Must: Maximizing Latent Capacity of Spatial Transcriptomics Data

TL;DR

MuST addresses modality bias in spatial transcriptomics by learning a uniform latent space that coherently fuses transcriptomic, morphological, and spatial information. It introduces topology discovery to capture local/global structure per modality and a topology fusion loss to align them, paired with augmentation and reconstruction objectives. Across mouse and human datasets from 10x Visium, Slide-seqV2, and Stereo-seq, MuST achieves superior tissue-structure delineation, laminar organization, and spatial trajectories, while balancing modality contributions as shown by Shapley analyses and improved ARI/MRRE metrics. The approach enables accurate downstream analyses (clustering, visualization, deconvolution, marker analysis, trajectory inference) with robust cross-platform performance, offering a versatile foundation for complex tissue biology analyses.

Abstract

Spatial transcriptomics (ST) technologies have revolutionized the study of gene expression patterns in tissues by providing multimodality data in transcriptomic, spatial, and morphological, offering opportunities for understanding tissue biology beyond transcriptomics. However, we identify the modality bias phenomenon in ST data species, i.e., the inconsistent contribution of different modalities to the labels leads to a tendency for the analysis methods to retain the information of the dominant modality. How to mitigate the adverse effects of modality bias to satisfy various downstream tasks remains a fundamental challenge. This paper introduces Multiple-modality Structure Transformation, named MuST, a novel methodology to tackle the challenge. MuST integrates the multi-modality information contained in the ST data effectively into a uniform latent space to provide a foundation for all the downstream tasks. It learns intrinsic local structures by topology discovery strategy and topology fusion loss function to solve the inconsistencies among different modalities. Thus, these topology-based and deep learning techniques provide a solid foundation for a variety of analytical tasks while coordinating different modalities. The effectiveness of MuST is assessed by performance metrics and biological significance. The results show that it outperforms existing state-of-the-art methods with clear advantages in the precision of identifying and preserving structures of tissues and biomarkers. MuST offers a versatile toolkit for the intricate analysis of complex biological systems.

Figures (20)

• Figure 1: Illustration of MuST.A Morphology (Mor.) modality data of the coronal dataset. B Illustration of the modality bias phenomenon. In the case of Mor. modality, there is a significant inconsistency in the contribution of Mor. data to the final result (detected by Shapley value winter2002shapley). This inconsistency causes many methods to discard weaker modalities such as Mor. and Spa. resulting in a loss of information. MuST improves the retention of weaker modalities by mapping the Mor modality to the Transcriptomic (Tra.) modality. The results of the interpretable analyses show that the phenomenon of modality bias is mitigated by the Mor. Encoder processing. C The statistical evidence of the modality bias phenomenon is mitigated. D The ST data and its topologies of Tra. modalities, Mor. modalities and Spa. modalities are fused into a universal latent feature space (universal latent space) in which the information loss problems caused by the modality bias phenomenon are mitigated and the topologies of multimodal data are well represented. Specifically, the unit of ST data is called a 'spot', and each spot contains Tra. data, Spa. data and Mor. data. MuST aims to map each spot into a vector of universal latent space in order to perform various downstream tasks. The Mor. and Tra. modality data are mapped to the single modal latent space feature by two graph encoders, with the spatial modal information used to construct the edges of the graph. The graph encoders represent each point of the ST data as a vector. The individual modal latent space features are then mapped to the universal latent space using a fusion encoder. The topology discovery operation extracts topological structures from the embedding by constructing KNN graphs (see Sec. \ref{['Graph_Construction_method']}), which express the topological relationships of the spot embedding. The topology fusion loss guides the network to learn a unified representation capable of fusing multimodal data, thereby improving performance in several downstream tasks. E Multiple downstream tasks. MuST incorporates information from multiple modalities and better supports multiple downstream tasks. These include (a) spatial clustering, (b) spot visualisation, (c) spatial deconvolution, (d) marker gene analysis, and (e) spatial trajectory inference. The mutual support of multiple downstream tasks allows for a more comprehensive understanding of ST data.
• Figure 2: MuST explores more biologically complex tissues in adult mouse brain section profiled by 10x Visium (https://www.10xgenomics.com/resources/datasets). A The annotation of hippocampus structures from the Allen Reference Atlas of an adult mouse brain. B H&E image of mouse brain coronal section. C Contribution of the morphology (Mor.) modality, transcriptome (Tra.) modality and spatial (Spa.) modality input data or MuST embedding to data labeling. D Clustering results by spatial methods, stSME, SpaGCN, MUSE, DeepST, GraphST, and MuST. E Spot visualization generated by the spatial methods. F Single cluster visualization of spatial domains identified by MuST and the corresponding marker gene expressions. G Allen Brain Institute reference atlas diagram of the mouse sagittal. H H&E image of mouse sagittal posterior brain section. I Contribution of the Mor. modality, Tra. modality and Spa. modality input data or MuST embedding to data labeling. J Clustering results by spatial methods, stSME, SpaGCN, MUSE, DeepST, GraphST, and MuST. K Spot visualization generated by the spatial methods. L Single cluster visualization of spatial domains identified by MuST and the corresponding marker gene expressions and cluster-related spatial deconvolution.
• Figure 3: MuST enables accurate identification of different organs in the Stereo-seq mouse embryo (https://db.cngb.org/stomics/mosta/). A Tissue domain annotations obtained from the original Stereo-seq study and clustering results by GraphST and MuST on the E14.5 mouse embryo data. B Spot visualizations generated by MuST representations. C MRRE scores generated by MUSE, SpaGCN, DeepST, GraphST, and MuST representations. Among them, methods MUSE and DeepST have an out-of-memory (OOM) problem. D Contribution of the transcriptome (Tra.) modality and spatial (Spa.) modality input data or MuST embedding to data labeling. E Single cluster visualization of selected spatial domains identified by the original Stereo-seq study, GraphST and MuST, and the corresponding spatial deconvolution and marker genes of selected spatial domains identified by MuST. F Tissue domain annotations obtained from the original Stereo-seq study and clustering results by GraphST and MuST on the E9.5 mouse embryo data. G Spot visualizations generated by MuST representations. H MRRE scores generated by MUSE, SpaGCN, DeepST, GraphST, and MuST representations. I Contribution of the Tra. modality and Spa. modality input data or MuST embedding to data labeling. J Single cluster visualization of selected spatial domains identified by the original Stereo-seq study, GraphST and MuST, and the corresponding spatial deconvolution and marker genes of selected spatial domains identified by MuST.
• Figure 4: MuST discerns relevant anatomical regions more accurately in the SlideseqV2 mouse hippocampus data (https://portals.broadinstitute.org/single_cell/study/slide-seq-study). A Allen Brain Institute reference atlas diagram of the mouse cortex. B Clustering results by spatial methods, STAGATE, GraphST and MuST on the mouse hippocampus data. C Spot visualizations generated by SpaGCN, STAGATE, GraphST and MuST representations. D Single cluster visualization of the tissue structures identified by MuST and the corresponding marker gene expressions. E Cluster-related spatial deconvolution analysis of the tissue structures identified by MuST. F Visualization of the celltype identified by MuST and the corresponding marker gene expressions. G Cluster-related spatial deconvolution analysis. H MRRE scores generated by SpaGCN, GraphST, STAGATE, and MuST representations.
• Figure 5: MuST identifies the laminar organization in the mouse olfactory bulb tissue sections profiled by Stereo-seq (https://github.com/JinmiaoChenLab/SEDR_analyses) and Slide-seqV2 (https://singlecell.broadinstitute.org/single_cell/study/SCP815/highly-sensitive-spatial-transcriptomics-at-near-cellular-resolution-with-slide-seqv2) respectively. A Laminar organization of mouse olfactory bulb annotated in the DAPI-stained image generated by Stereo-seq. B Clustering results by spatial methods, STAGATE, GraphST and MuST on the Stereo-seq mouse olfactory bulb tissue section. C Single cluster visualization of the spatial domains identified by MuST and the corresponding marker gene expressions. D Laminar organization of mouse olfactory bulb annotated by the Allen Reference Atlas. E Clustering results by spatial methods, STAGATE, GraphST and MuST on Slide-seqv2 mouse olfactory bulb tissue section. F Single cluster visualization of spatial domains identified by MuST and the corresponding marker gene expressions. G Spot visualizations and PAGA graphs generated by the representations of MuST. H MRRE scores generated by MUSE, SpaGCN, DeepST, GraphST, and MuST representations.