. By integrating transcriptomic similarity clustering and spatial co‐localization analysis, our data revealed an intimate association between PEComas and endothelial cells. Trajectory inference modeling identified a progressive transcriptional continuum between endothelial‐associated and PEComa‐associated cellular states, seemingly mimicking the EndMT process [21]. During this putative transition, we observed a downregulation of classical endothelial markers concomitant with the upregulation of specific transcription factors and mesenchymal traits. However, it must be robustly acknowledged that these findings are primarily based on computational trajectory inferences and structural co‐localization. While transcriptomically intriguing, these data are correlative. The mechanisms by which these cells acquire melanocytic features during an EndMT‐like event remain elusive, and definitive lineage‐tracing models (e.g., in vivo genetic fate mapping) are absolutely required to substantiate this proposed endothelial‐to‐PEComa lineage hierarchy. A major question raised by our study is how the observed PDGF‐associated signaling relates to the canonical TSC–mTOR pathway in PEComas. Alterations in TSC1/TSC2 drive robust mTORC1 activation, which typically suppresses autophagy via the inhibition of the ULK complex [27, 28]. Our findings do not challenge this model. Rather, the enrichment of PDGF‐associated signaling and autophagy‐related programs suggests that microenvironmental interactions may provide additional regulatory inputs to PEComa cells. We postulate that in the uniquely structured HAML microenvironment, prolonged Endothelial‐PEComas crosstalk bypasses the canonical mTOR suppression, utilizing Ras hyperactivation to promote ULK complex formation and sustain autophagy [29]. This compensatory autophagy may facilitate nutrient recycling and metabolic adaptation, allowing sub‐clonal PEComas to survive hypoxic or therapeutic stressors. Nevertheless, this proposed PDGF‐Ras‐autophagy axis remains computationally inferred. Future functional assays—incorporating spatial proteomics and in vitro receptor‐knockdown models—are paramount to validate whether this specific intercellular cross‐talk directly dictates autophagic flux and tumor progression. The interpretation of this study must be viewed in the context of several inherent limitations. First, due to the extreme clinical rarity of HAML, our study was constrained by a small sample cohort. Consequently, detailed clinicopathological correlations—such as the patients' systemic TSC genetic status, long‐term prognostic outcomes, and comprehensive clinical history—could not be fully stratified, limiting our ability to link specific novel PEComa subpopulations directly to clinical aggressiveness. Second, the technical challenges in processing fibrotic or lipid‐rich tissues may have resulted in preferential loss of certain vulnerable stromal or neoplastic cells during dissociation, necessitating the integration of ST data to partially correct this bias. Most importantly, the lack of immortalized, validated human HAML cell lines restricts the current feasibility of molecularly perturbing the implicated transcription factors and the PDGFD–PDGFRB axis. In conclusion, by characterizing HAML at a single‐nucleus and spatial resolution, this study maps the intra‐tumoral heterogeneity of PEComas and their specialized microenvironmental interactions. Although currently limited strictly to computational inference and correlative spatial data, it suggests an auxiliary pathogenic mode—linking putative endothelial derivation, Ras signaling, and paradoxical autophagic activation—that complements the traditional TSC/mTOR paradigm (Figure 7). Ultimately, validating these compensatory microenvironmental mechanisms may pave the way for dual‐targeted therapeutic strategies for HAML patients resistant to standard mTOR inhibition. Spatial multi‐omics mapping of hepatic angiomyolipoma reveals that PEComas tightly co‐localize with endothelial cells, which drive autophagic flux in adjacent tumor cells via a paracrine PDGFD–PDGFRB–Ras axis, establishing a compensatory survival network that operates in parallel to canonical TSC–mTOR defects. (This figure is used for the Graphical Table of Contents only). Author Contributions Xinyang Liu: writing – original draft, methodology, formal analysis, conceptualization. Yang Guo: visualization. Xuehong Tie: methodology. Yu Liu: validation. Fuyu Li: validation. Jingkai Liu: methodology. Xiaohang Sun: methodology. Jiajun Zhao: validation. Feng Geng: writing – review and editing. Yingjian Liang: funding acquisition, writing – review and editing. Funding This work was supported by National Natural Science Foundation of China, 61372142. Ethics Statement The study protocol was reviewed and approved by the research ethics committee of the First Affiliated Hospital of Harbin Medical University (Harbin, China) (Approval No. IRB‐AF/SC‐04/02.0). Consent Written informed consent was obtained from all patients prior to specimen collection in accordance with the Declaration of Helsinki. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Data S1: Supporting Information and Methods. Figure S1: Single‐nucleus RNA sequencing quality control metrics and cell‐type annotations. Figure S2: Spatial transcriptomics depth, unsupervised spatial clustering, and protein‐level validation. Figure S3: Intratumoral heterogeneity, marker gene identification, and cross‐organ comparative analysis with RAML and LAM. Figure S4: Endothelial‐mesenchymal transitional marker profiles and AMPK/PPARG activation. Figure S5: Copy number variation (CNV) landscapes and spatially resolved tumor‐stroma communication pathways. Table S1: Clinicopathological characteristics of the enrolled HAML patients. Table S2: Spatial cluster‐specific marker gene signatures for specimen HAML#1. Table S3: Spatial cluster‐specific marker gene signatures for specimen HAML#2. Table S4: Spatial cluster‐specific marker gene signatures for specimen HAML#3. Table S5: Canonical diagnostic and newly identified marker genes of PEComas. Table S6: Statistical overlap analysis between PEComa signatures and human CellMarker databases. Table S7: Significant ligand‐receptor interacting pairs inferred across distinct cell clusters by CellChat. Acknowledgments The authors have nothing to report. Liu X., Guo Y., Tie X., et al., “Mapping the Tissue Architecture and Endothelial‐PEC Network in HAML Using Single‐Cell and Spatial Transcriptomics,” Cancer Science (2026): 1–13, 10.1111/cas.70550. Lead Contact: Yingjian Liang. Contributor Information Feng Geng, Email: gengfeng@hrbmu.edu.cn. Yingjian Liang, Email: yingjianliang@hrbmu.edu.cn. Data Availability Statement The datasets generated and/or analyzed during the current study are available in the NCBI Gene Expression Omnibus (GEO) repository and can be accessed via the accession number GSE344843. All other supporting data or custom analytical scripts are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Data S1: Supporting Information and Methods. Figure S1: Single‐nucleus RNA sequencing quality control metrics and cell‐type annotations. Figure S2: Spatial transcriptomics depth, unsupervised spatial clustering, and protein‐level validation. Figure S3: Intratumoral heterogeneity, marker gene identification, and cross‐organ comparative analysis with RAML and LAM. Figure S4: Endothelial‐mesenchymal transitional marker profiles and AMPK/PPARG activation. Figure S5: Copy number variation (CNV) landscapes and spatially resolved tumor‐stroma communication pathways. Table S1: Clinicopathological characteristics of the enrolled HAML patients. Table S2: Spatial cluster‐specific marker gene signatures for specimen HAML#1. Table S3: Spatial cluster‐specific marker gene signatures for specimen HAML#2. Table S4: Spatial cluster‐specific marker gene signatures for specimen HAML#3. Table S5: Canonical diagnostic and newly identified marker genes of PEComas. Table S6: Statistical overlap analysis between PEComa signatures and human CellMarker databases. Table S7: Significant ligand‐receptor interacting pairs inferred across distinct cell clusters by CellChat. Data Availability Statement The datasets generated and/or analyzed during the current study are available in the NCBI Gene Expression Omnibus (GEO) repository and can be accessed via the accession number GSE344843. All other supporting data or custom analytical scripts are available from the corresponding author upon reasonable request.