Journal Club: Spatial Developmental Biology Meets Generative AI

· 10 min read · Paper Feed

Journal Club: Spatial Developmental Biology Meets Generative AI

Conceptual illustration: Spatially Resolved Developmental Biology Meets Generative AI—Three New Takes on Where Cells Are and What They Become

Introduction

For decades, developmental biology has been haunted by a simple question: where exactly is each cell, and what is it doing? Dissociative single-cell RNA sequencing gave us a good answer to the second part, but only by destroying the first: the moment you dissociate a tissue, you lose the map. The past few years have brought a wave of methods designed to put the cell back into the tissue, and 2026 is shaping up to be a strong year for spatially resolved developmental biology.

The three papers in this journal club all orbit the same idea: that spatial context is a primary axis of biological identity, not a nice-to-have. Paper 1 maps the spatiotemporal emergence of rare deep-cortical neurons in the developing mouse brain. Paper 2 does the same for the embryonic mouse lung, while also pushing the sensitivity limits of DBiT-seq. Paper 3 takes a different route: instead of measuring more, it builds a generative model—CoxFormer—that can infer the spatial expression of genes that were never directly profiled. Together, they form a triangle spanning wet-lab atlasing, technology optimization, and computational imputation.

What unites them is the conviction that current measurement technologies are too patchy or too coarse to capture development on their own. What differs is how each group proposes to escape those limits: more cells, more sensitive chemistries, or smarter models.

Paper 1: A Spatiotemporal Atlas of Sparse Deep-Cortical Neurons in the Mouse

Iyer & Cembrowski (2026), iScience — DOI: 10.1016/j.isci.2026.117558

Key Findings

  • The authors profiled >1.5 million brain cells across four developmental time points spanning embryonic to early postnatal stages, including both sexes, using single-cell spatial transcriptomics.
  • They focus on two deep-cortical populations that are often conflated: subplate and claustrum neurons. Despite shared adult molecular features, these populations have distinct developmental origins, transcriptomic trajectories, and spatial maturation programs.
  • The work establishes that rare excitatory types in the deep cortex follow non-overlapping spatiotemporal maturation curves, rather than the parallel trajectories often assumed.
  • The dataset is released as a community resource for both male and female samples, making it a practical reference for downstream studies of cortical patterning.

Methodology & Study Design

Iyer and Cembrowski used single-cell spatial transcriptomics, which keeps both cell identity and anatomical coordinates, at four developmental time points spanning embryonic and postnatal stages. That design lets them follow when and where the subplate and claustrum populations arise and how they mature. Sex was included as a biological variable from the start, which remains uncommon in single-cell atlases.

Why does this matter?

Previous developmental atlases of the mouse cortex (e.g., the Allen Brain Atlas and various scRNA-seq studies) have tended to emphasize the dominant layer-specific excitatory classes. By zooming in on the sparse populations that sit beneath layer 6, this study fills a gap. Subplate neurons are transient but functionally important—they guide thalamocortical wiring and largely disappear in adult circuits. The claustrum persists into adulthood and is implicated in salience and cross-modal integration. Showing that these two populations diverge in development, not just adulthood, reframes how we think about deep-cortical evolution and motivates functional follow-up with electrophysiology and connectomics.

Paper 2: An Optimized DBiT-seq Workflow for Embryonic Mouse Lung

Zhang, Law, Huang, Goodwin, Nelson et al. (2026), Advanced Science — DOI: 10.1002/advs.77678

Key Findings

  • A microfluidics-enabled DBiT-seq workflow was optimized to yield up to a twofold increase in average transcript recovery compared with prior implementations, on embryonic mouse lung tissue.
  • Sox9+ epithelial progenitors were found to occupy the entire airway epithelial tree at early stages, becoming restricted to distal tips only later, suggesting a proximal-to-distal specification wave.
  • A previously unrecognized mesenchymal cluster was identified, with a consistent spatial position and high expression of extracellular-matrix (ECM) genes. This cluster appears to be associated with embryonic lung innervation, implicating ECM remodeling in neural patterning.

Methodology & Study Design

DBiT-seq uses deterministic barcoding delivered via parallel microfluidic channels printed directly onto tissue sections. Zhang et al. optimized this workflow until it recovered up to twice as many transcripts on average as earlier implementations, then applied it to early embryonic mouse lung sections to follow how epithelial and mesenchymal populations take up their positions.

Why does this matter?

The lung is a classic model for branching morphogenesis, and the Sox9 gradient has been studied extensively. What this paper adds is a clean spatial story: progenitor identity is not regionally locked from the outset, but refines progressively as the tree branches. The discovery of an ECM-rich mesenchyme tied to innervation is a particularly satisfying result because it links two classically separate fields—matrix biology and neurodevelopment—into a single spatial hypothesis. The up-to-2× gain in transcript recovery is also notable for rare populations, which is exactly where the field has been struggling.

Paper 3: CoxFormer—Generative AI for Spatial Omics

Yang, Liao, Zhang, Wu, Jiao et al. (2026), Nature Communications — DOI: 10.1038/s41467-026-76404-8

Key Findings

  • CoxFormer integrates literature-derived gene knowledge, bulk co-expression networks, and large-scale single-cell atlases to learn 512-dimensional embeddings for 32,016 human genes.
  • Without a matched scRNA-seq reference, it supports four distinct inference tasks:
    1. Histology-based expression imputation for unassayed genes from H&E or other stains.
    2. Gene-activity prediction from chromatin accessibility (e.g., spatial ATAC-seq).
    3. Subcellular super-resolution inference of gene expression below the native spot/cell resolution.
    4. Pathological region detection in diseased tissue.
  • Because the embeddings are trained transcriptome-wide, the model effectively extends the reach of any spatial platform from a few hundred measured genes to 32,016 genes.

Methodology & Study Design

CoxFormer's gene embeddings act as a generative prior for spatial inference. Rather than training on a single matched dataset, it pulls from three complementary sources: curated gene–gene relationships from the literature, bulk-tissue co-expression, and large scRNA-seq atlases. The 512-D embeddings then serve as a shared latent space in which different spatial modalities can be queried, predicted, and aligned. The framework is reference-free in the spatial sense—you do not need a matched scRNA-seq dataset for your tissue of interest.

Why does this matter?

Every spatial transcriptomics platform is a trade-off: Visium and DBiT-seq offer whole-transcriptome coverage at low cellular resolution; MERFISH and seqFISH offer subcellular resolution but only for a few hundred genes. CoxFormer offers a way to bridge that gap computationally. If it generalizes, the field could shift from designing ever-larger targeted panels to building better priors over gene behavior. The four downstream applications show that one embedding can power a small ecosystem of spatial analyses, which is appealing for tool-builders.

The obvious caveat is that generative models are not measurements. Predictions inherit the biases of training data, and rare or tissue-specific genes may be poorly represented. The authors will need to convince the community with rigorous benchmarks, especially on tissues like those in Papers 1 and 2 where cell types are rare and developmental context matters.

Synthesis & Discussion

Taken together, these three papers sketch a coherent trajectory for spatial omics in 2026:

  1. The atlas problem is becoming tractable. Paper 1 demonstrates that even sparse neuronal populations can be mapped with high cellular and temporal resolution if you commit to a large enough dataset (>1.5M cells, four time points, both sexes). Paper 2 shows the same for a non-neural organ, with the added twist of a workflow upgrade that recovers up to twice as many transcripts.

  2. Spatial context is now the primary axis of analysis. Both experimental papers emphasize where a cell is, not just what it expresses, when interpreting developmental transitions. This is no longer an afterthought in scRNA-seq—it is the main event.

  3. Generative models are the next battleground. Paper 3 sits in stark contrast to the other two: where Iyer and Zhang measure more, Yang et al. infer more. The question for the field is not whether generative models will be used, but how we will validate them. CoxFormer's reference-free design is appealing but also means errors are harder to localize.

Open Questions Raised

  • Can CoxFormer accurately reconstruct the sparse populations highlighted by Iyer and Cembrowski, or will rare deep-cortical neurons be smoothed away by the prior?
  • Does the Sox9+ spatial restriction pattern observed in the lung generalize to other branching organs (kidney, mammary, salivary)?
  • The lung mesenchyme–innervation link invites a causal test: what happens to nerve migration when the ECM cluster is ablated?
  • How do we benchmark inferred spatial transcriptomes against measured ones without circularity?
  • Will the community settle on a unified file format for multi-modal spatial data, or will each platform continue to bring its own zoo?

For Your Lab Meeting

  1. Methodology: Iyer & Cembrowski sampled both sexes across four time points. For a study of your own, at what point does the added cost of including sex as a variable pay off, and when is it wasted power?
  2. Generalizability: The DBiT-seq sensitivity boost is shown in lung. Would you expect a similar gain in brain tissue, which is more lipid-rich? What controls would you demand before adopting the protocol?
  3. Validation: CoxFormer imputes expression without a matched scRNA-seq reference. Propose a concrete experimental design that could falsify one of its predictions in a tissue of your choice.
  4. Future directions: If you had to choose between a 10× larger spatial atlas and CoxFormer-style imputation for your next project, which would you pick, and why?
  5. Broader trends: These three papers span wet lab, technology, and AI. Is the spatial-omics community at risk of over-relying on computational imputation before the underlying chemistries have matured? Or is the AI work forcing the chemists to raise their game?

Key Terms

  • Spatial transcriptomics: A family of methods that retain the x, y (and sometimes z) coordinates of gene expression within a tissue section, as opposed to dissociated cells.
  • DBiT-seq (Deterministic Barcoding in Tissue sequencing): A spatial transcriptomics method that delivers barcodes via parallel microfluidic channels printed onto a slide, enabling sensitive, platform-flexible profiling.
  • Subplate neurons: A transient, early-born population of cortical excitatory neurons that guides thalamocortical axon targeting and largely disappears in the adult cortex.
  • Co-expression network: A graph in which genes are connected by edges representing correlated expression across samples, often used to infer functional relatedness.
  • Generative prior: A probabilistic model that defines a plausible distribution over data; in CoxFormer's case, a prior over gene expression patterns used to impute unmeasured values.

References

  1. Iyer S & Cembrowski MS (2026). Spatiotemporal development of sparse excitatory neuronal types within the deep mouse cortex. iScience. https://doi.org/10.1016/j.isci.2026.117558

  2. Zhang P, Law BK, Huang N, Goodwin K, Nelson CM, & Chan MM (2026). Mapping Embryonic Mouse Lung Development Using Enhanced Spatial Transcriptomics. Advanced Science. https://doi.org/10.1002/advs.77678

  3. Yang Y, Liao X, Zhang H, Wu Y, Jiao Y, Sun X, et al. (2026). CoxFormer enables spatial omics inference with multimodal generative modeling. Nature communications. https://doi.org/10.1038/s41467-026-76404-8

Single-Cell Transcriptomics scRNA-seq Genomics