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

Introduction: Spatial transcriptomics enables molecular mapping of ischemic stroke tissue, but spatial domain identification is challenging when injury disrupts normal tissue geometry. Methods relying on a single spatial-proximity graph cannot connect physically distant spots that share damage-associated transcriptional programs. Methods: We developed SpatialDomainAE, an unsupervised dual-graph attention autoencoder that constructs separate spatial-neighbor and expression-similarity graphs, processes each using graph attention, and combines their embeddings through learned per-spot fusion weights. The method was evaluated on a mouse middle cerebral artery occlusion 10× Visium dataset comprising control, 1-, 3-, and 7-day post-injury sections, totaling 10,173 spots and 22 annotated domains. All methods were evaluated over 10 random seeds using a shared Leiden-resolution protocol. Results: At 3 days post-injury, SpatialDomainAE achieved an ARI of 0.700 ± 0.025 and significantly exceeded all external baselines, including the dual-view Spatial-MGCN. Across all four samples, its performance was competitive rather than uniformly superior, and it was robust under a fixed clustering resolution. No comparable advantage was observed on an external human dorsolateral prefrontal cortex benchmark. Controlled experiments showed that the long-range transcriptomic content of the feature graph, rather than edge length alone, accounted for the improvement. Fusion weights separated lesion-associated domains at the region level, while differential expression and pathway enrichment recovered inflammatory, complement, gliosis, and proliferative programs. Discussion: SpatialDomainAE is particularly useful in disrupted pathological tissue. Its fusion weights should be interpreted as an exploratory region-level model diagnostic rather than a validated spot-level biomarker.