Background

Spatial omics measures gene expression, proteins, chromatin accessibility, histone modifications, metabolites, and imaging data at specific coordinates within tissues. Overlaying multiple molecular layers from the same region allows for a three-dimensional understanding of cell states and tissue architecture, but differences in signal distribution and noise levels across modalities often lead to loss of unique signals during integration.

Existing single-cell integration methods such as Seurat, MOFA+, and MultiVI do not fully account for spatial neighbor relationships. The spatial multimodal tool SpatialGlue is limited to a maximum of three molecular layers and incurs high computational costs when handling high-resolution imaging or datasets with millions of coordinates. Forcing dissimilar data into alignment can flatten biological differences, a problem known as 'over-alignment.' Additionally, these methods lack the ability to identify and flag unstable or unreliable locations in the analysis.

Key Discovery

SCIGMA, developed by researchers at Brown University, is an unsupervised deep learning model that combines a multi-view graph neural network with uncertainty-aware contrastive learning. The model merges spatial graphs connecting physically close locations and feature graphs linking locations with similar molecular profiles. A graph attention network (GAT) generates latent representations for each omics layer, and cross-attention integrates these representations into a shared space. A decoder reconstructs the original features to minimize information loss.

A key innovation is the uncertainty-aware contrastive loss, which learns location-specific temperature parameters. This ensures that shared representations of the same coordinate are close to modality-specific representations without uniformly mixing signals unique to each omics layer. Coordinates with poor alignment are assigned high uncertainty, allowing researchers to identify complex regions such as tumor heterogeneity or immune cell compartments, as well as areas suspected of technical noise.

The research team evaluated the model across 19 datasets from 10 tissue types and 9 platforms, including eight data types such as gene expression, protein, chromatin accessibility, histone modification, and metabolites. In comparable mouse brain spatial ATAC–RNA and spleen SPOTS data, SCIGMA outperformed Seurat v5, MOFA+ 1.13.0, MultiVI v1, and SpatialGlue v1 in spatial domain detection, modality-specific signal preservation, feature reconstruction, and reproducibility across repeated runs. Other methods failed to produce results in some large datasets even with over 400GB of CPU memory and two days of processing time, according to the researchers.

SCIGMA processed over a million spatial locations, including 2-micrometer resolution data from 10x Visium HD, using graph sampling. In a 5-month-old mouse brain Spatial-Mux-seq dataset, SCIGMA integrated RNA, protein, ATAC, H3K27ac, and H3K27me3 data simultaneously. Previously indistinct anatomical regions such as the cerebral cortex layers, hippocampal dentate gyrus, caudate putamen, and striatum were clearly separated in the shared representation. The Nature Genetics paper notes that this framework is extensible to future platforms beyond five modalities.

Implications and Outlook

SCIGMA addresses three major challenges in spatial multimodal analysis—preservation of modality-specific signals, computational scalability for large datasets, and uncertainty quantification—within a single framework. The ability to assign uncertainty to individual coordinates allows researchers to selectively validate clusters that may require pathological review or further experimentation. The software and reproducible analysis code are publicly available.

However, uncertainty scores reflect a mixture of biological heterogeneity and technical variability. High uncertainty alone cannot confirm new cell states or disease boundaries; tissue staining and independent marker validation are still required. Performance advantages are also limited to datasets with annotated ground truth or where competing methods are computationally feasible. Future validation is needed to assess how well SCIGMA handles differences in clinical sample layouts, missing modalities, and varying resolutions.