STADIA: A Spatio-Temporal Deep-Learning Framework for Cross-Stage Integration and Batch Correction of Plant...

Spatial transcriptomics enables the spatial architecture and transcriptional dynamics of plant organ development to be resolved in situ. However, integrating plant spatial transcriptomic data across developmental stages is challenging because genuine developmental variation can be confounded with technical variation arising from section source, capture region and sequencing batch, making it difficult to attenuate technical effects without compromising developmental signals. To address this challenge, we developed STADIA, a spatio-temporal deep-learning framework for plant developmental spatial transcriptomics. STADIA combines a graph-attention autoencoder with structural-context-conditioned batch-predictability suppression, stage-order constraints, PCGrad-based gradient coordination and hierarchical unbalanced optimal transport to learn an integrated latent representation that attenuates batch-associated variation while preserving spatial organization and developmental continuity and accommodating non-equivalent tissue transitions. In a semi-synthetic benchmark, STADIA maintained balanced performance across batch mixing, anatomical-domain preservation and developmental-information preservation under strong batch effects and high sparsity. In the same-stage replicate sections from the soybean early-maturation stage (EMS), STADIA increased local between-section mixing while retaining major anatomical compartments. In a poplar stem developmental series spanning the shoot apex to internode 7, the integrated representation showed increased cross-stage continuity while retaining stage-specific organization and spatially coherent tissue compartments, and supported exploratory identification of cambium-associated trajectory candidates. Together, these results support STADIA as a framework for cross-stage integration of plant spatial transcriptomic data and for downstream analysis of spatial domains and developmentally ordered transcriptional patterns.