ACT-ST: Design-Corrected Active Measurement for Sample-Efficient Spatial Transcriptomics

Spatial transcriptomics can reveal how gene expression varies across a tissue, but assaying every spatial location over large areas and many patients remains expensive. We study whether one can use auxiliary prediction models to reduce the number of sampling spots, while remaining unbiased. ACT-ST is a two-stage active-measurement framework inspired by active estimation and prediction powered inference. It first conduct a uniform pilot measurement and uses those to learn proxies and, when available, auxiliary models like scGPT. The proxies then score unmeasured spots by novelty, pilot error, transcriptomic rarity, and spatial coverage. ACT-ST converts these scores into randomized inclusion probabilities. For distance-binned gene-expression summaries, a model-assisted estimator then combines proxy totals with probability-weighted residuals from measured spots, correcting for nonuniform acquisition. In a retrospective two-section study, this design-corrected estimator reduces median two-point error by 42–45\% relative to uniform sampling across budgets. We also experiment on HER2 sections from eight patients, and demonstrate improvement over other active sampling baselines.