Tuesday, Aug 4: 10:30 AM - 12:20 PM
6426
Contributed Papers
Thomas M. Menino Convention & Exhibition Center
Room: CC-254A
Main Sponsor
Section on Statistics in Genomics and Genetics
Presentations
Recent advancements in technology have made it possible to profile spatial transcriptomics (ST), which captures genome-wide gene expression while maintaining the spatial arrangement of cells within tissues. This progress has led to a better understanding of cell communication and tissue organization. However, current techniques face a trade-off between experimental throughput and spatial resolution. Sequencing-based approaches tend to prioritize higher throughput, which results in lower resolution, producing multicellular pixel data. This type of data necessitates innovative computational methods to disentangle cell classes within a pixel and address potential confounding issues. In this study, we are interested in learning the cell class-specific information from low-resolution ST data, where each pixel contains a mixture of several different cell classes. We develop a framework leveraging topic modeling, gene network, and machine-learning to infer the distribution of cell classes across regions and genes.
Keywords
Spatial transcriptomic data
Low-resolution ST data
Topic modeling
Gene network
Cell class distribution
Cell-cell communication
Spatial omics technologies are transforming biomedical research by enabling genome-wide measurement of molecular activity while preserving the spatial context within tissues. These advances create unprecedented opportunities to uncover cell–cell interactions, tissue organization, and disease mechanisms. A crucial step in realizing this potential is identifying spatial domain markers, which are essential for defining tissue architecture and understanding disease progression. However, using the same data for both clustering and marker detection creates the problem of "double dipping", which can lead to inflated false discoveries, particularly when domain boundaries are poorly defined. To address this challenge, we develop SpaDS and SpaMDS, data splitting based approaches for differential expression testing after clustering for spatial omics, enabling robust spatial domain marker discovery with controlled false discovery rate and high power. Through extensive simulations and analyses of spatial omics datasets, we demonstrate that the data-splitting methods are easy to implement, adaptable to existing spatial omics data analysis pipelines, and often outperform other approaches.
Keywords
Statistical genomics
Spatial transcriptomics
data splitting
FDR control
We introduce MIRAGE, a statistical framework using cell-cell manifolds to test whether gene sets encode the same or distinct cellular dynamics. MIRAGE constructs manifolds for many gene sets, quantifies their geometric similarity, and uses hypothesis testing to merge only sets sharing common cell–cell geometry, identifying independent or co-occurring cellular dynamics. Applied to mouse pancreatic endocrinogenesis, MIRAGE separates developmental and cell-cycle programs, and in fibroblast reprogramming datasets, yields reproducible gene modules tracking shared intermediate states. In human Alzheimer's disease microglia, where heterogeneous stimuli produce overlapping homeostatic, inflammatory, and metabolic states, MIRAGE identifies one module capturing a within-donor continuum in both AD and non-AD brains, and a second branching module with one arm enriched for hypoxia and lysosomal stress genes, preferentially occupied by donors with elevated amyloid/tau burden and cognitive impairment. MIRAGE offers a principled approach for decomposing single-cell data into pathway-level manifolds, clarifying how co-occurring cellular programs contribute to disease-relevant heterogeneity.
Keywords
Gene clustering
Geometric structure
Nearest neighbor graph
Multivariate hypothesis testing
Identifying cellular neighborhoods is essential for understanding cell–cell interactions in a spatial context. However, existing approaches often overlook the complexity of multicellular interactions and the organization of spatial domains. We present POLYspace, a general and efficient framework for discovering and analyzing cellular neighborhoods of arbitrary topology while accounting for spatial domains. POLYspace formulates neighborhood identification as a subgraph searching problem and leverages C3G, a fast graph canonization algorithm we developed, to achieve scalability. Applied to one in-house dataset and three publicly available datasets spanning diverse platforms and tissues, POLYspace uncovers domain-specific cellular neighborhoods that are not captured by existing methods. These neighborhoods reveal key biological mechanisms and improve phenotype prediction.
Keywords
Spatial gene profiling
Graph theory
Graph canonization
Statistical genomics
Subgraph mining
Cellular microenvironments
Speaker
Huimin Wang, Duke University School of Medicine Dept. of Biostatistics & Bioinformation
Co-Author(s)
Roger McLendon, Department of Pathology, Duke University School of Medicine
Simon Gregory, Department of Neurosurgery, Duke University School of Medicine
Zhicheng Ji, Duke University
We propose a statistical framework for simultaneous clustering and deconvolution of spatial transcriptomics (ST) that is reference-free. Our method jointly estimates the reference cell-type-specific gene expression profiles and the cluster assignments and their cell-type composition of ST spots through solving an well-defined optimization problem. We introduce a regularization term that makes the problem scale invariant which separates compositional signatures from library‑size effects and enabling clustering that groups spots by true compositional similarity even when counts differ. Furthermore, we leverage spatial information to encourage neighboring spots to have similar cell-type compositions. We develop a computationally efficient algorithm to solve the optimization problem and establish theoretical properties of our estimator. Through extensive simulations and applications to real ST data, we demonstrate that our method outperforms existing reference-free methods and is able to uncover biologically meaningful clusters and accurately estimate cell-type compositions without relying on external single-cell references.
Keywords
Clustering
Spatial Transcriptomics
Scale invariant
Post selection inference
Optimization
Reference‑free deconvolution
Speaker
Hyun Jung Koo, School of Statistics, University of Minnesota - Twin Cities
Co-Author
Aaron Molstad, University of Minnesota
Spatial transcriptomics (ST) enables genome-wide expression profiling with preserved spatial context. However, integrating multiple tissue sections remains challenging due to the lack of direct spatial correspondence across different samples. We present SpatialCCA, a method for joint dimension reduction of two ST tissue sections that embeds them into a shared latent space while incorporating spatial information. SpatialCCA extends canonical correlation analysis through a kernel-based spatial regularization and models tissue-specific variation to mitigate batch effects. The resulting embedding captures both cross-section transcriptional similarity and within-tissue spatial organization. Using simulation generated by randomly partitioning real ST tissue data into two sets of spatial spots, we show that SpatialCCA yields embeddings with stronger correlation with the original coordinates compared to embeddings derived from gene profiles alone. By improving estimation of cross-tissue relationships, SpatialCCA provides a principled framework for integrating two ST slices and can serve as an effective pre-processing step for downstream analysis.
Keywords
Spatial Transcriptomics
Canonical Correlation Analysis
Dimension Reduction
Cross-tissue Integration
Speaker
Jingxian Tang, Boston University School of Public Health
Co-Author(s)
Ching-Ti Liu, Boston University School of Public Health
Lukas Weber, Boston University, Department of Biostatistics
Single-cell high-throughput chromatin conformation capture (scHi-C) profiles 3D genome architecture at cellular resolution. While recent frameworks use spatial patterns for dissimilarity measures in clustering, the inherent sparsity and high dimensionality of scHi-C matrices pose challenges. Crucially, existing measures often fail to distinguish biologically meaningful structural zeros (SZs) from technical dropouts.
We introduce ssJSD (spatial and structural-zero-aware Jensen-Shannon Divergence), a framework explicitly accounting for scHi-C sparsity. By integrating band-wise contact profiles with SZ-induced sparsity matrices, ssJSD leverages both spatial patterns and biological absence of contacts. We adopted two integration strategies: early fusion, concatenating information into a single representation, and late fusion, integrating JSD-based dissimilarities via diverse averaging. Through simulations and applications to human cell lines and prefrontal cortex data, we demonstrate that ssJSD improves clustering accuracy and effectively distinguishes cell types. Our findings highlight that integrating SZ patterns is important for accurately quantifying cell-to-cell variability.
Keywords
Single cell Hi-C
Single-cell clustering
Contact distance profile
Structural zeros
Jensen-Shannon divergence
Data integration