Hi everyone,

I'd like to share FlashDeconv, a high-performance spatial transcriptomics deconvolution tool we recently developed.

Motivation

With the advent of Visium HD and other high-resolution spatial platforms generating millions of spots per sample, computational efficiency becomes increasingly important. While excellent tools like cell2location, RCTD, and CARD have established strong foundations for spatial deconvolution, we saw an opportunity to develop a complementary approach optimized specifically for ultra-large-scale datasets.

Our Approach: FlashDeconv

FlashDeconv uses structure-preserving randomized sketching to achieve:

Quick Start

Installation:

pip install flashdeconv

Usage:

import flashdeconv as fd
import scanpy as sc

# Load your spatial and single-cell data
adata_spatial = sc.read_h5ad("spatial.h5ad")
adata_sc = sc.read_h5ad("reference.h5ad")

# Run deconvolution
fd.tl.deconvolve(adata_spatial, adata_sc, cell_type_key="cell_type")

# Results stored in adata_spatial.obsm["deconvolution"]

When to Use FlashDeconv

FlashDeconv is particularly useful when:

• Working with Visium HD or other high-resolution platforms with >100k spots
• Running large-scale analyses across many samples
• Computational resources are limited

For smaller datasets, established tools like cell2location, RCTD, and CARD remain excellent choices with their own unique strengths.

Links

• GitHub: https://github.com/cafferychen777/flashdeconv
• Paper (bioRxiv): https://doi.org/10.64898/2025.12.22.696108
• PyPI: https://pypi.org/project/flashdeconv/

We'd love to hear your feedback and are happy to answer any questions!