Integrative spatiotemporal transcriptomics identifies a liver metastasis-related initial cell population associated with the SEMA3A-NRP1 axis in pancreatic ductal adenocarcinoma

Abstract

Background

Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal malignancy in which liver metastasis represents the principal determinant of poor prognosis. Although metastatic dissemination is thought to be driven by highly plastic tumor cells, the transcriptional features of liver metastasis–related initial cell (LMIC) and its spatial crosstalk with the metastatic microenvironment during PDAC progression remain incompletely defined.

Methods

We integrated bulk transcriptomic, single-cell RNA sequencing, and 10x Genomics spatial transcriptomic datasets from multiple clinical cohorts. Malignant epithelial cells were accurately identified using a machine learning–based approach (scMalignantFinder). Developmental trajectory inference algorithms (VECTOR, scTour, and Monocle2) were applied to define LMIC. Intercellular communication was interrogated using CellChat and CellPhoneDB, while spatial ecological niches were resolved using RCTD and MISTy. Functional validation of key signaling axes was performed by immunohistochemistry (IHC), wound-healing assays, and Transwell invasion assays.

Results

We identified a TSPAN1_MC2 subpopulation of malignant epithelial cells with high differentiation potential and further characterized LMIC located at the origin of the developmental trajectory, characterized by concurrent epithelial–mesenchymal transition (EMT) and stemness features. The LMIC population exhibited elevated expression of stemness-associated genes, including NRP1 and NRP2, and was significantly associated with poor clinical outcomes and resistance to immunotherapy. Cell–cell communication and spatial analyses revealed that cancer-associated fibroblasts (CAFs) secrete SEMA3A, which engages NRP1 on LMIC, mediating their spatial co-localization and functional reprogramming. In vitro assays supported that over-expression of SEMA3A and NRP1 could enhance PDAC cell migration and invasion. The transcription factor YY1 was nominated as a potential regulatory associated with the LMIC transcriptional program, showing structural alignment with NRP1 expression and the activation of pathways such as WNT and PI3K–AKT signaling. In addition, the LMIC phenotype displayed pronounced lipid metabolic reprogramming, with up-regulation of rate-limiting enzymes including FASN and ACACA. Pharmacogenomic simulations further suggested a potential correlation with heightened sensitivity of the LMIC population to FAK inhibitors (PF-562271) and HDAC inhibitors (Mocetinostat), suggesting hypothetical therapeutic vulnerabilities.

Conclusions

At single-cell resolution, this study characterizes the LMIC population in PDAC and implicates a CAF-associated SEMA3A–NRP1 signaling axis within their spatial and functional microenvironmental niches. These findings provide a refined conceptual framework for the development of prospective targeted strategies aimed at disrupting early microenvironmental cross-talk associated with PDAC liver metastasis.

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-026-08451-0.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is among the most lethal solid malignancies, characterized by an insidious onset and late clinical presentation. As a result, the majority of patients are diagnosed at advanced stages and are ineligible for curative resection, contributing to a dismal 5-year overall survival rate of approximately 12% [1]. At the time of diagnosis, nearly 40% of patients present with liver metastasis [2, 3]. More concerningly, even after curative-intent resection of the primary tumor, up to 56% of patients subsequently develop hepatic recurrence [4]. Despite these clinical observations, the molecular mechanisms that govern PDAC cell colonization of the liver and the establishment of a permissive metastatic microenvironment remain incompletely understood. This critical knowledge gap continues to hinder the development of effective, targeted therapeutic strategies for metastatic PDAC.

Metastasis is a multi-step and inherently inefficient cascade, increasingly recognized to be driven by a rare subset of tumor cells characterized as metastasis-related initial cells (MICs), which exhibit pronounced developmental plasticity, self-renewal capacity, and invasive potential [5]. During metastatic colonization, these malignant clones undergo dynamic epithelial–mesenchymal transition (EMT) and extensive metabolic reprogramming, enabling adaptation to the distinct biochemical and biophysical constraints of distant organs [6, 7]. Concurrently, stromal components within the target organ actively participate in extracellular matrix remodeling and engage in complex paracrine signaling with tumor cells, thereby establishing a supportive metastatic niche [8]. However, in the context of PDAC progression, the precise transcriptional identity of liver metastasis–related initial cells (LMICs), their spatial communication with stromal populations in the metastatic microenvironment, and the core transcriptional circuits that sustain their stem-like and metabolically active states remain poorly characterized.

A major obstacle to dissecting the metastatic cascade lies in the profound intratumoral heterogeneity of PDAC. Conventional bulk RNA sequencing approaches, constrained by signal averaging, often obscure the molecular features of rare but biologically consequential subpopulations and fail to capture their dynamic evolutionary trajectories. In contrast, scRNA-seq has emerged as a transformative technology for resolving tumor ecosystems at single-cell resolution [9]. By leveraging high-dimensional clustering and developmental trajectory inference algorithms, scRNA-seq enables the identification of rare MIC populations within heterogeneous primary and metastatic lesions, as well as the reconstruction of their transcriptional reprogramming during intravasation, circulation, and distal colonization. Although LMICs have been identified in colorectal cancer, with evidence supporting functional crosstalk with cancer-associated fibroblasts (CAFs) [10], the identity and mechanistic roles of LMICs in PDAC liver metastasis remain to be systematically elucidated. This gap significantly limits our understanding of metastatic evolution in PDAC and impedes the discovery of actionable therapeutic targets.

To address these challenges, we integrated bulk RNA-seq, scRNA-seq, and spatial transcriptomic data from multiple clinical cohorts to construct a high-resolution spatiotemporal atlas of PDAC primary tumors and matched liver metastasis. By applying complementary trajectory inference frameworks and spatial co-localization models, we identified a distinct LMIC subpopulation within malignant epithelial cells, characterized by high malignant potential, differentiation capacity, and dynamic EMT features. Further integration of single-cell communication analyses with multi-view spatial modeling uncovered intricate stromal–tumor interaction networks within the metastatic microenvironment, thereby delineating a pathogenic spatial niche associated with clinical recurrence and potential tumor colonization. Collectively, this study provides a comprehensive, multi-dimensional view of PDAC liver metastasis and nominates candidate therapeutic targets with hypothetical translational relevance for intercepting metastatic progression.

Materials and methods

Data acquisition and preprocessing

Data sources and study design

Transcriptomic datasets related to PDAC, including bulk RNA-seq, scRNA-seq, and ST-RNAseq, were comprehensively collected from multiple public repositories, including TCGA, GEO, ICGC, CPTAC, and ArrayExpress. Detailed information on all datasets and their corresponding clinicopathological characteristics is summarized in Supplementary Table 1. In addition, to eliminate potential confounding artifacts driven by therapeutic interventions on malignant cell plasticity and microenvironmental compositions, a strict clinical exclusion criterion was applied: all samples from patients who underwent post-chemotherapy, radiotherapy, or neo-adjuvant treatments were systematically excluded. Consequently, 100% of the single-cell and spatial transcriptomic cohorts analyzed in this study were derived exclusively from treatment-naive PDAC specimens.

Bulk transcriptomic data preprocessing

For bulk transcriptomic analyses, a total of 13 independent cohorts were included (TCGA-PAAD, E-MTAB-6134, GSE21501, GSE28735, GSE57495, GSE62452, GSE71729, GSE78229, GSE79668, GSE85916, CPTAC-PDAC, ICGC-AU-Array, ICGC-AU-Seq, and ICGC-CA-Seq), comprising 2,136 samples, including 1,868 PDAC tumor tissues and 268 adjacent or normal pancreatic tissues. Corresponding clinical data were collected for downstream prognostic analyses. For microarray datasets, standardized preprocessing procedures were applied. Probe identifiers were mapped to official gene symbols according to the corresponding platform annotation files. When multiple probes mapped to the same gene, the average expression value was used to represent gene-level expression. Expression matrices were subsequently normalized using log2 transformation [log2(expression + 1)]. To mitigate technical biases arising from cross-platform integration and multi-cohort heterogeneity, batch effects were corrected using the removeBatchEffect() function implemented in the limma package (v3.62.1) [11]. For sequencing datasets, we converted the raw count matrix to transcripts per million (TPM) to ensure standardized comparisons between samples. Subsequently, a Log2 transformation log2(TPM + 1) was applied to the TPM matrix to stabilize the variance and optimize the data distribution for downstream statistical analysis.

Single-cell transcriptomic data preprocessing

To characterize the tumor microenvironment at single-cell resolution, raw count matrices from 54 samples were obtained from four independent public datasets (GSE205013, GSE111672, GSE197177, and CRA001160), encompassing primary PDAC, liver metastatic PDAC, and normal pancreatic tissues. Rigorous quality control was performed using the Seurat package (v5.3.0) [12]. First, potential doublets were identified and removed using scDblFinder (v1.23.4) with default parameters [13]. Cells were further filtered based on the following criteria: (i) retaining cells with 200–5,000 detected genes to exclude low-quality droplets and multiplets; (ii) excluding cells with >10% mitochondrial gene expression, indicative of stressed or dying cells; and (iii) removing cells with >5% hemoglobin-related gene expression or >5% ribosomal gene expression to minimize technical noise. Following quality control, normalization was performed using the NormalizeData() function in Seurat with the default log-normalization method (normalization.method = “LogNormalize”, scale factor = 10,000). Highly variable genes (n = 3,000) were identified using variance-stabilizing transformation. After centering and scaling the data, principal component analysis (PCA) was conducted based on these features. To determine the optimal number of principal components (PCs), we implemented a dynamic selection strategy that integrates multiple criteria: cumulative variance explained (>90%), individual PC contribution (<5%), and identification of the inflection point in the variance decay curve. The minimum number of PCs satisfying these conditions was retained for downstream analyses. To correct for batch effects arising from multiple datasets, the Harmony algorithm (v1.2.3) was applied in the PCA space [14], using sample identity as a covariate and a maximum of 20 iterations to ensure convergence. Based on the Harmony-corrected embeddings, a shared nearest neighbor (SNN) graph was constructed using FindNeighbors(), followed by Louvain clustering via FindClusters() with a resolution parameter of 1.2, which was empirically determined to achieve biologically meaningful cluster granularity. Finally, uniform manifold approximation and projection (UMAP) was applied for nonlinear dimensionality reduction and visualization.

Spatial transcriptomic data preprocessing

For spatial transcriptomic analyses, 10x Genomics Visium datasets of primary and liver metastatic PDAC were obtained from GSE274557 and GSE235315. Spatial expression matrices were loaded using the Load10X_Spatial() function in Seurat, enabling integration of gene expression profiles with corresponding spatial coordinates. To account for sequencing depth variation while preserving spatial biological heterogeneity, normalization and variance stabilization were performed using the SCTransform() method based on a regularized negative binomial model. PCA was subsequently conducted on highly variable genes, and the top 30 principal components were used to construct the SNN graph. Unsupervised clustering was performed using FindClusters(), followed by UMAP for low-dimensional visualization of spatial expression patterns.

Differential expression and functional enrichment analysis

Identification of differentially expressed genes (DEGs)

For bulk RNA-seq data, differential expression analysis was conducted using the R package limma. Differentially expressed genes (DEGs) were identified based on strict thresholds of |log2(fold change)| >0.585 and an adjusted p < 0.05. For scRNA-seq data, cluster-specific marker genes were identified using the FindAllMarkers function in the Seurat package, utilizing the Wilcoxon rank-sum test. Genes with an average log2(fold change) >0.585 and an adjusted p < 0.05 were considered significant DEGs and retained for downstream functional annotation.

Gene set variation analysis and pathway scoring

To quantify pathway activation at single-cell resolution, the irGSEA package (v3.3.2) was employed [15]. Specifically, the AUCell algorithm (seed = 123) was used to calculate enrichment scores for individual cells across the Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and HALLMARK gene sets. For bulk RNA-seq cohorts, gene set variation analysis was performed using the GSVA package (v2.2.0) with default parameters to derive enrichment scores for each independent sample [16].

Biological function and pathway enrichment analysis

To elucidate the biological functions and signaling pathways associated with the identified DEGs, both over-representation analysis (ORA) and gene set enrichment analysis (GSEA) were performed using the GO, KEGG, and HALLMARK databases. Pathway enrichment calculations and statistical evaluations were conducted utilizing the clusterProfiler package (v4.16.0) [17]. High-quality, customized visualizations of the significant enrichment results were generated using ggplot2 (v3.5.2).

Identification of malignant cells and genomic variation analysis

Identification of malignant epithelial cells

To accurately distinguish between normal and malignant epithelial cell populations, we utilized the scMalignantFinder software (v1.1.9) [18], a machine-learning-based framework designed for single-cell classification. This tool leverages prior features derived from large-scale pan-cancer single-cell reference datasets to infer the malignancy probability of target cells. Using the normalized expression matrix as input, we performed model fitting and prediction via a logistic regression algorithm integrated with an internal pre-trained dataset. Based on the algorithm-derived prediction thresholds, all epithelial cells were strictly categorized into either normal or malignant epithelial cell groups.

Inference and quantitative scoring of single-cell copy number variations

To validate the genomic alteration landscape of the identified malignant epithelial cells, we inferred large-scale chromosomal copy number variations from the scRNA-seq data using the infercnv package (v1.22.0) (https://github.com/broadinstitute/infercnv). Given their stable diploid genomic background within the tumor microenvironment, plasma cells were selected as the normal reference population. For the implementation, the denoising mode was enabled, and the Hidden Markov Model (HMM) prediction was deactivated. Malignant cells were then subjected to hierarchical clustering using the “ward.D2” method. To further quantify the overall burden of copy number variations at the individual cell level, we developed a customized quantitative scoring strategy (TotalCNV score) based on the expression distribution of the reference population. First, the mean expression and standard deviation (SD) of the plasma cell reference group were calculated. A normal diploid range was defined as the mean ±2 SD (assigned a variation weight of 0). Loci deviating from this baseline were categorized and weighted as follows: partial loss or amplification (weight = 1), and complete loss or extreme amplification (weight = 2). Finally, the total CNV score for each cell was calculated by aggregating the absolute weight scores across all genomic loci. This continuous score was integrated as metadata into the Seurat object to quantitatively represent the degree of single-cell genomic instability.

Inference of cell differentiation trajectories

Assessment of developmental potency and differentiation direction

To systematically characterize the differentiation hierarchy and cellular plasticity of malignant epithelial subclusters, developmental potency scores were quantified using the CytoTRACE2 package (v1.0.0) [19] with the species parameter set to “human.” This algorithm integrates single-cell mRNA abundance and distribution patterns to yield a potency index ranging from 0 to 1, where higher scores reflect a more primordial, stem-like cellular state.

Streamlined trajectory inference using complementary frameworks

To ensure maximum robustness and mitigate algorithm-specific biases in tracing the developmental trajectory of LMIC, we implemented a multi-algorithmic validation framework using three independent computational platforms without predefining a developmental root:

VECTOR (v0.0.4): Applied to infer directional transcriptional flow grids and vector networks within the UMAP embedding space based on automated center-finding coordinates, thereby objectively mapping potential state transition peak points.

scTour (v0.1.2): Introduced as a deep learning alternative [20]. Trained on 3000 highly variable genes (HVGs), scTour combines a variational autoencoder (VAE) with ordinary differential equations (ODEs) to simultaneously resolve pseudotime vector fields within a high-dimensional latent space.

Monocle2 (v2.36.0): Utilized to reconstruct branching differentiation topologies via the Discriminative Dimensionality Reduction with Trees (DDRTree) algorithm [21]. Under a negative binomial distribution framework, feature-selected HVGs were utilized to dynamically position cellular subclusters and evaluate their lineage-transition characteristics along the pseudotime axis using the orderCells() function.

Inference of inter-cellular communication

Reconstruction of communication networks based on CellChat

To systematically quantify the interaction probabilities between LMIC and other stromal or immune cell types, we performed inter-cellular communication network inference using the CellChat package (v2.1.2) [22]. CellChat models the strength of cellular communication based on the law of mass action by integrating gene expression levels with prior ligand-receptor interaction knowledge. For this analysis, we utilized the CellChatDB.human database, which covers three major signaling categories: Secreted Signaling, ECM-Receptor, and Cell-Cell Contact. To address the inherent sparsity of single-cell sequencing data, the smoothData() function was applied to smooth the expression profiles by incorporating a human protein-protein interaction network. Signaling centrality, specifically incoming and outgoing flow, for each cell subcluster was calculated using the netAnalysis_computeCentrality() function. The contribution of specific communication patterns was visualized using hierarchical clustering heatmaps, bubble plots, and chord diagrams.

Statistical inference and validation via CellPhoneDB

To further validate key communication events, we applied the CellPhoneDB algorithm (v5.0.0) [23]. This method constructs a null distribution by performing 1000 random permutations of cell-type labels to infer the statistical significance of ligand-receptor pair expression within specific cell-cell dyads. Interactions were considered significant if the p value was < 0.05 and the corresponding genes were expressed in at least 10% of the cells within the cluster. A comprehensive communication landscape was established by calculating the total number and strength of interactions between LMIC and the tumor microenvironment. For candidate signaling pairs, fine-grained visualizations of spatial co-expression trends and interaction networks were generated using the plot_cpdb() and plot_cpdb4() functions from the ktplots package (v2.4.3) [23].

Bayesian deconvolution and clinical prognostic validation

Deconvolution of bulk transcriptomes using BayesPrism

To evaluate the infiltration characteristics of single-cell-identified subpopulations, such as LMIC, within large-scale clinical cohorts, we performed in silico deconvolution using the BayesPrism (v2.1.1) algorithm, which is based on a joint Bayesian model [24]. First, a high-resolution reference map was constructed using our finely annotated scRNA-seq expression matrix. To ensure the robustness of the reference and mitigate potential biases arising from unbalanced cell-type distributions, we performed random downsampling, retaining 500 cells per cell type. Prior to model construction, we employed the cleanup.genes() function to strictly filter feature genes, excluding mitochondrial genes (MT-), ribosomal genes (RPS/RPL), sex chromosome-linked genes (chrX/Y), and non-coding RNA interference. Furthermore, the select.gene.type() function was utilized to restrict the analysis to protein-coding genes to enhance the signal-to-noise ratio. The deconvolution object was initialized using new.Prism() with an outlier truncation parameter (outlier.cut = 0.01). Finally, absolute abundance scores for each cell subpopulation across the TCGA-PAAD cohort were inferred using the run.Prism() function in multi-threading mode.

Prognostic evaluation and risk assessment

Following the inference of cellular infiltration abundances, these scores were integrated with patient clinical follow-up data to assess the prognostic value of specific subclusters. For prognostic stratification, the survminer package (v0.5.0) was used to determine the optimal cutoff values for categorizing patients into high- and low-infiltration groups. Survival assessment was conducted using the survival package (v3.8–3). Kaplan-Meier curves were generated, and the statistical significance of differences in overall survival (OS) and progression-free survival (PFS) between groups was evaluated using the log-rank test. To quantify the independent predictive power of specific cellular abundances, we employed a univariate Cox proportional hazards regression model to calculate the Hazard Ratio (HR) and the 95% confidence interval (95% CI).

Inference of gene regulatory networks and functional essentiality analysis

Transcription factor (TF) activity inference

To resolve the core regulatory mechanisms driving specific cellular lineage evolution, we systematically inferred TF activities at single-cell resolution using the DoRothEA (v1.18.0) framework [25, 26]. DoRothEA constructs high-confidence TF-target gene interaction networks (regulons) by integrating manually curated databases, ChIP-seq evidence, and transcription factor binding motif predictions. In this framework, the biological activity of a TF is quantified as a weighted enrichment score based on the transcriptional states of its direct downstream targets. Specifically, we filtered the dorothea_regulon_human library to include only regulons with confidence levels A (highest) and B (high). TF activity scores for individual cells were then calculated using the VIPER (Virtual Inference of Protein-activity by Enriched Regulon analysis) algorithm. The resulting VIPER scores were centered, scaled, and integrated into the Seurat object. To identify lineage-specific regulatory features, TFs were ranked based on their variance across subclusters, allowing for the precise identification of master regulators associated with specific cellular markers.

Functional essentiality validation via DepMap

To address the functional limitations of purely transcriptomic-based inferences, we utilized the Cancer Dependency Map (DepMap) portal to access genome-wide CRISPR-Cas9 screening data (25Q3 Release). This analysis aimed to validate the essentiality of core TFs for tumor survival in vitro. Focusing on a panel of pancreatic ductal adenocarcinoma (PDAC)-derived cell lines, we extracted gene dependency and gene effect scores for the identified candidate TFs. The gene effect score reflects the impact of gene knockout on cell proliferation, where more negative values indicate a higher degree of growth inhibition. The dependency score represents the probability that a cell line is dependent on a specific gene for survival. By assessing the dependency profiles of key TFs inferred from our single-cell atlas, we objectively corroborated the critical biological roles of these core regulators in maintaining tumor cell viability and growth.

Single-cell drug sensitivity prediction

Drug response scoring based on transcriptomic signatures

To identify potential therapeutic vulnerabilities and assess the pharmacological sensitivity of specific cell subclusters, we employed the beyondcell framework (v2.2.2) [27]. This algorithm integrates single-cell transcriptomic data with large-scale pharmacogenomics databases to infer drug efficacy at single-cell resolution. Specifically, we utilized the pre-defined pharmacological signature collection (PSc) and calculated enrichment scores for the single-cell expression matrix using the bcScore() function. To mitigate the technical challenges posed by “dropout” effects inherent in scRNA-seq data, a stringent expression threshold was applied (expr.thres = 0.1), ensuring that signature genes were incorporated into the sensitivity scoring only if they were stably expressed in at least 10% of the cells within a cluster.

Quality control and normalization of the drug sensitivity matrix

Following the calculation of initial sensitivity scores, a secondary quality control (QC) and normalization pipeline were implemented to ensure the statistical robustness of downstream comparative analyses. We calculated the effective scoring coverage for each candidate drug across the entire cell population. Drugs with an effective coverage of less than 10% of total cells were excluded from the analysis to minimize stochastic noise and lack of specificity. For the high-confidence candidate drugs retained, a conservative imputation strategy was adopted, where missing values in non-responsive subclusters were assigned a score of 0, representing a biological “baseline” or non-responsive state. Finally, the filtered and normalized drug sensitivity matrix was re-aligned with the Seurat object metadata to facilitate the identification of differentially sensitive drugs across malignant subclusters.

Spatial deconvolution and co-localization analysis

Spatial cell-type decomposition using RCTD

To map single-cell-resolved cell-type labels onto spatial transcriptomics loci, we employed the Robust Cell Type Decomposition (RCTD) algorithm within the spacexr package (v2.2.1) [28]. First, a reference matrix was constructed using the previously annotated scRNA-seq data, excluding rare subclusters with fewer than 25 cells to ensure the robustness of the reference signatures. For deconvolution, the raw count matrix and corresponding tissue coordinates of the ST data were extracted, and an analysis object was initialized using the create.RCTD() function. Given the spot-level resolution characteristic of 10x Visium technology, the algorithm was executed in “doublet mode.” This mode assumes that each spot primarily consists of two major cell types, providing more precise cell proportion estimates than traditional full-mixture models. The resulting proportions were retained for downstream spatial association analyses.

Multi-view spatial dependency modeling via MISTy

To quantitatively evaluate the spatial dependencies and potential interaction strengths between cell subclusters within the tumor microenvironment, we utilized the

MISTy framework (v1.16.0) [29]. Specifically, the cell-type proportion matrix derived from RCTD was used as input features. Based on the tissue coordinates, four spatial views were constructed to capture interactions across different scales: Intraview: To assess the intrinsic cell-type associations within individual spots. Paraview: To model diffusive interactions in the immediate neighborhood using a Gaussian kernel, with the spatial radius defined as the mean distance to the nearest neighbor plus one standard deviation. Exponential and Linear views: To capture long-range spatial signals using exponential decay and linear kernels, respectively, representing different signaling attenuation rates. The multi-view models were executed using the run_misty() function to calculate the contribution (Importance scores) of various “predictors” to the distribution of “target” cell subclusters. We focused on interactions demonstrating a significant R² gain and an importance score. Finally, a network clustering algorithm based on importance weights was applied to identify distinct cellular interaction communities within the spatial niches.

Heterotypic cellular networks and spatial niche identification

Analysis of heterotypic cellular connectivity (CellDegree)

To quantitatively evaluate the physical proximity between different cell types, we developed a spatial connectivity scoring algorithm based on the semla framework (v1.3.2) [30]. First, spatial neighborhoods were constructed by extracting spot centroid coordinates via the Staffli coordinate system. A K-nearest neighbor (KNN) graph was generated based on Euclidean distances, capturing 6 nearest neighbors to perfectly match the intrinsic hexagonal grid geometry layout of the 10x Visium platform. Subsequently, spatial binarization was performed using a proportion-based thresholding approach: a spot was defined as an “active site” for a specific cell type only if its inferred proportion exceeded 0.3. Finally, heterotypic connectivity was calculated using the adjacency matrix. For a spot identified as cell type A, the connectivity score was defined as the number of its 6 immediate spatial neighbors belonging to cell type B. This analysis aimed to identify specific microenvironmental components with which target cells preferentially maintain physical contact.

Spatial niche decomposition via non-negative matrix factorization (NMF)

To identify co-localized functional units (niches) within the tissue, we performed non-negative matrix factorization (NMF) on the cell-type proportion matrix derived from RCTD. Using the RcppML package (v0.3.7), the matrix was decomposed into a basis matrix representing spatial patterns and a coefficient matrix representing cell-type contributions. The optimal number of factors was determined based on the saturation of the reconstruction error and biological interpretability. The resulting factor weights were max-normalized to quantify the contribution of each cell subcluster to specific spatial niches, facilitating the characterization of organized cellular communities across the tumor landscape.

Inference of single-cell metabolic landscapes

Metabolic pathway activity assessment using scMetabolism

To characterize metabolic reprogramming across diverse cell populations at single-cell resolution, we quantified metabolic activities using the scMetabolism package (v0.2.1) [31]. Utilizing the KEGG and REACTOME databases as references, the VISION algorithm was employed as the underlying enrichment engine. To ensure data integrity, scores were calculated under stringent conditions without data imputation (imputation = FALSE). Core metabolic pathways significantly upregulated in LMIC compared to primary tumor cells, metastatic malignant cells, and normal epithelial cells were identified based on a threshold of fold change > 1. Differential metabolic signatures were visualized using hierarchical clustering heatmaps.

Metabolic flux dynamics simulation via scFEA

To further resolve the dynamic metabolic requirements of LMIC from a flux-centric perspective, we applied the scFEA (v1.1) framework, which is based on a deep-learning-driven graph neural network [32]. Using the raw single-cell expression matrix as input and the human metabolic flux model (M168 module) as a template, we constructed non-linear kinetic systems to infer the reaction fluxes of 168 metabolic modules and the mass balance of 70 core metabolites. The resulting flux and balance matrices were integrated into the Seurat object, and cell-type-specific flux markers were identified using the FindAllMarkers() function. We prioritized metabolic reactions with significant flux gain in LMIC (average log2FC > 0.585 and adjusted p < 0.05). Heatmaps were generated to illustrate the specific metabolic advantages of LMIC in the uptake and transformation of key compounds, such as lipid precursors and amino acid derivatives.

Cell migration and invasion assays

Wound healing assay

To evaluate horizontal cell migration capacity, transfected AsPC-1 cells (categorized into NC, OE-NRP1, and OE-SEMA3A groups) were seeded into 6-well plates. Upon reaching approximately 100% confluence, a vertical scratch was made on the cell monolayer using a 200 µL pipette tip. The cells were washed three times with PBS to remove debris and subsequently cultured in low-serum medium containing 1% FBS. The denuded area was observed and photographed at 0, 6, 24, 48, 72, and 96 hours post-scratching using an inverted microscope. The wound width was measured using ImageJ software, and the healing rate and migration distance were calculated to quantify the migration velocity.

Transwell invasion assay

The in vitro invasive capacity of cells was assessed using Transwell chambers with 8 µm pore size filters. Prior to the experiment, the lower surface of the upper chamber was pre-coated with diluted Matrigel to simulate the extracellular matrix environment. Harvested cells from each treatment group were resuspended in serum-free medium and seeded into the upper chamber (approximately cells/well). The lower chamber was filled with 600 µL of complete medium containing 10% FBS to serve as a chemoattractant. After 24 hours of incubation, non-invaded cells on the upper surface of the membrane were removed. The invaded cells on the lower surface were fixed with 4% paraformaldehyde for 20 minutes and stained with 0.1% crystal violet for 15 minutes.

Statistical analysis

All computational analyses, bioinformatic modeling, and statistical inferences in this study were conducted using R (v4.5.1) and Python (v3.10). For comparisons of continuous variables between groups, the non-parametric Wilcoxon rank-sum test was employed. Associations between genes, cell-type abundances, or metabolite balances were evaluated using Spearman’s rank correlation coefficients to capture potential non-linear regulatory relationships. All statistical tests were two-sided. A p value < 0.05 was considered statistically significant and is indicated in the figures as: p < 0.05*, p < 0.01**, and p < 0.001***.

Results

Single-cell transcriptomic landscape reveals cellular heterogeneity and microenvironmental remodeling in primary and metastatic PDAC

To systematically investigate cellular heterogeneity and microenvironmental evolution during the liver metastasis of PDAC, we performed scRNA-seq on paired primary PDAC tumors and their corresponding liver metastasis. After stringent quality control and doublet removal (Fig. 1A and Supplementary Fig. 1A–C), we successfully identified 13 major cell lineages based on the expression patterns of canonical marker genes (Fig. 1B). Compositional analysis revealed striking differences in the cellular architecture between primary and metastatic sites (Fig. 1C, D). Notably, proliferative epithelial cells, monocytes/macrophages, and endothelial cells were significantly expanded in liver metastasis, suggesting a highly proliferative niche and profound immune remodeling. Conversely, neutrophils, B cells, mast cells, and fibroblasts were predominantly enriched in the primary tumors. Each cell lineage exhibited highly specific marker gene profiles—such as MKI67 for proliferative cells and VWF for endothelial cells—further validating the fidelity of our cellular annotations (Fig. 1E, F). Utilizing the scMalignantFinder algorithm, we abandoned a purely cluster-based approach to precisely distinguish malignant epithelial cells from their normal counterparts (Fig. 1G). Malignant cells were markedly enriched in liver metastasis, whereas normal epithelial cells were primarily confined to the primary tissue (Fig. 1H), indicating that the metastatic liver environment is dominated by expanded malignant clones. Functional characterization further demonstrated that malignant cells derived from liver metastasis exhibited significantly elevated scores for EMT, Total CNV, and upregulated stemness features (Stemness_UP) compared to those from primary sites (Fig. 1I). We subsequently observed that ANXA1, KRT7, and LDHB were significantly overexpressed in liver metastasis (Fig. 1J). The up-regulation of ANXA1 is closely associated with enhanced cell motility and EMT initiation, while high KRT7 expression, a key marker of ductal malignancy, reflects a superior capacity for tumor progression. Furthermore, the enrichment of LDHB suggests that malignant cells undergo metabolic reprogramming to adapt to the hepatic metabolic niche, providing bioenergetic support for metastatic colonization. To evaluate the intra-tumoral heterogeneity of the malignant population, we employed the ROGUE purity metric. The malignant cells yielded a ROGUE score of approximately 0.8 (Fig. 1K), notably lower than the benchmark for highly homogeneous populations (0.9), revealing substantial transcriptomic diversity within the malignant compartment. Collectively, these findings provide a comprehensive landscape of cellular heterogeneity across PDAC metastatic organs and identify a subset of metastasis-associated malignant cells characterized by high metabolic activity and invasive potential.

Fig. 1.

Single-cell transcriptomic landscape of primary PDAC and paired liver metastasis reveals intratumoral heterogeneity. Schematic of experimental design and clinical sample collection. (B) UMAP plot showing cell clusters from primary PDAC and paired liver metastasis, identifying 13 major cell lineages. (C) stacked bar plot showing the cellular composition across primary and metastatic sites. (D) Heatmap demonstrating the tissue preference (ROI/Tissue preference) of various cell lineages. (E) Heatmap of lineage-specific marker genes. (F) UMAP feature plots showing expression patterns of canonical markers for each lineage (MKI67, VWF, TPSAB1, etc.). (G) sub-clustering of epithelial cells, colored by cell identity (normal vs. Malignant). (H) Heatmap comparing the distribution of normal and malignant epithelial cells. (I) raincloud plots comparing EMT scores, total CNV burden, and Stemness_up scores between primary and metastatic malignant cells. (J) violin plots showing expression heterogeneity of functional genes (ANXA1, KRT7, and LDHB). (K) boxplot comparing transcriptomic purity (ROGUE score) between malignant and normal epithelial populations

Identification of a pro-metastatic malignant subcluster associated with poor clinical prognosis

To resolve the observed transcriptomic heterogeneity within the malignant compartment, we performed high-resolution re-clustering, which identified four functionally distinct subclusters. Based on their respective signature genes, these were designated as SRGN_MC1, TSPAN1_MC2, AKR1C2_MC3, and SERPINA3_MC4 (Fig. 2A and Supplementary Table 2). Correlation analysis confirmed significant divergence in their global expression profiles (Supplementary Fig. 2B). Notably, the TSPAN1_MC2 subcluster exhibited unique biological properties; it comprised cells derived from both primary and metastatic sites and demonstrated superior developmental potency (Fig. 2B, C and Supplementary Fig. 2A). To further elucidate the evolutionary hierarchy and lineage dynamics within the malignant population, we performed pseudotemporal trajectory analysis using the Monocle2 algorithm (Fig. 2D). The results showed that TSPAN1_MC2 cells were distributed across all stages of the trajectory and were significantly enriched at key lineage bifurcations. This pattern suggests that TSPAN1_MC2 represents a highly dynamic evolutionary state, potentially serving as an evolutionary intermediate linked with prospective cross-organ dissemination themes. Furthermore, this subcluster displayed remarkable transcriptional plasticity and stemness features, characterized by the specific enrichment of EGFR, hypoxia, and MAPK signaling pathways, alongside high expression of key stemness regulators such as SOX2 and BMP4 (Fig. 2E). Remarkably, TSPAN1_MC2 specifically overexpressed genes encoding the secretable proteins TFF3 and LGALS4, a trend not observed in other subclusters (Fig. 2G). In gastrointestinal malignancies, TFF3 is traditionally associated with mucosal repair and enhanced epithelial motility, with its upregulation often predicting increased aggressiveness. Meanwhile, LGALS4 is involved in the modulation of cell adhesion and polarity. The specific high expression of these genes not only serves as a robust marker for the invasive TSPAN1_MC2 state but also implies that this cluster might leverage these secretable factors to modulate local physical properties during distant dissemination. Clinical relevance analysis further substantiated the pathogenic potential of this subcluster (Fig. 2F, H and Supplementary Fig. 2C, D). In the TCGA cohort, patients with higher TSPAN1_MC2 signature scores exhibited significantly worse prognosis, showing a positive correlation with clinical metastasis. Additionally, the proportion of this subcluster was significantly elevated in elderly patients (>65 years), suggesting a potential association with prolonged tumor evolution and immune evasion. To refine the functional characterization of TSPAN1_MC2, we performed ORA and GSEA on its differentially expressed genes (Fig. 2I–K). These analyses revealed significant activation of the EGFR pathway, cell-cell adhesion, and pancreatic development pathways. Crucially, signaling pathways related to cell migration, locomotion, and the regulation of localization were significantly positively enriched in TSPAN1_MC2 (NES > 1, adjusted p < 0.05). Collectively, these findings identify TSPAN1_MC2 as a malignant subcluster with high developmental potency, robust stemness features, and significant prognostic value. The highly motile and invasive molecular signature of this subcluster positions it as a critical candidate population driving liver metastasis in PDAC.

Fig. 2.

High-resolution sub-clustering of malignant cells delineates functional evolution and clinical relevance. UMAP plots of malignant epithelial sub-clustering, colored by subcluster (left) and tissue origin (right). (B) developmental potency and evolutionary trends inferred by the VECTOR algorithm. (C) UMAP plot showing differentiation maturity scores calculated by CytoTRACE2. (D) pseudotemporal trajectory analysis using monocle 2, showing cell distribution along pseudotime (top) and the positioning of the four malignant subclusters (bottom). (E) Heatmap of stemness markers and AUCell activity scores for stem cell-related pathways. (F) kaplan-meier curves evaluating the impact of TSPAN1_MC2 abundance on patient OS and PFS. (G) violin plots of signature genes TFF3 and LGALS4 across subclusters. (H) ridge plots showing the clinical correlation of TSPAN1_MC2 enrichment with M-stage and age. (I) Heatmap of marker genes defining the four malignant subclusters. (J) GO enrichment analysis of biological processes for each subcluster. (K) GSEA plots revealing the upregulation of invasion-related pathways in TSPAN1_MC2

Spatial transcriptomics reveals the distinct niches and cellular proximities of the pro-metastatic subcluster

Given the prominent malignant features of the TSPAN1_MC2 cluster in both primary and metastatic sites, we employed ST-RNAseq to investigate its organizational patterns within the tissue microenvironment. Spatial interaction network analysis identified TSPAN1_MC2 as a central hub of inter-cellular communication across both tissue types, exhibiting high spatial co-localization with endothelial cells and fibroblasts (Fig. 3A, B, E, F). Utilizing heterotypic cellular network association analysis, we further quantified the robust physical proximity between TSPAN1_MC2 and mesenchymal components (Fig. 3C, G). Notably, compared to primary tumors, the co-localization clusters formed by TSPAN1_MC2 with fibroblasts and endothelial cells were more dense and contiguous in liver metastasis. This topological enrichment coincides with extensive stromal cell remodeling and localized structural reorganization within distant organs, indicating the establishment of a permissive microenvironmental niche for metastatic seeding and sub-clonal growth. Subsequent tissue domain segmentation revealed that TSPAN1_MC2 was significantly enriched within the tumor core regions of both primary and metastatic lesions (Fig. 3D, H). To further resolve the functional attributes of these spatial niches, we performed non-negative matrix factorization (NMF) (Fig. 3I, J). In liver metastasis, the “Factor 2” niche—characterized by the co-localization of TSPAN1_MC2 and fibroblasts—was significantly enriched in biological processes related to cell-cell adhesion and junctions. This suggests high physical integrity within this niche, potentially facilitating collective cell migration or providing defense against shear stress. Additionally, the enrichment of chemokine receptor binding implies that this area serves as an active center for immune modulation and cellular recruitment. In contrast, the corresponding “Factor 1” niche in primary tumors exhibited a distinct desmoplastic signature, primarily enriched in extracellular matrix (ECM) reorganization, reflecting the typical collagen-dense microenvironment of PDAC, and blood vessel development, which provides essential blood supply for tumor growth. Collectively, these spatial transcriptomic findings delineate the organizational characteristics and functional evolution of the TSPAN1_MC2 subcluster during cross-organ progression. Within the primary site, this cluster is closely aligned with gene expression signatures governing interstitial fibrosis and extracellular matrix deposition to support local tissue expansion. In liver metastasis, its function repertoire shifts toward strengthening cell adhesion and chemotactic interactions to optimize metastatic colonization. This dynamic adaptation of the spatial niche highlights a core mechanism by which TSPAN1_MC2 drives the transition of PDAC from local progression to distant metastasis.

Fig. 3.

Spatial transcriptomics resolves the spatial architecture and microenvironmental crosstalk of TSPAN1_MC2. (A, E) dot plots showing spatial co-localization probabilities in liver metastasis (A) and primary tumors (E). (B, F) visualization of cellular interaction networks in metastatic (B) and primary (F) tissues. (C, G) heterotypic network analysis demonstrating the spatial proximity of TSPAN1_MC2 with CAFs and endothelial cells. (D, H) boxplots comparing TSPAN1_MC2 enrichment between malignant and adjacent normal domains. (I, J) spatial niche analysis using NMF, showing cellular composition (left) and KEGG enrichment of niche-specific genes (right)

Evolutionary trajectory analysis defines liver metastasis-related initial cell

To resolve the developmental dynamics of the aggressive TSPAN1_MC2 subcluster, we inferred differentiation trajectories using three independent computational frameworks: VECTOR, scTour, and Monocle. These algorithms yielded highly convergent results, consistently positioning Cluster 1 at the developmental root of the malignant cells (Fig. 4A, Supplementary Fig. 3A-E). Accordingly, we defined the liver-metastatic cells within Cluster 1 as liver metastasis-related initial cell (LMIC) (Fig. 4B). Biologically, the positioning of LMIC at the trajectory apex reflects their superior developmental plasticity and primordial status as the clonal seeds for hepatic colonization. Functional characterization revealed that LMIC are defined by the concurrent activation of EMT, ECM remodeling, and focal adhesion, coupled with a downregulation of oxidative phosphorylation (Fig. 4C). Notably, LMIC undergoes complex molecular remodeling along the differentiation trajectory. The trajectory root (LMIC state) is characterized by the enrichment of gene modules governing stemness maintenance (NRP1/2, HMGA2), metabolic rewiring (FASN), and ECM sensing (FN1, THBS1, TNC). Progressing toward the trajectory terminus, a prominent upregulation of mature metastatic markers (TFF3, LGALS4, AGR3) and phenotypic evolution drivers (CDCA7, VIM) is observed (Fig. 4D). In addition, compared with other cells, LMIC significantly upregulates the scores of Invasion, Metastasis, and CytoTRACE2 Score (Fig. 4E). GSEA further illustrated that LMIC are highly active in spatial sensing and signal transduction, with marked upregulation of ECM-receptor interactions, focal adhesions, and oncogenic signaling pathways. Conversely, protein synthesis and canonical glandular secretory functions were significantly suppressed (Fig. 4F, J), indicating a shift toward a dedifferentiated malignant state.

Fig. 4.

Identification and phenotypic characterization of LMIC. UMAP plots showing differentiation potency inferred by VECTOR. (B) UMAP plot identifying LMIC and other malignant populations. (C) UMAP feature plots of AUCell scores for oxidative phosphorylation, EMT, focal adhesion, and ECM-receptor interaction. (D) Heatmap of key genes across the LMIC. HES1 and TFF3 increasingand others like NRP1 and HMGA2 decreasing as cells progress through the differentiatio. (E) box plots compared invasion metastasis and CytoTRACE scores between LMIC and other cells. (F) Bar plot of metabolic and signaling pathways enriched in LMIC. (G) venn diagram defining the LMIC signature from the intersection of up-regulated DEGs. (H) correlation between the LMIC signature score and stem cell-related GO terms. (I) correlation of the LMIC signature with CDH1, VIM, and CDCA7. (J) integrated heatmap profiling the functional landscape of LMIC vs. primary malignant and normal epithelial cells

To evaluate the clinical significance of LMIC, we derived an 84-gene LMIC signature (LMICs) (Fig. 4G and Supplementary Table 3). Correlation analysis demonstrated that the LMICs score positively associates with stemness and mesenchymal markers (VIM) while negatively correlating with epithelial markers (CDH1) and proliferation indices (CDCA7) (Fig. 4H, I). This suggests that LMICs score prioritize invasion and survival maintenance over rapid proliferation. High LMICs score served as an independent risk factor for poor prognosis and correlated with advanced clinical stages (Supplementary Fig. 3F). Regarding the immune landscape, LMICs score significantly co-expressed checkpoints such as HAVCR2 and PDCD1LG2 specifically in liver metastasis but not in primary tumors, indicating the induction of an organ-specific immunosuppressive barrier during hepatic colonization (Supplementary Fig. 3 G). TIP scoring and immune infiltration analysis confirmed that high LMICs score positively correlates with the recruitment of immunosuppressive populations, including Tregs, MDSCs, and M2 macrophages (Supplementary Fig. 3 H, K). Furthermore, we acknowledge that PDAC-specific immunotherapy transcriptomic datasets are critically unavailable at present. As an exploratory alternative to generate hypothetical insights, we interrogated multiple cross-cancer ICB cohorts as surrogate, concept-validating models to explore the potential clinical relevance of LMICs score. Across these datasets, non-responders exhibited higher LMICs scores, and high-score patients showed significantly reduced survival following treatment (Supplementary Fig. 3I-J). However, given the distinct tumor microenvironments across different malignancies, extrapolating these cross-cancer findings directly to PDAC is inherently premature. These results should be interpreted strictly as preliminary computational evidence that the LMIC-driven immune niche may associate with therapeutic resistance, and these purely exploratory associations warrant further validation in future, stringently designed PDAC-specific clinical trials. To ensure these prognostic associations were not confounded by bulk tissue heterogeneity in PDAC, we performed multivariate Cox proportional hazards regression analysis across nine independent validation cohorts, systematically adjusting for estimated tumor purity (via the ESTIMATE algorithm) alongside standard clinical covariates. Remarkably, a high LMICs score remained a robust and independent unfavorable prognostic factor for PDAC patients (Supplementary Table 4, Supplementary Fig. 3 L), indicating that its clinical relevance is independent of tumor purity fluctuations. In addition, we observed a obvious co-localization and co-expression between the LMICs score and fibroblasts in the spatial transcriptomes related to primary and liver-metastatic PDAC (Supplementary Fig. 4A-B and Supplementary Table 5) which is consistent with the results we observed in TSPAN2_MC2 in the previous results (Fig. 3A,E). Collectively, this study defines the developmental seeds of PDAC liver metastasis at single-cell resolution, establishes the LMICs score as a robust clinical prognostic risk factor independent of tumor purity across multi-center cohorts, and reveals the core characteristics by which these plastic cells maintain colonization flexibility via hybrid EMT phenotypes and evade immune clearance through organ-specific immunosuppression.

Spatial transcriptomic and single-cell communication landscapes present the SEMA3A-NRP1 signaling axis as a prominent stromal-epithelial crosstalk feature

To systematically explore the communication networks operating between LMIC and other components of the TME, we inferred the inter-cellular interaction landscape based on our single-cell data. Analyses using two independent algorithms, CellphoneDB and CellChat, consistently revealed a robust signaling crosstalk between LMIC and surrounding stromal cells, particularly cancer-associated fibroblasts (CAFs) (Fig. 5A and Supplementary Fig. 5A). These computational inferences align with the spatial co-localization of LMIC and CAFs previously observed via spatial transcriptomics, collectively pointing toward a potential paracrine-like associative network. Given the central transcriptomic position of LMIC within the simulated cellular network, we subsequently screened for specialized ligand-receptor pairs potentially characterizing these microenvironmental interactions. Ligand-receptor mapping nominated the SEMA3A-NRP1 signaling axis as an important feature connecting CAFs, endothelial cells, and LMIC (Supplementary Fig. 5B, D). Expression profiling indicated that the ligand SEMA3A primarily originates from CAFs, while its receptor, NRP1, is specifically enriched in LMIC (Fig. 5B). Notably, both the ligand and receptor were significantly up-regulated in liver metastasis compared to primary tumors (Fig. 5C and Supplementary Fig. 5C). Furthermore, SEMA3A exhibited a strong mesenchymal signature across multiple PDAC cohorts, showing tight co-expression with stromal remodeling genes (Supplementary Fig. 5E–G). To elucidate the functional consequences of this signaling axis, we analyzed its correlation with markers of tumor progression (Fig. 5D). Both SEMA3A and NRP1 displayed significant positive correlations with mesenchymal markers (VIM, CDH2), invasion-driving factors (SNAI1, SERPINE1, COL1A1), and the TGF-β signaling pathway (SMAD2/4), while negatively correlating with epithelial markers (CDH1) and proliferation indices (CDCA7). This expression signature suggests that the SEMA3A-NRP1 axis does not primarily drive clonal expansion; instead, it promotes metastatic colonization and survival in distant organs by inducing mesenchymal phenotypic transition and aggressive remodeling in LMIC. Clinical evaluation further substantiated the pathogenic significance of this axis. High expression of both SEMA3A and NRP1 predicted poor patient prognosis and advanced clinical stages (Fig. 5E, F, H and Supplementary Tables 6, 7). At single-cell resolution, this axis demonstrated stable lineage-restricted expression across various cancer types and PDAC specifically, being predominantly localized to malignant cells, fibroblasts, and endothelial cells (Fig. 5G, I and Supplementary Fig. 6B, D). Importantly, spatial transcriptomics sections provided direct visual evidence of the regional co-localization of SEMA3A and NRP1 (Fig. 5J). This spatial proximity suggests that CAFs may sustain the stemness and metastatic potential of LMIC within the liver microenvironment through the continuous activation of the NRP1 receptor by locally high concentrations of SEMA3A. Finally, pathway correlation analysis revealed that SEMA3A expression positively correlates with thr 14 oncogenic hallmarks. Similarly, NRP1 exhibited broad activation of oncogenic pathways, despite its negative correlation with the cell cycle (Supplementary Fig. 5I, J). Sex-stratified pan-cancer analysis further showed an up-regulation of SEMA3A and NRP1 in malignancies of the brain, pancreas, liver, and stomach (Supplementary Fig. 6A, C). Notably, this axis outlines a computationally nominated and provisionally characterized candidate cross-talk pathway that theoretically associates with heightened in vitro cell motility and aligns with adverse clinical survival outcomes, though its causal requirement in physiological metastasis remains to be fully uncoupled. Collectively, these multi-omic data imply that CAF-derived SEMA3A and LMIC-localized NRP1 constitute a tightly coordinated spatial expression axis, outlining a conserved, microenvironment-dependent paracrine correlation network that correlates with the metastatic progression of PDAC.

Fig. 5.

Cafs promote PDAC liver metastasis via the SEMA3A-NRP1 signaling axis. TME communication maps inferred by CellphoneDB (top) and CellChat (bottom). (B) UMAP feature plots of SEMA3A (CAFs) and NRP1 (LMIC). (C) ridge plots comparing NRP1 and SEMA3A expression between primary and metastatic PDAC. (D) correlation heatmap between SEMA3A/NRP1 and markers of EMT, and TGF-β. (E) meta-analysis of multiple clinical cohorts assessing the prognostic value of NRP1 and SEMA3A. (F) ridge plots showing NRP1 expression across pathological grades and T-stages. (G, I) validation of NRP1 (G) and SEMA3A (I) lineage localization in independent datasets. (H) SEMA3A expression across clinical stages. (J) spatial transcriptomic validation of SEMA3A-NRP1 co-localization in metastatic sections

The transcriptional factor YY1 correlates with the regulatory circuitry and oncogenic programs characterizing LMIC

To define the molecular mechanisms sustaining the biological identity of LMIC, we performed GO and KEGG enrichment analyses on their signature genes (Fig. 6A). The results revealed a significant enrichment of pathways related to microenvironmental interactions (ECM-receptor interaction, focal adhesion), directed cell locomotion (positive regulation of cell motility, anatomical structure morphogenesis), and canonical oncogenic signaling. Among these, we prioritized the WNT, PI3K-AKT, and TGF-β pathways, as they exhibited both high and concentrated AUCell activity scores within the LMIC population (Fig. 6B). These pathways are recognized as central signaling hubs that drive EMT, maintain stemness, and mediate metastatic colonization. GSEA further confirmed the robust up-regulation of these three pathways in LMIC (WNT: NES = 1.527; PI3K-AKT: NES = 1.402; TGF-β: NES = 1.562; p < 0.05; Fig. 6C), with activities significantly exceeding those in primary malignant and normal epithelial cells (Fig. 6D). Notably, we observed a distinct hierarchical divergence in the association between the receptor gene NRP1 and these oncogenic hallmarks across different cell populations (Fig. 6E). In LMIC, NRP1 displayed a robust positive correlation with all three pathways and stemness features, whereas this co-expression pattern was significantly attenuated or absent in primary PDAC and normal epithelial cells. Crucially, the correlation with stemness scores was uniquely significant within the LMIC subcluster. This phenomenon implies that NRP1 represents a highly integrated component of LMIC-specific regulatory network, wherein its transcriptomic alignment with stemness and oncogenic signals coincides with the metastatic-initiating potential attributed to this subpopulation. To screen for candidate master TFs potentially coordinating this regulatory architecture, we quantified TF activities at single-cell resolution using the DoRothEA algorithm (Fig. 6F), identifying PAX6 and YY1 as candidate factors specifically activated in LMIC. To validate the functional essentiality of these candidates, we leveraged genome-wide CRISPR screening data from the DepMap database. In silico analysis corroborated that YY1 exhibits significantly lower gene effect scores and high dependency probabilities across multiple PDAC cell lines, underscoring its important role in sustaining PDAC cell viability and growth (Fig. 6G). This functional genomic positioning supports the relevance of YY1 within the framework of the LMIC-associated phenotype. Correlation analysis further demonstrated a high degree of positive association between YY1 activity and NRP1 expression (Fig. 6H). Across different groups, both the transcriptional activity (DoRothEA score) and mRNA levels of YY1 were significantly higher in LMIC compared to primary malignant subclusters and normal epithelium (Fig. 6I, J). Within the LMIC population, YY1 activity positively correlated with various LMIC-characteristic pathways (Supplementary Fig. 7A). Furthermore, JASPAR motif analysis identified potential binding sites for YY1 within the NRP1 promoter region, with significant binding scores of 8.55 and 6.67 (Supplementary Fig. 7B and Supplementary Table 8). In addition, we also observed a significant correlation between the expression of NRP1 and the transcriptional regulation score and expression of YY1 at the transcriptome level (Supplementary Fig. 7C-D). Collectively, this integrated computational evidence layer proposes a speculative model wherein YY1 represents a candidate transcriptional node linked with LMIC identity, potentially participating in the transcriptional maintenance of NRP1. Mechanistically, such conceptual regulation of the receptor, if integrated with the upstream SEMA3A-NRP1 spatial communication coordinates, might form a hypothetical feed-forward expression network that correlates with the malignant phenotype and metastatic-associated transcriptomic architecture of LMIC. However, we explicitly acknowledge that these entirely computational alignments—relying strictly on single-cell transcriptomic correlations and JASPAR matrix sequence matching—remain preliminary and descriptive at present. To definitively establish the physical transactivation kinetics and validate this proposed regulatory feedback loop, stringent empirical validations, such as dual-luciferase reporter assays and chromatin immunoprecipitation (ChIP) sequencing platforms, are critically required and represent an essential avenue for future independent mechanistic investigations.

Fig. 6.

Regulatory circuitry and transcriptional drivers of the LMIC phenotype. (A) KEGG and GO enrichment analysis of LMIC signature genes. (B) UMAP feature plots of WNT, PI3K-AKT, and TGF-β signaling activities. (C) GSEA curves confirming the up-regulation of these pathways in LMIC. (D) comparison of pathway scores between LMIC, primary malignant cells, and normal epithelium. (E) correlation networks between NRP1 expression and oncogenic pathway activities across cell groups. (F) Heatmap of TF activities inferred by DoRothEA, identifying YY1 and PAX6. (G) DepMap dependency and gene effect scores for YY1 and PAX6 in PDAC lines. (H) correlation between YY1 activity and NRP1 mRNA levels. (I-J) comparison of YY1 DoRothEA activity (I) and mRNA levels (J) across groups

Multidimensional metabolic reprogramming presents potential therapeutic vulnerabilities and computational intervention strategies for LMIC

To meet the bioenergetic demands of metastatic progression, LMIC undergo profound metabolic rewiring. Using scFEA to quantify metabolic flux (Fig. 7A), we observed that LMIC maintain expanded anabolic pools, characterized by significantly higher steady-state levels of biosynthetic precursors (e.g., Acetyl-CoA, cholesterol) and essential amino acids (e.g., glutamine, arginine) compared to primary malignant cells. This high-throughput metabolic signature underscores the plasticity of LMIC in exploiting the hepatic niche for energy and biosynthesis. Integration with scMetabolism nominated three core metabolic axes in LMIC (Fig. 7B): Lipid Remodeling: LMIC displayed an enrichment of de novo lipogenesis programs, evidenced by the up-regulation of ChREBP-mediated pathways and rate-limiting enzymes such as FASN, ACACA, and ACLY (Fig. 7C). Glycan Modification: Significant activation of glycosaminoglycan and sialic acid biosynthesis implies that LMIC might modulate cell-surface glycocalyx properties, potentially correlating with the SEMA3A-NRP1 signaling perception and evade immune surveillance. Redox Maintenance: Enhanced one-carbon and glutamine metabolism, associated with MAT2A and GLS expression, point toward a prospective coordination of antioxidant defense and methyl donors for YY1-mediated epigenetic regulation. Validation of the necessity of the above key genes in PDAC in vitro cell lines showed that knockout of MAT2A, ACLY, ACACA, and FASN in most cell lines led to obvious growth inhibition (Supplementary Fig. 8C-D). Leveraging these identified vulnerabilities, we utilized the BeyondCell framework to nominate potential precision intervention strategies, hypothesizing that targeting these specific nodes could selectively impair LMIC fitness (Fig. 7D, E and Supplementary Table 9). Screening identified three potent therapeutic classes (Supplementary Fig. 8A, B): Adhesion blockade: In silico analysis nominated potential sensitivity to FAK (PF-562271) and Src (WH-4025) inhibitors, aligning with the proposed therapeutic relevance of the SEMA3A–NRP1–FAK axis. Metabolic targeting: A candidate dependence on fatty acid oxidation was suggested by the inferred sensitivity to etomoxir (a CPT1A inhibitor). Epigenetic reversal: The model-predicted effectiveness of mocetinostat (an HDAC inhibitor) hypothesizes that epigenetic intervention might serve as a strategy to dismantle the YY1-maintained LMIC state. To further assess the computational robustness of these therapeutic leads, we performed a correlation analysis between the transcriptomic profiles of LMICs score (alongside discrete NRP1 and SEMA3A expression) and predicted drug sensitivity scores. Overall, this in silico modeling indicated that the LMICs score positively correlate with the inferred sensitivity to FR-139317, WH-4025, PF-562271 and mocetinostat, while demonstrating a negative association with predicted responses to cyanocobalamin and etomoxir (Supplementary Fig. 8E). In summary, our multi-omic integration suggests a model wherein LMIC programmatically coordinate lipid adaptation, glycan remodeling, and redox balance pathways that correlate with their fitness features. Consequently, this computational framework nominates a combined candidate therapeutic strategy targeting FAK signaling, lipid metabolism, and epigenetic modifications as prospective leads for further independent evaluation in intercepting PDAC liver metastasis.

Fig. 7.

Metabolic reprogramming and pharmacological landscape of LMIC. (A) metabolic flux and balance profiles inferred by scFEA. (B) reactome and KEGG metabolic pathway enrichment using scMetabolism. (C) expression of rate-limiting metabolic enzymes across cell subclusters. (D) volcano plot of differential drug sensitivity predicted by BeyondCell. (E) UMAP projections of sensitivity scores for candidate drugs (e.g., Mocetinostat, Etomoxir)

The SEMA3A-NRP1 axis promotes PDAC cell migration and invasion and correlates with poor clinical prognosis

To validate the functional impact of the bioinformatically inferred SEMA3A-NRP1 axis on PDAC progression, we performed gain-of-function assays in AsPC-1 cells by over-expressing the receptor NRP1 (OE-NRP1) or the ligand SEMA3A (OE-SEMA3A). Wound healing assays demonstrated that cells over-expressing either NRP1 or SEMA3A exhibited significantly accelerated closure rates at 24, 48, 72, and 96 hours compared to the empty vector control (NC), indicating enhanced horizontal migration capacity (Fig. 8A). Quantitative analysis confirmed that both healing rates and migration distances were significantly higher in the OE-NRP1 and OE-SEMA3A groups within the initial 0–24 hour period (Fig. 8D). Subsequently, Transwell invasion assays revealed a marked increase in the number of cells penetrating the Matrigel in the over-expression groups (Fig. 8B), with statistically significant differences in invaded cell counts (p < 0.001) (Fig. 8C). Complementary microscopy at various magnifications further corroborated that activation of the SEMA3A-NRP1 axis substantially enhances the in vitro migratory and invasive potential of PDAC cells (Supplementary Fig. 9A, B). At the clinical level, IHC staining revealed that NRP1 protein levels were significantly higher in PDAC tissues compared to normal pancreas, with the most intense signals observed in metastatic lesions, suggesting a close association with distant dissemination (Fig. 8E). Survival analysis using the CPTAC cohort further revealed that patients with high NRP1 expression had significantly shorter overall survival (p = 0.02) (Fig. 8F). Similarly, SEMA3A expression was significantly up-regulated in tumor tissues (Fig. 8G), with its mean expression levels showing a highly significant difference between tumor and normal samples (p < 0.001) (Fig. 8H). Clinical prognosis evaluation confirmed that high SEMA3A expression serves as a significant indicator of poor patient outcome (p = 0.05) (Fig. 8H). Collectively, these in vitro functional assays and clinical cohort data support that the SEMA3A-NRP1 signaling axis is an important molecular mechanism driving PDAC malignancy and influencing patient survival.

Fig. 8.

Functional and clinical validation of the SEMA3A-NRP1 axis in PDAC progression. (A) wound healing assays of AsPC-1 cells (NC, OE-NRP1, OE-SEMA3A) from 0 h to 96 h. Scale bar, 500 µm. (B) Transwell invasion assays showing increased invasiveness in OE groups. Scale bar, 200 µm. (C) quantitative analysis of invaded cell counts. (D) healing rates and migration distances over 0–24 h. (E) IHC staining of NRP1 in normal, primary tumor, and metastatic samples. Scale bars, 20 µm and 100 µm. (F) kaplan-meier survival curves of NRP1 from the CPTAC cohort. (G-H) IHC of SEMA3A (G) and its clinical assessment in the CPTAC cohort (H)

Discussion

The high mortality rate of PDAC is primarily attributed to its early systemic dissemination, particularly liver metastasis. While chemotherapy remains the backbone of clinical management, its efficacy in controlling metastatic initiation and distant colonization is profoundly limited, necessitating the identification of early intervention targets [33]. Emerging evidence suggests that metastatic seeding is often driven by specific cell states with stem-like properties, a paradigm observed in colorectal, ovarian, and lung cancers [10, 34, 35]. However, the identity and functional architecture of these metastasis-related initial cells in PDAC remain poorly defined. By integrating single-cell RNA sequencing with spatial transcriptomics and pseudotemporal trajectories, our study identifies LMIC state as a pivotal developmental seed driving PDAC progression and highlights the SEMA3A-NRP1 axis as a vulnerable node for therapeutic intervention.

The definition of LMIC as a critical state in PDAC liver metastasis is supported by its unique multidimensional molecular profile rather than its lineage position alone. Functionally, LMIC are characterized by the coordinated activation of EMT, adhesion remodeling, and stemness maintenance, suggesting they represent a specialized functional state optimized for environmental adaptation and colonization rather than mere rapid proliferation. Metabolically, LMIC exhibit significant reprogramming, likely to navigate the nutrient competition and oxidative stress inherent in the hepatic niche. This metabolic plasticity—encompassing membrane biogenesis, redox homeostasis, and epigenetic fuel supply—may be fundamental to the persistent survival of disseminated clones [36, 37]. Furthermore, the correlation between LMICs score and immunosuppressive phenotypes in the metastatic microenvironment suggests that this state not only survives the journey but actively remodels the local immune ecology to evade clearance. Clinically, the robust, multi-center prognostic association between an elevated LMIC score and dismal patient survival, independent of tissue tumor purity variations, underscores the clinical relevance of this cell state as a high-fidelity risk indicator of metastatic potential.

Our findings further implicate the CAF-mediated SEMA3A-NRP1 axis as an important paracrine circuit sustaining LMIC aggressiveness. Although SEMA3A, as a microenvironmental ligand, plays a role in niche remodeling [38], its systemic targeting is challenged by its broad tissue distribution and lack of signaling specificity. In contrast, the enrichment of its receptor, NRP1, specifically within the LMIC population makes it a more precise and druggable candidate for disrupting CAF-LMIC crosstalk. Our analysis suggests that NRP1 may function as a master scaffold, integrating EMT, WNT, PI3K-AKT, and TGF-β signaling to enhance metastatic fitness. In addition, our upstream regulatory modeling positions the transcription factor YY1 as a candidate master regulator of this circuit, exhibiting prominent binding matrix affinity for the NRP1 promoter region and strong co-expression with the LMIC malignant phenotype. Previous studies have documented the role of NRP1 in cell migration, angiogenesis, and therapy resistance across various malignancies [39–42]. By combining our drug sensitivity predictions with this mechanistic framework, we propose that blocking microenvironmental communication via NRP1, particularly when combined with inhibitors of its downstream oncogenic effectors, represents a potent strategy to intercept the PDAC metastatic cascade.

Despite these insights, several prominent technical limitations in our current mechanistic characterization and translational modeling must be explicitly noted to guide prospective validation. In vitro gain-of-function simulations within the metastasis-derived AsPC-1 cell line indicate that forced over-expression of SEMA3A or NRP1 substantially augments cellular migration and invasion. However, no loss-of-function experiments or in vivo metastasis models were conducted to validate the functional necessity of this signaling axis in liver colonization. In addition, positioning the LMIC population at the computational root of a developmental trajectory reflects an evolutionary transcriptomic alignment rather than a fully verified metastatic causality, and characterizing these candidate cells as definitive functional ‘initiators’ at this stage requires tempered translational caution. Furthermore, while our analysis prioritizes a CAF-to-LMIC paracrine directionality, we explicitly acknowledge that SEMA3A and NRP1 expression can display localized, promiscuous distribution across diverse tissue elements. We therefore cannot entirely exclude the biological possibility of an alternative or concurrent epithelial-derived autocrine loop, wherein LMIC intrinsically generate SEMA3A pools to complement microenvironmental stromal supplies, reinforcing their survival fitness during remote dissemination. Additionally, although our BeyondCell pharmacological scoring and public DepMap in silico screening nominated key metabolic enzymes (FASN, ACACA, ACLY, and MAT2A) and targeted candidates (such as the FAK inhibitor PF-562271 and HDAC inhibitor mocetinostat) as theoretical vulnerabilities, we note the current absence of prospective, in-house wet-lab experimental validation testing whether LMIC are indeed selectively sensitive to these specific small-molecule blocks.

Future investigations must prioritize a multi-pronged validation roadmap to translate these computational discoveries. Mechanistically, subsequent functional assays require the implementation of high-resolution functional perturbations, dual-luciferase reporter assays, and chromatin immunoprecipitation (ChIP) sequencing to definitively establish the physical transactivation kinetics between YY1 and the NRP1 promoter region, as well as cell-type-specific knockouts to rigorously dissect the relative functional contributions of CAF-derived versus tumor-derived SEMA3A ligands. Additionally, while our spot-level computational deconvolution provides statistical cross-validation of the geographic coupling between LMIC and CAFs at the transcriptomic level, validating individual candidate protein features within human tissue architectures remains a critical necessity. Because the 84-gene LMIC signature represents a highly coordinated, multi-layered molecular program that may elude traditional single-marker immunohistochemistry (IHC), subsequent high-priority validation cohorts should leverage multiplexed immunofluorescence (mIF) panels focusing on canonical cluster markers, such as TSPAN1, to visually and topologically confirm the complex LMIC-CAF physical co-localization patterns at single-cell histological resolution. Finally, stringent wet-lab empirical testing is critically mandatory. To definitively prove metastatic initiation, future studies must employ fluorescence-activated cell sorting (FACS) to isolate specific sub-populations (such as NRP1+ malignant cells) followed by in vivo lineage-tracing assays, limiting-dilution transplantation, and organ-specific seeding evaluations. Such prospective validations should leverage patient-derived organoid (PDO) or patient-derived xenograft (PDX) preclinical metastatic models to cross-examine the real-world anti-metastatic efficacy of blocking this microenvironmental loop, alongside testing our computationally nominated small-molecule leads (such as PF-562271, mocetinostat, and lipogenic enzyme inhibitors) in preventing niche formation and metastatic outgrowth. In conclusion, by integrating single-cell, spatial, and multi-center clinical evidence, this study defines a biologically significant cell state in PDAC metastasis and provides a solid, highly calibrated theoretical foundation for the subsequent development of anti-metastatic therapeutic strategies.

Conclusion

In summary, by integrating scRNAseq, ST-RNAseq, and multicenter clinical verification cohorts, this study systematically characterizes the LMIC population as a primordial developmental reservoir driving downstream PDAC progression. We define LMIC as a high-risk, clinically unfavorable malignant state coordinated by precise metabolic adaptations, localized organ-specific immunosuppressive barriers, and aggressive hybrid mesenchymal phenotypes. Furthermore, our spatial alignment models nominate the CAF-derived SEMA3A-NRP1 signaling framework as a highly conserved stromal-epithelial crosstalk axis closely associated with LMIC transcriptomic fitness and prospective liver colonization. Collectively, these multi-scale computational and histochemical insights expand our conceptual understanding of the spatial microenvironmental cues governing the initiation of PDAC systemic dissemination, providing a robust theoretical and computational foundation for the future validation of early anti-metastatic interception strategies.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Hong-xiang Lu and Zhi-ying Jiang jointly conceived and designed the research protocol. Yang Li and Ya-die Liu were responsible for the implementation of the data analysis process and the writing of the paper. All authors reviewed and edited the manuscript.

Funding

This work was supported by General Research Topics on the Development of Medical Science and Technology in Nanjing (No. YKK24226) and National Natural Science Foundation of China (No. 82101851).

Data availability

Raw data supporting the obtained results are available at the corresponding author.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors agreed on the manuscript.

Competing interests

The authors declare no competing interests.