Journal Club: From Mixed Signals to Single-Cell Maps — Three Roads to Spatial Deconvolution

· 9 min read · Paper Feed

Journal Club: From Mixed Signals to Single-Cell Maps — Three Roads to Spatial Deconvolution

Conceptual illustration: From Mixed Signals to Single-Cell Maps — Three Roads to Spatial Deconvolution

Introduction

If you've ever stared at a Visium spot containing 5–20 overlapping cells and wondered which gene belongs to which cell type, this journal club is for you. The three papers we've selected attack the same bottleneck from different angles: how do we recover cell-type-specific and single-cell-resolved information from inherently aggregated or low-resolution molecular measurements? This is the central challenge of modern spatial omics and bulk tumor transcriptomics, where the raw data are mixtures—mixtures of cells, of states, and of signals from the tumor, stroma, and immune compartments.

The unifying theme is deconvolution and single-cell reconstruction in space. Paper 1 (Enninful et al., 2026, Nature Methods) introduces DBiTplus, an experimental workflow that co-captures the transcriptome and proteome on the same tissue section using deterministic barcoding followed by multiplexed protein imaging. Paper 2 (Dai et al., 2025, Nature Reviews Cancer) provides a review of 43 transcriptomic deconvolution methods with cancer applications in mind. Paper 3 (Bourgain-Chang et al., 2026, Nature Communications) presents CellRefiner, a physics-inspired computational tool that reconstructs single-cell-resolution spatial maps from non-single-cell data using paired scRNA-seq.

Together, these papers show how the field is moving from crude proportion estimation toward physically grounded, multimodal single-cell reconstruction—yet they also expose persistent challenges around tumor plasticity, reference dependence, and clinical translation.

Paper 1: DBiTplus — Same Slide, Two Languages (RNA + Protein)

Key Findings

  • DBiTplus combines sequencing-based spatial transcriptomics with multiplexed protein imaging (CODEX-style) on the same tissue section, with no need to split samples.
  • Demonstrated across diverse and clinically challenging specimens: frozen mouse embryos, FFPE human lymph nodes, and FFPE human lymphoma tissues, recovering spatial organization of germ layers, immune zones, and lymphoma architecture.
  • Applied to human lymphoma, the authors uncovered mechanisms of lymphomagenesis, progression, and transformation, using protein-defined cell types to anchor deconvolution of mixed transcriptomic pixels.
  • Achieved single-cell-resolved spatial transcriptome atlases via imaging-guided deconvolution, using protein markers as ground-truth anchors for cell-type assignment.

Methodology & Study Design

DBiTplus extends the original DBiT-seq platform (microfluidic deterministic barcoding of fixed tissue) with a twist: after in-situ reverse transcription and spatial barcode delivery, RNase H–mediated cDNA retrieval liberates the library without destroying the section. The intact tissue is then handed off for high-plex protein imaging (e.g., CODEX or equivalent cyclic immunofluorescence). The team built computational pipelines that use protein images to deconvolve mixed transcriptomic pixels into their constituent cells—essentially borrowing the spatial precision of imaging to refine the cell-type resolution of sequencing data.

Significance

Before DBiTplus, users had to choose: genome-wide RNA or high-resolution protein maps—and often on different sections, raising alignment and biological-variability issues. DBiTplus delivers both modalities from one slide and is compatible with FFPE, the universal currency of clinical pathology. This reframes how we should think about integrating spatial omics in tumor microenvironment studies.

Paper 2: A Practitioner's Guide to Transcriptomic Deconvolution in Cancer

Key Findings

  • Systematic review of 43 deconvolution methods, organized by data type (bulk vs spatial), reference availability (reference-based, semi-reference, reference-free), and underlying statistical framework (regression, matrix factorization, deep learning, Bayesian).
  • Highlights recurrent failure modes: misclassification of malignant vs. normal epithelial cells at high tumor purity, sensitivity to reference-cell purity, and instability across patient cohorts.
  • Maps method choice to use case: immune infiltration estimation, subtype discovery, prognostic biomarker identification, spatial architecture inference, and tumor–immune surveillance analysis.
  • Identifies a major frontier: capturing tumor cell plasticity and dynamic cell states, which static reference matrices fundamentally cannot represent.

Methodology & Study Design

This is a narrative and methodological review—not a benchmarking study—but it synthesizes existing benchmarks and translates them into actionable guidance. Each method is categorized by assumptions, input requirements (signature matrix, scRNA-seq reference, spatial coordinates), and output type (proportions, cell-type expression, full decomposition). The authors walk through concrete workflows for typical cancer research scenarios, including which tools hold up under high tumor purity (a regime where many methods break down).

Significance

With thousands of bulk RNA-seq samples already banked in resources like TCGA, and a growing wave of spatial cohorts, deconvolution is the bridge that turns these averaged profiles into per-cell-type and per-state readouts. Dai et al. provide a cancer-focused framework that explicitly acknowledges tumor-specific challenges—intra-tumoral heterogeneity, plasticity, atypical expression programs—that generic deconvolution guides often ignore. The paper is, in effect, a shopping list with warnings.

Paper 3: CellRefiner — Physics-Inspiring Cell Mapping

Key Findings

  • CellRefiner integrates a paired scRNA-seq dataset with non-single-cell-resolution spatial data (e.g., Visium 55 µm spots, Slide-seqV2 10 µm beads, or imaging-based platforms like MERFISH, seqFISH, STARmap) to reconstruct single-cell positions within each capture site.
  • Models cells as particles connected by forces, optimizing cell locations under three constraints: (1) spatial proximity, (2) gene expression similarity to the local expression profile, and (3) known ligand–receptor interactions between neighboring cells.
  • Benchmarking across simulated and real datasets spanning five platforms shows recovery of spatial patterns and improved downstream cell–cell communication analysis, demonstrated on mouse cortex and lymph node tissues.
  • Improves contact-based cell–cell communication inference because it produces a candidate list of plausible cell–cell contacts rather than spot-level aggregates.

Methodology & Study Design

CellRefiner treats deconvolution as a physical inverse problem. Given a reference scRNA-seq atlas and spatial expression, it iteratively refines cell positions and their assigned transcriptomes to minimize a composite energy function. The inclusion of ligand–receptor interactions as a force term pushes the algorithm toward configurations consistent with known signaling biology—a feature absent from most matrix-factorization-based deconvoluters. The method is benchmarked on Visium (10x Genomics), Slide-seqV2, MERFISH, seqFISH, and STARmap, spanning both sequencing- and imaging-based spatial technologies.

Significance

Where DBiTplus uses protein imaging as the anchor and deconvolution guides use reference matrices, CellRefiner uses biophysical prior knowledge to impute single-cell locations. This positions it as a hybrid approach—part deconvolution, part spatial reconstruction—that performs especially well for downstream analyses (e.g., cell–cell communication) that require knowing which specific cells touch each other.

Synthesis & Discussion

How the Papers Complement Each Other

These three papers form a methodological triangle:

  • DBiTplus = experimental foundation (RNA + protein, same section).
  • Dai et al. = strategic overview (which computational tool for which cancer question).
  • CellRefiner = spatial refinement (single-cell reconstruction under biophysical constraints).

Stacking them illustrates a modern spatial omics pipeline: one could use Dai's framework to pick a baseline deconvoluter, use DBiTplus to generate matched RNA+protein data, and apply CellRefiner-style refinement when single-cell positions are needed.

Agreements

All three converge on a few principles:

  1. Reference data is essential—whether single-cell atlases (CellRefiner), signature matrices (Dai), or protein-imaging maps (DBiTplus).
  2. Tumor heterogeneity is the test case. The cancer microenvironment—tumor, stromal, and immune cells coexisting in gradients—is exactly where all three approaches try to shine.
  3. Simple matrix unmixing is no longer enough. Each paper pushes beyond regression-style deconvolution toward more integrative, context-aware inference.

Tensions and Open Questions

  • Reference bias: Dai et al. repeatedly flag how reference mismatch (e.g., a healthy reference applied to tumor samples) leads to systematic errors. Both DBiTplus (using a CODEX panel) and CellRefiner (using external scRNA-seq) implicitly assume their reference captures the relevant biology—something the Nature Reviews paper warns against.
  • What is "ground truth"?: DBiTplus treats protein images as anchors; CellRefiner treats ligand–receptor priors as anchors; Dai et al. emphasize that no anchor is unbiased. How do we reconcile them?
  • Spatial rigor: CellRefiner's "forces" model is elegant but currently lacks rigorous comparison to classical deconvolution in real tumor data—a target for the future benchmarking studies Dai et al. call for.
  • Plasticity and rare states: All three acknowledge but none fully solve the dynamic-cell-state problem—a frontier explicitly identified by Dai et al. that demands new statistical frameworks.

Emerging Patterns

  • Multimodal integration is becoming table stakes. RNA alone is insufficient; protein imaging, spatial priors, and signaling biology are all being added to the inference pipeline.
  • Movement from proportions to positions. The community is shifting from "what fraction of cells are T cells?" to "where are the T cells, and which tumor cell are they touching?"
  • Physics- and constraint-based models are entering the deconvolution space, complementing matrix factorization and deep learning.

For Your Lab Meeting

  1. Choosing your anchor: DBiTplus uses protein imaging as the cell-typing anchor, CellRefiner uses scRNA-seq + ligand–receptor priors, and classical deconvoluters use signature matrices. Which anchor is most robust in your tissue of interest, and how would you test that empirically?
  2. Reference bias in tumor deconvolution: Dai et al. note systematic errors when normal-tissue references are applied to tumors. If you only have a diseased reference, what downstream analyses become suspect, and what QC steps would you add?
  3. Single-cell resolution vs. single-cell accuracy: CellRefiner produces imputed single-cell positions, not directly measured ones. How do we communicate this uncertainty to collaborators (especially clinicians)?
  4. Modality matchmaking: When does adding a protein channel (DBiTplus) pay off more than adding a better scRNA-seq reference (CellRefiner)? Can we think about cost per added biological insight?
  5. The plasticity problem: All three papers treat cell identity as relatively static. What would a "plasticity-aware" deconvolution or reconstruction method even look like? Sketch the conceptual pieces.

Key Terms

  • DBiT-seq: A sequencing-based spatial transcriptomics method that uses microfluidic channels to deliver spatial barcodes to a tissue section in a deterministic pattern.
  • CODEX (CO-Detection by indEXing): A multiplexed immunofluorescence technique that iteratively cycles antibodies and reporters to image 50+ proteins on the same section at subcellular resolution.
  • FFPE (Formalin-Fixed Paraffin-Embedded): The standard preservation method for clinical pathology samples; notoriously challenging for RNA-based assays due to crosslinking and degradation.
  • Deconvolution: Computational unmixing of a mixed signal (e.g., bulk RNA-seq, spatial spot) into its constituent cell types and/or cell-type-specific expression profiles.
  • Ligand–receptor interaction: A pair of molecules (one secreted by cell A, one expressed on cell B) that mediates cell–cell signaling; used as a spatial prior in methods like CellRefiner.
  • Visium / Slide-seqV2: Spot-based spatial transcriptomics platforms where each capture site contains RNA from multiple cells (e.g., Visium spots are 55 µm and typically cover 1–10 cells; Slide-seqV2 beads are ~10 µm).

References

  1. Archibald Enninful, Zhaojun Zhang, D. Klymyshyn, Matthew Ingalls, Mingyu Yang, Hailing Zong, et al. (2026). Integration of imaging-based and sequencing-based spatial omics mapping on the same tissue section via DBiTplus.. Nature Methods. https://doi.org/10.1038/s41592-025-02948-0

  2. Yaoyi Dai, Shuai Guo, Yidan Pan, Carla Castignani, Matthew D. Montierth, P. van Loo, et al. (2025). A guide to transcriptomic deconvolution in cancer. Nature Reviews. Cancer. https://doi.org/10.1038/s41568-025-00886-9

  3. Bourgain-Chang E, Kuang X, Cang Z, & Nie Q (2026). Reconstructing single-cell resolution from spatial transcriptomics with CellRefiner.. Nature communications. https://doi.org/10.1038/s41467-026-70090-2

Spatial Transcriptomics Spatial Omics Single-Cell