StaR: A Stability-Aware Representation Learning Framework for Spatial Domain Identification

Spatial domain identification is a fundamental task in spatial transcriptomics, yet state-of-the-art methods exhibit pronounced sensitivity to random initialization. Evaluating four representative methods—STAGATE, GraphST, SpaceFlow, and STcluster—across 1,000 random seeds on all 12 sections of the DLPFC benchmark reveals worst-to-best ARI ranges of up to 0.396, a gap that rivals reported inter-method differences in published benchmarks. We propose StaR (Stability-Aware Representation Learning), a plug-in training framework that addresses this limitation without modifying the base encoder architecture. As a proof of concept, our first instantiation P-StaR-A (Prototype-Anchored Adaptive Fine-Tuning) achieves a 55% relative CV reduction on STAGATE across all 12 DLPFC sections (mean CV: 11.6% → 5.2%), with the most volatile sample improving from CV = 28.4% to 1.9%. The OSRE project will extend StaR to the remaining three backbones, develop a learned cross-backbone stability gate, and release a scanpy-compatible library.