NicheIB: Targeted Spatial Bottlenecks for Predictive Microenvironment Discovery

Spatial transcriptomics methods typically cluster spots into coherent tissue domains, but many biological questions are target-specific: which spatial context predicts a particular cell state, gene program, or clinical phenotype? We propose NicheIB, a Targeted Spatial Bottleneck (TSB) that learns which spatial edges matter for a user-specified target via input-dependent Hard Concrete edge gates with controllable sparsity budgets. Training the same model on the same tissue with different targets—cortical-layer annotations, oligodendrocyte gene programs, excitatory neuron markers—produces distinct edge retention patterns whose correlation mirrors the biological similarity of the targets, confirmed across human cortex, mouse brain, and cancer tissues. The bottleneck also reveals which programs are spatially structured at a given resolution and which are not, serving as a diagnostic tool for spatial biology. On mouse brain region prediction, TSB retains only 20% of spatial edges while approaching the full-graph baseline ($F_1 = 0.915$ vs. $0.926$), identifying a minimal informative subgraph from a mostly redundant spatial graph. External validation confirms that the retained subgraph selectively enhances spatial autocorrelation for target-relevant genes and is enriched for known ligand–receptor pairs, independently of the model.