SPARTA enables integrative 3D spatial domain identification and gene prioritization across tissue sections

Spatial transcriptomics provides insights into tissue organization by preserving spatial gene expression patterns. However, integrating multiple tissue sections remains challenging due to slice-specific variations, while conventional spatial domain identification offers limited insight into domain-associated molecular features. We introduce SPARTA (Spatial Pattern Analysis with Reconstruction, Triplet Alignment, and Attribution), a framework combining Moran-guided masked graph learning, cross-slice representation alignment, and spatial domain-conditioned gene attribution. A shared graph autoencoder operates on spatial graphs constructed independently for each slice. Its masking distribution combines uniform sampling with dynamically updated weights derived from local Moran's I, giving greater sampling weight to spots with lower local spatial consistency. Cross-slice triplet constraints and alternating slice-confusion training encourage comparable latent representations, while graph message passing remains restricted to each slice. Model-based clustering identifies spatial domains from representations computed using unmasked inputs. To associate the identified domains with gene expression features, SPARTA employs five independently initialized surrogate classifiers trained to reproduce the inferred domain labels with the encoder fixed. Signed gradients of target-domain logits with respect to PCA inputs are averaged across classifiers and projected into standardized gene space using the fitted PCA component matrix. Gene attribution scores are aggregated within slices and averaged equally across slices, then combined with domain-versus-rest expression fold changes from raw counts to prioritize candidate genes. Evaluations across five human and mouse datasets comprising 59 slices show that SPARTA recovers major tissue structures and generates domain-associated gene rankings, although performance and surrogate fidelity vary across datasets. The prioritized genes remain model-derived candidates whose biological specificity and function require independent validation.