LATTICE demonstrated stable optimization behavior, reproducible embeddings across analysis seeds, and complete multimodal integration across all samples, position LATTICE as a practical and empirically grounded framework for multimodal spatial omics integration, while also highlighting the need for stronger supervision and broader external benchmarking.
Abstract
Spatially resolved omics studies increasingly combine transcriptomic and epigenomic assays, yet downstream analysis is often still performed using single-modality pipelines. We present LATTICE (Latent Alignment of Tissue-level and Transcriptomic Information for Cross-modal Embedding), a graph-based self-supervised framework that learns spot-level representations from harmonized multimodal features. LATTICE integrates five aligned modality blocks per Visium spot: Visium RNA, scMultiome RNA, scMultiome ATAC, spatial ATAC, and spatial CUT\&Tag. These modalities capture spatial transcriptomic measurements, single-cell inferred regulatory activity, and in situ chromatin and histone states within a unified lattice representation. LATTICE constructs a spatial neighborhood graph and trains a TransformerConv encoder using masked reconstruction, cross-modal alignment, and spatial smoothness objectives. On a private 11-sample melanoma cohort from an anonymized clinical collaborator comprising 54{,}912 total spots, LATTICE demonstrated stable optimization behavior, reproducible embeddings across analysis seeds, and complete multimodal integration across all samples. Adding scMultiome RNA to Visium RNA alone substantially improved concordance with Space Ranger clusters across 11 runs (adjusted Rand index [ARI] +0.157, normalized mutual information [NMI] +0.143, and spatial contiguity +0.174). Additional modalities further improved spatial contiguity and multimodal utility score (MUS), although they sometimes reduced agreement with RNA-derived reference labels, likely because the learned embeddings captured chromatin and regulatory structure beyond transcriptomic similarity alone. These results position LATTICE as a practical and empirically grounded framework for multimodal spatial omics integration, while also highlighting the need for stronger supervision and broader external benchmarking.
Abstract Motivation Spatial transcriptomics (ST) enables molecular profiling within native tissue architecture, yet accurate delineation of spatial domains in ST data is challenging, as it demands the coordinated integration of transcriptomic, spatial, and tissue histological information. Results We present SRLST, an unsupervised representation learning framework that holistically harmonize these three complementary data modalities to precisely uncover tissue organization. SRLST employs a dual-graph variational autoencoding strategy to jointly model spatial proximity and morphological relations, fusing these with gene-expression embeddings into a unified latent space. Across distinct experimental datasets, SRLST consistently outperforms existing methods in delineating cortical organization, identifying small discontinuous tissue compartments, and capturing complex intratumor heterogeneity. Availability and implementation The code implementation of the SRLST algorithm is available at https://github.com/lanbiolab/SRLST.
Multimodal spatial omics integration offers a powerful paradigm to decipher the hierarchical regulatory mechanisms underlying cellular function and tissue architecture. In this study, a novel method of multimodal spatial omics fusion, named mmspao, is proposed to obtain cross-modal interactive features and combine them with the features of each modality to obtain fusion results at the spatial resolution. This method integrates the data of spatial transcriptomics, epigenomics, and proteomics. It further integrates information from each modality using adjacency graph modeling and latent spatial representation. We demonstrate the effectiveness of this method on simulated and real multiomics data. By leveraging adjacency graph modeling and latent spatial representation, mmspao effectively aligns multiomics modalities within shared spatial domains, preserving individual gene expression profiles while generating globally integrated fusion maps that advance the decoding of tissue architecture and cellular regulatory hierarchies.
Cross-domain analysis of spatial transcriptomics is challenging because tissues from different organs, diseases and experimental platforms exhibit distinct cellular compositions, spatial organisations and technical biases, making direct comparison of tissue states difficult. Existing methods primarily focus on domain integration or batch correction but generally do not explicitly model the intrinsic geometry underlying tissue-state organisation across biological systems. This paper presents recurrence geometric artificial intelligence (RGAI), a geometric deep-learning framework for discovering and aligning latent tissue states across heterogeneous spatial transcriptomic domains. RGAI first learns domain-specific latent representations using variational graph autoencoders while simultaneously estimating a Riemannian metric tensor that captures the local geometry of each latent manifold. Geodesic distances induced by the learned metric are used to construct multiscale recurrence graphs that characterise intrinsic tissue-state organisation independently of the original measurement space. Cross-domain manifold correspondence is then established through entropy-regularised Gromov–Wasserstein alignment, after which fuzzy clustering identifies latent tissue states and optimal transport aligns tissue-state signatures across domains. Evaluation on six human spatial transcriptomic datasets spanning wound healing, periodontitis, oral squamous cell carcinoma, head and neck squamous cell carcinoma, cardiac tissue and colorectal cancer shows that RGAI automatically determines biologically meaningful latent tissue-state complexity and identifies coherent recurrence-based tissue states within each domain. The learned geometric representations enable cross-domain alignment of latent manifolds while preserving biologically interpretable tissue-state correspondences despite substantial differences in cellular composition and tissue architecture, demonstrating that integrating learned Riemannian geometry, recurrence analysis and optimal transport provides a robust and interpretable framework for cross-domain tissue-state discovery and comparison in spatial transcriptomics.
Abstract Motivation Spatial transcriptomics (ST) technologies profile gene expression while preserving tissue architecture, enabling the study of spatial cellular organization and microenvironmental interactions. However, raw ST data are heavily affected by technical noise, sparsity, and dropout events, which obscure true biological signals and hinder downstream analyses. While recent denoising methods incorporate spatial neighborhood information, they lack explicit supervision due to missing cell-type annotations in ST data. To address this challenge, we introduce SPIDER, a semi-supervised framework that leverages independently generated, annotated single-cell RNA-seq (scRNA-seq) references to guide ST denoising. SPIDER synthesizes pseudo-ST data from scRNA-seq to inject cell-type information without requiring paired measurements. The method constructs three graphs capturing spatial proximity, transcriptional similarity in real-ST data, and transcriptional structure in pseudo-ST data. Graph encoders map these representations into a shared latent space, and a domain-alignment module transfers biologically meaningful structure from pseudo-ST to real-ST embeddings. A graph-attention decoder with a zero-inflated negative binomial objective reconstructs denoised ST expression profiles. Results We benchmark SPIDER on human dorsolateral prefrontal cortex and breast cancer datasets. SPIDER consistently enhances spatial gene expression patterns, recovers known tissue structures, and achieves superior clustering performance compared to existing approaches. Marker gene analyses demonstrate improved spatial continuity and clearer anatomical organization. By directly producing denoised expression matrices, SPIDER improves both accuracy and interpretability, providing a generalizable solution for robust ST data denoising. Availability and implementation The source code and dataset is available at: (https://github.com/compbiolabucf/SPIDER) and archived on Zenodo (https://doi.org/10.5281/zenodo.20613921).
Experiments demonstrate that the GraphFusionNN fusion strategy— integrating graph topology, spatial context, and external embeddings—significantly improves classification accuracy, macro-F1, and robustness compared to single-modality models.
StKAN is introduced, a novel framework integrating Kolmogorov-Arnold Network with variational autoencoder to effectively model spatially resolved gene expression with graph attention network and shows strong potential for downstream analyses, offering deeper insights into disease pathology and tumor invasion.
Jing Lin, Aijing Feng, Yankun Cao et al.· Computational biology and ch...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
