A complete guide to spatial transcriptomics cell type deconvolution using FlashDeconv and the Python ecosystem.

Atlas-scale spatial transcriptomics requires deconvolution methods that preserve rare biological signals without prohibitive computational costs. Here, we introduce FlashDeconv, a framework built on structure-preserving randomized sketching. Unlike variance-based methods that conflate biological information with population abundance, FlashDeconv employs leverage-score importance sampling to prioritize transcriptomically distinct markers—preserving rare cell type signals that standard feature selection discards. Benchmarking demonstrates accuracy comparable to top-tier Bayesian methods while accelerating inference by orders of magnitude. Applied to human ovarian cancer cohorts, FlashDeconv reproduces clinical response signatures in seconds, enabling rapid patient stratification. This throughput also enables systematic scale-space exploration: we define a “resolution horizon” (8–16 µm) beyond which cellular co-localization signals undergo sign inversion due to geometric mixing. Operating below this horizon, FlashDeconv uncovers cryptic Tuft cell niches enriched for intestinal stem cells—biological architecture obscured by both variance-based feature selection and coarse spatial binning. FlashDeconv provides a scalable, mathematically grounded framework for atlas-scale spatial discovery.

The is the second in a two-part series of using publicly available data to replicate a published work in public health.

The is the first in a two-part series of using publicly available data to replicate a published work in public health.