X-SPEAR: An Extensible Framework for Spatial Panel Evaluation and Refinement

Abstract

Targeted spatial transcriptomics platforms such as Xenium require a predefined gene panel, and the genes in that panel are detected less efficiently, and to gene-specific degrees, in the spatial assay than in the single-cell reference used to choose them. We present X-SPEAR, a framework that couples DART marker selection with probe-level off-target assessment. DART learns short, interpretable Boolean gene rules and assembles several gene-disjoint rules per cell type. Its detection-aware training thins the reference to per-gene detection rates measured in a different, already completed experiment, so no Xenium data from the target tissue are needed before the panel is ordered. On held-out Xenium data from ovarian cancer, mouse brain and cervical cancer, DART substantially outperformed existing marker-selection methods at matched panel size, by 0.04-0.11 macro-F1 over Wilcoxon rank-sum, 0.04-0.08 over COSG and 0.10-0.18 over NS-Forest, and it reached Wilcoxon's best accuracy with 42-69% fewer genes. Ablations show that most of this gain comes from detection-aware training, which also improves other rank-based selectors and requires a detection profile measured with the same assay chemistry. At similar accuracy, DART's rule structure adds rules that faithfully summarise the downstream classifier, robustness to single-probe loss across 45 retinal ganglion cell subtypes, and better resolution of closely related cell states. We are applying X-SPEAR to design a Xenium panel for mouse retinal ganglion cell subtypes, and its in situ evaluation is in progress. X-SPEAR is available as an open-source Python package.