Abstract
Spatial transcriptomic super-resolution recovers tissue organization obscured by coarse sampling, but fitting a model to each section greatly increases computational cost and reduces practicality. We introduce ASTRA, a cross-platform pretrained model that reconstructs expression from histology and measured RNA by zero-shot inference. Training with arbitrarily shaped count-bearing regions, termed parents, enables one frozen model to accommodate dense bins and separated spots. Across six held-out Visium HD sections, ASTRA achieved higher correlations under dense-bin and pseudo-Visium sampling, lower dense-bin errors and substantially shorter processing times than existing methods. It improved colorectal tissue segmentation and tumor-boundary recovery. In real colorectal Visium, reconstruction spatially resolved a stromal transcriptional state supported by measured RNA. The same checkpoint supported regional and nucleus-associated analyses across 43 invasive lobular carcinoma sections and within-spot interpretation across 92 triple-negative breast cancer sections. ASTRA establishes a reusable framework for practical, cohort-scale spatial transcriptomic super-resolution across platforms.