Much like how a student synthesises lessons from different subjects to tackle complex real-world problems, scientists can acquire data from different levels of molecular biology to study the mechanics of health and disease. Thanks to experimental tools, the various fields of ‘omics’ today— epigenomics, transcriptomics, proteomics and more—can capture vast landscapes of molecular data that reveal every change to a cell or tissue’s DNA; every message it writes in RNA; and every protein it expresses to adapt to its environment.
Ideally, such shifts are best studied in their spatial context—which genes in which part of the brain misfire in dementia, for example? However, today’s computational tools typically analyse omics data without the spatial component, or process only a single omic modality at a time, potentially leaving out crucial information.
“The preservation of spatial information allows more accurate inferences on cell-cell interactions and localised changes in cell composition,” said Jinmiao Chen, a Principal Investigator at the A*STAR Bioinformatics Institute (A*STAR BII). “We need tailored tools that can integrate complementary information from all omics layers while being spatially aware to extract systems- level features or signatures.”
Aiming to fill that gap, Chen and A*STAR BII colleagues teamed up with researchers from the A*STAR Institute of Molecular and Cell Biology (A*STAR IMCB), A*STAR Singapore Immunology Network (A*STAR SIgN), National University of Singapore, and institutes in China and the US. Together, they created SpatialGlue, an analytical tool that uses an artificial intelligence (AI) model that integratively analyses spatial data from the epigenome-transcriptome and the transcriptome-proteome.
As omics data can be high-dimensional—a transcriptome, for example, can have over 10,000 measurable features—SpatialGlue first simplifies such data while ensuring it still captures the relevant biological variations. In parallel, the model constructs graphs that map the proximity of different molecule and cell types to each other.
“SpatialGlue then trains a deep learning AI model based on the simplified and mapped omics data, which adaptively combines the different modalities to learn an integrated representation of them,” said Chen. “The model also learns the relative importance of each of these modalities to develop a final representation that fits different analytical needs, such as the clustering of certain cells with shared functions.”
SpatialGlue was tested against a suite of five simulated datasets and 12 experimental datasets, which included molecular data from two species (mice and human) and four tissue types derived from four technology platforms.
The team found that SpatialGlue successfully captured more anatomical details and correctly distinguished tissue regions—such as brain cortex layers—at a higher resolution than existing methods. SpatialGlue also identified subtypes of immune cells in three spatial zones of the spleen, uncovering new information from the original dataset.
To enhance their model, the team plans to add imaging data from immunostaining experiments and other platforms that captures additional information about the cells, such as cell size and shape. They are also exploring collaborations to apply SpatialGlue for more in-depth studies of clinical samples.
“Analysis of spatial multi-omics data can deepen our understanding of the molecular underpinnings of diseases, aiding the discovery of biomarkers and therapeutic targets,” said Chen.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR A*STAR Institute of Molecular and Cell Biology (A*STAR IMCB) and A*STAR Singapore Immunology Network (A*STAR SIgN).
