FlashDeconv enables atlas-scale, multi-resolution spatial deconvolution via structure-preserving sketching

Coarsening Visium HD resolution from 8 to 64 <i>μ</i>m can flip cell-type co-localization from negative to positive ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>r</mml:mi> <mml:mo>=</mml:mo> <mml:mo>-</mml:mo> <mml:mn>0.12</mml:mn> <mml:mo>→</mml:mo> <mml:mo>+</mml:mo> <mml:mn>0.80</mml:mn></mml:math> ), yet investigators are routinely forced to coarsen because current deconvolution methods cannot scale to million-bin datasets. Here we introduce FlashDeconv, which combines leverage-score importance sampling with sparse spatial regularization to match top-tier Bayesian accuracy while processing 1.6 million bins in 153 seconds on a standard laptop. Systematic multi-resolution analysis of Visium HD mouse intestine reveals a tissue-specific <i>resolution horizon</i> (8-16 <b><i>μ</i></b> m)-the scale at which this sign inversion occurs-validated by Xenium ground truth. Below this horizon, FlashDeconv provides the first sequencing-based quantification of Tuft cell chemosensory niches (15.3-fold stem cell enrichment). In a 1.6-million-bin human colorectal cancer cohort, FlashDeconv further uncovers neutrophil inflammatory microdomains in low-UMI regions that classification-based methods discard, recovering spatially organized biology from measurements previously considered uninformative.