henhaobio
Sign in

bioRxivBioinformaticsPreprint

FlashDeconv reveals resolution horizons in atlas-scale spatial transcriptomics

Chen Yang, Xianyang ZhangiD, Jun Chen

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Coarsening Visium HD resolution from 8 to 64 m can flip cell-type co-localization from negative to positive (r = -0.12 [->] +0.80), yet many widely used compositional deconvolution workflows require coarsening or subsampling at million-bin scale. Here we introduce FlashDeconv, which combines leverage-score importance sampling with sparse spatial regularization to achieve competitive benchmark accuracy while processing 1.6 million bins in 153 seconds on commodity hardware. Systematic multi-resolution analysis of Visium HD mouse intestine reveals a tissue-specific resolution horizon (8-16 m), the scale at which this sign inversion occurs, validated by Xenium ground truth. Below this horizon, FlashDeconv provides, to our knowledge, 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 uncovers neutrophil inflammatory microdomains co-localized with immunoregulatory dendritic cells (mRegDC) at the tumor-stroma interface, spatial niches largely missed by discrete-label summaries, with RCTD doublet mode labeling only 2.3% of hotspot bins as neutrophil singlets.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 5 (11 Sept 2026) · CC BY 4.0

Leverage-weighted sketching plus sparse spatial regularization deconvolves 1.6 million Visium HD bins in 153 seconds on commodity hardware, revealing tissue-specific resolution limits where cell-type co-localization can reverse sign.

Many compositional deconvolution workflows coarsen or sub-sample million-bin Visium HD data, and spatially aware methods such as CARD need dense O(N^2) covariance, yet coarsening can flip inferred co-localization. FlashDeconv instead compresses genes with a leverage-weighted CountSketch, argued to retain rare cell-type markers that PCA or highly variable gene (HVG) selection underweight, and adds a sparse graph Laplacian penalty that scales as O(N).

Key results

  1. ShownRuntime and memory scale linearly with spot number, so million-bin Visium HD sections are deconvolved in minutes on consumer-grade hardware without a GPU.

    1,595,565 colorectal cancer bins (3 patients, 38 cell types) in 153 s (~10,400 bins/s); 351,817 mouse intestine bins at 8 µm in 12.0 s; 100,000 spots in under 4 s and a simulated 1 million spots in about 3 min on a MacBook Pro M2 Max (32GB, no GPU). Competitor scaling is cited or extrapolated, e.g. RCTD over 30 h for 1 million spots from its documentation. · Linear scalability enables atlas-scale…; Atlas-scale deconvolution of human colorectal…; Scale-space analysis of Visium HD reveals…; Benchmarking

  2. ShownOn the Spotless benchmark, accuracy ranked first of 13 methods on synthetic mixtures and was competitive but mixed on real data; in ablation, leverage weighting accounted for the rare-cell gains.

    54 Silver Standard datasets: Pearson 0.944, Jensen-Shannon divergence (JSD) 0.047 vs RCTD 0.934/0.050, Cell2Location 0.918/0.061, NNLS 0.850/0.118; rare-cell r 0.35 with leverage sketching vs 0.14 uniform, 0.16 PCA, 0.12 HVG. Real data: STARMap Pearson 0.665 (3rd), lowest JSD 0.252; seqFISH+ olfactory bulb 7th; liver Visium JSD 0.0561 (3rd) but area under the precision-recall curve (AUPR) 0.66 (7th). · Downstream impact on deconvolution accuracy; Cross-tissue generalization; FlashDeconv achieves competitive accuracy on…; Real-world validation on Visium case studies

  3. ShownBin size changes biological conclusions. In mouse small intestine, single-cell-type purity collapses between 8 and 16 µm and the Paneth–Goblet correlation reverses sign, while mouse brain cortex degrades gradually. Imaging-based ground truth reproduces both patterns.

    Mouse small intestine bins >80% one type: 61.5% (8 µm), 13.3% (16 µm), 3.9% (128 µm); entropy 0.23 to 0.48. Paneth–Goblet r = −0.12 at 8 µm (N = 351,817) vs +0.80 at 64 µm; Xenium mouse colon without deconvolution: r = −0.06 (8 µm) vs +0.68 (128 µm). Brain cortex purity 0.54 to 0.36 over 8–128 µm, 12 of 28 lineage pairs flip sign, confirmed by MERFISH (115,772 cells). · The resolution horizon; Cross-tissue validation: mouse brain cortex; Coarse binning induces spurious colocalization

  4. ShownIn a 3-patient human colorectal cancer Visium HD cohort, continuous proportions exposed neutrophil microdomains at the tumor–stroma interface co-localized with LAMP3+ mature regulatory dendritic cells (mRegDC). RCTD doublet-mode labels largely missed them; independent Xenium data supported both.

    Of 16,827 neutrophil hotspot bins, RCTD called 2.3% neutrophil singlets and withheld 61%. Xenium pseudo-bins (Patient 1): lineage r > 0.88 at 8–128 µm; 4 µm AUPRC 0.87 vs 0.62 NNLS and 0.60 marker scoring; 38-type proportions r = 0.78 vs −0.02 for RCTD singlets. Marteau Xenium atlas: 10 of 15 patients showed mRegDC enrichment (median 1.49-fold) in a 50 µm radius around neutrophil clusters. · Atlas-scale deconvolution of human colorectal…; Discussion

  5. SuggestedAt 8 µm in mouse small intestine, focal Tuft cell niches (whose markers variance-based selection underranks) were strongly enriched for intestinal stem cells. The signal fades at coarse resolution and was weaker and more fragmented with RCTD.

    2,244 Tuft niches reaching 61% Tuft proportion; 16.2-fold stem cell and 2.8-fold enteroendocrine enrichment, enterocytes 0.20×, goblet 0.14×, Paneth 0.15× (permutation p < 10^−3). Tuft markers rank 21 percentile points lower by variance than by leverage. Maximum Tuft proportion 61% at 8 µm vs 4% at 128 µm. At 16 µm, stem enrichment at hotspots 7.7-fold vs 3.0-fold with RCTD. · FlashDeconv resolves the Tuft-Stem chemosensory…; Cross-method comparison confirms niche…

Methods, data and reproducibility

Graph-regularized non-negative least squares in a 512-dimensional sketch: HVGs plus reference markers (Log-CPM) are projected by a leverage-weighted CountSketch, with ℓ1 and k-nearest-neighbour Laplacian penalties, solved by block coordinate descent with Laplace-approximation intervals. Visium HD was deconvolved mainly at 8 to 128 µm bins, with 2–4 µm bins characterized for data quality (Fig. 6 spans 2–128 µm).

  • Spotless benchmark: Silver Standard synthetic pseudo-spot datasets (6 tissues × 9 abundance patterns) and Gold Standard STARMap and seqFISH+ data · mouse (stated for Gold Standards and kidney) · n = 54 Silver Standard datasets; STARMap 108 spots; seqFISH+ 7 fields of view · Zenodo 10277187
  • Spotless Visium case studies: mouse liver (zonation annotations, snRNA-seq reference of 133,779 cells) and melanoma (Molecular Cartography ground truth) · mouse (liver); melanoma not stated · n = liver: 4 slides, 5,762 spots; melanoma: 3 slides, 7,557 spots · Zenodo 10277187 (Spotless benchmark datasets)
  • Mouse brain single-cell reference (leverage analyses) and mouse brain Visium sections from the cell2location data portal · mouse · n = reference: 40,532 cells, 59 cell types, 31,053 genes · E-MTAB-11114 (Visium); E-MTAB-11115 (snRNA-seq reference)
  • Visium HD mouse small intestine (FFPE, 10x Genomics) with Haber et al. scRNA-seq reference; Xenium fresh frozen mouse colon (10x Genomics) as cell-level ground truth · mouse · n = 351,817 bins at 8 µm; reference 10,896 cells · GSE92332 (Haber reference)
  • Visium HD mouse brain cortex, with Allen Brain Cell Atlas MERFISH as ground truth · mouse · n = 453,820 bins; MERFISH 115,772 cells
  • Visium HD human colorectal cancer cohort (Oliveira et al.) with Chromium Flex scRNA-seq reference and adjacent-section Xenium data (Patient 1) · human · n = 3 patients; 1,595,565 bins at 8 µm; 38 cell types; 289,352 Xenium cells · 10.5281/zenodo.15042463
  • Colorectal cancer validation cohorts: Marteau et al. Xenium atlas, Pelka scRNA-seq atlas, TCGA-COAD · human · n = 15 patients (3.7 million cells); 60 patients; 592 patients
  • Human ovarian cancer Visium data (Denisenko et al.) · human · GSE211956
Methods & tools
FlashDeconv, Numba, Spotless benchmark, RCTD, Cell2Location, SpatialDWLS, NNLS, Enrichr, Seurat v3 HVG selection, Visium HD, Xenium, MERFISH
Software & environment
Open-source Python package; solver inner loops JIT-compiled with Numba. No package versions, containers or requirement files are stated.
Assay conditions & compute
Runtime and memory benchmarks on an Apple MacBook Pro, M2 Max, 32GB unified memory, macOS, no GPU (8 CPU cores in the scaling figure). Examples: 1,595,565 bins in 153 s; 351,817 bins in 12.0 s; Laplace uncertainty added 1.1 s to a 7.6 s run on 241,351 bins.
Reproducibility gaps
No access links are given for the Visium HD mouse brain cortex, Allen Brain Cell Atlas MERFISH, Marteau Xenium, Pelka scRNA-seq and TCGA-COAD data used for validation; several analyses are documented only in Supplementary Notes not included in this text.
Limitations (6)
  • With small datasets (under 50 spots) sketching and spatial smoothing add little: on seqFISH+ olfactory bulb (fewer than 10 spots per field of view) the method ranked 7th; on seqFISH+ cortex all methods, including FlashDeconv, had Pearson below 0.30. · stated by the authors
  • Nearly collinear reference signatures limit spot-level attribution: liver endothelial subtypes (cosine similarity >0.975) gave AUPR rank 7 despite JSD rank 3, and on melanoma the preprocessing choice trades rare-type fidelity against the dominant melanocytic proportion. · stated by the authors
  • On single-cell-resolution platforms such as Xenium, spatial regularization can spread signal across cell boundaries; the authors recommend dedicated classification or doublet models for closely related subtypes. · stated by the authors
  • Cell types absent from the reference bias estimates, and cross-modality differences in marker ranking can degrade leverage weighting; with dominant batch effects, explicit noise models may be more robust, and wider platform-mismatch testing is left for future work. · stated by the authors
  • Accuracy for the 12 competing methods comes from the published Spotless results rather than reruns, and large-scale RCTD and Cell2Location runtimes are cited or extrapolated rather than measured here. · noted in this summary
  • The resolution-horizon and Tuft-niche results each rest on one Visium HD mouse small intestine dataset (brain cortex likewise one dataset), and the human colorectal cancer Visium HD cohort has 3 patients. · noted in this summary

Read this if you deconvolve Visium HD or other million-bin spatial data, need rare cell types preserved, or must choose a bin size without creating spurious co-localization.

Citations
2

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLopez '22Gortana '26Taguchi '25Luo '24Swain '23Radkevich '24
66% semantic matchBioinformatics
Romain Lopez, Baoguo Li, Hadas Keren‐Shaul+14 more·Nature Biotechnology·Apr 21, 2022· 275 citations