Journal Club: Denoising, Deconvolution, and Graph Models in Spatial Transcriptomics

· 9 min read · Paper Feed

Journal Club: Denoising, Deconvolution, and Graph Models in Spatial Transcriptomics

Conceptual illustration: Denoising, Deconvolution, and Graph Models in Spatial Transcriptomics

Introduction

Spatial transcriptomics (ST) has moved from a niche curiosity into a routine tool in molecular biology. By preserving tissue architecture while quantifying gene expression, ST lets us ask where a gene is active, not just whether it is active. That positional information has reshaped how we think about tumor microenvironments, brain regionalization, and developmental patterning. The field still wrestles with three stubborn problems: noisy and sparse data, the resolution-versus-coverage trade-off, and the difficulty of pulling interpretable biology out of high-dimensional spatial measurements.

The three papers in this journal club tackle these problems from different angles. Lee et al. (2025) provides a broad clinical-research review framing the spatial omics landscape, from technologies to translational applications. Yang et al. (2025) introduces Spotiphy, a computational toolkit that deconvolves sequencing-based ST spots into single-cell-resolved whole-transcriptome images. Chitra et al. (2025) presents GASTON, an interpretable deep learning method that models spatial gene expression as a topographic landscape via a quantity called isodepth. Read together, they show a field consolidating its tools while pushing toward single-cell resolution, mechanistic interpretation, and clinical relevance.

This post summarizes each paper, highlights its technical contributions, and proposes discussion questions for your next lab meeting.

Paper 1: Spatial Omics in Clinical Research — A Field-Wide Review

Key Findings

  • Multi-modal coverage: The review systematically catalogues spatial transcriptomics, proteomics (e.g., multiplexed ion-beam imaging, CODEX), and epigenomics (e.g., spatial ATAC-seq, spatial CUT&Tag), emphasizing that each modality answers distinct biological questions.
  • Disease applications converge on three areas: cancer (tumor–stroma boundaries, immune infiltration), neurology (cell-type localization in Alzheimer's, Parkinson's), and immunology (lymphoid tissue organization, infection response).
  • Validation gap identified: The authors stress that without orthogonal validation — paired scRNA-seq, immunohistochemistry, or in situ hybridization — spatial findings risk being artefactual, especially at sub-spot resolution.
  • Translational emphasis: They argue that standardization of sample handling, benchmarking of computational pipelines, and integration with traditional omics are the rate-limiting steps for clinical adoption, not the assays themselves.

Methodology & Study Design

This is a narrative review rather than a primary research paper. The authors synthesize recent (2020–2025) literature on spatial omics platforms — 10x Visium, Slide-seq, Stereo-seq, MERFISH, and imaging-based chemistries. They evaluate computational pipelines for cell-type decomposition (cell2location, Tangram, Stereoscope), domain detection (SpaGCN, BayesSpace), and multi-modal data fusion. Sample sizes are not the focus, but the review explicitly compares trade-offs, such as 10x Visium HD's 2 µm resolution versus Stereo-seq's sub-cellular spot density.

Significance

The review functions as a roadmap rather than a single advance. Its value lies in crystallizing where the field stands: which technologies are clinically ready, which remain research-grade, and where the methodological bottlenecks live. For labs considering adopting spatial omics, this is a pragmatic entry point — particularly the discussion of reproducibility, batch effects, and the lack of community-wide analytical standards.

Paper 2: Spotiphy — Pseudo-Single-Cell Whole-Transcriptome Imaging

Key Findings

  • Benchmark performance: Spotiphy delivers the most precise cellular proportions in extensive benchmarking across multiple platforms (10x Visium HD, Stereo-seq, Slide-seqV2), outperforming existing deconvolution tools.
  • Alzheimer's disease insight: Applied to mouse brain ST data, Spotiphy-derived single-cell profiles revealed regional specifications of astrocytes and disease-associated microglia (DAMs) in Alzheimer's models — patterns that were invisible at spot-level resolution.
  • Tumor microenvironment remodeling: In human breast cancer ST data, Spotiphy identified multiple spatial domains and alterations in tumor–TME interactions that bulk-style analyses would miss.
  • Whole-section capability: Unlike imaging-based ST methods limited by gene panel size (often a few hundred to a few thousand genes), Spotiphy produces whole-transcriptome images across the entire tissue section.

Methodology & Study Design

Spotiphy (spot imager with pseudo-single-cell-resolution histology) is a computational toolkit, not a new wet-lab chemistry. It takes sequencing-based ST outputs, where each spot captures multiple cells, and uses a probabilistic model that combines histological image features with expression-based priors to infer single-cell-resolved transcriptomes. The "pseudo-single-cell" qualifier matters: Spotiphy estimates what each individual cell's transcriptome would look like, but does not physically isolate single cells. Key inputs include H&E or high-resolution histology images co-registered with ST spots. Datasets span healthy and Alzheimer's mouse brain plus human breast cancer samples.

Significance

Spotiphy attacks the central resolution-versus-coverage dilemma. Imaging-based methods (MERFISH, seqFISH) give single-cell resolution but only profile a few hundred to a few thousand genes, while sequencing-based methods (Visium, Stereo-seq) cover whole transcriptomes but at multi-cell resolution. By bridging these, Spotiphy makes single-cell whole-transcriptome spatial analysis tractable on commodity sequencing hardware — a meaningful democratization of the technology.

Paper 3: GASTON — Topographic Maps of Spatial Gene Expression

Key Findings

  • The isodepth concept: GASTON introduces isodepth — analogous to elevation on a topographic map — as a 1D coordinate summarizing a cell's spatial position. Contours of constant isodepth enclose domains with distinct cell-type composition.
  • Brain gradients recovered: Applied to mouse brain ST data, GASTON identified gradients of neuronal differentiation and firing activity that align with known anatomical axes.
  • Tumor microenvironment gradients: In tumor data, it uncovered gradients of metabolism and immune activity across the TME, identifying spatial transitions invisible to clustering-based methods.
  • Interpretable by design: Unlike black-box deep learning models, GASTON learns piecewise linear expression functions and explicit spatial gradients — meaning each output is a literal mathematical model of gene behavior across space.

Methodology & Study Design

GASTON (Gradient Analysis of Spatial Transcriptomics Organization with Neural networks) is an unsupervised, interpretable deep learning algorithm. It takes ST count data plus spatial coordinates and learns:

  1. An isodepth function mapping spatial location → 1D topographic coordinate.
  2. Spatial gradients (vectors pointing in the direction of maximum expression change).
  3. Piecewise linear functions modeling how each gene's expression varies along the topography — capturing both smooth gradients and abrupt domain boundaries.

The neural network architecture is constrained so that its outputs correspond directly to these mathematical objects, rather than being latent features that require post-hoc interpretation. Tested across multiple tissues and species, including tumor and brain datasets.

Significance

GASTON reframes spatial gene expression not as a clustering problem (find discrete domains) but as a landscape problem (find continuous gradients and abrupt transitions simultaneously). Most existing tools excel at one or the other; GASTON aims at both. The interpretability-by-design architecture is also a methodological statement: deep learning in genomics does not have to be opaque.

Synthesis & Discussion

Read together, these papers trace a clear trajectory. Lee et al. tells us what spatial omics can do clinically and where the bottlenecks are. Spotiphy addresses the resolution gap, pushing sequencing-based ST toward single-cell granularity without abandoning whole-transcriptome coverage. GASTON attacks the interpretation gap, giving us mathematical objects — gradients, isodepth contours — that map directly onto biological hypotheses.

A few cross-cutting themes:

  • Computational tooling is the limiting reagent. All three papers, especially Spotiphy and GASTON, argue that wet-lab spatial technologies have outpaced our ability to analyze the data. The field is now in a tools-building phase.
  • Interpretability matters for clinical translation. GASTON's piecewise linear functions and Spotiphy's whole-transcriptome outputs are both designed to produce biologically legible results. Black-box outputs are increasingly seen as insufficient.
  • Validation remains the weakest link. Lee et al. flag this most explicitly, but Spotiphy's claims of single-cell resolution still rely on computational inference rather than direct measurement, and GASTON's isodepth must be verified against ground-truth anatomy.
  • The tumor microenvironment is the common proving ground. All three papers cite cancer applications, suggesting TME biology — with its mixture of malignant, stromal, and immune cells — is where spatial methods deliver the highest return on investment.

Open questions: how well do Spotiphy's inferred single-cell profiles hold up against matched scRNA-seq? Can GASTON's isodepth concept generalize to 3D spatial datasets (MERFISH stacks, Stereo-seq thick sections)? And will standardized benchmarking, as Lee et al. advocate, become community practice?

For Your Lab Meeting

  1. Resolution vs. coverage: Spotiphy claims to achieve single-cell whole-transcriptome resolution computationally. Under what biological conditions would you expect this inference to break down (e.g., highly overlapping cell types, low RNA content cells)?
  2. Isodepth as abstraction: GASTON's isodepth is a learned 1D coordinate, not a physical axis. What experiments could you design to test whether isodepth corresponds to a true biological axis (e.g., developmental time, differentiation state)?
  3. Clinical readiness: Lee et al. argue standardization is the rate-limiting step. Which of the bottlenecks they list — sample handling, batch correction, benchmarking — would you tackle first in your own work?
  4. Method benchmarking: Both Spotiphy and GASTON outperform competitors on benchmarks. What gold-standard dataset would you build to fairly compare these tools head-to-head?
  5. Integration with single-cell atlases: How could these spatial tools be combined with reference scRNA-seq atlases (e.g., Tabula Sapiens, HCA) to refine cell-type annotations in spatial data?

Key Terms

  • Spatial transcriptomics (ST): A family of technologies that measure gene expression while preserving the spatial location of each measurement within a tissue section.
  • Deconvolution: A computational process that estimates the proportions (or profiles) of individual cell types contributing to a mixed signal — in this case, multiple cells captured within a single ST spot.
  • Isodepth: A learned 1D coordinate (introduced by GASTON) that summarizes a spatial location's position along a tissue's expression landscape, analogous to elevation on a topographic map.
  • Tumor microenvironment (TME): The ecosystem of cancer cells, stromal cells, immune cells, vasculature, and extracellular matrix surrounding a tumor — a frequent application area for spatial methods.
  • Disease-associated microglia (DAM): A transcriptionally distinct microglial state implicated in neurodegenerative diseases such as Alzheimer's.

References

  1. Yoonji Lee, Mingyu Lee, Yoojin Shin, Kyuri Kim, & Taejung Kim (2025). Spatial Omics in Clinical Research: A Comprehensive Review of Technologies and Guidelines for Applications. International Journal of Molecular Sciences. https://doi.org/10.3390/ijms26093949

  2. Jiyuan Yang, Ziqian Zheng, Yun Jiao, Kaiwen Yu, Sheetal Bhatara, Xu Yang, et al. (2025). Spotiphy enables single-cell spatial whole transcriptomics across an entire section. Nature Methods. https://doi.org/10.1038/s41592-025-02622-5

  3. Uthsav Chitra, Brian J. Arnold, Hirak Sarkar, Kohei Sanno, Cong Ma, Sereno Lopez-Darwin, et al. (2025). Mapping the topography of spatial gene expression with interpretable deep learning. Nature Methods. https://doi.org/10.1038/s41592-024-02503-3

Spatial Transcriptomics Spatial Omics Single-Cell