Trimodal, uncertainty-guided whole-slide framework for genome-scale spatial expression and image-only virtual perturbation in cancer cohorts

Spatial transcriptomics is powerful but costly; hematoxylin and eosin (H&E) images are routine. We present Coladan-human3K, the largest human spatial transcriptomics resource (~ 3,000 profiles), and Coladan, a trimodal (image, language, spatial-gene) whole-slide framework predicting genome-wide genes per spot with calibrated uncertainty while preserving foundation-model representations. Across 32 Visium datasets, Coladan improves Pearson correlation from 0.230 to 0.431 (~ 1.9 ×), shows pathway-level enrichment consistency, and transfers zero-shot to VisiumHD and spot-level Xenium. Classification token (CLS) embedding-only perturbation performs on par with expression-based baselines, enabling image-only virtual perturbation without measured expression, illustrated on normal and cancer prostate sections for in-situ hypothesis generation.