Dual-graph attention autoencoder for spatial domain identification in ischemic stroke

Area of Science:

• Computational biology
• Genomics
• Neuroscience

Background:

• Spatial transcriptomics is key for mapping tissue but struggles with disrupted geometry in conditions like ischemic stroke.
• Existing methods using single spatial graphs fail to link physically distant but transcriptionally similar damaged tissue regions.

Purpose of the Study:

• To develop an unsupervised deep learning method, SpatialDomainAE, for accurate spatial domain identification in pathologically altered tissue.
• To overcome limitations of single-graph approaches in capturing long-range transcriptional similarities in damaged tissue.

Main Methods:

• Developed SpatialDomainAE, an unsupervised dual-graph attention autoencoder using separate spatial-neighbor and expression-similarity graphs.
• Integrated embeddings from both graphs using learned per-spot fusion weights.
• Evaluated on mouse middle cerebral artery occlusion 10× Visium data across multiple injury time points.

Main Results:

• SpatialDomainAE achieved a high Adjusted Rand Index (ARI) of 0.700 ± 0.025 at 3 days post-injury, outperforming baseline methods.
• Demonstrated robustness across samples and under fixed clustering resolution.
• Controlled experiments confirmed that transcriptomic content, not just proximity, drove performance improvements.

Conclusions:

• SpatialDomainAE effectively identifies spatial domains in disrupted pathological tissue, particularly beneficial for ischemic stroke research.
• Fusion weights offer exploratory region-level insights into tissue pathology, aiding in understanding inflammatory and proliferative responses.