Deciphering Microenvironmental Heterogeneity by Scalable Niche Guided Module Discovery

Deciphering Microenvironmental Heterogeneity by Scalable Niche Guided Module Discovery | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Deciphering Microenvironmental Heterogeneity by Scalable Niche Guided Module Discovery Xun Lan, Chang Liu, Yuze Zhou, Longchen Xu, Xianhan Qin, Tianhao Liu, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7243521/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Spatial transcriptomics provides high-dimensional gene expression data while preserving spatial context, offering novel insights into tissue composition and heterogeneity. Each spot or cell in the spatial transcriptome could be reflected as gene modules influenced by its surrounding microenvironment, with module interactions vital for tissue architecture and function. Here, we present Scalable Niche Guided Module Discovery (SIGMOD), a method that integrates prior constructed microenvironment information with gene expression decompositions to uncover gene modules, enabling a deeper understanding of crosstalk within the microenvironment. SIGMOD identifies cell type–specific and cell state–specific, clinically relevant gene modules, uncovering gene module–module interactions in 10X ST, Visium, Xenium, and CosMX data, demonstrating its effectiveness and broad applicability. Biological sciences/Computational biology and bioinformatics/Computational models Biological sciences/Computational biology and bioinformatics/Gene regulatory networks Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 INTRODUCTION Structurally organization of different cell types within tissues is crucial for their biological functions. Spatially proximal cells can influence the composition and activity of intracellular gene modules through various mechanisms, thereby shaping cellular states and functions. Recent advances in spatial transcriptomics (ST) have opened new avenues for the systematic and comprehensive analysis of gene module interactions within tissue contexts. Current ST technology could be categorized into two groups 1 : sequencing-based approaches (for example, Visium 2 , DBiT-Seq 3 , Slide-Seq 4 , Slide-tags 5 and Stereo-seq 6 ) and image-based approaches (for example, SeqFISH 7 , RNAScope 8 , MERFISH 9 , Xenium 10 and CosMx 11 ), each offering unique advantages and complementary insights for studying spatially regulated gene activity. Sequencing based methods rely on predefined capture area, or spot and measure unbiased genome-wide transcriptomic profiles. Specifically, when integrated with single-cell transcriptomic data, each spot can be modeled as a mixture of cell type–specific gene modules, analogous to a deconvolution process in this context. Image-based spatial transcriptomics approaches, which leverage high-resolution fluorescent imaging, achieve single-cell resolution but are typically restricted to predefined gene panels. In this context, each cell can be modeled as a mixture of gene modules that are specific to distinct cellular states and are often inferred de novo. Many computational algorithms have recently been proposed, with some focusing on the deconvolution of spatial transcriptomics data using single-cell RNA sequencing as a reference 12 – 17 , while others aim to infer gene modules de novo to capture underlying transcriptional programs 18 – 22 . However, these methods primarily rely on spatial coordinates and often overlook the broader microenvironmental context, limiting their ability to uncover interactions between gene modules that may drive complex cellular behaviors. Here, we present SIGMOD (Scalable Niche-Guided Module Discovery), a computational framework designed to dissect gene modules and uncover interactions between them within the spatial microenvironment. SIGMOD frames gene module discovery as a gene expression reconstruction problem, incorporating the microenvironmental context as a prior to guide the underlying generative process. It not only reveals the composition and activity of gene modules but also elucidates how the microenvironment influences these modules, thereby uncovering interactions between them. This framework enhances the interpretability of spatial transcriptomic data and facilitates a deeper understanding of the regulatory architecture within complex tissue environments. SIGMOD effectively deconvolved cell type–specific gene modules within each spot using 10x ST and Visium data, uncovering region-specific modules and inter-module interactions that drive disease processes such as Alzheimer’s disease and metastasis. In Xenium data, SIGMOD identified interpretable gene modules and characterized interactions among immunosuppressive programs. When applied to CosMX data at single-cell resolution, SIGMOD revealed how neighboring cell types influence the activity of distinct gene modules, highlighting neutrophils as key contributors to the hypoxic adaptation of tumor cells. RESULTS Overview of SIGMOD SIGMOD comprises two key steps: niche construction and gene module inference (Fig. 1 ). In this study, the niche is defined as the local microenvironment surrounding each cell or spot. SIGMOD offers two primary approaches for niche construction: spatially variable ligand-receptor interaction analysis and proximity-based cell enrichment. To construct niche defined with spatially variable ligand-receptor interaction, SIGMOD identifies co-localized ligand-receptor pairs by assuming that relatively high expression of a receptor in a specific spatial region, coupled with relatively high expression of its corresponding ligand in the same region, indicating co-localization of a pair of ligand and receptor and potential biologically relevant interactions. Then SIGMOD partitions the space into a grid of unannotated regions or uses annotated regions if provided. SIGMOD evaluates simultaneously enriched ligand-receptor pairs across regions and calculates the probability of ligand-receptor interactions across regions, identifying interactions with significant spatial specificity. The interaction of a pair of ligand and receptor is characterized by both co-localization enrichment and spatial specificity. To perform niche construction using proximity-based cell enrichment, SIGMOD defines cell’s niche based on the number and types of neighboring cells within a specified distance. Specifically, for imaging-based data with single-cell resolution (e.g., Xenium), SIGMOD quantifies the interaction strength of specific ligand–receptor pairs within a defined spatial radius and integrates this interaction information into the modeling of the niche to guide gene module discovery. For the next step, gene module inference, SIGMOD integrates structural topic modeling 23 and deep generative modeling 24 , using gene expression and niche information as inputs to predict module activity in spots or cells, the quantifiable effect of niche on module activity, and gene activity within modules. In SIGMOD, gene expression is modeled using a Gamma-Poisson distribution, with its parameters determined by module activity—guided by niche information—and gene activity, informed either by predefined cell type–specific gene modules from single-cell RNA-seq data or inferred directly from the data in the case of de novo module discovery. SIGMOD outputs offer insights into gene module activities and cross–cell-type module–module interactions. Together, these capabilities make SIGMOD a powerful tool for decoding spatially resolved cellular niches and intercellular communication networks. Accurate Deconvolution of Regional Cell Type–Specific Gene Module Distributions and Their Associated Ligand–Receptor Interactions in Pancreatic Ductal Adenocarcinoma We first applied SIGMOD to a human pancreatic ductal adenocarcinoma (PDAC) spatial transcriptomic dataset (denotated as PDAC-A) 25 , which contained annotated single cell RNA transcriptomics with manually annotated histological regions (Fig. 2 A). This dataset was widely used for benchmarking deconvolution methods 26 . SIGMOD identified spatial variable and regional specific ligand receptor interactions and demarcated the spatial structural heterogeneity of the ST data, which can be used for subsequent niche construction (Supplementary Fig. 1A-C). Guided by niche information, SIGMOD well characterized pancreatic and tumoral cell type-specific gene modules into different tissue regions (Fig. 2 B). SIGMOD demonstrated statistically significant correlations between the inferred cell type-specific gene modules proportions and canonical marker genes for 19 cell types. Its performance was comparable to that of Cell2location and RCTD, while outperforming other existing methods (Fig. 2 C, Supplementary Fig. 1D, E). SIGMOD correctly identified region-specific cell type-specific gene modules, such as the localization of the two clones of cancer cells, ductal high hypoxic cells, and fibroblast cells in tumor region, as well as acinar cells and endocrine cells in pancreatic region (Supplementary Fig. 1F). The dominant cell type-specific modules proportions inferred by SIGMOD effectively captured the segregation between pancreatic, cancer and duct epithelium regions (Supplementary Fig. 1G). SIGMOD also distinguishes the enrichment of ductal high hypoxic cell-specific gene module within ductal epithelium, further validated by the specific gene expression of ductal high hypoxic cells, which was ignored by other methods (Supplementary Fig. 1F, H). SIGMOD also uncovered cell type–specific gene module–module interactions by integrating the effects of the niche on gene module proportions, the composition of ligand–receptor pairs within niches, and the ligand/receptor expression profiles of annotated cell types derived from paired single-cell data (Fig. 2 D). For example, two cancer clone–specific gene modules and a fibroblast-specific gene module exhibited strong mutual interactions (Fig. 2 E). Among the identified ligand–receptor pairs, LAMB3–CD151 and FGG–ITGAV mainly affected those associated with cancer clone A. Additionally, THBS2–NOTCH3 and FN1–ITGA5 predominantly impacted fibroblast-specific gene modules (Fig. 2 F). SIGMOD also identified CXCL12–CD4 as a key interaction mediating effects on mDCs B–specific gene modules (Supplementary Fig. 2I). These results demonstrate that SIGMOD not only accurately deconvolves cell type–specific gene module proportions but also reveals the module–module interactions that influence the spatial distribution and organization of distinct cell types. Excels in cell type–specific gene module deconvolution and interaction inference in Alzheimer’s disease To further evaluate the accuracy of SIGMOD in identifying cell type–specific gene modules, we applied it to 10x Visium data from an Alzheimer's disease (AD) sample 17 , using Xenium data from a matched AD sample as the ground truth (Fig. 2 G). SIGMOD identified distinct cellular niches in AD samples, defined by spatially variable and region-specific patterns of ligand–receptor interactions, highlighting localized microenvironmental signaling landscapes (Supplementary Fig. 2A–C). For quantitative assessment, we computed the correlation for each cell type across all spots, as well as the correlation and mean squared error (MSE) for each spot across all cell types. SIGMOD achieved the highest correlations for individual cell types across spots and performed particularly well for well-organized excitatory glutamatergic neurons (Fig. 2 H). Additionally, it showed comparable performance to Spotiphy and outperformed other methods in both correlation and MSE when evaluated across all cell types per spot (Supplementary Fig. 2D, E). In addition, we found that the cell type–specific gene module priors derived from niche information were well aligned with the final inferred module proportions. This concordance was particularly evident for major inhibitory GABAergic interneurons and glial cells, which correspond closely to well-defined niches (e.g., niche 4, 7, and 10) (Supplementary Fig. 2C, F).SIGMOD further inferred cell type–specific gene module–module interaction networks by integrating the effects of niche context on gene module proportions, the composition of ligand–receptor pairs within niches, and ligand/receptor expression profiles from annotated cell types derived from paired single-cell data. Prominent interactions were observed among excitatory neurons (e.g., L5 PT CTX, L6 IT CTX), glial populations (e.g., microglia, astrocytes), and immune-related cells (e.g., macrophages, neutrophils). Notably, several cell types—such as L5 IT CTX, LAMP5, and Pvalb-positive interneurons—exhibited strong intra- and inter-type interactions, suggesting central roles in the cellular communication network (Fig. 2 I). We further focused on microglia-centered interactions, revealing strong self-interaction of microglia and notable crosstalk with macrophages, subiculum (SUB), and neutrophils. The analysis also identified key ligand–receptor pairs mediating these interactions. In particular, Apoe–Trem2 and Sema6d–Trem2 emerged as major contributors to microglia-associated signaling, implicating these axes in the regulatory roles of microglia within the Alzheimer’s disease microenvironment 27 , 28 (Fig. 2 J). Together, these results demonstrate SIGMOD’s capability to uncover cell type–specific gene modules and their associated interaction networks. Comparison of Primary Colorectal Tumors and Liver Metastases Reveals Metastasis-Associated Gene Modules We next applied SIGMOD to 4 specimens of colorectal cancer (CRC) primary sites (C1 to C4) and 2 specimens of CRC liver metastatic tumors (L1 and L2) from a previous study using 10x Visium spatial platform 29 . Cells in the single cell transcriptome reference dataset from the original study were annotated with 9 cell types, which was further classified into 64 distinct, refined cell states. Using SIGMOD, we estimated the proportions of 64 refined cell states- specific gene modules of each spot. We first identified spatially variable ligand–receptor (LR) interactions in each sample using SIGMOD. A substantial number of unique LR interactions were observed in individual samples, and the overlap of LR interactions across all six samples was minimal, indicating that each sample harbors distinct LR interaction patterns (Supplementary Fig. 3A). To incorporate LR interactions across the six samples, we retained interaction values only for spots exhibiting high LR interaction activity, setting the remaining values to zero. SIGMOD then performed non-negative matrix factorization on the LR interaction scores to derive niche information. This niche information was subsequently used for dimensionality reduction and clustering of spots. While some spots clustered homogeneously based on niche similarity, others exhibited substantial heterogeneity, reflecting diverse microenvironmental contexts (Fig. 3 A, Supplementary Fig. 3B). SIGMOD utilized LR-based niche information to infer cell type–specific gene modules. By leveraging these derived gene modules, SIGMOD successfully integrated all samples into a shared latent space (Fig. 3 B, C). Niches exhibited distinct distributions within the latent space, indicating strong niche specificity and their association with cell type–specific gene modules (Supplementary Fig. 3C). To elucidate similarities and distinctions in cell type module composition across samples, we identified spatial hubs within the integrated space (Fig. 3 B, C). Most hubs were detected in multiple samples (Supplementary Fig. 3B, C), with their composition dominated by tumor cells. However, these hubs reflected distinct tumor cell subtypes, indicating that the latent space preserved tumor heterogeneity (Supplementary Fig. 3D, E). We also found that the cell type–specific gene module priors derived from niche information were well aligned with the final inferred module proportions, indicating that niche information provides valuable guidance for gene module inference (Supplementary Fig. 3F). In general, spatial hubs exhibited distinct combinations of cell type–specific gene modules (Fig. 3 D). For example, Hub 1 was enriched for Endo_CLEC4G, Mac_CXCL9, CD8_TXNIP, and CD4_SELL, which were predominantly derived from metastatic tumors and was validated in matched single-cell datasets (Supplementary Fig. 3G). In contrast, Hub 7 was mainly composed of Tumor_AREG, which was specifically associated with primary tumors (Fig. 3 D, E). However, we also observed that some hubs from different tissue sections shared similar gene module compositions. For instance, Hub 3, enriched in liver metastatic tumors, and Hub 10, enriched in primary tumor sites, both showed high levels of Tumor_MKI67, suggesting that the key differences between primary and metastatic tumors may lie not only in the presence of specific gene modules, but in the patterns of interactions among them (Fig. 3 D, E). To further validate this observation, we first identified differentially expressed genes across spatial hubs and performed gene set enrichment analysis. Hub 3 showed enrichment for biological processes such as cholesterol metabolism and complement and coagulation cascades—well-established hallmarks of tumor metastasis 30 , 31 . In contrast, Hub 10 was primarily enriched for cell cycle–related pathways. These results suggest that spatial hubs reflect coordinated activity of multiple gene modules, rather than being driven by a single module, in shaping distinct tumor microenvironments (Fig. 3 F). Because SIGMOD could reveal gene module-module interactions analysis, that inspired us to perform differential gene module-module interactions analysis between liver metastasis and primary tumors. We found that interactions between Endo_CLEC4G, Fib_CXCL12, Fib_C3, Mac_CXCL9 and Tumor_GNG13 contribute most to the differential gene module-module interactions (Fig. 3 G). These interactions involved adhesion-related pairs such as COL1A1–ITGB1 and THBS1–ITGA3, phagocytic interactions such as CD34–SELL and CD99–CD81, and APP family members such as APP–SORL1, suggesting a metastasis-associated microenvironment characterized by enhanced cell–matrix adhesion and tumor–immune crosstalk 32 (Fig. 3 H, Supplementary Fig. 3H). Collectively, SIGMOD integrated primary and metastatic samples, uncovering distinct cell type–specific gene modules, spatial hub architectures, and context-dependent ligand–receptor interactions that drive cellular heterogeneity. Unveiling Immunosuppressive Gene Modules Interactions in the Breast Cancer Microenvironment We next applied SIGMOD to a high-resolution ST dataset from 10x Xenium of human breast cancer (BC) 33 . BC exhibited heterogeneity with significant variations in molecular characteristics and cell type distributions (Fig. 4 A). Here, known ligand–receptor interactions were used to construct niche definitions as prior for gene modules, and SIGMOD was configured to infer 20 gene modules, in accordance with the annotations reported in the original study 33 . We first evaluated the performance of SIGMOD and competing methods in identifying gene modules using five quantitative metrics involved module coherence, module diversity, and topic sparsity 34 – 36 (Fig. 4 B, Supplementary Fig. 4A). These metrics consistently ranked SIGMOD among the top-performing methods, and often as the best. SIGMOD achieved the highest scores in both module coherence and module diversity, indicating its strong ability to capture tightly co-expressed gene sets within modules and to generate distinct, non-redundant gene modules. Visual inspection of SIGMOD’s output revealed well-defined gene modules that spatially corresponded to distinct tissue regions (Fig. 4 C). Some gene modules exhibited strong activity across multiple cell types, suggesting potential shared functions, while others were specifically active in distinct cell types (Fig. 4 D). By assigning each cell to the gene module with the highest activity, the resulting confusion matrix showed that most gene modules were highly specific to individual cell types, although some modules captured shared transcriptional features among related cell types (Supplementary Fig. 4B). For example, gene module 5 was associated with both T cells and B cells, reflecting common immune-related programs, whereas gene module 15 was specifically restricted to KRT15⁺ myoepithelial cells. Most cell type markers were enriched within their corresponding gene modules, although some modules encompassed markers from multiple cell types in alignment with the gene modules’ activities within cells (Supplementary Fig. 4C). Each gene module was characterized by representative marker genes and enriched for specific biological functions (Fig. 4 E, Supplementary Fig. 4D). For example, gene modules 4, 8, 12, and 17 exhibited similar spatial patterns and were all associated with tumor-related processes; however, they differed in their marker genes and enriched pathways, pointing to distinct biological roles. Gene module 4 was represented by GATA3, an estrogen associated marker 37 and was enriched in estrogen response early. Gene module 8, marked by CCND1 38 , involved in androgen-mediated DNA damage repair, was associated with Notch signaling and androgen response. Gene module 12 was defined by FASN 33 , an established marker of tumor invasiveness, and was enriched for cholesterol homeostasis and mTORC1 signaling. In contrast, gene module 16 was characterized by cell cycle–related genes such as PCLAF, TOP2A, and CENPF, and was enriched for the G2/M checkpoint pathway. These results demonstrate that while spatially co-expressed, gene modules can reflect distinct molecular programs, providing fine-grained insights into tumor heterogeneity. We further investigated cross-cell-type gene module interactions. Spatial correlation analysis revealed three distinct correlation patterns among gene modules: (1) modules active within the same cell type, such as tumor-associated modules 4, 8, 12, 18, and 20; (2) DCIS- and myoepithelial-related modules, including 1, 11, 15, and 17; and (3) additional patterns involving other gene modules (Supplementary Fig. 4E). Niche effect analysis showed that specific ligand–receptor interactions regulate the activity of distinct gene modules. For example, S100A4–EGFR significantly influenced the activity of modules 10 and 13, while CD274–CD80 prominently affected module 2 (Supplementary Fig. 4F). These ligand–receptor pairs were also selectively enriched within corresponding gene modules (Supplementary Fig. 4G). By integrating spatial correlations, niche-derived influences, and ligand assignment in modules, SIGMOD reconstructed a network of cross-cell-type gene module interactions in breast cancer (Fig. 4 F). Notably, SIGMOD identified crosstalk between myeloid-related modules 2, 6, and 7 and lymphoid-associated module 5, mediated by immune checkpoint interactions such as CD86–CTLA4, PDCD1LG2–PDCD1, and CD274–PDCD1, highlighting immunosuppressive signaling within the breast cancer microenvironment (Fig. 4 G). Clinically relevant Cross-Cell-Type Gene Module Interactions in Non-Small Cell Lung Cancer Finally, we evaluated SIGMOD’s ability to identify cell states and their corresponding gene modules within single-cell resolution spatial transcriptomics data, as well as to characterize how these states are influenced by the surrounding microenvironment. To this end, we applied SIGMOD to a human non-small cell lung cancer (NSCLC) dataset generated using the CosMx SMI platform 11 . The selected sample, Lung #5 − 1, comprised 87,606 spots with expression profiles for 960 genes. (Fig. 5 A). Analysis of the heterogeneity in surrounding cells revealed a preferential enrichment of certain cell types in the vicinity of specific cell populations (Fig. 5 B). For example, tumor cells exhibited substantial heterogeneity in their neighboring composition, with varying degrees of association with fibroblasts and neutrophils. These findings suggest that SIGMOD can be effectively used to uncover distinct cellular states within a given cell type. In this analysis, neighboring cell types were treated as spatial niches to guide gene module discovery. The cell type–specific gene module distributions inferred by SIGMOD further indicated that different cellular states exhibit distinct spatial preferences, reflecting underlying microenvironmental influences (Fig. 5 C, Supplementary Fig. 5A). For example, the spatial distributions of the three-tumor cell–associated gene modules varied markedly: tumor module 1 was enriched in the tumor core, module 2 was concentrated within the tumor interior, and module 3 was primarily localized at the tumor margins (Fig. 5 C). Based on the influence of surrounding cell types on gene module activity and the spatial correlations between different gene modules, SIGMOD further constructed a cross-cell-type gene module interaction network, in which each node represents a cell type–associated gene module and each edge denotes the activation of one gene module by another (Fig. 5 D). As shown in the figure, the neutrophil-associated gene module 2 exhibited strong interactions with multiple other gene modules, particularly those associated with tumor cells. Analysis of the incoming and outgoing interaction strengths of each module revealed that neutrophil-associated gene modules exhibited both high incoming and outgoing connectivity. Similarly, tumor-associated, and fibroblast-associated gene modules also showed strong bidirectional interaction strengths, indicating their central roles in cross-cell-type communication within the tumor microenvironment (Fig. 5 E). Neutrophil-associated, tumor-associated, and fibroblast-associated gene modules also exhibited similar spatial distributions (Fig. 5 C). Cross-cell-type gene module interaction analysis of tumor-associated gene modules reveals that different tumor modules were influenced by distinct modules from other cell types (Supplementary Fig. 5B). For example, SIGMOD inferred that the activity of tumor gene module 1 was upregulated by neutrophil-associated gene module 2, while tumor gene module 2 was influenced by fibroblast-associated gene module 3. Additionally, tumor gene module 3 was regulated by multiple sources, including neutrophil-associated gene modules 1 and 2, fibroblast-associated gene module 2, and a T cell–associated gene module (Supplementary Fig. 5B). These inferred regulatory relationships are consistent with the observed spatial co-localization of tumor cells with fibroblasts and neutrophils, reflecting varying degrees of microenvironmental influence across tumor cell states (Fig. 5 B). Gene set enrichment analysis revealed that tumor gene module 1 was associated with hypoxia, tumor gene module 2 with the cell cycle, and tumor gene module 3 with the inflammatory response (Fig. 5 F). These results suggest that neutrophil-associated gene module 2 promotes hypoxic adaptation, fibroblast-associated gene module 3 stimulates cell proliferation, and neutrophil, fibroblast, and T cell–associated gene modules are involved in driving the inflammatory response. Indeed, neutrophil gene module 1 was enriched in inflammatory pathways, consistent with tumor gene module 3, while fibroblast gene module showed enrichment in angiogenesis and E2F target genes, both of which may contribute to tumor proliferation (Fig. 5 F). Additionally, these modules exhibited distinct representative genes (Supplementary Fig. 5C-E). In an independent non-small cell lung cancer dataset 39 (Supplementary Fig. 5F), the scores for different tumor, neutrophil, and fibroblast modules inferred by SIGMOD successfully stratified the corresponding cells into distinct subgroups. This demonstrates the broad applicability, representativeness, and heterogeneity of these modules (Supplementary Fig. 5G). Neutrophil gene module 2 also showed a significant Pearson correlation with tumor gene module 1 when calculated across all patients (Supplementary Fig. 5H), indicating a strong association between neutrophil gene module 2 and tumor gene module 1. We next validated the clinical relevance of these modules in multiple independent LUAD cohorts 40 – 53 . We found that hypoxia-related tumor gene module 1 and its associated neutrophil gene module 2 were linked to poor prognosis. In contrast, the inflammatory gene module, especially neutrophil gene module 1 and fibroblast gene module 2, was associated with better prognosis (hypoxia-related tumor gene module 1 was associated with poor prognosis in 6/15 datasets, neutrophil gene module 2 in 6/15, neutrophil gene module 1 in 3/15, and fibroblast gene module 2 in 7/15 datasets) (Fig. 5 G). These findings suggest that the interaction between specific tumor-associated modules and their microenvironmental counterparts could serve as valuable prognostic markers. SIGMOD further investigated the ligands that influence tumor module activity. For example, the hypoxia-related tumor gene module 1 identified neutrophil-derived CXCL8 as a potential regulatory factor, particularly targeting NDRG1 (Fig. 5 H). CXCL8 has previously been implicated in hypoxia adaptation 54 . Additionally, TIMP1 from CD4 T memory gene module 2 may modulate MHC class II expression in tumor gene module 3 (Fig. 5 I). Moreover, multiple ligands from fibroblast gene module 3, such as LGALS1, were found to contribute to the proliferation properties of tumor gene module 2 (Supplementary Fig. 5I). These results demonstrate that SIGMOD can resolve cross-cell-type gene module interactions with clinical relevance. DISCUSSION In spatial transcriptomics, genes expressed in each spot or cell are organized into coherent modules that reflect structured patterns of gene expression. Understanding module–module interactions is crucial for elucidating tissue architectures and their functional roles. Gene modules can be classified as either cell type-specific or cell state-specific, depending on the situations, especially the spatial transcriptomics technology used. The former is typically associated with sequencing-based technologies such as 10X ST or 10X Visium, which have lower resolution, and module discovery in this context is analogous to cell type deconvolution. In contrast, cell state-specific modules are usually derived from high-resolution, imaging-based technologies like Nanostring CosMx or Xenium, where modules represent diverse gene expression patterns, and their activity reflects the functional state of specific cell types. Moreover, in multicellular organisms, cells do not function in isolation; their abundance and state are influenced by the surrounding microenvironment. To consider this, we developed SIGMOD, a method designed to identify functional modules within spatial transcriptomics data. To capture the impact of the microenvironment on these modules, SIGMOD first constructs niche information for each spot or cell based on ligand–receptor interactions or proximity-based cell enrichment. We hypothesize that ligand–receptor interactions play a crucial role in processes such as cell migration and downstream signaling, while neighboring cells also influence the state of the central cell. Using this niche information, SIGMOD infers gene modules either from reference single-cell RNA-seq data or via de novo discovery. This approach not only reveals gene module activity and assignments but also explores the spatial factors driving these changes, including cross-cell-type module–module interactions. SIGMOD’s versatility across sequencing-based (10X ST/Visium) and imaging-based (Xenium/CosMx) platforms highlights its ability to deconvolve modules at varying resolutions, offering insights into both cell type–specific (e.g., cancer clone modules in PDAC) and state-specific (e.g., hypoxia-responsive modules in lung cancer) patterns. There are several limitations for SIGMOD, including its current inability to incorporate additional data types like H&E histopathology images or pathology-defined lesion regions for capturing transcriptomic features linked to phenotypic traits. It also lacks integration of prior knowledge, such as gene sets, pathways, or cell types, to guide module construction, potentially limiting accuracy and relevance. Additionally, SIGMOD does not yet support analysis of temporal data or spatial multi-omics datasets, restricting its utility in fully understanding complex biological systems. The core concept of SIGMOD, integrating microenvironment data for gene module discovery and interaction analysis, makes it adaptable to diverse spatial omics datasets. As spatial transcriptomics evolves, SIGMOD’s flexibility and integration with emerging technologies will ensure its ongoing impact. METHODS SIGMOD model We developed SIGMOD, a Bayesian model that integrates spatial transcriptomics with local microenvironmental context to infer cell- or gene module–specific expression patterns across tissues. SIGMOD operates in two key stages: niche construction and gene module deconvolution. Niche features are derived from spatially informative signals such as ligand–receptor interactions or neighboring cell enrichment, depending on the resolution and platform of the dataset. These niche features guide the inference of module proportions through a sparse softmax transformation, and spatial coherence is enforced by a graph Laplacian regularization term constructed from spatial adjacency. Inference is performed using Automatic Differentiation Variational Inference (ADVI) implemented in Pyro. Niche Construction To incorporate local microenvironmental context, SIGMOD constructs niche features for each spatial location or cell. Depending on the resolution and data type, niche construction is based on either (i) spatially variable ligand–receptor interactions, identified through co-localization and region-specific enrichment statistics followed by dimensionality reduction using non-negative matrix factorization (NMF); (ii) neighboring cell-type enrichment, where the frequency of adjacent cell types is used as niche features; or (iii) cell–cell ligand–receptor interactions calculated across spatially adjacent cells. Spatial adjacency graphs are constructed using k-nearest neighbors for regular grids or Delaunay triangulation for irregular cell positions. Module Inference SIGMOD performs module inference using a Bayesian framework that models gene expression as a mixture of gene modules modulated by local niche features. It supports both reference-based and reference-free modes: the former leverages prior single-cell transcriptomic profiles to define module signatures, while the latter infers them directly from the data. Module proportions for each spot or cell are computed via a sparse softmax transformation applied to a niche-weighted linear predictor, regularized using a structured horseshoe prior. To ensure spatial coherence, SIGMOD incorporates a graph Laplacian–based regularization term, which penalizes sharp transitions in module proportions across neighboring locations. Module Interaction Analysis Based on the inferred module proportions and niche influences, SIGMOD enables downstream analysis of gene module–module interactions. In ligand–receptor–based niches, the contribution of each interaction to module expression is quantified by combining learned niche weights with ligand and receptor expression levels from reference scRNA-seq data. In cell–type–based niches, cross-cell-type module dependencies are inferred by correlating module proportions between neighboring cells and linking them through niche-to-module influence weights. This framework allows SIGMOD to reveal spatial communication patterns and biologically meaningful interactions across cell types and gene programs. The details of SIGMOD are provided in Supplementary Note. Benchmarking and Comparison Methods To evaluate the performance of SIGMOD, we compared it against several existing methods for spatial transcriptomics analysis. For reference-based gene module discovery, we included Tangram, SPOTlight, Cell2location, CARD, RCTD, and Spotiphy, all of which utilize single-cell RNA-seq as reference to infer cell-type compositions across spatial locations. For reference-free gene module discovery, we compared SIGMOD with STAMP, NMF, LDA, and LDVAE, which infer spatial gene programs without requiring external reference data. All methods were run using their recommended parameters and official tutorials to ensure fair comparison. The details of comparison are provided in Supplementary Note. Datasets and Preprocessing We applied SIGMOD to multiple spatial transcriptomics datasets across diverse platforms and biological contexts, including human pancreatic ductal adenocarcinoma (PDAC, 10x ST), colorectal cancer with liver metastasis (CRC, 10x Visium), breast cancer (Xenium), and non-small cell lung cancer (CosMx SMI), as well as mouse Alzheimer’s disease tissue (10x Visium). For each dataset, standard preprocessing steps were applied, including gene and spot/cell filtering based on expression thresholds. When available, matched single-cell RNA-seq data were used for reference-based gene module inference. Niche construction strategies were selected according to data resolution, using ligand–receptor interactions, neighboring cell enrichment, or cell–cell interactions depending on the dataset. The details of real datasets analysis in Supplementary Note. Declarations SUPPLEMENTARY DATA Supplementary Data are available at NAR online. CONFLICT OF INTEREST The authors declare no competing interests. FUNDING This work was partially supported by the National Natural Science Foundation of China (81972680 to X.L.), the Tsinghua University Peking University Joint Center for Life Sciences (61020100119 to X.L.; postdoctoral fellowship to L.X.), the Beijing Natural Science Foundation (20201100463 to X.L.), the Damo Academy through the Damo Academy Innovative Research Program, and the Dushi Program. AUTHOR CONTRIBUTIONS X.L. and C.L. conceived the project. C.L. designed the models. C.L. and Y.Z. implemented the code. C.L., Y.Z., L.X. and X.Q. performed the analyses. T.L., C.T., and J.L. helped in designing the analyses and explaining the results. X.L. supervised the project. All authors contributed to the writing of the manuscript. ACKNOWLEDGEMENTS We acknowledge the High-Performance Computing Core Facility of School of Basic Medical Sciences, Tsinghua University, for computational support. DATA AVAILABILITY This paper analyzes existing, publicly available data. The PDAC data was downloaded from the Gene Expression Omnibus under accession number GSE111672. The AD data was downloaded from https://doi.org/10.5281/zenodo.10520022 . The CRC liver metastasis data was downloaded from the Gene Expression Omnibus under accession number GSE225857. The BRCA data was accessed from https://www.10xgenomics.com/products/xenium-in-situ/preview-dataset-human-breast . The NSCLC data was from https://nanostring.com/products/cosmx-spatial-molecular-imager/nsclc-ffpe-dataset . The single-cell lung cancer atlas (LuCA) was accessed from CELLxGENE. Code availability The source code for SIGMOD is freely available online at GitHub at https://github.com/Boxedpig/SIGMOD and https://doi.org/10.5281/zenodo.16350291 . References Cheng, M., et al.: Spatially resolved transcriptomics: a comprehensive review of their technological advances, applications, and challenges. J. Genet. Genomics. 50 , 625–640 (2023) Visium Spatial Platform: 10x Genomics https://www.10xgenomics.com/platforms/visium Liu, Y., et al.: High-Spatial-Resolution Multi-Omics Sequencing via Deterministic Barcoding in Tissue. Cell. 183 , 1665–1681e18 (2020) Rodriques, S.G., et al.: Slide-seq: A scalable technology for measuring genome-wide expression at high spatial resolution. Science. (2019). 10.1126/science.aaw1219 Russell, A.J.C., et al.: Slide-tags enables single-nucleus barcoding for multimodal spatial genomics. Nature. 625 , 101–109 (2024) Chen, A., et al.: Spatiotemporal transcriptomic atlas of mouse organogenesis using DNA nanoball-patterned arrays. Cell. 185 , 1777–1792e21 (2022) Eng, C.-H.L., et al.: Transcriptome-scale super-resolved imaging in tissues by RNA seqFISH+. Nature. 568 , 235–239 (2019) Wang, F., et al.: RNAscope: A Novel in Situ RNA Analysis Platform for Formalin-Fixed, Paraffin-Embedded Tissues. J. Mol. Diagn. 14 , 22–29 (2012) Chen, K.H., Boettiger, A.N., Moffitt, J.R., Wang, S., Zhuang, X.: Spatially resolved, highly multiplexed RNA profiling in single cells. Science. 348 , aaa6090 (2015) Xenium In Situ Platform: 10x Genomics (2024). https://www.10xgenomics.com/platforms/xenium He, S., et al.: High-plex imaging of RNA and proteins at subcellular resolution in fixed tissue by spatial molecular imaging. Nat. Biotechnol. 40 , 1794–1806 (2022) Kleshchevnikov, V., et al.: Cell2location maps fine-grained cell types in spatial transcriptomics. Nat. Biotechnol. 40 , 661–671 (2022) Ma, Y., Zhou, X.: Spatially informed cell-type deconvolution for spatial transcriptomics. Nat. Biotechnol. 40 , 1349–1359 (2022) Cable, D.M., et al.: Robust decomposition of cell type mixtures in spatial transcriptomics. Nat. Biotechnol. 40 , 517–526 (2022) Elosua-Bayes, M., Nieto, P., Mereu, E., Gut, I., Heyn, H.: SPOTlight: seeded NMF regression to deconvolute spatial transcriptomics spots with single-cell transcriptomes. Nucleic Acids Res. 49 , e50 (2021) Biancalani, T., et al.: Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram. Nat. Methods. 18 , 1352–1362 (2021) Yang, J., et al.: Spotiphy enables single-cell spatial whole transcriptomics across an entire section. Nat. Methods. 1–13 (2025). 10.1038/s41592-025-02622-5 Zhong, C., Ang, K.S., Chen, J.: Interpretable spatially aware dimension reduction of spatial transcriptomics with STAMP. Nat. Methods. 21 , 2072–2083 (2024) Kim, H., et al.: CellNeighborEX: deciphering neighbor-dependent gene expression from spatial transcriptomics data. Mol. Syst. Biol. 19 , e11670 (2023) He, S., et al.: Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs. Nat. Biotechnol. 1–13 (2024). 10.1038/s41587-024-02173-8 Fischer, D.S., Schaar, A.C., Theis, F.J.: Modeling intercellular communication in tissues using spatial graphs of cells. Nat. Biotechnol. 41 , 332–336 (2023) Svensson, V., Gayoso, A., Yosef, N., Pachter, L.: Interpretable factor models of single-cell RNA-seq via variational autoencoders Chen, L., Mankad, S.A.: Structural Topic Sentiment Model for Text Analysis. SSRN Scholarly Paper at https:// (2022). doi.org/10.2139/ssrn.4020651 Bingham, E., et al.: Pyro: Deep Universal Probabilistic Programming. Preprint at (2018). https://doi.org/10.48550/arXiv.1810.09538 Moncada, R., et al.: Integrating microarray-based spatial transcriptomics and single-cell RNA-seq reveals tissue architecture in pancreatic ductal adenocarcinomas. Nat. Biotechnol. 38 , 333–342 (2020) Li, H., et al.: A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics. Nat. Commun. 14 , 1548 (2023) Krasemann, S., et al.: The TREM2-APOE Pathway Drives the Transcriptional Phenotype of Dysfunctional Microglia in Neurodegenerative Diseases. Immunity. 47 , 566–581e9 (2017) Albanus, R.D., et al.: Systematic analysis of cellular crosstalk reveals a role for SEMA6D-TREM2 regulating microglial function in Alzheimer’s disease. 11.11.516215 Preprint at (2022). https://doi.org/10.1101/2022.11.11.516215 (2023) Wang, F., et al.: Single-cell and spatial transcriptome analysis reveals the cellular heterogeneity of liver metastatic colorectal cancer. Sci. Adv. 9 , eadf5464 (2023) Zhao, H., et al.: Inflammation and tumor progression: signaling pathways and targeted intervention. Sig Transduct. Target. Ther. 6 , 1–46 (2021) Luo, X., et al.: Emerging roles of lipid metabolism in cancer metastasis. Mol. Cancer. 16 , 76 (2017) Liu, Z., et al.: Focal adhesion in the tumour metastasis: from molecular mechanisms to therapeutic targets. Biomark. Res. 13 , 38 (2025) Janesick, A., et al.: High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis. Nat. Commun. 14 , 8353 (2023) Bouma, G.: Normalized (Pointwise) Mutual Information in Collocation Extraction WebberWilliam, M.A.: ZobelJustin. A similarity measure for indefinite rankings. ACM Trans. Inform. Syst. (TOIS). (2010). 10.1145/1852102.1852106 Hoyer, P.: O. Non-negative matrix factorization with sparseness constraints Asch-Kendrick, R., Cimino-Mathews, A.: The role of GATA3 in breast carcinomas: a review. Hum. Pathol. 48 , 37–47 (2016) Comstock, C.E.S., et al.: Cyclin D1 Is a Selective Modifier of Androgen-dependent Signaling and Androgen Receptor Function. J. Biol. Chem. 286 , 8117–8127 (2011) Salcher, S., et al.: High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Cancer Cell. 40 , 1503–1520e8 (2022) Fouret, R., et al.: A Comparative and Integrative Approach Identifies ATPase Family, AAA Domain Containing 2 as a Likely Driver of Cell Proliferation in Lung Adenocarcinoma. Clin. Cancer Res. 18 , 5606–5616 (2012) Botling, J., et al.: Biomarker Discovery in Non–Small Cell Lung Cancer: Integrating Gene Expression Profiling, Meta-analysis, and Tissue Microarray Validation. Clin. Cancer Res. 19 , 194–204 (2013) Collisson, E.A., et al.: Comprehensive molecular profiling of lung adenocarcinoma. Nature. 511 , 543–550 (2014) Schabath, M.B., et al.: Differential association of STK11 and TP53 with KRAS mutation-associated gene expression, proliferation and immune surveillance in lung adenocarcinoma. Oncogene. 35 , 3209–3216 (2016) Wilkerson, M.D., et al.: Differential Pathogenesis of Lung Adenocarcinoma Subtypes Involving Sequence Mutations, Copy Number, Chromosomal Instability, and Methylation. PLoS One. 7 , e36530 (2012) Rousseaux, S., et al.: Ectopic Activation of Germline and Placental Genes Identifies Aggressive Metastasis-Prone Lung Cancers. Sci. Transl. Med. 5 (186ra66–186ra), 66 (2013) Takeuchi, T., et al.: Expression Profile–Defined Classification of Lung Adenocarcinoma Shows Close Relationship With Underlying Major Genetic Changes and Clinicopathologic Behaviors. JCO 24, 1679–1688 (2006) Larsen, J.E., et al.: Gene Expression Signature Predicts Recurrence in Lung Adenocarcinoma. Clin. Cancer Res. 13 , 2946–2954 (2007) Shedden, K., et al.: Gene Expression-Based Survival Prediction in Lung Adenocarcinoma: A Multi-Site, Blinded Validation Study. Nat. Med. 14 , 822–827 (2008) Sato, M., et al.: Human lung epithelial cells progressed to malignancy through specific oncogenic manipulations. Mol. Cancer Res. 11 , 638–650 (2013) Okayama, H., et al.: Identification of Genes Upregulated in ALK-Positive and EGFR/KRAS/ALK-Negative Lung Adenocarcinomas. Cancer Res. 72 , 100–111 (2012) Tomida, S., et al.: Relapse-Related Molecular Signature in Lung Adenocarcinomas Identifies Patients With Dismal Prognosis. JCO 27, 2793–2799 (2009) Der, S.D., et al.: Validation of a Histology-Independent Prognostic Gene Signature for Early-Stage, Non–Small-Cell Lung Cancer Including Stage IA Patients. J. Thorac. Oncol. 9 , 59–64 (2014) Bild, A.H., et al.: Oncogenic pathway signatures in human cancers as a guide to targeted therapies. Nature. 439 , 353–357 (2006) Cambier, S., Gouwy, M., Proost, P.: The chemokines CXCL8 and CXCL12: molecular and functional properties, role in disease and efforts towards pharmacological intervention. Cell. Mol. Immunol. 20 , 217–251 (2023) Additional Declarations There is NO Competing Interest. Supplementary Files SuppleInfo.docx Supplementary material SuppleTable.zip Supplementary table Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7243521","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":501310182,"identity":"6f75630e-4e52-4ade-a32d-20746d816b54","order_by":0,"name":"Xun Lan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYDACdiD+wMDAA0ZAwNhAUAszUNWMBFK1MPMkgFjEajE4zPzsse2POhlz/rXHJH4w2MhuOMD87AE+LZLNbObGOQmHeSxnvEuT7GFIM95wgM3cAJ8WfmYGM+mchAM8BjfOmEnwMBxO3HCAh00CnxY2ZvZv0hYJdWAtkn8Y/hPWws/MYybNkMDMY3C+x0yah+EAYS2SzTxlkj1ph4G28BhbyxgkG888zGaGV4vB8fZtEj9s6uwNzp8xvPmmwk6273jzM7xaEEAiAWQCAzimiAT8B4hWOgpGwSgYBSMMAAByND/ms7Y4dQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-6523-046X","institution":"Tsinghua University","correspondingAuthor":true,"prefix":"","firstName":"Xun","middleName":"","lastName":"Lan","suffix":""},{"id":501310183,"identity":"3ee2f9d2-81ec-45ba-b5e3-6ad054217b80","order_by":1,"name":"Chang Liu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Chang","middleName":"","lastName":"Liu","suffix":""},{"id":501310184,"identity":"b3a5b45d-2306-4c55-9b97-05b48af5310c","order_by":2,"name":"Yuze Zhou","email":"","orcid":"https://orcid.org/0009-0004-2938-5802","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Yuze","middleName":"","lastName":"Zhou","suffix":""},{"id":501310185,"identity":"35ab9d26-2265-41b0-9b0f-fe23abadc726","order_by":3,"name":"Longchen Xu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Longchen","middleName":"","lastName":"Xu","suffix":""},{"id":501310186,"identity":"c766c5a3-075a-4154-aa25-efe623c33ff6","order_by":4,"name":"Xianhan Qin","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Xianhan","middleName":"","lastName":"Qin","suffix":""},{"id":501310187,"identity":"d6576fff-3a40-443b-b18e-b8df931604d8","order_by":5,"name":"Tianhao Liu","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Tianhao","middleName":"","lastName":"Liu","suffix":""},{"id":501310188,"identity":"831368a5-f87c-4290-a5ad-9edd2e5e4513","order_by":6,"name":"Chen Tian","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Chen","middleName":"","lastName":"Tian","suffix":""},{"id":501310189,"identity":"87254d6b-ad81-45cd-a0fe-23dee7ce1450","order_by":7,"name":"Jie Li","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Jie","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2025-07-29 12:55:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7243521/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7243521/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89805160,"identity":"2c7d25c3-c270-46d0-8286-68ab045261e2","added_by":"auto","created_at":"2025-08-25 08:58:53","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":902457,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e1 Overview of SIGMOD.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSIGMOD integrates two core computational steps: niche construction and gene expression generation. For niche construction, it employs two approaches—ligand-receptor (LR) interaction-based niches (top left), which account for both co-localization (spatial overlap of ligand-receptor pairs) and region-specific enrichment of interactions, and proximity-based niches (top right), defined by enumerating neighboring cell types within a spatial distance threshold. During gene expression generation, SIGMOD leverages a structural topic modeling (middle) framework to simultaneously infer cell type modules along with LR interactions driving their spatial abundance and gene modules, enabling cross-cell-type analysis of module-module interactions (bottom). This dual-output framework systematically decodes spatial cellular organization and intercellular communication networks.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/527a4cb5e46cedaf06f49bad.jpeg"},{"id":89805162,"identity":"ed399728-8af8-4d29-8976-6ab0080177ba","added_by":"auto","created_at":"2025-08-25 08:58:53","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1302827,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAnalyzing of PDAC and AD data.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA, Annotated spatial regions in the PDAC dataset.\u003c/p\u003e\n\u003cp\u003eB, Pie charts depicting inferred cell-type composition at each spatial spot, generated by SIGMOD.\u003c/p\u003e\n\u003cp\u003eC, Dot plot showing correlations between inferred cell type-specific modules proportions and corresponding cell-type-specific marker genes across spatial locations for each algorithm. The colors represent correlations for each cell type over all algorithms. The p-values indicate the significance level under one-sided Pearson rank test for positive correlation.\u003c/p\u003e\n\u003cp\u003eD, Heatmap displaying the influence of niche composition on cell type enrichment. Values indicate the change in cell type abundance per unit increase in niche activity.\u003c/p\u003e\n\u003cp\u003eE, Circle plot showing cell type-specific gene module interactions where each node represents a gene module and edge represent influence from one gene module to another. Edge width corresponding to interaction strength.\u003c/p\u003e\n\u003cp\u003eF, Dot plot visualizing the impact of LR interactions among cell type specific gene modules. Dot size and color intensity reflect interaction strength, calculated as the product of niche-cell type association (Figure 2E), niche-LR interaction relationships (Figure S2C), and mean ligand/receptor gene expression across cell types from single-cell RNA-seq data.\u003c/p\u003e\n\u003cp\u003eG, Annotated mainly cell types in the AD dataset.\u003c/p\u003e\n\u003cp\u003eH, Table displaying the correlation for each cell types across all spots computed for each method.\u003c/p\u003e\n\u003cp\u003eI, Circle plot showing cell type-specific gene module interactions in AD data.\u003c/p\u003e\n\u003cp\u003eJ, Dot plot visualizing the impact of LR interactions among cell type specific gene modules in AD data.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/e7c27ec4818852847c772b1f.png"},{"id":89805820,"identity":"dd96782b-98c3-4844-b9c5-54a12a4f196b","added_by":"auto","created_at":"2025-08-25 09:06:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1057084,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAnalyzing of CRC liver metastasis data.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA, UMAP visualization of the CRC liver metastasis spatial transcriptomics data, reduced using niche information and colored by sample labels.\u003c/p\u003e\n\u003cp\u003eB, UMAP visualization of the same dataset, reduced using SIGMOD inferred cell type proportions and colored by spatial hub labels.\u003c/p\u003e\n\u003cp\u003eC, UMAP visualization of the dataset reduced by SIGMOD inferred cell type proportions, colored by sample labels.\u003c/p\u003e\n\u003cp\u003eD, Dot plot illustrating enrichment of cell type-specific gene modules across spatial hubs. Dot size represents the frequency of cell type-specific gene module presence in each hub, and color indicates mean module abundance.\u003c/p\u003e\n\u003cp\u003eE, Heatmap showing the proportion of each spatial hub in primary (CRC) and liver metastasis samples.\u003c/p\u003e\n\u003cp\u003eF, Dot plot of KEGG pathways enriched in different spatial hubs.\u003c/p\u003e\n\u003cp\u003eG, Circle plot showing differential gene module-module interactions compared between liver metastasis with primary colorectal cancer.\u003c/p\u003e\n\u003cp\u003eH, Dot plot visualizing the impact of LR interactions among cell type specific gene modules in Figure 3G for L2 samples.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/30363840b5a878c10b395a33.png"},{"id":89805165,"identity":"d19f552e-a35c-4964-8e68-15f12276610e","added_by":"auto","created_at":"2025-08-25 08:58:53","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2486340,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAnalyzing of BRCA data.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA, Spatial plot of BRCA sample data acquired using Xenium, annotated with original cell type labels.\u003c/p\u003e\n\u003cp\u003eB, Boxplots of module coherence and module diversity scores of SIGMOD and the four competing methods obtained over five different runs with different seeds. In the box plot, the center line denotes the median, box limits denote the upper and lower quartiles and whiskers denote 1.5 × interquartile range.\u003c/p\u003e\n\u003cp\u003eC, Spatial plots showing spatial patterns of different gene modules’ activity.\u003c/p\u003e\n\u003cp\u003eD, Dot plot showing gene module activity across original annotated cell types. Dot size represents the percentage of cells within each annotated type assigned to the module, and color indicates mean module activity intensity.\u003c/p\u003e\n\u003cp\u003eE, Dot plot of Hallmark Pathways GSEA results enriched across modules. Dot color indicates normalized enrichment score (NES), and size represents −log₁₀(p-value).\u003c/p\u003e\n\u003cp\u003eF, Network plot of cross-cell-type gene module interactions. Nodes represent gene modules from different cell types; edge thickness reflects interaction strength.\u003c/p\u003e\n\u003cp\u003eG, Network plot highlighting cross-cell-type interactions mediated by specific ligand-receptor pairs.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/98cea1596b6b91d11f062e4b.png"},{"id":89805166,"identity":"ccbd0847-3c15-4647-a756-7414572e58f1","added_by":"auto","created_at":"2025-08-25 08:58:53","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1165837,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAnalyzing of NSCLC dataset.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA, Spatial plot of NSCLC sample data using the CosMx Spatial Molecular Imager (SMI), annotated with original cell type labels.\u003c/p\u003e\n\u003cp\u003eB, Heatmap showing neighborhood preference of different cell type.\u003c/p\u003e\n\u003cp\u003eC, Spatial plot showing tumor, neutrophil and fibroblasts associated gene module activity in corresponding cells. Color intensity reflects the strength of module activity.\u003c/p\u003e\n\u003cp\u003eD, Circle plot showing cross-cell-type gene module interactions. Nodes represent gene modules from different cell types and edge reflects interaction.\u003c/p\u003e\n\u003cp\u003eE, Scatter plot showing incoming and outgoing interaction strength for each gene module.\u003c/p\u003e\n\u003cp\u003eF, Dot plot showing GSEA enriched results of Hallmark Pathways across tumor, neutrophil and fibroblasts associated gene module. Dot color indicates normalized enrichment score (NES), and size represents −log₁₀(p-value).\u003c/p\u003e\n\u003cp\u003eG, Forest plot summarizing hazard ratios (HRs) and p-values for associations between different gene modules and survival in multiple LUAD cohorts. HRs are shown as means with 95% confidence intervals.\u003c/p\u003e\n\u003cp\u003eH, Line plot visualizing the impact of signature ligands from neutrophil gene module 2 on tumor cell gene module 1. Line thickness and color intensity represent interaction strength.\u003c/p\u003e\n\u003cp\u003eI, Line plot visualizing the impact of signature ligands from CD4 T memory gene module 2 and fibroblast gene module 2 on tumor cell gene module 3. Line thickness and color intensity represent interaction strength.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/459789f6b5406fb7ffb205e8.png"},{"id":89806988,"identity":"7ade07ee-b50f-4bde-83b7-6a9a3f4e70ef","added_by":"auto","created_at":"2025-08-25 09:22:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6230793,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/be01a0c2-11ba-4493-9508-c0940395d81a.pdf"},{"id":89805167,"identity":"4562ab7e-a232-4694-8ffa-9fc0bee577ae","added_by":"auto","created_at":"2025-08-25 08:58:53","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":5647051,"visible":true,"origin":"","legend":"Supplementary material","description":"","filename":"SuppleInfo.docx","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/1e633e18ba68aa34ffd9868e.docx"},{"id":89805176,"identity":"97f750bf-d93c-4f60-a6ac-bd174ce8cef6","added_by":"auto","created_at":"2025-08-25 08:58:54","extension":"zip","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":50069447,"visible":true,"origin":"","legend":"Supplementary table","description":"","filename":"SuppleTable.zip","url":"https://assets-eu.researchsquare.com/files/rs-7243521/v1/4aa26723d4d1d31b99c79d99.zip"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Deciphering Microenvironmental Heterogeneity by Scalable Niche Guided Module Discovery","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eStructurally organization of different cell types within tissues is crucial for their biological functions. Spatially proximal cells can influence the composition and activity of intracellular gene modules through various mechanisms, thereby shaping cellular states and functions. Recent advances in spatial transcriptomics (ST) have opened new avenues for the systematic and comprehensive analysis of gene module interactions within tissue contexts. Current ST technology could be categorized into two groups\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e: sequencing-based approaches (for example, Visium\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e, DBiT-Seq\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, Slide-Seq\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, Slide-tags\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e and Stereo-seq\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e) and image-based approaches (for example, SeqFISH\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e, RNAScope\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e, MERFISH\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, Xenium\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e and CosMx\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e), each offering unique advantages and complementary insights for studying spatially regulated gene activity. Sequencing based methods rely on predefined capture area, or spot and measure unbiased genome-wide transcriptomic profiles. Specifically, when integrated with single-cell transcriptomic data, each spot can be modeled as a mixture of cell type\u0026ndash;specific gene modules, analogous to a deconvolution process in this context. Image-based spatial transcriptomics approaches, which leverage high-resolution fluorescent imaging, achieve single-cell resolution but are typically restricted to predefined gene panels. In this context, each cell can be modeled as a mixture of gene modules that are specific to distinct cellular states and are often inferred de novo. Many computational algorithms have recently been proposed, with some focusing on the deconvolution of spatial transcriptomics data using single-cell RNA sequencing as a reference\u003csup\u003e\u003cspan additionalcitationids=\"CR13 CR14 CR15 CR16\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, while others aim to infer gene modules de novo to capture underlying transcriptional programs\u003csup\u003e\u003cspan additionalcitationids=\"CR19 CR20 CR21\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. However, these methods primarily rely on spatial coordinates and often overlook the broader microenvironmental context, limiting their ability to uncover interactions between gene modules that may drive complex cellular behaviors.\u003c/p\u003e\u003cp\u003eHere, we present SIGMOD (Scalable Niche-Guided Module Discovery), a computational framework designed to dissect gene modules and uncover interactions between them within the spatial microenvironment. SIGMOD frames gene module discovery as a gene expression reconstruction problem, incorporating the microenvironmental context as a prior to guide the underlying generative process. It not only reveals the composition and activity of gene modules but also elucidates how the microenvironment influences these modules, thereby uncovering interactions between them. This framework enhances the interpretability of spatial transcriptomic data and facilitates a deeper understanding of the regulatory architecture within complex tissue environments. SIGMOD effectively deconvolved cell type\u0026ndash;specific gene modules within each spot using 10x ST and Visium data, uncovering region-specific modules and inter-module interactions that drive disease processes such as Alzheimer\u0026rsquo;s disease and metastasis. In Xenium data, SIGMOD identified interpretable gene modules and characterized interactions among immunosuppressive programs. When applied to CosMX data at single-cell resolution, SIGMOD revealed how neighboring cell types influence the activity of distinct gene modules, highlighting neutrophils as key contributors to the hypoxic adaptation of tumor cells.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cb\u003eOverview of SIGMOD\u003c/b\u003e\u003c/p\u003e\u003cp\u003eSIGMOD comprises two key steps: niche construction and gene module inference (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). In this study, the niche is defined as the local microenvironment surrounding each cell or spot. SIGMOD offers two primary approaches for niche construction: spatially variable ligand-receptor interaction analysis and proximity-based cell enrichment. To construct niche defined with spatially variable ligand-receptor interaction, SIGMOD identifies co-localized ligand-receptor pairs by assuming that relatively high expression of a receptor in a specific spatial region, coupled with relatively high expression of its corresponding ligand in the same region, indicating co-localization of a pair of ligand and receptor and potential biologically relevant interactions. Then SIGMOD partitions the space into a grid of unannotated regions or uses annotated regions if provided. SIGMOD evaluates simultaneously enriched ligand-receptor pairs across regions and calculates the probability of ligand-receptor interactions across regions, identifying interactions with significant spatial specificity. The interaction of a pair of ligand and receptor is characterized by both co-localization enrichment and spatial specificity. To perform niche construction using proximity-based cell enrichment, SIGMOD defines cell\u0026rsquo;s niche based on the number and types of neighboring cells within a specified distance. Specifically, for imaging-based data with single-cell resolution (e.g., Xenium), SIGMOD quantifies the interaction strength of specific ligand\u0026ndash;receptor pairs within a defined spatial radius and integrates this interaction information into the modeling of the niche to guide gene module discovery.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFor the next step, gene module inference, SIGMOD integrates structural topic modeling\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e and deep generative modeling\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, using gene expression and niche information as inputs to predict module activity in spots or cells, the quantifiable effect of niche on module activity, and gene activity within modules. In SIGMOD, gene expression is modeled using a Gamma-Poisson distribution, with its parameters determined by module activity\u0026mdash;guided by niche information\u0026mdash;and gene activity, informed either by predefined cell type\u0026ndash;specific gene modules from single-cell RNA-seq data or inferred directly from the data in the case of de novo module discovery. SIGMOD outputs offer insights into gene module activities and cross\u0026ndash;cell-type module\u0026ndash;module interactions. Together, these capabilities make SIGMOD a powerful tool for decoding spatially resolved cellular niches and intercellular communication networks.\u003c/p\u003e\u003cp\u003e\u003cb\u003eAccurate Deconvolution of Regional Cell Type\u0026ndash;Specific Gene Module Distributions and Their Associated Ligand\u0026ndash;Receptor Interactions in Pancreatic Ductal Adenocarcinoma\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe first applied SIGMOD to a human pancreatic ductal adenocarcinoma (PDAC) spatial transcriptomic dataset (denotated as PDAC-A)\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, which contained annotated single cell RNA transcriptomics with manually annotated histological regions (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). This dataset was widely used for benchmarking deconvolution methods\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. SIGMOD identified spatial variable and regional specific ligand receptor interactions and demarcated the spatial structural heterogeneity of the ST data, which can be used for subsequent niche construction (Supplementary Fig.\u0026nbsp;1A-C). Guided by niche information, SIGMOD well characterized pancreatic and tumoral cell type-specific gene modules into different tissue regions (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). SIGMOD demonstrated statistically significant correlations between the inferred cell type-specific gene modules proportions and canonical marker genes for 19 cell types. Its performance was comparable to that of Cell2location and RCTD, while outperforming other existing methods (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC, Supplementary Fig.\u0026nbsp;1D, E). SIGMOD correctly identified region-specific cell type-specific gene modules, such as the localization of the two clones of cancer cells, ductal high hypoxic cells, and fibroblast cells in tumor region, as well as acinar cells and endocrine cells in pancreatic region (Supplementary Fig.\u0026nbsp;1F). The dominant cell type-specific modules proportions inferred by SIGMOD effectively captured the segregation between pancreatic, cancer and duct epithelium regions (Supplementary Fig.\u0026nbsp;1G). SIGMOD also distinguishes the enrichment of ductal high hypoxic cell-specific gene module within ductal epithelium, further validated by the specific gene expression of ductal high hypoxic cells, which was ignored by other methods (Supplementary Fig.\u0026nbsp;1F, H).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSIGMOD also uncovered cell type\u0026ndash;specific gene module\u0026ndash;module interactions by integrating the effects of the niche on gene module proportions, the composition of ligand\u0026ndash;receptor pairs within niches, and the ligand/receptor expression profiles of annotated cell types derived from paired single-cell data (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD). For example, two cancer clone\u0026ndash;specific gene modules and a fibroblast-specific gene module exhibited strong mutual interactions (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eE). Among the identified ligand\u0026ndash;receptor pairs, LAMB3\u0026ndash;CD151 and FGG\u0026ndash;ITGAV mainly affected those associated with cancer clone A. Additionally, THBS2\u0026ndash;NOTCH3 and FN1\u0026ndash;ITGA5 predominantly impacted fibroblast-specific gene modules (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eF). SIGMOD also identified CXCL12\u0026ndash;CD4 as a key interaction mediating effects on mDCs B\u0026ndash;specific gene modules (Supplementary Fig.\u0026nbsp;2I). These results demonstrate that SIGMOD not only accurately deconvolves cell type\u0026ndash;specific gene module proportions but also reveals the module\u0026ndash;module interactions that influence the spatial distribution and organization of distinct cell types.\u003c/p\u003e\u003cp\u003e\u003cb\u003eExcels in cell type\u0026ndash;specific gene module deconvolution and interaction inference in Alzheimer\u0026rsquo;s disease\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo further evaluate the accuracy of SIGMOD in identifying cell type\u0026ndash;specific gene modules, we applied it to 10x Visium data from an Alzheimer's disease (AD) sample\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, using Xenium data from a matched AD sample as the ground truth (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eG). SIGMOD identified distinct cellular niches in AD samples, defined by spatially variable and region-specific patterns of ligand\u0026ndash;receptor interactions, highlighting localized microenvironmental signaling landscapes (Supplementary Fig.\u0026nbsp;2A\u0026ndash;C). For quantitative assessment, we computed the correlation for each cell type across all spots, as well as the correlation and mean squared error (MSE) for each spot across all cell types. SIGMOD achieved the highest correlations for individual cell types across spots and performed particularly well for well-organized excitatory glutamatergic neurons (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eH). Additionally, it showed comparable performance to Spotiphy and outperformed other methods in both correlation and MSE when evaluated across all cell types per spot (Supplementary Fig.\u0026nbsp;2D, E). In addition, we found that the cell type\u0026ndash;specific gene module priors derived from niche information were well aligned with the final inferred module proportions. This concordance was particularly evident for major inhibitory GABAergic interneurons and glial cells, which correspond closely to well-defined niches (e.g., niche 4, 7, and 10) (Supplementary Fig.\u0026nbsp;2C, F).SIGMOD further inferred cell type\u0026ndash;specific gene module\u0026ndash;module interaction networks by integrating the effects of niche context on gene module proportions, the composition of ligand\u0026ndash;receptor pairs within niches, and ligand/receptor expression profiles from annotated cell types derived from paired single-cell data. Prominent interactions were observed among excitatory neurons (e.g., L5 PT CTX, L6 IT CTX), glial populations (e.g., microglia, astrocytes), and immune-related cells (e.g., macrophages, neutrophils). Notably, several cell types\u0026mdash;such as L5 IT CTX, LAMP5, and Pvalb-positive interneurons\u0026mdash;exhibited strong intra- and inter-type interactions, suggesting central roles in the cellular communication network (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eI). We further focused on microglia-centered interactions, revealing strong self-interaction of microglia and notable crosstalk with macrophages, subiculum (SUB), and neutrophils. The analysis also identified key ligand\u0026ndash;receptor pairs mediating these interactions. In particular, Apoe\u0026ndash;Trem2 and Sema6d\u0026ndash;Trem2 emerged as major contributors to microglia-associated signaling, implicating these axes in the regulatory roles of microglia within the Alzheimer\u0026rsquo;s disease microenvironment\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eJ). Together, these results demonstrate SIGMOD\u0026rsquo;s capability to uncover cell type\u0026ndash;specific gene modules and their associated interaction networks.\u003c/p\u003e\u003cp\u003e\u003cb\u003eComparison of Primary Colorectal Tumors and Liver Metastases Reveals Metastasis-Associated Gene Modules\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe next applied SIGMOD to 4 specimens of colorectal cancer (CRC) primary sites (C1 to C4) and 2 specimens of CRC liver metastatic tumors (L1 and L2) from a previous study using 10x Visium spatial platform\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Cells in the single cell transcriptome reference dataset from the original study were annotated with 9 cell types, which was further classified into 64 distinct, refined cell states. Using SIGMOD, we estimated the proportions of 64 refined cell states- specific gene modules of each spot. We first identified spatially variable ligand\u0026ndash;receptor (LR) interactions in each sample using SIGMOD. A substantial number of unique LR interactions were observed in individual samples, and the overlap of LR interactions across all six samples was minimal, indicating that each sample harbors distinct LR interaction patterns (Supplementary Fig.\u0026nbsp;3A). To incorporate LR interactions across the six samples, we retained interaction values only for spots exhibiting high LR interaction activity, setting the remaining values to zero. SIGMOD then performed non-negative matrix factorization on the LR interaction scores to derive niche information. This niche information was subsequently used for dimensionality reduction and clustering of spots. While some spots clustered homogeneously based on niche similarity, others exhibited substantial heterogeneity, reflecting diverse microenvironmental contexts (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, Supplementary Fig.\u0026nbsp;3B). SIGMOD utilized LR-based niche information to infer cell type\u0026ndash;specific gene modules. By leveraging these derived gene modules, SIGMOD successfully integrated all samples into a shared latent space (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB, C). Niches exhibited distinct distributions within the latent space, indicating strong niche specificity and their association with cell type\u0026ndash;specific gene modules (Supplementary Fig.\u0026nbsp;3C). To elucidate similarities and distinctions in cell type module composition across samples, we identified spatial hubs within the integrated space (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB, C). Most hubs were detected in multiple samples (Supplementary Fig.\u0026nbsp;3B, C), with their composition dominated by tumor cells. However, these hubs reflected distinct tumor cell subtypes, indicating that the latent space preserved tumor heterogeneity (Supplementary Fig.\u0026nbsp;3D, E). We also found that the cell type\u0026ndash;specific gene module priors derived from niche information were well aligned with the final inferred module proportions, indicating that niche information provides valuable guidance for gene module inference (Supplementary Fig.\u0026nbsp;3F).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn general, spatial hubs exhibited distinct combinations of cell type\u0026ndash;specific gene modules (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD). For example, Hub 1 was enriched for Endo_CLEC4G, Mac_CXCL9, CD8_TXNIP, and CD4_SELL, which were predominantly derived from metastatic tumors and was validated in matched single-cell datasets (Supplementary Fig.\u0026nbsp;3G). In contrast, Hub 7 was mainly composed of Tumor_AREG, which was specifically associated with primary tumors (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD, E). However, we also observed that some hubs from different tissue sections shared similar gene module compositions. For instance, Hub 3, enriched in liver metastatic tumors, and Hub 10, enriched in primary tumor sites, both showed high levels of Tumor_MKI67, suggesting that the key differences between primary and metastatic tumors may lie not only in the presence of specific gene modules, but in the patterns of interactions among them (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD, E). To further validate this observation, we first identified differentially expressed genes across spatial hubs and performed gene set enrichment analysis. Hub 3 showed enrichment for biological processes such as cholesterol metabolism and complement and coagulation cascades\u0026mdash;well-established hallmarks of tumor metastasis\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. In contrast, Hub 10 was primarily enriched for cell cycle\u0026ndash;related pathways. These results suggest that spatial hubs reflect coordinated activity of multiple gene modules, rather than being driven by a single module, in shaping distinct tumor microenvironments (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eF). Because SIGMOD could reveal gene module-module interactions analysis, that inspired us to perform differential gene module-module interactions analysis between liver metastasis and primary tumors. We found that interactions between Endo_CLEC4G, Fib_CXCL12, Fib_C3, Mac_CXCL9 and Tumor_GNG13 contribute most to the differential gene module-module interactions (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eG). These interactions involved adhesion-related pairs such as COL1A1\u0026ndash;ITGB1 and THBS1\u0026ndash;ITGA3, phagocytic interactions such as CD34\u0026ndash;SELL and CD99\u0026ndash;CD81, and APP family members such as APP\u0026ndash;SORL1, suggesting a metastasis-associated microenvironment characterized by enhanced cell\u0026ndash;matrix adhesion and tumor\u0026ndash;immune crosstalk\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eH, Supplementary Fig.\u0026nbsp;3H). Collectively, SIGMOD integrated primary and metastatic samples, uncovering distinct cell type\u0026ndash;specific gene modules, spatial hub architectures, and context-dependent ligand\u0026ndash;receptor interactions that drive cellular heterogeneity.\u003c/p\u003e\u003cp\u003e\u003cb\u003eUnveiling Immunosuppressive Gene Modules Interactions in the Breast Cancer Microenvironment\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe next applied SIGMOD to a high-resolution ST dataset from 10x Xenium of human breast cancer (BC)\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. BC exhibited heterogeneity with significant variations in molecular characteristics and cell type distributions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). Here, known ligand\u0026ndash;receptor interactions were used to construct niche definitions as prior for gene modules, and SIGMOD was configured to infer 20 gene modules, in accordance with the annotations reported in the original study\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. We first evaluated the performance of SIGMOD and competing methods in identifying gene modules using five quantitative metrics involved module coherence, module diversity, and topic sparsity\u003csup\u003e\u003cspan additionalcitationids=\"CR35\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB, Supplementary Fig.\u0026nbsp;4A). These metrics consistently ranked SIGMOD among the top-performing methods, and often as the best. SIGMOD achieved the highest scores in both module coherence and module diversity, indicating its strong ability to capture tightly co-expressed gene sets within modules and to generate distinct, non-redundant gene modules. Visual inspection of SIGMOD\u0026rsquo;s output revealed well-defined gene modules that spatially corresponded to distinct tissue regions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC). Some gene modules exhibited strong activity across multiple cell types, suggesting potential shared functions, while others were specifically active in distinct cell types (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eD). By assigning each cell to the gene module with the highest activity, the resulting confusion matrix showed that most gene modules were highly specific to individual cell types, although some modules captured shared transcriptional features among related cell types (Supplementary Fig.\u0026nbsp;4B). For example, gene module 5 was associated with both T cells and B cells, reflecting common immune-related programs, whereas gene module 15 was specifically restricted to KRT15⁺ myoepithelial cells. Most cell type markers were enriched within their corresponding gene modules, although some modules encompassed markers from multiple cell types in alignment with the gene modules\u0026rsquo; activities within cells (Supplementary Fig.\u0026nbsp;4C). Each gene module was characterized by representative marker genes and enriched for specific biological functions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eE, Supplementary Fig.\u0026nbsp;4D). For example, gene modules 4, 8, 12, and 17 exhibited similar spatial patterns and were all associated with tumor-related processes; however, they differed in their marker genes and enriched pathways, pointing to distinct biological roles. Gene module 4 was represented by GATA3, an estrogen associated marker\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e and was enriched in estrogen response early. Gene module 8, marked by CCND1\u003csup\u003e38\u003c/sup\u003e, involved in androgen-mediated DNA damage repair, was associated with Notch signaling and androgen response. Gene module 12 was defined by FASN\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e, an established marker of tumor invasiveness, and was enriched for cholesterol homeostasis and mTORC1 signaling. In contrast, gene module 16 was characterized by cell cycle\u0026ndash;related genes such as PCLAF, TOP2A, and CENPF, and was enriched for the G2/M checkpoint pathway. These results demonstrate that while spatially co-expressed, gene modules can reflect distinct molecular programs, providing fine-grained insights into tumor heterogeneity.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe further investigated cross-cell-type gene module interactions. Spatial correlation analysis revealed three distinct correlation patterns among gene modules: (1) modules active within the same cell type, such as tumor-associated modules 4, 8, 12, 18, and 20; (2) DCIS- and myoepithelial-related modules, including 1, 11, 15, and 17; and (3) additional patterns involving other gene modules (Supplementary Fig.\u0026nbsp;4E). Niche effect analysis showed that specific ligand\u0026ndash;receptor interactions regulate the activity of distinct gene modules. For example, S100A4\u0026ndash;EGFR significantly influenced the activity of modules 10 and 13, while CD274\u0026ndash;CD80 prominently affected module 2 (Supplementary Fig.\u0026nbsp;4F). These ligand\u0026ndash;receptor pairs were also selectively enriched within corresponding gene modules (Supplementary Fig.\u0026nbsp;4G). By integrating spatial correlations, niche-derived influences, and ligand assignment in modules, SIGMOD reconstructed a network of cross-cell-type gene module interactions in breast cancer (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eF). Notably, SIGMOD identified crosstalk between myeloid-related modules 2, 6, and 7 and lymphoid-associated module 5, mediated by immune checkpoint interactions such as CD86\u0026ndash;CTLA4, PDCD1LG2\u0026ndash;PDCD1, and CD274\u0026ndash;PDCD1, highlighting immunosuppressive signaling within the breast cancer microenvironment (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eG).\u003c/p\u003e\u003cp\u003e\u003cb\u003eClinically relevant Cross-Cell-Type Gene Module Interactions in Non-Small Cell Lung Cancer\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFinally, we evaluated SIGMOD\u0026rsquo;s ability to identify cell states and their corresponding gene modules within single-cell resolution spatial transcriptomics data, as well as to characterize how these states are influenced by the surrounding microenvironment. To this end, we applied SIGMOD to a human non-small cell lung cancer (NSCLC) dataset generated using the CosMx SMI platform\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. The selected sample, Lung #5\u0026thinsp;\u0026minus;\u0026thinsp;1, comprised 87,606 spots with expression profiles for 960 genes. (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). Analysis of the heterogeneity in surrounding cells revealed a preferential enrichment of certain cell types in the vicinity of specific cell populations (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). For example, tumor cells exhibited substantial heterogeneity in their neighboring composition, with varying degrees of association with fibroblasts and neutrophils. These findings suggest that SIGMOD can be effectively used to uncover distinct cellular states within a given cell type. In this analysis, neighboring cell types were treated as spatial niches to guide gene module discovery. The cell type\u0026ndash;specific gene module distributions inferred by SIGMOD further indicated that different cellular states exhibit distinct spatial preferences, reflecting underlying microenvironmental influences (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC, Supplementary Fig.\u0026nbsp;5A). For example, the spatial distributions of the three-tumor cell\u0026ndash;associated gene modules varied markedly: tumor module 1 was enriched in the tumor core, module 2 was concentrated within the tumor interior, and module 3 was primarily localized at the tumor margins (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). Based on the influence of surrounding cell types on gene module activity and the spatial correlations between different gene modules, SIGMOD further constructed a cross-cell-type gene module interaction network, in which each node represents a cell type\u0026ndash;associated gene module and each edge denotes the activation of one gene module by another (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eD). As shown in the figure, the neutrophil-associated gene module 2 exhibited strong interactions with multiple other gene modules, particularly those associated with tumor cells. Analysis of the incoming and outgoing interaction strengths of each module revealed that neutrophil-associated gene modules exhibited both high incoming and outgoing connectivity. Similarly, tumor-associated, and fibroblast-associated gene modules also showed strong bidirectional interaction strengths, indicating their central roles in cross-cell-type communication within the tumor microenvironment (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eE). Neutrophil-associated, tumor-associated, and fibroblast-associated gene modules also exhibited similar spatial distributions (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eCross-cell-type gene module interaction analysis of tumor-associated gene modules reveals that different tumor modules were influenced by distinct modules from other cell types (Supplementary Fig.\u0026nbsp;5B). For example, SIGMOD inferred that the activity of tumor gene module 1 was upregulated by neutrophil-associated gene module 2, while tumor gene module 2 was influenced by fibroblast-associated gene module 3. Additionally, tumor gene module 3 was regulated by multiple sources, including neutrophil-associated gene modules 1 and 2, fibroblast-associated gene module 2, and a T cell\u0026ndash;associated gene module (Supplementary Fig.\u0026nbsp;5B). These inferred regulatory relationships are consistent with the observed spatial co-localization of tumor cells with fibroblasts and neutrophils, reflecting varying degrees of microenvironmental influence across tumor cell states (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Gene set enrichment analysis revealed that tumor gene module 1 was associated with hypoxia, tumor gene module 2 with the cell cycle, and tumor gene module 3 with the inflammatory response (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eF). These results suggest that neutrophil-associated gene module 2 promotes hypoxic adaptation, fibroblast-associated gene module 3 stimulates cell proliferation, and neutrophil, fibroblast, and T cell\u0026ndash;associated gene modules are involved in driving the inflammatory response. Indeed, neutrophil gene module 1 was enriched in inflammatory pathways, consistent with tumor gene module 3, while fibroblast gene module showed enrichment in angiogenesis and E2F target genes, both of which may contribute to tumor proliferation (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eF). Additionally, these modules exhibited distinct representative genes (Supplementary Fig.\u0026nbsp;5C-E). In an independent non-small cell lung cancer dataset\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e (Supplementary Fig.\u0026nbsp;5F), the scores for different tumor, neutrophil, and fibroblast modules inferred by SIGMOD successfully stratified the corresponding cells into distinct subgroups. This demonstrates the broad applicability, representativeness, and heterogeneity of these modules (Supplementary Fig.\u0026nbsp;5G). Neutrophil gene module 2 also showed a significant Pearson correlation with tumor gene module 1 when calculated across all patients (Supplementary Fig.\u0026nbsp;5H), indicating a strong association between neutrophil gene module 2 and tumor gene module 1. We next validated the clinical relevance of these modules in multiple independent LUAD cohorts\u003csup\u003e\u003cspan additionalcitationids=\"CR41 CR42 CR43 CR44 CR45 CR46 CR47 CR48 CR49 CR50 CR51 CR52\" citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. We found that hypoxia-related tumor gene module 1 and its associated neutrophil gene module 2 were linked to poor prognosis. In contrast, the inflammatory gene module, especially neutrophil gene module 1 and fibroblast gene module 2, was associated with better prognosis (hypoxia-related tumor gene module 1 was associated with poor prognosis in 6/15 datasets, neutrophil gene module 2 in 6/15, neutrophil gene module 1 in 3/15, and fibroblast gene module 2 in 7/15 datasets) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eG). These findings suggest that the interaction between specific tumor-associated modules and their microenvironmental counterparts could serve as valuable prognostic markers. SIGMOD further investigated the ligands that influence tumor module activity. For example, the hypoxia-related tumor gene module 1 identified neutrophil-derived CXCL8 as a potential regulatory factor, particularly targeting NDRG1 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eH). CXCL8 has previously been implicated in hypoxia adaptation\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e. Additionally, TIMP1 from CD4 T memory gene module 2 may modulate MHC class II expression in tumor gene module 3 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eI). Moreover, multiple ligands from fibroblast gene module 3, such as LGALS1, were found to contribute to the proliferation properties of tumor gene module 2 (Supplementary Fig.\u0026nbsp;5I). These results demonstrate that SIGMOD can resolve cross-cell-type gene module interactions with clinical relevance.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eIn spatial transcriptomics, genes expressed in each spot or cell are organized into coherent modules that reflect structured patterns of gene expression. Understanding module–module interactions is crucial for elucidating tissue architectures and their functional roles. Gene modules can be classified as either cell type-specific or cell state-specific, depending on the situations, especially the spatial transcriptomics technology used. The former is typically associated with sequencing-based technologies such as 10X ST or 10X Visium, which have lower resolution, and module discovery in this context is analogous to cell type deconvolution. In contrast, cell state-specific modules are usually derived from high-resolution, imaging-based technologies like Nanostring CosMx or Xenium, where modules represent diverse gene expression patterns, and their activity reflects the functional state of specific cell types.\u003c/p\u003e\u003cp\u003eMoreover, in multicellular organisms, cells do not function in isolation; their abundance and state are influenced by the surrounding microenvironment. To consider this, we developed SIGMOD, a method designed to identify functional modules within spatial transcriptomics data. To capture the impact of the microenvironment on these modules, SIGMOD first constructs niche information for each spot or cell based on ligand–receptor interactions or proximity-based cell enrichment. We hypothesize that ligand–receptor interactions play a crucial role in processes such as cell migration and downstream signaling, while neighboring cells also influence the state of the central cell. Using this niche information, SIGMOD infers gene modules either from reference single-cell RNA-seq data or via de novo discovery. This approach not only reveals gene module activity and assignments but also explores the spatial factors driving these changes, including cross-cell-type module–module interactions. SIGMOD’s versatility across sequencing-based (10X ST/Visium) and imaging-based (Xenium/CosMx) platforms highlights its ability to deconvolve modules at varying resolutions, offering insights into both cell type–specific (e.g., cancer clone modules in PDAC) and state-specific (e.g., hypoxia-responsive modules in lung cancer) patterns.\u003c/p\u003e\u003cp\u003eThere are several limitations for SIGMOD, including its current inability to incorporate additional data types like H\u0026amp;E histopathology images or pathology-defined lesion regions for capturing transcriptomic features linked to phenotypic traits. It also lacks integration of prior knowledge, such as gene sets, pathways, or cell types, to guide module construction, potentially limiting accuracy and relevance. Additionally, SIGMOD does not yet support analysis of temporal data or spatial multi-omics datasets, restricting its utility in fully understanding complex biological systems.\u003c/p\u003e\u003cp\u003eThe core concept of SIGMOD, integrating microenvironment data for gene module discovery and interaction analysis, makes it adaptable to diverse spatial omics datasets. As spatial transcriptomics evolves, SIGMOD’s flexibility and integration with emerging technologies will ensure its ongoing impact.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cp\u003e\u003cb\u003eSIGMOD model\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe developed SIGMOD, a Bayesian model that integrates spatial transcriptomics with local microenvironmental context to infer cell- or gene module–specific expression patterns across tissues. SIGMOD operates in two key stages: niche construction and gene module deconvolution. Niche features are derived from spatially informative signals such as ligand–receptor interactions or neighboring cell enrichment, depending on the resolution and platform of the dataset. These niche features guide the inference of module proportions through a sparse softmax transformation, and spatial coherence is enforced by a graph Laplacian regularization term constructed from spatial adjacency. Inference is performed using Automatic Differentiation Variational Inference (ADVI) implemented in Pyro.\u003c/p\u003e\u003cp\u003e\u003cb\u003eNiche Construction\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo incorporate local microenvironmental context, SIGMOD constructs niche features for each spatial location or cell. Depending on the resolution and data type, niche construction is based on either (i) spatially variable ligand–receptor interactions, identified through co-localization and region-specific enrichment statistics followed by dimensionality reduction using non-negative matrix factorization (NMF); (ii) neighboring cell-type enrichment, where the frequency of adjacent cell types is used as niche features; or (iii) cell–cell ligand–receptor interactions calculated across spatially adjacent cells. Spatial adjacency graphs are constructed using k-nearest neighbors for regular grids or Delaunay triangulation for irregular cell positions.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModule Inference\u003c/b\u003e\u003c/p\u003e\u003cp\u003eSIGMOD performs module inference using a Bayesian framework that models gene expression as a mixture of gene modules modulated by local niche features. It supports both reference-based and reference-free modes: the former leverages prior single-cell transcriptomic profiles to define module signatures, while the latter infers them directly from the data. Module proportions for each spot or cell are computed via a sparse softmax transformation applied to a niche-weighted linear predictor, regularized using a structured horseshoe prior. To ensure spatial coherence, SIGMOD incorporates a graph Laplacian–based regularization term, which penalizes sharp transitions in module proportions across neighboring locations.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModule Interaction Analysis\u003c/b\u003e\u003c/p\u003e\u003cp\u003eBased on the inferred module proportions and niche influences, SIGMOD enables downstream analysis of gene module–module interactions. In ligand–receptor–based niches, the contribution of each interaction to module expression is quantified by combining learned niche weights with ligand and receptor expression levels from reference scRNA-seq data. In cell–type–based niches, cross-cell-type module dependencies are inferred by correlating module proportions between neighboring cells and linking them through niche-to-module influence weights. This framework allows SIGMOD to reveal spatial communication patterns and biologically meaningful interactions across cell types and gene programs.\u003c/p\u003e\u003cp\u003eThe details of SIGMOD are provided in Supplementary Note.\u003c/p\u003e\u003cp\u003e\u003cb\u003eBenchmarking and Comparison Methods\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo evaluate the performance of SIGMOD, we compared it against several existing methods for spatial transcriptomics analysis. For reference-based gene module discovery, we included Tangram, SPOTlight, Cell2location, CARD, RCTD, and Spotiphy, all of which utilize single-cell RNA-seq as reference to infer cell-type compositions across spatial locations. For reference-free gene module discovery, we compared SIGMOD with STAMP, NMF, LDA, and LDVAE, which infer spatial gene programs without requiring external reference data. All methods were run using their recommended parameters and official tutorials to ensure fair comparison. The details of comparison are provided in Supplementary Note.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDatasets and Preprocessing\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe applied SIGMOD to multiple spatial transcriptomics datasets across diverse platforms and biological contexts, including human pancreatic ductal adenocarcinoma (PDAC, 10x ST), colorectal cancer with liver metastasis (CRC, 10x Visium), breast cancer (Xenium), and non-small cell lung cancer (CosMx SMI), as well as mouse Alzheimer’s disease tissue (10x Visium). For each dataset, standard preprocessing steps were applied, including gene and spot/cell filtering based on expression thresholds. When available, matched single-cell RNA-seq data were used for reference-based gene module inference. Niche construction strategies were selected according to data resolution, using ligand–receptor interactions, neighboring cell enrichment, or cell–cell interactions depending on the dataset. The details of real datasets analysis in Supplementary Note.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eSUPPLEMENTARY DATA\u003c/h2\u003e\u003cp\u003eSupplementary Data are available at NAR online.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003ch2\u003eCONFLICT OF INTEREST\u003c/h2\u003e\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFUNDING\u003c/h2\u003e\u003cp\u003eThis work was partially supported by the National Natural Science Foundation of China (81972680 to X.L.), the Tsinghua University Peking University Joint Center for Life Sciences (61020100119 to X.L.; postdoctoral fellowship to L.X.), the Beijing Natural Science Foundation (20201100463 to X.L.), the Damo Academy through the Damo Academy Innovative Research Program, and the Dushi Program.\u003c/p\u003e\u003ch2\u003eAUTHOR CONTRIBUTIONS\u003c/h2\u003e\u003cp\u003eX.L. and C.L. conceived the project. C.L. designed the models. C.L. and Y.Z. implemented the code. C.L., Y.Z., L.X. and X.Q. performed the analyses. T.L., C.T., and J.L. helped in designing the analyses and explaining the results. X.L. supervised the project. All authors contributed to the writing of the manuscript.\u003c/p\u003e\u003ch2\u003eACKNOWLEDGEMENTS\u003c/h2\u003e\u003cp\u003eWe acknowledge the High-Performance Computing Core Facility of School of Basic Medical Sciences, Tsinghua University, for computational support.\u003c/p\u003e\u003ch2\u003eDATA AVAILABILITY\u003c/h2\u003e\u003cp\u003eThis paper analyzes existing, publicly available data. The PDAC data was downloaded from the Gene Expression Omnibus under accession number GSE111672. The AD data was downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.10520022\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.10520022\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The CRC liver metastasis data was downloaded from the Gene Expression Omnibus under accession number GSE225857. The BRCA data was accessed from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.10xgenomics.com/products/xenium-in-situ/preview-dataset-human-breast\u003c/span\u003e\u003cspan address=\"https://www.10xgenomics.com/products/xenium-in-situ/preview-dataset-human-breast\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The NSCLC data was from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://nanostring.com/products/cosmx-spatial-molecular-imager/nsclc-ffpe-dataset\u003c/span\u003e\u003cspan address=\"https://nanostring.com/products/cosmx-spatial-molecular-imager/nsclc-ffpe-dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The single-cell lung cancer atlas (LuCA) was accessed from CELLxGENE.\u003c/p\u003e\u003ch2\u003eCode availability\u003c/h2\u003e\u003cp\u003eThe source code for SIGMOD is freely available online at GitHub at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Boxedpig/SIGMOD\u003c/span\u003e\u003cspan address=\"https://github.com/Boxedpig/SIGMOD\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.16350291\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.16350291\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCheng, M., et al.: Spatially resolved transcriptomics: a comprehensive review of their technological advances, applications, and challenges. J. Genet. Genomics. \u003cb\u003e50\u003c/b\u003e, 625\u0026ndash;640 (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVisium Spatial Platform: \u003cem\u003e10x Genomics\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.10xgenomics.com/platforms/visium\u003c/span\u003e\u003cspan address=\"https://www.10xgenomics.com/platforms/visium\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, Y., et al.: High-Spatial-Resolution Multi-Omics Sequencing via Deterministic Barcoding in Tissue. Cell. \u003cb\u003e183\u003c/b\u003e, 1665\u0026ndash;1681e18 (2020)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRodriques, S.G., et al.: Slide-seq: A scalable technology for measuring genome-wide expression at high spatial resolution. Science. (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1126/science.aaw1219\u003c/span\u003e\u003cspan address=\"10.1126/science.aaw1219\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRussell, A.J.C., et al.: Slide-tags enables single-nucleus barcoding for multimodal spatial genomics. Nature. \u003cb\u003e625\u003c/b\u003e, 101\u0026ndash;109 (2024)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, A., et al.: Spatiotemporal transcriptomic atlas of mouse organogenesis using DNA nanoball-patterned arrays. Cell. \u003cb\u003e185\u003c/b\u003e, 1777\u0026ndash;1792e21 (2022)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEng, C.-H.L., et al.: Transcriptome-scale super-resolved imaging in tissues by RNA seqFISH+. Nature. \u003cb\u003e568\u003c/b\u003e, 235\u0026ndash;239 (2019)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, F., et al.: RNAscope: A Novel in Situ RNA Analysis Platform for Formalin-Fixed, Paraffin-Embedded Tissues. J. Mol. Diagn. \u003cb\u003e14\u003c/b\u003e, 22\u0026ndash;29 (2012)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, K.H., Boettiger, A.N., Moffitt, J.R., Wang, S., Zhuang, X.: Spatially resolved, highly multiplexed RNA profiling in single cells. Science. \u003cb\u003e348\u003c/b\u003e, aaa6090 (2015)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXenium In Situ Platform: \u003cem\u003e10x Genomics\u003c/em\u003e (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.10xgenomics.com/platforms/xenium\u003c/span\u003e\u003cspan address=\"https://www.10xgenomics.com/platforms/xenium\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe, S., et al.: High-plex imaging of RNA and proteins at subcellular resolution in fixed tissue by spatial molecular imaging. Nat. Biotechnol. \u003cb\u003e40\u003c/b\u003e, 1794\u0026ndash;1806 (2022)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKleshchevnikov, V., et al.: Cell2location maps fine-grained cell types in spatial transcriptomics. Nat. Biotechnol. \u003cb\u003e40\u003c/b\u003e, 661\u0026ndash;671 (2022)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMa, Y., Zhou, X.: Spatially informed cell-type deconvolution for spatial transcriptomics. Nat. Biotechnol. \u003cb\u003e40\u003c/b\u003e, 1349\u0026ndash;1359 (2022)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCable, D.M., et al.: Robust decomposition of cell type mixtures in spatial transcriptomics. Nat. Biotechnol. \u003cb\u003e40\u003c/b\u003e, 517\u0026ndash;526 (2022)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eElosua-Bayes, M., Nieto, P., Mereu, E., Gut, I., Heyn, H.: SPOTlight: seeded NMF regression to deconvolute spatial transcriptomics spots with single-cell transcriptomes. Nucleic Acids Res. \u003cb\u003e49\u003c/b\u003e, e50 (2021)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBiancalani, T., et al.: Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram. Nat. Methods. \u003cb\u003e18\u003c/b\u003e, 1352\u0026ndash;1362 (2021)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang, J., et al.: Spotiphy enables single-cell spatial whole transcriptomics across an entire section. Nat. Methods. \u003cb\u003e1\u0026ndash;13\u003c/b\u003e (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41592-025-02622-5\u003c/span\u003e\u003cspan address=\"10.1038/s41592-025-02622-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhong, C., Ang, K.S., Chen, J.: Interpretable spatially aware dimension reduction of spatial transcriptomics with STAMP. Nat. Methods. \u003cb\u003e21\u003c/b\u003e, 2072\u0026ndash;2083 (2024)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, H., et al.: CellNeighborEX: deciphering neighbor-dependent gene expression from spatial transcriptomics data. Mol. Syst. Biol. \u003cb\u003e19\u003c/b\u003e, e11670 (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe, S., et al.: Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor\u0026ndash;immune hubs. Nat. Biotechnol. \u003cb\u003e1\u0026ndash;13\u003c/b\u003e (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41587-024-02173-8\u003c/span\u003e\u003cspan address=\"10.1038/s41587-024-02173-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFischer, D.S., Schaar, A.C., Theis, F.J.: Modeling intercellular communication in tissues using spatial graphs of cells. Nat. Biotechnol. \u003cb\u003e41\u003c/b\u003e, 332\u0026ndash;336 (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSvensson, V., Gayoso, A., Yosef, N., Pachter, L.: Interpretable factor models of single-cell RNA-seq via variational autoencoders\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, L., Mankad, S.A.: Structural Topic Sentiment Model for Text Analysis. SSRN Scholarly Paper at https:// (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003edoi.org/10.2139/ssrn.4020651\u003c/span\u003e\u003cspan address=\"10.2139/ssrn.4020651\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBingham, E., et al.: Pyro: Deep Universal Probabilistic Programming. Preprint at (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.1810.09538\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1810.09538\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMoncada, R., et al.: Integrating microarray-based spatial transcriptomics and single-cell RNA-seq reveals tissue architecture in pancreatic ductal adenocarcinomas. Nat. Biotechnol. \u003cb\u003e38\u003c/b\u003e, 333\u0026ndash;342 (2020)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, H., et al.: A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics. Nat. Commun. \u003cb\u003e14\u003c/b\u003e, 1548 (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKrasemann, S., et al.: The TREM2-APOE Pathway Drives the Transcriptional Phenotype of Dysfunctional Microglia in Neurodegenerative Diseases. Immunity. \u003cb\u003e47\u003c/b\u003e, 566\u0026ndash;581e9 (2017)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlbanus, R.D., et al.: Systematic analysis of cellular crosstalk reveals a role for SEMA6D-TREM2 regulating microglial function in Alzheimer\u0026rsquo;s disease. 11.11.516215 Preprint at (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1101/2022.11.11.516215\u003c/span\u003e\u003cspan address=\"10.1101/2022.11.11.516215\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, F., et al.: Single-cell and spatial transcriptome analysis reveals the cellular heterogeneity of liver metastatic colorectal cancer. Sci. Adv. \u003cb\u003e9\u003c/b\u003e, eadf5464 (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao, H., et al.: Inflammation and tumor progression: signaling pathways and targeted intervention. Sig Transduct. Target. Ther. \u003cb\u003e6\u003c/b\u003e, 1\u0026ndash;46 (2021)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLuo, X., et al.: Emerging roles of lipid metabolism in cancer metastasis. Mol. Cancer. \u003cb\u003e16\u003c/b\u003e, 76 (2017)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, Z., et al.: Focal adhesion in the tumour metastasis: from molecular mechanisms to therapeutic targets. Biomark. Res. \u003cb\u003e13\u003c/b\u003e, 38 (2025)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJanesick, A., et al.: High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis. Nat. Commun. \u003cb\u003e14\u003c/b\u003e, 8353 (2023)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBouma, G.: Normalized (Pointwise) Mutual Information in Collocation Extraction\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWebberWilliam, M.A.: ZobelJustin. A similarity measure for indefinite rankings. ACM Trans. Inform. Syst. (TOIS). (2010). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/1852102.1852106\u003c/span\u003e\u003cspan address=\"10.1145/1852102.1852106\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHoyer, P.: O. Non-negative matrix factorization with sparseness constraints\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAsch-Kendrick, R., Cimino-Mathews, A.: The role of GATA3 in breast carcinomas: a review. Hum. Pathol. \u003cb\u003e48\u003c/b\u003e, 37\u0026ndash;47 (2016)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eComstock, C.E.S., et al.: Cyclin D1 Is a Selective Modifier of Androgen-dependent Signaling and Androgen Receptor Function. J. Biol. Chem. \u003cb\u003e286\u003c/b\u003e, 8117\u0026ndash;8127 (2011)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSalcher, S., et al.: High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Cancer Cell. \u003cb\u003e40\u003c/b\u003e, 1503\u0026ndash;1520e8 (2022)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFouret, R., et al.: A Comparative and Integrative Approach Identifies ATPase Family, AAA Domain Containing 2 as a Likely Driver of Cell Proliferation in Lung Adenocarcinoma. Clin. Cancer Res. \u003cb\u003e18\u003c/b\u003e, 5606\u0026ndash;5616 (2012)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBotling, J., et al.: Biomarker Discovery in Non\u0026ndash;Small Cell Lung Cancer: Integrating Gene Expression Profiling, Meta-analysis, and Tissue Microarray Validation. Clin. Cancer Res. \u003cb\u003e19\u003c/b\u003e, 194\u0026ndash;204 (2013)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCollisson, E.A., et al.: Comprehensive molecular profiling of lung adenocarcinoma. Nature. \u003cb\u003e511\u003c/b\u003e, 543\u0026ndash;550 (2014)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchabath, M.B., et al.: Differential association of STK11 and TP53 with KRAS mutation-associated gene expression, proliferation and immune surveillance in lung adenocarcinoma. Oncogene. \u003cb\u003e35\u003c/b\u003e, 3209\u0026ndash;3216 (2016)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWilkerson, M.D., et al.: Differential Pathogenesis of Lung Adenocarcinoma Subtypes Involving Sequence Mutations, Copy Number, Chromosomal Instability, and Methylation. PLoS One. \u003cb\u003e7\u003c/b\u003e, e36530 (2012)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRousseaux, S., et al.: Ectopic Activation of Germline and Placental Genes Identifies Aggressive Metastasis-Prone Lung Cancers. Sci. Transl. Med. \u003cb\u003e5\u003c/b\u003e(186ra66\u0026ndash;186ra), 66 (2013)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTakeuchi, T., et al.: Expression Profile\u0026ndash;Defined Classification of Lung Adenocarcinoma Shows Close Relationship With Underlying Major Genetic Changes and Clinicopathologic Behaviors. \u003cem\u003eJCO\u003c/em\u003e 24, 1679\u0026ndash;1688 (2006)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLarsen, J.E., et al.: Gene Expression Signature Predicts Recurrence in Lung Adenocarcinoma. Clin. Cancer Res. \u003cb\u003e13\u003c/b\u003e, 2946\u0026ndash;2954 (2007)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShedden, K., et al.: Gene Expression-Based Survival Prediction in Lung Adenocarcinoma: A Multi-Site, Blinded Validation Study. Nat. Med. \u003cb\u003e14\u003c/b\u003e, 822\u0026ndash;827 (2008)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSato, M., et al.: Human lung epithelial cells progressed to malignancy through specific oncogenic manipulations. Mol. Cancer Res. \u003cb\u003e11\u003c/b\u003e, 638\u0026ndash;650 (2013)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOkayama, H., et al.: Identification of Genes Upregulated in ALK-Positive and EGFR/KRAS/ALK-Negative Lung Adenocarcinomas. Cancer Res. \u003cb\u003e72\u003c/b\u003e, 100\u0026ndash;111 (2012)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTomida, S., et al.: Relapse-Related Molecular Signature in Lung Adenocarcinomas Identifies Patients With Dismal Prognosis. \u003cem\u003eJCO\u003c/em\u003e 27, 2793\u0026ndash;2799 (2009)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDer, S.D., et al.: Validation of a Histology-Independent Prognostic Gene Signature for Early-Stage, Non\u0026ndash;Small-Cell Lung Cancer Including Stage IA Patients. J. Thorac. Oncol. \u003cb\u003e9\u003c/b\u003e, 59\u0026ndash;64 (2014)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBild, A.H., et al.: Oncogenic pathway signatures in human cancers as a guide to targeted therapies. Nature. \u003cb\u003e439\u003c/b\u003e, 353\u0026ndash;357 (2006)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCambier, S., Gouwy, M., Proost, P.: The chemokines CXCL8 and CXCL12: molecular and functional properties, role in disease and efforts towards pharmacological intervention. Cell. Mol. Immunol. \u003cb\u003e20\u003c/b\u003e, 217\u0026ndash;251 (2023)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7243521/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7243521/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSpatial transcriptomics provides high-dimensional gene expression data while preserving spatial context, offering novel insights into tissue composition and heterogeneity. Each spot or cell in the spatial transcriptome could be reflected as gene modules influenced by its surrounding microenvironment, with module interactions vital for tissue architecture and function. Here, we present Scalable Niche Guided Module Discovery (SIGMOD), a method that integrates prior constructed microenvironment information with gene expression decompositions to uncover gene modules, enabling a deeper understanding of crosstalk within the microenvironment. SIGMOD identifies cell type\u0026ndash;specific and cell state\u0026ndash;specific, clinically relevant gene modules, uncovering gene module\u0026ndash;module interactions in 10X ST, Visium, Xenium, and CosMX data, demonstrating its effectiveness and broad applicability.\u003c/p\u003e","manuscriptTitle":"Deciphering Microenvironmental Heterogeneity by Scalable Niche Guided Module Discovery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-25 08:58:49","doi":"10.21203/rs.3.rs-7243521/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"communications-biology","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"commsbio","sideBox":"Learn more about [Communications Biology](http://www.nature.com/commsbio/)","snPcode":"","submissionUrl":"","title":"Communications Biology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Communications Series","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"7a846c08-d031-4977-89c6-c95691ef661e","owner":[],"postedDate":"August 25th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":53249704,"name":"Biological sciences/Computational biology and bioinformatics/Computational models"},{"id":53249705,"name":"Biological sciences/Computational biology and bioinformatics/Gene regulatory networks"}],"tags":[],"updatedAt":"2026-04-30T05:35:23+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-25 08:58:49","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7243521","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7243521","identity":"rs-7243521","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}