Prioritizing perturbation-responsive gene patterns using interpretable deep learning.

Spatially resolved transcriptomics enables mapping of multiplexed gene expression within tissue contexts. While existing methods prioritize spatially variable genes within a single slice, few address identifying genes with differential spatial expression patterns (DSEPs) across multiple conditions-an critical need for complex experimental designs. Challenges include modeling cross-slice spatial variation, scalability to large datasets, and disentangling inter-slice heterogeneity. We introduce DSEP gene prioritization as a new analytical task and present River, an interpretable deep learning framework that identifies genes exhibiting condition-relevant spatial changes. River features a two-branch predictive architecture and a post hoc attribution strategy to rank genes (or other features) by their contribution to condition differences. Its spatially-informed modeling ensures scalability to large spatial datasets, and we further decouple spatial and non-spatial components to enhance interpretability. We evaluate River on simulations and apply it to diverse biological contexts, including embryogenesis, diabetes-affected spermatogenesis, and lupus-associated splenic changes. In triple-negative breast cancer, River prioritizes survival-associated spatial patterns that generalize across patients. River is distribution-agnostic and compatible with diverse spatial data types, offering a flexible and scalable solution for analyzing tissue-wide expression dynamics across multiple biological conditions.