The challenge of reading data from different tissues together

INSPIRE, a computational method designed to reduce technical differences in spatial transcriptomics (ST) data while preserving tissue-specific biological signals, was reported in Nature Genetics. The researchers combined graph neural networks, adversarial learning, and non-negative matrix factorization to jointly interpret spatial structure and associated gene programs across multiple tissue sections.

ST links gene expression to locations within tissue. However, technologies differ in the genes they measure and their spatial resolution, while differences in samples, experimental conditions, and developmental stages also affect analysis. Simply combining datasets can lead researchers to mistake technical differences for tissue characteristics or, conversely, erase biological signals found only in a particular section. Existing spatial analysis methods, including SpiceMix and NSFH, are useful for finding interpretable spatial patterns, but have limited means of handling complex, unwanted variation across multiple sections.

Jointly learning spatial neighborhoods and gene programs

INSPIRE takes gene expression counts and spatial coordinates from multiple sections as input. A graph neural network (GNN) incorporates information from neighboring cells or measurement points, while adversarial learning encourages shared signals across sections to align. The design also limits the discriminator’s scope so that section-specific cell populations are not indiscriminately mixed.

Non-negative matrix factorization (NMF) represents expression signals across sections using spatial factors and gene weights. Spatial factors show where a particular pattern occurs in the tissue; gene weights show which genes characterize that pattern. The approach aims to interpret the expression patterns of co-occurring genes in relation to tissue structure and biological processes.

The researchers assessed the distinction between cortical layers and white matter in 12 sections of human dorsolateral prefrontal cortex from three donors. They reported better performance than comparator methods on the adjusted Rand index (ARI) and normalized mutual information (NMI), which measure agreement with manual annotations, and on average silhouette width (ASW), which assesses separation between regions. The reported advantage applies to the data and evaluation conditions used in the paper.

For cross-technology integration, the team jointly analyzed Slide-seqV2 and MERFISH data from mouse brain. The former measured more than 20,000 genes and the latter 1,122; the actual integration used 1,077 genes shared by both datasets. INSPIRE used information from MERFISH to distinguish cortical layers in Slide-seqV2. In human breast cancer, the researchers analyzed 283,063 cells and 313 genes remaining after filtering data from two sections measured with Xenium, capturing differences in spatial factors associated with tumor regions and myoepithelial and endothelial cells.

Potential and limitations for research tissue maps

These results support reusing heterogeneous public datasets to compare tissue structure and the tumor microenvironment. The researchers also applied INSPIRE to mouse embryos at different developmental stages and to 3D reconstruction from adjacent sections. However, completing a precise map of the entire human body and enabling patient-specific diagnosis or optimal drug selection were not evaluated in this study.

Joint interpretation currently depends on genes shared between datasets, so signals from genes present in only some datasets may be excluded. Interpolated gene expression values should be distinguished from actual measurements, and quantifying uncertainty associated with data integration itself remains a challenge. Follow-up studies will need to validate inferred spatial patterns and gene programs using independent data and experiments.

The study and its analytical conditions are described in the Nature Genetics paper and the open full text.