Journal Club: From Spot-Level Signals to Cell-Type-Resolved 3D Tissue Maps in Spatial Transcriptomics

Introduction

Spatial transcriptomics (ST) has changed how we study tissue biology by linking gene activity with spatial coordinates. Whether you're probing the tumor microenvironment, charting organ development, or mapping cell–cell communication, ST can answer questions that bulk or single-cell RNA-seq alone can't. But the technology has outpaced its own ecosystem. Hundreds of computational tools now exist for denoising, clustering, deconvolution, alignment, and domain identification, and benchmarks are scattered across dozens of papers with conflicting conclusions. Method choice can noticeably alter which tissue domains are detected, which spatially variable genes (SVGs) rise to the top, and which ligand–receptor interactions get inferred, shaping the biological conclusions downstream.

The three papers in this journal club address a single overarching question: how do we extract biologically meaningful, high-resolution structure from spatial transcriptomics data, and how do we choose the right tools for the job? Gillespie et al. (2025) offer a meta-review that synthesizes benchmarking evidence across four core ST tasks, giving practitioners an evidence-based decision framework. Tu et al. (2025) introduce STged, a deconvolution framework that recovers cell-type-specific gene expression rather than just cell-type proportions. Zhang et al. (2026) push the field into continuous three-dimensional tissue reconstruction with SINTER3D, an implicit neural representation (INR) framework. Together, these works chart a trajectory from spot-level summaries to cell-type-resolved and ultimately continuous 3D molecular atlases.

Paper 1: Navigating the Spatial Transcriptomics Software Maze

Key Findings

• Synthesizes benchmarking results across four core ST analysis tasks: tissue architecture identification, spatially variable gene discovery, cell–cell communication, and deconvolution.
• Provides practical metrics — accuracy, robustness, scalability, interpretability, and hardware requirements — that practitioners can match against dataset and resource constraints.
• Highlights where consensus is strong across benchmarks (e.g., certain deconvolution and clustering tools consistently outperform others) versus where results are method- or dataset-dependent.
• Emphasizes that the field suffers from fragmented benchmarking: individual studies test a handful of tools on one or two datasets, making cross-study comparison difficult.

Methodology & Study Design

This is a meta-review, not an original benchmarking study. The authors systematically surveyed existing benchmarking papers that evaluated ST software across the four tasks. Rather than re-running benchmarks, they aggregated evidence into a unified comparative framework, an approach aligned with the broader trend toward systematizing computational biology through meta-analyses and best-practice recommendations.

Significance

Gillespie et al. distinguish themselves from previous ST reviews that either catalog tools or benchmark a narrow task. By synthesizing multiple benchmarks, they provide a defensible, transparent framework for pipeline design. As the technology moves from boutique to routine in developmental biology, neuroscience, and cancer research, this kind of consolidated guidance matters for reproducibility and clinical translation.

Paper 2: STged — Deconvolving Expression, Not Just Cell Types

Key Findings

• STged reconstructs cell-type-specific gene expression profiles from mixed spots, going beyond traditional deconvolution that only estimates cell-type proportions.
• In simulations, STged consistently outperforms existing deconvolution methods in accuracy and robustness.
• Applied to human pancreatic ductal adenocarcinoma (PDAC) and human squamous cell carcinoma (SCC), STged identifies microenvironment-specific highly variable genes and reconstructs spatial cell–cell communication networks.
• In mouse kidney tissue, STged uncovers dynamic spatial gene expression patterns and distinct gene programs, enabling near-single-cell resolution analysis of tissue heterogeneity.

Methodology & Study Design

STged uses a non-negative least-squares (NNLS) regression framework that integrates:

1. Graph-based spatial correlations — leveraging the spatial neighborhood structure of spots.
2. Reference-derived gene signatures — typically from matched single-cell RNA-seq data.

The NNLS formulation ensures that reconstructed expression profiles remain biologically interpretable (non-negative, additive contributions). By modeling expression as a function of both reference signatures and spatial context, STged solves a deeper inverse problem: what is each cell type expressing in its local tissue neighborhood? rather than merely which cell types are present?

Significance

This work represents a clear step forward from composition-only deconvolution. Traditional methods answer "who is here?"; STged answers "what are they doing here?" That shift enables downstream analyses — spatially aware differential expression, microenvironment-specific signaling inference, and near-single-cell resolution domain mapping — that were previously inaccessible from spot-level data alone. It positions graph-based, reference-guided regression as a strong contender in the growing arsenal of learning-based and autoencoder-based ST methods.

Paper 3: SINTER3D — Continuous 3D Reconstruction with Implicit Neural Representations

Key Findings

• SINTER3D models gene expression as continuous functions of 3D coordinates, enabling joint interpolation across multiple genes simultaneously.
• Demonstrates superior performance across five diverse datasets: adult mouse brain, human dorsolateral prefrontal cortex (DLPFC), developing human heart, Drosophila embryo, and breast cancer tissue.
• Enables three downstream applications: (1) virtual section generation, (2) spatial-domain identification, and (3) cell-type deconvolution in 3D.
• Reconstructs biologically meaningful 3D molecular structures that capture tissue architecture across organs and species.

Methodology & Study Design

The core innovation is the use of implicit neural representations (INRs), a class of neural networks that parameterize continuous functions. In SINTER3D, each gene's expression is modeled as a function ( f(x, y, z) ) that takes 3D coordinates as input and outputs expression intensity. Unlike traditional approaches that interpolate each gene independently and struggle with large inter-section gaps, SINTER3D learns a shared neural representation that captures spatial smoothness across all genes simultaneously. This allows the model to fill in missing tissue between serial sections with biologically coherent expression patterns.

Significance

SINTER3D pushes spatial transcriptomics beyond the 2D slide paradigm. Current ST protocols capture thin tissue sections, leaving large 3D gaps between sampled slices. SINTER3D's continuous representation lets researchers generate virtual sections at arbitrary depths, effectively turning a stack of 2D snapshots into a navigable 3D volume. The INR paradigm trades some interpretability for flexibility and smoothness, and its generalization across five tissue types and species suggests a robustness that many ST methods lack.

Synthesis & Discussion

How Do These Papers Complement Each Other?

These three works occupy complementary positions on the ST analysis spectrum:

• Gillespie et al. (2025) answers the meta-question: which tools should I use, and what evidence supports my choice? It is a prescriptive map of the existing ecosystem.
• Tu et al. (2025) answers a methodological gap: how do I move from spot-level composition to cell-type-specific expression? It is a technical innovation in 2D deconvolution.
• Zhang et al. (2026) answers a dimensionality gap: how do I move from discrete 2D sections to continuous 3D tissue volumes? It is a technical innovation in spatial reconstruction.

Together, they describe a clear trajectory: the field is moving from spot-level → cell-type-resolved → continuous 3D representations of tissue biology.

Agreements and Tensions

Agreements:

• All three recognize deconvolution and spatial domain identification as central analytic challenges.
• All emphasize using spatial structure — graph neighborhoods in STged, coordinate-based continuity in SINTER3D, and task-specific method selection in the meta-review.
• All implicitly support the trend toward best-practice guidelines for ST analysis.

Tensions:

• Interpretability vs. flexibility: STged's NNLS framework is relatively transparent; SINTER3D's INR is a black box that sacrifices interpretability for expressive power.
• 2D vs. 3D: STged operates on individual tissue sections; SINTER3D explicitly transcends 2D into continuous volumes. How STged's cell-type-specific deconvolution would perform in a 3D SINTER3D-reconstructed volume is an open question.
• Synthesis vs. innovation: Gillespie et al. consolidate existing evidence; STged and SINTER3D propose algorithmic advances. The meta-review does not yet evaluate methods like STged or SINTER3D, leaving a gap that future benchmarks must fill.

Emerging Patterns

1. From coarse to fine resolution: ST is transitioning from composition maps (cell types per spot) to expression maps (genes per cell type per spot) and further to continuous fields (gene expression as a function of space).
2. Stronger spatial modeling: Graph-based (STged) and coordinate-based (SINTER3D) models are becoming standard, replacing purely expression-based methods.
3. Maturation toward standardization: The meta-review reflects a field ready for consensus guidelines, not just tool proliferation.
4. Cross-organ generalization: SINTER3D's validation across five tissue types signals that the field is moving beyond single-organ demos.

What Questions Remain Unanswered?

• How do STged and SINTER3D perform against each other on the same datasets, and how should they be integrated into a unified pipeline?
• Can INR-based continuous models be made interpretable enough for clinical or regulatory use?
• How do we validate inferred 3D structures and cell–cell communication networks experimentally (e.g., multiplex imaging, lineage tracing)?
• How do these methods scale to whole-organ or whole-body atlases, and what are the computational costs?
• Can models trained on one species transfer to another, and how do we quantify uncertainty in deconvolved or interpolated expression?

For Your Lab Meeting

1. Methodology trade-offs: STged uses interpretable NNLS regression; SINTER3D uses black-box INRs. For your next ST project, would you prioritize interpretability (STged) or flexibility and 3D continuity (SINTER3D)? What factors would tip the balance — dataset size, biological question, downstream analysis, or regulatory requirements?

2. Benchmarking gaps: Gillespie et al. note that benchmarks are fragmented. If you were designing a new benchmarking study, what metrics and datasets would you include to fairly evaluate STged and SINTER3D alongside traditional methods? How would you handle the "ground truth" problem when ground truth is unknown?

3. Generalizability: SINTER3D is validated on five tissue types. Is five enough to claim a general-purpose tool? What tissue types, species, or disease models would you want to see before adopting it for your own work?

4. Integration potential: Could STged's cell-type-specific deconvolution be applied within a SINTER3D-reconstructed 3D volume? What would the computational pipeline look like, and what new biological questions could it answer?

5. Future of ST analysis: Looking five years ahead, will the field converge on a few dominant tools, or will fragmentation persist? What role should community benchmarks, containerization (Docker, Singularity), and workflow languages (Snakemake, Nextflow) play in standardization?

Key Terms

• Spatial Transcriptomics (ST): A class of technologies that measure gene expression while preserving the spatial location of each measurement in a tissue section. Popular platforms include 10x Genomics Visium and Slide-seq.
• Deconvolution: A computational process that estimates the composition of cell types within a measurement. Traditional deconvolution outputs cell-type proportions; advanced methods like STged output cell-type-specific expression profiles.
• Spatially Variable Gene (SVG): A gene whose expression varies significantly across spatial locations in a tissue, often indicating biological patterning, signaling gradients, or tissue domains.
• Implicit Neural Representation (INR): A neural network that parameterizes a continuous function (e.g., gene expression as a function of 3D coordinates). INRs enable smooth interpolation and resolution-independent queries, making them powerful for 3D tissue reconstruction.
• Non-Negative Least Squares (NNLS): A constrained regression method that ensures all predicted values are non-negative. In ST, NNLS enforces that reconstructed expression profiles represent additive, biologically interpretable contributions from cell-type signatures.

---

References

1. Jessica Gillespie, Maciej Pietrzak, Min-Ae Song, & Dongjun Chung (2025). A Meta-Review of Spatial Transcriptomics Analysis Software. Cells. https://doi.org/10.3390/cells14141060

2. Jia-Juan Tu, Hong Yan, Xiao-Fei Zhang, & Zhixiang Lin (2025). Precise gene expression deconvolution in spatial transcriptomics with STged. Nucleic Acids Research. https://doi.org/10.1093/nar/gkaf087

3. Zhang T, Li S, Zhang H, Zhang R, Zhao Z, Wang R, et al. (2026). SINTER3D: continuous 3D reconstruction of spatial transcriptomics via implicit neural representations.. Genome biology. https://doi.org/10.1186/s13059-026-04160-5