Journal Club: Pushing Spatial Transcriptomics Toward Single-Cell Resolution—Two Computational Strategies and an Epigenetic Reality Check

· 9 min read · Paper Feed

Journal Club: Pushing Spatial Transcriptomics Toward Single-Cell Resolution—Two Computational Strategies and an Epigenetic Reality Check

Conceptual illustration: Pushing Spatial Transcriptomics Toward Single-Cell Resolution—Two Computational Strategies and an Epigenetic Reality Check

Introduction

Spatial transcriptomics has changed how we look at tissue biology, letting us ask not just which genes are expressed, but where they are expressed within intact tissue architecture. Every spatial platform comes with a tradeoff, though: spot-based methods like 10x Visium mix multiple cells into each capture area, subcellular platforms like Visium HD or Stereo-seq measure only a fraction of the transcriptome per cell, and even the best assays suffer from dropout. So the single-cell-resolution story we want to tell is always partly an inference story.

The three papers in this journal club approach that inference problem from complementary angles. CoexpressDeconvolve (Perik-Zavodskaia et al., iScience, 2026) tackles spot-based mixing head-on, attempting reference-free, single-cell–resolution deconvolution directly from Visium-style data. stPainter (Yang et al., Nature Communications, 2026) takes a generative-AI approach, using a pan-cancer scRNA-seq foundation model to "paint in" missing genes on sparse spatial maps. Dionisi et al. (Nature Communications, 2026) provide a complementary biological anchor by mapping the epigenetic signatures that distinguish early human B-cell lineage branches—a reminder that chromatin state, not just RNA abundance, defines cellular fate.

Read together, these papers sketch a coherent picture: computational methods can substantially upgrade the resolution of spatial data, but only if we keep refining the molecular and mechanistic ground truth those methods are meant to recover.

Paper 1: CoexpressDeconvolve — Reference-Free Deconvolution from Spot Data

Key Findings

  • Integer cell counts per spot, no reference required. CoexpressDeconvolve reports the lowest cell-count error among tested tools on synthetic Visium benchmarks, beating Tangram, cell2location, and STdeconvolve.
  • Highest per-slide cell-type concordance. On the same synthetic data, the framework achieved the top agreement with ground-truth cell-type composition per slide.
  • Competitive expression reconstruction. Gene-level reconstruction fidelity was on par with reference-based baselines, despite using no external scRNA-seq atlas.
  • Output mimics Space Ranger. The tool returns a feature-barcode matrix compatible with standard scRNA-seq pipelines, which the authors demonstrate on human breast cancer and tongue squamous cell carcinoma (TSCC) datasets, resolving tumor microenvironment composition and identifying malignant progression axes.

Methodology & Study Design

  • Combines a hybrid housekeeping-library-size calibration with topic modeling on a spatial gene co-expression manifold.
  • Genes are clustered by co-expression, each cluster is fit to a distribution, and a latent topic model estimates how many "pseudo-cells" of each program occupy each spot.
  • Benchmarked on synthetic Visium data (where ground truth is known), then applied to two real cancer datasets (breast cancer and TSCC) for qualitative validation.
  • Compared head-to-head against Tangram, cell2location, and STdeconvolve, three widely used deconvolution tools representing deep-learning cell-to-spot mapping, Bayesian, and topic-modeling approaches, respectively.

Significance

Most deconvolution tools return fractions, not counts, and many require matched scRNA-seq references that are often unavailable or biased. CoexpressDeconvolve's combination of reference-free operation with integer cell counts and Space-Ranger-compatible output addresses a practical bottleneck: it can be applied to legacy Visium datasets where no tissue-matched scRNA-seq exists. That makes spatial analysis more reproducible across labs and less dependent on bespoke reference panels.

Paper 2: stPainter — A Generative Foundation Model for Sparse Spatial Data

Key Findings

  • Pan-cancer pretrained model with zero-shot transfer. stPainter, built on a latent diffusion architecture guided by Stochastic Differential Equations (SDEs), was pretrained on a large pan-cancer scRNA-seq atlas and applied to six different spatial transcriptomics datasets across distinct cancer types without retraining.
  • Improved downstream clustering and pathway analysis. Imputed expression led to finer-grained subpopulation discovery and stronger pathway enrichment than raw sparse data.
  • Cross-modality validation with CODEX. Comparison with spatially resolved proteomics (CODEX) provided independent support that imputed cellular compositions agree with protein-level tissue organization.

Methodology & Study Design

  • A conditional generative model (a latent diffusion architecture guided by stochastic differential equations, per the abstract) that takes sparse spatial measurements as conditioning input and outputs expanded, near-single-cell expression profiles plus latent embeddings for clustering.
  • Pretrained once on a pan-cancer scRNA-seq atlas, then applied zero-shot across six tumor spatial datasets.
  • Validated against CODEX spatial proteomics, an orthogonal modality measuring dozens of proteins in the same tissue regions.
  • Evaluated on fine-grained subpopulation detection and pathway enrichment, two common downstream tasks in spatial tumor biology.

Significance

Where CoexpressDeconvolve emphasizes interpretability and reference independence, stPainter bets on scale and generalization. This is the emerging "foundation model" paradigm for spatial omics: one large pretrained model that handles many tissues without dataset-specific tuning. For tumor microenvironment studies, where matched references are scarce and tissues are heterogeneous, this is a meaningful step toward plug-and-play spatial analysis.

Paper 3: Dionisi et al. — Epigenetic Bifurcation in Human B-Cell Lineages

Key Findings

  • T2 stage splits into IgMhi and IgMlo branches. Using bulk and single-cell ATAC-seq, CUT&RUN, RNA-seq, and CITE-seq on T1, T2Mhi, and T2Mlo cells from adult and cord blood, the authors map chromatin accessibility distinguishing the two T2 trajectories.
  • Epigenetic signatures persist into mature subsets. Accessible chromatin domains set in T2Mlo cells are retained in naive B cells, while T2Mhi signatures persist in memory and marginal zone B cells.
  • Spatial validation with imaging mass cytometry + RNAscope. Spleen, appendix, and tonsil tissue confirms the in situ expression of marker genes identified in the sequencing data.
  • Transcriptional diversity within mature populations. Memory and marginal zone subsets show transcriptional heterogeneity that reflects their earlier T2 origins.

Methodology & Study Design

  • Multi-modal profiling of rare sorted human B-cell subsets: ATAC-seq (chromatin accessibility), CUT&RUN (targeted chromatin marks), RNA-seq (transcriptome), and CITE-seq (joint RNA + surface protein).
  • Cells isolated from human adult and cord blood, sorted into T1, T2Mhi, and T2Mlo fractions.
  • Validated spatially in spleen, appendix, and tonsil using imaging mass cytometry (IMC) combined with RNAscope in situ hybridization.

Significance

Dionisi et al. do not propose a new computational method, but they lay out the kind of mechanistic, epigenetic ground truth that deconvolution and imputation tools assume but rarely verify. By showing that subtle chromatin differences correspond to durable, transcriptionally distinct lineages, they highlight that RNA-only inference of cell state may miss the molecular foundations of lineage commitment—an important caveat for any spatial method that reconstructs expression from mixed signals.

Synthesis & Discussion

Complementary, Not Competing, Approaches

CoexpressDeconvolve and stPainter sit at two philosophical poles in modern spatial transcriptomics. CoexpressDeconvolve says: don't trust a reference you don't have—let gene co-expression structure reveal cell programs. stPainter says: a massive pretrained atlas is the reference—zero-shot generalization is more valuable than purity. Both push toward single-cell resolution, but through very different inductive biases. Notably, neither is fully validated against the kind of multi-omic, lineage-resolved ground truth that Dionisi et al. provide.

What Patterns Emerge?

  1. Single-cell resolution is becoming the expectation, not the exception for spatial data, even when the raw measurements are multi-cell or sparse.
  2. The field is bifurcating into reference-free and reference-heavy camps, echoing the long-running debate in scRNA-seq annotation (clustering vs. label transfer).
  3. Epigenetic context is increasingly treated as essential for interpreting inferred cell states, particularly for lineage commitment and memory, as the B-cell study demonstrates.

Open Questions

  • How well do deconvolved or imputed cell assignments hold up against true joint RNA+ATAC spatial measurements, a technology that is now emerging?
  • Does a large pretrained atlas (stPainter) help or constrain discovery of novel or rare cell states compared to reference-free methods (CoexpressDeconvolve)?
  • Can epigenetic priors, such as the T2 bifurcation signatures identified by Dionisi et al., be incorporated into spatial deconvolution models to improve lineage resolution?
  • How well do these spatial methods generalize beyond cancer to immune, developmental, or inflammatory tissues where lineage decisions are central?

For Your Lab Meeting

  1. Methodological choice: If you had a Visium dataset from a tissue with no matched scRNA-seq, would you reach for CoexpressDeconvolve's reference-free approach or stPainter's pan-cancer atlas first? What are the failure modes of each?
  2. Validation gap: All three papers rely on different forms of validation (synthetic ground truth, CODEX proteomics, IMC + RNAscope). What single experiment would most rigorously test whether inferred single-cell states from spatial data are biologically real?
  3. Reference dependence: stPainter is pretrained on pan-cancer scRNA-seq. Could this atlas bias imputation away from non-malignant or tissue-specific cell states, especially in developmental or inflammatory contexts?
  4. Multi-omic integration: Should future spatial deconvolution methods incorporate chromatin accessibility or histone-mark data as priors, as suggested by Dionisi et al.'s epigenetic signatures? What would the data-architecture work look like?
  5. Causality vs. description: These papers mostly describe cell states. How would you design a study that moves from descriptive spatial maps to causal models linking epigenetic programs, spatial niches, and functional outcomes?

Key Terms

  • Spot-based spatial transcriptomics: Technologies like 10x Visium where each capture area (spot) contains RNA from multiple adjacent cells, producing mixed expression signals.
  • Deconvolution: Computational separation of a mixed signal into its component sources—in spatial omics, inferring which cell types contributed to each spot.
  • Topic modeling: A statistical method (borrowed from natural language processing) that discovers latent "topics" (here, gene programs or cell states) from co-occurrence patterns in a dataset.
  • Latent diffusion model: A generative deep learning architecture that learns to reverse a gradual noising process, used here (in stPainter) to impute missing gene expression from sparse measurements.
  • ATAC-seq / CUT&RUN: Epigenetic assays profiling chromatin accessibility (ATAC-seq) or specific histone modifications (CUT&RUN), used to map regulatory landscapes that precede transcriptional changes.

References

  1. Perik-Zavodskaia O, Perik-Zavodskii R, Alrhmoun S, & Sennikov S (2026). CoexpressDeconvolve enables reference-free single-cell-resolution deconvolution from spot-based spatial transcriptomics. iScience. https://doi.org/10.1016/j.isci.2026.116824

  2. Yang Y, Luo Y, Zhang K, Zhang Z, Peng H, Cao C, et al. (2026). Enhancing pan-cancer spatial transcriptomics at single-cell resolution with stPainter. Nature communications. https://doi.org/10.1038/s41467-026-76552-x

  3. Dionisi C, Kelly A, Pitcher MJ, Kottoor SH, Dalton A, Montorsi L, et al. (2026). Epigenetic signatures mark early peripheral human B lineage bifurcation and differential transcriptional profiles in mature populations. Nature communications. https://doi.org/10.1038/s41467-026-75364-3

Single-Cell Transcriptomics scRNA-seq Genomics