Journal Club: Single-Cell and Spatial Multi-Omics in Plant Biology
Introduction
Plants are not the simple, passive organisms they're sometimes made out to be. Behind the apparent stillness sits a tightly coordinated cast of cell types, developmental programs, and stress responses, from a leaf yellowing in autumn to a callus rebuilding an entire embryo. For decades, plant biologists studied these processes in bulk tissue, averaging signals across dozens of cell types and smearing out the heterogeneity that actually drives biology. Single-cell sequencing, and its spatial cousin, now lets us watch that choreography step by step.
This week we walk through three papers showing how scRNA-seq, snRNA-seq, spatial transcriptomics, and spatial metabolomics are reshaping plant biology. Guo et al. (2025, Cell) build a million-nucleus Arabidopsis atlas of leaf senescence. Li et al. (2025, Nature Communications) present SpaIM, a framework that borrows information from scRNA-seq to fill in sparse spatial data. Ge et al. (2025, Nature Communications) combine scRNA-seq, spatial transcriptomics, and spatial metabolomics to dig into cotton somatic embryogenesis. Together, they sketch a clear direction for the field.
Below, we go through each paper, point out what is technically distinctive, and discuss what they collectively suggest about computational plant biology.
Paper 1: A Million-Nucleus Atlas of Arabidopsis Leaf Senescence and Nutrient Allocation
Citation: Guo X., Wang Y., Zhao C., et al. An Arabidopsis single-nucleus atlas decodes leaf senescence and nutrient allocation. Cell, 2025. DOI: 10.1016/j.cell.2025.03.024
Key Findings
- Unprecedented scale: The authors profiled over 1 million nuclei from 20 Arabidopsis tissues spanning multiple developmental stages, one of the largest plant snRNA-seq efforts to date.
- New cell types uncovered: By profiling intact nuclei instead of protoplasts (which bias against large, fragile cell types), they found previously uncharacterized populations, including rare vascular and epidermal subtypes missed in earlier single-protoplast studies.
- Coordinated senescence: Time-resolved sampling showed that mesophyll, epidermis, and vasculature initiate senescence in a highly coordinated, near-synchronous way, rather than each cell type aging on its own schedule.
- Molecular aging indices: Two new quantitative metrics score the "aging state" of individual leaf cells, enabling systematic comparisons across cell types and conditions.
- Hub gene network: Using WGCNA, the team nominated hundreds of candidate hub genes, several of which were functionally validated. These were integrated into a model of carbon and nitrogen allocation from source leaves to sink organs — essentially a cell-resolved view of whole-plant nutrient economy.
Methodology & Study Design
The team ran snRNA-seq on nuclei isolated from flash-frozen tissue, avoiding the enzymatic digestion that wipes out sensitive cell types. After strict QC, they clustered nuclei to define cell types, then used trajectory and pseudotime analyses to model senescence progression. Layering WGCNA on top of the snRNA-seq matrix let them move from individual marker genes to network-level regulators, a practical way to extract biology from a large but noisy dataset.
Why this matters
This work pushes plant single-cell biology from "what cell types exist" to "what those cell types are doing over time." The carbon/nitrogen allocation model is notable because it ties molecular data directly to agronomically meaningful traits like nutrient remobilization during grain filling. For crop scientists, the atlas offers a concrete starting point for translating single-cell insights into field-relevant phenotypes.
Paper 2: SpaIM — Style-Transfer Imputation for Spatial Transcriptomics
Citation: Li B., Tang Z., Budhkar A., et al. SpaIM: single-cell spatial transcriptomics imputation via style transfer. Nature Communications, 2025. DOI: 10.1038/s41467-025-63185-9
Key Findings
- A unified imputation framework: SpaIM uses style transfer learning to predict unmeasured gene expression in ST data by leveraging matched scRNA-seq references. The model separates "content" (shared biological signal) from "style" (modality-specific technical artifacts).
- Beats the competition: Benchmarked on 53 datasets spanning sequencing-based ST (e.g., 10x Visium, Stereo-seq) and imaging-based platforms (e.g., MERFISH, seqFISH) across multiple tissue types, SpaIM consistently outperformed 12 state-of-the-art imputation methods.
- Downstream gains: Beyond raw expression recovery, SpaIM improved ligand-receptor interaction inference, spatial domain identification, and differential gene expression analysis, so benefits compound across a typical ST analysis pipeline.
- Open and accessible: Released as open-source software, SpaIM lowers the barrier for labs without deep machine learning expertise.
Methodology & Study Design
Style transfer, borrowed from computer vision, is a natural fit for cross-modal integration. The intuition: scRNA-seq and ST share biological "content" but differ in technical "style" (sparsity patterns, capture biases, normalization). SpaIM uses an adversarial architecture to pull these apart, then predicts gene expression in ST spots using scRNA-seq as a style-aware reference. The result handles both sequencing- and imaging-based ST without bespoke retraining.
Why this matters
Most imputation methods were built for one platform in one tissue. SpaIM's platform-agnostic performance tackles a real bottleneck in plant ST, where researchers often toggle between Visium (low resolution, high gene coverage) and Stereo-seq (high resolution, sparse data). By unifying these modalities, SpaIM makes cross-technology and cross-species comparisons more straightforward.
Paper 3: Spatiotemporal Multi-omics of Cotton Somatic Embryogenesis
Citation: Ge X., Yu X., Liu Z., et al. Spatiotemporal transcriptome and metabolome landscapes of cotton somatic embryos. Nature Communications, 2025. DOI: 10.1038/s41467-025-55870-6
Key Findings
- Triple-modality integration: The authors combined scRNA-seq, spatial transcriptomics, and spatial metabolomics to profile cotton somatic embryogenesis (SE), a developmental process that has resisted clean mechanistic dissection for decades.
- Stage-specific gene programs: Single-cell and spatial data revealed distinct transcriptional programs marking early embryogenic cells, embryonic shoot/apical meristem precursors, and maturing somatic embryos.
- Two functional markers: AATP1 and DOX2 were identified as key markers and functionally validated, linking specific genes to SE progression.
- Metabolic coordination: Spatial metabolomics showed that changes in sugars, amino acids, and secondary metabolites align spatially with the transcriptional landscape, direct evidence that gene expression reshapes local biochemistry during regeneration.
- Community resource: All data are accessible through a dedicated web portal at https://cotton.cricaas.com.cn/somaticembryo/, lowering the entry barrier for the cotton community.
Methodology & Study Design
This is the most methodologically integrative paper of the three. The team aligned scRNA-seq clusters to ST spots (a workflow SpaIM would now improve), then layered in mass spectrometry imaging for metabolites. Functional validation in cotton is itself nontrivial — transformation is slow and genotype-dependent — so the experimental confirmation of AATP1 and DOX2 represents a real investment.
Why this matters
Cotton SE remains a bottleneck for crop breeding and genetic engineering. By mapping the cellular and metabolic transitions during SE, this study provides a molecular blueprint for improving regeneration rates, work with implications beyond cotton. The transcript-plus-metabolite-in-space design is also a methodological template for studying other recalcitrant plant developmental processes.
Synthesis & Discussion
Read individually, these three papers ask different biological questions. Read together, they reveal an emerging architecture of plant single-cell and spatial omics.
A methodological triangle
Each paper sits at a different vertex of a methodological triangle:
- Scale and depth of biology — Guo et al. push breadth (1M+ nuclei, 20 tissues).
- Cross-modal computational integration — Li et al. push methodology (style-transfer imputation).
- Multi-modal empirical integration — Ge et al. push depth (transcript + metabolite + space + function).
The plant biology community no longer has to choose between "lots of cells" and "lots of modalities." These papers show the field is converging on both.
Convergent themes
A few patterns stand out when the papers are read together:
- Nuclei beat protoplasts for plant scRNA-seq. Guo et al.'s use of snRNA-seq echoes a broader recognition that protoplasting biases against certain cell types. SpaIM and Ge et al. both benefit implicitly, since their scRNA-seq references are increasingly built from nuclei.
- Spatial context is no longer optional. All three papers either generate or rely on spatial data. SpaIM exists because ST data are sparse, and Ge et al. explicitly use ST to ground scRNA-seq findings.
- Functional validation closes the loop. All three studies move beyond descriptive transcriptomics: Guo et al. validate hub genes, Ge et al. mutate AATP1 and DOX2, and SpaIM demonstrates downstream analytical gains.
Open questions
- Will SpaIM's performance hold up on plant data, where cell walls and tissue heterogeneity challenge imputation assumptions validated mostly on mammalian datasets?
- How transferable are Guo et al.'s senescence hub genes to crops like maize or soybean, where senescence directly impacts yield?
- Can the cotton SE atlas be used to engineer improved regeneration protocols in other species?
- As datasets grow to millions of cells, are current WGCNA-style analyses still scalable, or do we need new network methods?
For Your Lab Meeting
Use these questions to spark discussion:
- Methodology trade-offs: Guo et al. used snRNA-seq rather than scRNA-seq. Given that SpaIM relies heavily on scRNA-seq references, do we need plant-specific benchmarking of how snRNA-seq vs scRNA-seq references affect ST imputation quality?
- Generalizability: SpaIM was validated on 53 datasets, but how many were plant-derived? Should plant labs run their own benchmarks before adopting it for, say, maize anther development?
- Hub gene interpretation: With "hundreds" of senescence hub genes nominated, how should the community prioritize functional follow-up? Are there network properties (centrality, condition-specificity) that should be weighed alongside expression change?
- Multi-modal integration limits: Ge et al. integrate three modalities, but the spatial resolution of metabolomics (~30 µm) doesn't match single-cell resolution. How should we interpret "co-localization" when modalities differ in scale?
- Future direction: If we had to pick one of these three papers to extend into your lab's organism of choice, which would be most tractable and why?
Key Terms
- snRNA-seq (single-nucleus RNA-seq): Profiling of isolated nuclei rather than whole cells. Avoids enzymatic protoplasting, capturing cell types that are otherwise lost, especially in plants with tough cell walls.
- WGCNA (Weighted Gene Co-expression Network Analysis): A method that groups genes into modules based on correlated expression across samples, then relates modules to traits. Useful for finding "hub" regulators.
- Style transfer learning: A machine learning approach, originally from computer vision, that separates content from style in a representation. Here, used to separate biological signal from technical noise across scRNA-seq and ST modalities.
- Somatic embryogenesis (SE): A plant tissue culture process where somatic cells develop into embryos without fertilization. A key tool for crop breeding and genetic engineering, but often inefficient and genotype-dependent.
- Spatial transcriptomics (ST): Methods that preserve tissue location while measuring gene expression, either via in situ hybridization or in situ capture. Currently limited by gene coverage (imaging) or resolution (capture-based).
References
Xing Guo, Yichuan Wang, Caiyao Zhao, Cong Tan, Wei Yan, Sunhuan Xiang, et al. (2025). An Arabidopsis single-nucleus atlas decodes leaf senescence and nutrient allocation.. Cell. https://doi.org/10.1016/j.cell.2025.03.024
Bo Li, Ziyang Tang, Aishwarya Budhkar, Xiang Liu, Tonglin Zhang, Baijian Yang, et al. (2025). SpaIM: single-cell spatial transcriptomics imputation via style transfer. Nature Communications. https://doi.org/10.1038/s41467-025-63185-9
Xiaoyang Ge, Xiaole Yu, Zhixin Liu, Jiachen Yuan, Ai-guo Qin, Ye Wang, et al. (2025). Spatiotemporal transcriptome and metabolome landscapes of cotton somatic embryos. Nature Communications. https://doi.org/10.1038/s41467-025-55870-6