Drought and salinity stress remodel Asian rice (Oryza sativa) leaf development through cell‐type‐specific regulatory programs

Summary

Osmotic stressors impair tissue expansion and slow crop development, which can ultimately reduce yield. The mechanisms through which dividing and differentiating cells perceive, integrate, and respond to stress are not well understood. To explore this knowledge gap, we assessed Asian rice (Oryza sativa) seedling responses to drought and salinity stress by performing single‐nucleus transcriptome analyses on developing leaves and by assaying ion distribution using ICP‐MS and CoroNa Green staining. With a dataset of > 125 000 high‐quality nuclei, we identified transcriptome signatures of developmental trajectories for each of the major tissue systems, from their origin as dividing cells in the shoot apical meristem to mature cells in epidermal, mesophyll, and vascular tissues. We explored the extensive heterogeneity between cell types with respect to their responses to stress by developing an ensemble‐based Perturbation Index and a tool for visualizing pseudobulk expression data and characterizing regulatory modules for different stress responses. Transcriptionally, moderate stress perturbs cells early in development and induces regulatory networks that inhibit cell division, whereas mild stressors primarily disrupt gene expression in differentiated tissues. Overall, these results indicate that leaf development occurring in salt and drought stress conditions is the product of a complex interplay of developmental, environmental, and cell‐type contexts.

Osmotic stressors induce developmental and physiological changes in rice (Oryza sativa) seedlings, which are modulated by cell‐type‐specific gene expression and regulatory networks. We exposed plants to two levels of drought and salinity stress and measured the effects of these stressors on developing leaves by measuring physiological traits, ion accumulation patterns, and single‐nucleus RNA‐sequencing. We built tools for assessing stress responses for each cell and for visualizing gene expression on tissue maps. We showed that stress responses differed between cell types and developmental states and that more severe stressors led to qualitatively different responses than milder stressors.

See also the Commentary on this article by Murillo & Conde, 252: 1351–1353.

	Contents
	Summary	1376
I.	Introduction	1376
II.	Materials and Methods	1377
III.	Results	1379
IV.	Discussion	1390
	Acknowledgements	1391
	References	1391

I. Introduction

Agricultural productivity is determined by the availability of sufficient clean, fresh water to sustain plant growth (Ali & Talukder, 2008). As an adaptive response to prolonged periods of mild water insufficiency, development may be altered to protect the plant from irreparable damage and to prepare it for continuing stress conditions (Skirycz & Inzé, 2010). For example, drought and salinity stresses that occur early in plant growth reduce the rate of leaf expansion, increase water use efficiency, reduce overall carbon assimilation capacity (Morales et al., 2012; Robertson et al., 2025), and reduce the ability to resume growth when water availability increases after drought (Van Hautegem et al., 2025), leading to a reduced plant stature and ultimately to reduced yield (Huanhe et al., 2024). Growth reduction occurs by controlled alterations to cell division and expansion in the meristematic tissue (Nelissen et al., 2012, 2018; Verelst et al., 2013). Moreover, these stressors affect the growth and expansion of different cell types differently. For example, in maize, leaf growth under salinity stress is restricted by the stiffening of epidermal cell walls which prevents the expansion of interior cell types (Zörb et al., 2015). Despite the detailed knowledge of the role of water limitations in reducing cell division and expansion in developing leaves, the cellular and transcriptional mechanisms through which developing leaf tissues respond to osmotic stressors are not well understood.

Drought and salinity stresses lead to similar cellular phenotypes in Oryza sativa (Asian rice). These include reduced cell size, impaired membrane integrity, and reduced turgor pressure (Chaves et al., 2009). However, some of the underlying causes of these stress phenotypes differ. While both drought and saline treatments create conditions of osmotic stress due to reduced water uptake, salinity stress also leads to ion toxicity caused by the accumulation of sodium ions in the tissues (Yan et al., 2020). Ion toxicity affects enzyme function through competitive inhibition of potassium as a co‐factor (Houmani et al., 2023), disruption of calcium signaling by altering calcium localization within cells (Kader et al., 2007), and reduction in cell wall integrity through microtubule depolymerization (Endler et al., 2015) and increased pH of the extracellular spaces (Li & Yang, 2023). Importantly, these cellular phenotypes are known to vary between cells. For example, mesophyll cells exhibit altered anatomy under salinity stress depending on their abaxial/adaxial position within the leaf (Ouk et al., 2024). Moreover, there is histological evidence that salt accumulation in rice leaves is restricted to specific cell types within vascular bundles, suggesting that the extent of ion toxicity may vary between cell types (Chakraborty et al., 2020).

Single‐cell and single‐nucleus RNA‐sequencing provides the resolution necessary to study osmotic stress responses of individual cell types and developmental states. Recent scRNA‐seq studies in plants have uncovered cell‐type‐specific responses to drought stress in Arabidopsis leaves (Berrío et al., 2021, 2025; Illouz‐Eliaz et al., 2025) and sorghum leaves (Stata et al., 2025), and salinity stress in cotton roots (Li et al., 2024), cabbage roots (Liu et al., 2025), Brassica species roots (Wang et al., 2026), and whole rice seedlings (Wang et al., 2021). Drought has been found to have the greatest impact on mesophyll cells in mature Arabidopsis leaves (Berrío et al., 2025), while salinity had the greatest effect on the endodermis and root hairs of rice roots (Wang et al., 2021), highlighting the need to profile distinct cell types. Additionally, profiling developing tissues allows for the discovery of key stress responses at specific developmental time points, such as alterations to carbon metabolism in differentiating cells experiencing hypoxia stress (Hill et al., 2026). These findings emphasize the value in studying cell‐specific responses to stress for formulating a model of how the tissue as a whole regulates development.

In this study, we generated a comprehensive single‐nucleus transcriptome atlas of developing rice leaves grown under drought and salinity stress at mild and moderate intensities. This design enabled a direct comparison between osmotic stress type, stress extent, and developmental state. We found extensive heterogeneity between cell types with respect to their responses to stress type and severity in terms of transcriptome changes and ion accumulation patterns. Transcriptionally, moderate stress perturbs cells early in development and induces ethylene response factor (ERF)‐mediated regulatory networks that inhibit cell division, whereas mild stress primarily disrupts transcription in differentiated tissues. Uniquely, mild salinity disrupts mesophyll development by altering cell expansion and shifting light capturing functions to large parenchyma and bundle sheath cells. Overall, these results indicate that osmotic stressors induce distinct responses in different cell populations thereby influencing leaf development.

II. Materials and Methods

II..1. 1. Plant growth conditions and stress treatments

Oryza sativa L. cv Nipponbare rice seeds (GSOR 100; Dale Bumpers National Rice Research Center, USDA, AR, USA) were germinated in moist, dark conditions at 30°C for 96 h before planting. Stable gravimetric soil water content (SWC) was achieved by oven drying Sungro Sunshine Mix 1 (Sun Gro Horticulture, Agawam, MA, USA) at 65°C for 48 h and then wetting with either fresh tap water or saline according treatment (Dowd et al., 2019): control (3 g water g⁻¹ soil media; electrical conductivity (EC) 0.5 dS m⁻¹), mild drought (1.5 g water g⁻¹ soil; EC 0.5 dS m⁻¹), moderate drought (1 g water g⁻¹ soil; EC 0.5 dS m⁻¹), mild salinity (3 g 100 mM NaCl g⁻¹ soil; EC 1.25 dS m⁻¹), and moderate salinity (3 g 200 mM NaCl g⁻¹ soil; EC 2.5 dS m⁻¹). Pots were weighed daily and re‐watered to their original weight to maintain their SWC and EC until the third leaves had extended c. 1 cm beyond the sheath collar. Seedlings were grown in a controlled environment chamber (Controlled Environments Ltd, Winnipeg, MB, Canada) with a 12 h : 12 h,  30°C : 25°C, day : night cycle, with 700 μmol m² s⁻¹ light intensity and 80% relative humidity.

II..2. 2. Division zone measurements

Division zone lengths were measured using methods described in Sprangers et al. (2016), which defined the division zone length as the distance between the leaf base and the furthest mitotic nucleus in the epidermal cells adjacent to the stomata. To aid in this analysis, cleared tissues were stained with 4',6‐Diamidino‐2‐phenylindole (DAPI), to facilitate visualization of cell margins and nuclei. Leaf sections were imaged using an LSM‐700 confocal microscope (Zeiss, Oberkochen, Germany). Cell lengths in the division zone were measured using internuclear distance in the epidermis using Fiji (Schindelin et al., 2012).

II..3. 3. Leaf water potential

Leaf water potential was measured with a PMS Model 600 Pressure Chamber (PMS Instrument Co., Albany, OR, USA). Seedlings at the third leaf stage (9–12 days after planting (DAP)) were cut with a razor blade c. 1 cm above the soil, placed in the pressure chamber, and an air‐tight seal was formed with clay. Nitrogen gas was slowly added until water could be seen bubbling from the cut site, at which point the gas pressure was recorded.

II..4. 4. Chlorophyll content

Leaf Chl content was measured following a protocol from Pocock et al. (2004). Chl was extracted in acetone and absorbance was measured at 663, 645, and 750 nm on a BioTek Synergy H1 plate reader (Agilent Technologies, Santa Clara, CA, USA). Chl concentrations were calculated using the Arnon method (Arnon, 1949).

II..5. 5. Light response curves

Gas exchange measurements were taken between ZT 4‐8 on fully expanded third leaves using an LI‐6800 Portable Photosynthesis System (LI‐COR, Lincoln, NE, USA). Light response curves were generated using the Light_Response program, and analysis was carried out with the photosynthesis R package (Stinziano et al., 2023) using the [fit_photosynthesis] and [marshall_biscoe_1980] functions.

II..6. 6. Sodium and potassium ion content

Sodium and potassium ion concentrations were measured in developing third leaves, fully expanded first and second leaf blades and sheaths. Ions were extracted using the protocol described by Hansen et al. (2013) and measured using inductively coupled plasma mass spectrometry (ICP‐MS). Sodium and potassium ions were quantified in the extracts using an Agilent 7850 ICP‐MS (Mississauga, ON, Canada) at the Richardson Center for Food Technology and Research (Winnipeg, MB, Canada).

II..7. 7. Sodium ion imaging in plant tissues

A 1‐cm section of tissue beginning 5 cm above the soil interface was excised from developing third leaves, and embedded in 3% agarose blocks. Sodium ions were stained with CoroNa Green (Thermo Fisher Scientific, Mississauga, ON, Canada), and cell walls were stained with Congo Red (Sigma‐Aldrich, Oakville, ON, Canada) following protocols described by Theerawitaya et al. (2020) and in personal communication from L. Shaar‐Moshe and S. M. Brady. Images were captured using an LSM‐700 confocal microscope (Zeiss, Oberkochen, Germany).

II..8. 8. Nuclear isolation and snRNA‐seq library generation

For each sample, one developing leaf, from the shoot apical meristem to the green leaf tip, was harvested at ZT 2. Nuclei were released from the tissue using a protocol modified from Guillotin et al. (2023). Approximately 16 000 nuclei per sample were prepared for sequencing using the 10X Genomics Chromium Next GEM Single Cell 3′ Reagents Kits v3.1 (Dual Index) protocol on the 10X Genomics Chromium X instrument (Pleasanton, CA, USA). Details of this procedure are included in Supporting Information Methods S1. Libraries were sequenced at Génome Québec (Montréal, QC, Canada) for sequencing of 100 bp paired‐end reads on a NovaSeq 6000 (Illumina, San Diego, CA, USA), which generated an average of 25 000 reads per nucleus, for a total of c. 4.3 billion reads across the 11 libraries (Table S1).

II..9. 9. snRNA‐seq data processing

Data processing was performed on HPC clusters supported by the Prairie DRI group and the Digital Research Alliance of Canada (https://alliancecan.ca). Reads were pseudo‐aligned to exons of the IRGSP1.0 rice genome (Kawahara et al., 2013), and high‐quality nucleus transcriptomes were determined on a distribution basis with a minimum of 600 UMIs per nucleus using scKB‐COPILOT (Hsu et al., 2022). Doublets were predicted and removed with DoubletFinder (McGinnis et al., 2019), and nuclei with > 10% chloroplast reads were removed; no mitochondrial read filtering was performed as all nuclei contained < 10% mitochondrial reads.

All analyses were carried out in R v.4.3.2 (R Core Team, 2023). Plots were made with Seurat (Hao et al., 2024), scCustomize (Samuel Marsh et al., 2024), and ggplot2 (Wickham, 2011). Developmental trajectories and pseudotime values were predicted using Monocle3 (Trapnell et al., 2017), using Uniform Manifold Approximation and Projection (UMAPs) and clusters defined in Seurat. Trajectories were inferred with [learn_graph] and the starting point was selected as the cell population containing mitotic cells. Gene Ontology (GO) term enrichment analyses were performed with the topGO R package (Alexa & Rahnenfuhrer, 2010), using Fisher's exact test. Transcription factor (TF)‐binding motif enrichment analysis was performed with Analysis of Motif Enrichment (AME; McLeay & Bailey, 2010) from MEME‐Suite (Bailey et al., 2015) using the memes R package (Nystrom & McKay, 2021); motif logos were obtained from Cis‐BP (Lambert et al., 2019). Gene symbols and conversions between Rice Annotation Project (RAP)/Michigan State University (MSU) locus identifiers were obtained from RAP‐database (DB) (Sakai et al., 2013). Statistical tests for leaf emergence, division zone length, cell length, Chl content, and ion contents were performed using ANOVA and Tukey's Honestly Significant Difference test, implemented in the multcompView R package (Graves et al., 2006).

II..10. 10. Annotation of nuclei

To annotate tissue systems, module scores of tissue‐specific genes were calculated using Seurat's [AddModuleScore] function. Genes were combined from predicted cell‐type marker lists from other rice single‐cell RNA‐seq atlases (Wang et al., 2021, 2025; Swift et al., 2024). Nuclear transcriptomes were then manually assigned to either the meristem, epidermis, mesophyll, or vasculature tissue systems based on module scores and developmental trajectory branch points. Cell‐type annotation was performed for cells in each tissue system by subsetting and reprojecting in UMAP space, and clustering cells into finer populations. Cluster‐specific markers were identified with Seurat's [FindAllMarkers] (false discovery rate (FDR) < 0.01, |LFC| > 0.5; including all growth conditions), and the top 200 markers, ranked by FDR, were used for GO term enrichment (P < 0.05). Cell types were annotated based on the expression of known marker genes and enriched GO terms. RNA in situ hybridization (ISH) to confirm cell annotations was performed following the procedure described by Regan et al. (1999), Mukherjee et al. (2015), Canton et al. (1999). Details of this protocol are included in Methods S1. Primer sequences are listed in Table S2.

II..11. 11. Quantitative ggplot map for visualizing pseudobulk sequencing data

We adapted a tool for viewing pseudobulk transcriptomes for different cell types based on micrographs of plant tissue using the ggPlantmap R package (Jo & Kajala, 2024). Briefly, a cross‐section of the sheath of a well‐watered plant was stained with Congo Red, such that the walls of a wide variety of cell types were clearly visible. The perimeter of each cell, defined by the cell wall, was traced using Icy (De Chaumont et al., 2012) and then annotated using the anatomical definitions from Zeng et al. (2016).

For the new Rice Single Cell Abiotic Stress Pseudobulk eFP Browser view, gene expression values representing pseudobulk normalized counts were databased on the BAR server at the University of Toronto (Toufighi et al., 2005; Sullivan et al., 2025). The ‘Rice Single Cell Abiotic Stress Pseudobulk eFP Browser’ view may be explored at https://bar.utoronto.ca/efp_rice/cgi-bin/efpWeb.cgi?dataSource=rice_single_cell.

II..12. 12. Quantitative RT‐PCR

To validate expression of stress‐responsive genes, the upper half (c. 3 cm) of developing third rice leaves in each growth condition were collected for quantitative reverse transcription polymerase chain reaction (qRT‐PCR) of select transcripts (n = 3). Primer sequences are listed in Table S2. qRT‐PCR was performed using PowerUp SYBR Green Master Mix (Applied Biosystems, Waltham, MA, USA). Relative expression was quantified using the pcr R package (Ahmed & Kim, 2018) using the delta–delta Ct method (Livak & Schmittgen, 2001).

II..13. 13. Perturbation Index development

Nuclei were clustered using Seurat's [FindClusters] at four resolutions (0.1, 0.2, 0.4, and 0.8), resulting in 9, 14, 24, and 33 clusters, respectively. All perturbation metrics were calculated at each clustering resolution. Differentially expressed genes (DEGs) were calculated using Seurat's [FindMarkers] with the ‘test.use’ argument set to ‘DESeq2’ (FDR < 0.05). For this analysis, all cells of each annotation were first pseudobulked, such that DEGs could be identified between cells of the same type between treatments. Augur (Skinnider et al., 2021) was run with default settings, using normalized counts from individual nuclei (no replicates; only condition labels). scDist (Nicol et al., 2024) was run with default settings, with ‘fixed.effects’ set to the growth condition, and ‘random.effects’ set to replicate. As Mixscape (Papalexi et al., 2021) runs on individual nuclei without cluster labeling, a single run per stressor was performed. This produced 13 scores for each nucleus in the dataset (four clustering resolutions for three tools, plus Mixscape), which were then averaged. Further details are included in the Methods S1.

II..14. 14. Gene regulatory network modelling

Gene regulatory network (GRN) prediction was performed with MINI‐EX v3.1 (Ferrari et al., 2022; Staut et al., 2023). Input to MINI‐EX consisted of cell‐type‐specific DEGs identified with Seurat's [FindAllMarkers] and pseudobulked counts calculated for each meristem cluster and growth condition. Regulons across the stressors were hierarchically clustered and plotted with the pheatmap R package (Kolde, 2019), and the ERF‐mediated subnetwork was plotted with CytoScape (Shannon et al., 2003; Gustavsen et al., 2019).

Extended methods have been included as Methods S1.

III. Results

III..1. 1. Osmotic stressors delay and dampen leaf growth and development in rice

To investigate the effects of osmotic stressors on rice leaf development, we grew Asian rice (O. sativa cv Nipponbare) seedlings in well‐watered soil media or in mild or moderate drought or saline conditions (Fig. 1a,b; Table 1). The seeds were pregerminated and then transferred into pots of prepared soil; these conditions were maintained until 1 d after the third leaves emerged beyond the sheath collar. The stress conditions used here were experimentally determined by testing a range of soil water and salinity conditions to identify treatments that supported a mild (10%) or moderate (20–30%) growth reduction without delaying the expansion of the third leaf by > 2 d relative to the control conditions without halting development or leading to tissue death (Fig. 1c). To ensure that the differences between the stress treatments were due to the severity of stress, rather than the duration for which the assayed tissue endured the stress, we confirmed that all the assayed leaves were harvested 2 d after they initiated (Fig. S1; Table 1).

Fig. 1.

Osmotic stressors induce variable responses in developing rice leaves. (a) Experimental design. (b) Images of developmentally matched seedlings at third leaf emergence. (c) Third leaf growth after emergence within each condition (n = 9). Error bars are ±SD. (d) Division zone lengths (in mm) of developing third leaves, defined as the distance from the shoot apical meristem to the last dividing epidermal cell (n = 9). (e) Water potential of developing third leaves (mega Pascals (MPa); n = 9). (f) Light response curves of carbon assimilation (A) in fully elongated third leaves (in μmol CO₂ m⁻² s⁻¹; n = 9) at different light levels reported as photosynthetic photon flux density (PPFD). The colored lines represent the Loess regression for each treatment, the gray bands surrounding the colored lines represent the 95% confidence interval. (g) Sodium ion (Na⁺) and potassium ion (K⁺) content within developing third leaves, and the mature leaf blades and sheaths of the first and second leaves (in mg g⁻¹ DW tissue; n = 4); error bars are + SD. (h) Representative cross‐sections of large vascular bundles in developing third leaves in well‐watered or salinity conditions, stained with CoroNa Green to visualize Na⁺ accumulation (green) and Congo Red to visualize cell walls and chloroplasts (red). Images taken at ×400 magnification. Bars, 50 μm. (i) Cross‐sections of whole third leaves taken at ×200 magnification. Bars, 50 μm. For (d, e, g) the statistically different values are indicated with lettered notation (ANOVA with Tukey's post hoc test).

Table 1.

Soil water content, electrical conductivity, and experiment duration.

Growth condition	Soil water content (g water g−1 soil)	Soil water electrical conductivity (dS m−1)	Days to third leaf initiation (DAP)	Days to tissue collection (DAP)
Well‐watered	3	0.5	4	6
Mild drought	1.5	na	5	7
Moderate drought	1	na	6	8
Mild salinity	3	1.25	5	7
Moderate salinity	3	2.5	6	8

Electrical conductivity is reported at dS m⁻¹; days to third leaf initiation and tissue collection are reported as days after planting (DAP). na, not applicable.

We measured the length of the division zone for each condition, defined by the distance from the shoot apical meristem to the last dividing epidermal cell (Sprangers et al., 2016) because it is directly related to leaf expansion. The lengths of the division zones for all stress treatments were reduced relative to the well‐watered condition with the greatest reduction in length in the moderate drought condition (Fig. 1d). Notably, the cell length within the division zone was not significantly different between treatments (Fig. S2). This supports the hypothesis that cells in the stressed tissues began expanding after fewer divisions than in control conditions, which is consistent with previous reports in maize (Nelissen et al., 2012) and wheat (Hu & Schmidhalter, 2008).

To ensure that the differences in soil water availability led to physiological differences in the plants, we measured leaf water potential and assayed the photosynthetic capacity of plants grown in the different stress conditions. The leaf water potential was reduced in all conditions, with the greatest reduction in the moderate drought and salinity treatments (Fig. 1e). Relative to well‐watered leaves, Chl content was slightly increased in mild drought and salinity and greatly increased in moderate drought and salinity (Fig. S3). This is consistent with findings of increased Chl content in the growth zone of developing maize leaves under severe drought stress (Avramova et al., 2015). To assess photosynthetic capacity of the leaves, we measured the carbon assimilation rate (A) at 10 light intensities ranging from 0 to 2100 μmol photons m⁻² s⁻¹, and fit models to the curves (Marshall & Biscoe, 1980; Fig. 1f). This analysis determined that the leaves experiencing mild drought and salinity had a small reduction in maximum photosynthetic rate relative to well‐watered leaves, while moderate drought and salinity had a much greater reduction, demonstrating greater disruption of energy production under moderate stress.

III..2. 2. Salt accumulation in tissues is heterogeneous – between organs, cells – and depends on stress extent

Previous studies have shown that developing tissues are protected from high levels of salt accumulation (Wang et al., 2012). As such, we wanted to determine whether the developing leaves grown in the two salinity intensities had elevated ion content relative to control leaves. We measured the accumulation of sodium and potassium ions in developing leaves (third leaves; emerged by 1–2 cm beyond the sheath collar) and in the blades and sheaths of fully expanded leaves (first and second leaves). Ion content of the blades and sheaths of the fully expanded leaves was measured separately because of demonstrated differences in sodium accumulation in these tissues in rice (Kobayashi et al., 2017; Neang et al., 2019). In all measured tissues, the sodium ion concentration was highest in the moderate salinity treatment and lowest in the control treatment (Fig. 1g). Notably, sodium ion accumulation was lower in developing leaves than in mature leaves, and as expected, highest in leaf sheaths. We also measured potassium ion concentration in these tissues, as Na⁺ : K⁺ ratios are critical for ion and nutrient balance in rice cells (Wang et al., 2012; Chen et al., 2015). The potassium ion concentration was consistent across tissues, and it decreased with increasing salinity extent (Fig. 1g), resulting in elevated Na⁺ : K⁺ ratios under salinity (Fig. S4), consistent with previous findings from salt‐susceptible rice varieties (Wangsawang et al., 2018; Liu et al., 2019).

Given the clear heterogeneity of salt accumulation between organs, we next examined the distribution of sodium ions across cells and cell types. For this, we examined sodium accumulation in developing rice leaves using CoroNa Green, a sodium ion indicator dye (Figs 1h, S5). In well‐watered conditions, sodium ions were detected only in the periphery of xylem and xylem parenchyma cells. In the mild salinity treatment, sodium ions were also detected in fiber and mestome sheath cells and extensively in the vacuoles of mesophyll cells. In moderate salinity treatment, the distribution of sodium ions was further expanded to include bundle sheath, epidermis, and bulliform cells. These accumulation patterns are similar to previous studies using CoroNa Green in salt treated rice leaves, which show localization to vascular bundles (Nath et al., 2015; Theerawitaya et al., 2020) and in wild and domesticated tomato leaves (Shaar‐Moshe et al., 2026). In addition to the wider distribution of sodium ions with increasing salinity, we note that there appear to be substantial differences in sodium ion accumulation between cells of the same type. For example, in the moderate salinity treatment, some bundle sheath cells had clear and abundant sodium ion accumulation, while other neighboring cells appeared to have none. We also note sodium appeared to be restricted to the large vascular bundles in the well‐watered leaves but also accumulated in the small and simple vascular bundles under mild and moderate salinity (Fig. 1i). The differences of sodium ion accumulation within and between cell types suggested to us that the stress responses might also vary across the tissue.

III..3. 3. Module scores place nuclear transcriptomes in known developmental contexts

To explore the cell scale heterogeneity of drought and salinity stress responses, we performed a single‐nucleus RNA‐sequencing (snRNA‐seq) experiment. The third leaves were dissected from the seedlings 1 d after they emerged beyond the sheath collar, nuclei were isolated, and snRNA‐seq libraries were synthesized. After filtering out doublets and low‐quality nuclei that included > 10% chloroplast reads or low unique molecular identifier (UMI) counts using a distribution‐based approach (Hsu et al., 2022), the final dataset was composed of 126 139 nuclei. A median of 1304 genes was detected per nucleus; across the population of nuclei, 25 985 genes were detected in at least 100 cells, representing 68.6% of nuclear‐encoded genes in the O. sativa genome (Table S1).

We used a data‐driven approach to annotate each nuclear transcriptome with a tissue system identity. To do this, we extracted marker genes from three single‐cell studies on developing rice coleoptiles (Swift et al., 2024), seedlings (Wang et al., 2021), and young leaves (Wang et al., 2025). Markers from each study were combined to calculate a module score for each nuclear transcriptome, based on marker gene expression in the meristem (631 genes), epidermis (1047 genes), vasculature (808 genes), and mesophyll (785 genes; Fig. 2a; Table S3). We reasoned that the marker genes we had used to define the tissue systems might be more predictive of fully differentiated cells than cells in transitional states, and so to annotate the immature cell states, we leveraged pseudotime analysis (Fig. 2b). By predicting developmental trajectories of cell states initiating among the meristematic cells through each tissue system with smaller trajectories branching from them, we could refine tissue level annotations. Nuclei were clustered and assigned a tissue system based on sharing the characteristics of either the meristem (c. 10%), epidermis (c. 25%), vasculature (c. 25%), or mesophyll (c. 40%; Fig. S6). The relative proportion of nuclei annotated to each tissue system was conserved across treatments and biological replicates (Fig. S7).

Fig. 2.

Nuclear transcriptomes recapitulate known tissue systems and cell types. (a) Uniform Manifold Approximation and Projections (UMAPs) containing transcriptomes from all conditions, colored by Seurat module scores of tissue system markers from three rice leaf single‐cell datasets. (b) Predicted pseudotime values showing developmental trajectories from dividing cells in the meristem (early) to mature cell states in each tissue system (late). Darker colors indicate earlier pseudotime. (c) Reprojected UMAP of a subset of cells from the vasculature tissue system, used to annotate individual cell types. CC, companion cella; SE, sieve element. (d) Dot plot of known marker genes used to assign vascular cell types. The size of the dot indicates the percentage of nuclei in each cell type expressing the gene, and color indicates the average scaled expression level. (e) Cell‐type annotations of all nuclear transcriptomes. All growth conditions are included. (f) ggPlantMap representation of a large vascular bundle highlighting annotated mature cell types in our dataset. (g) Example marker genes plotted on the plant map. Altered Phloem Development (APL) – Phloem marker. Cinnamoyl‐CoA Reductase 20 (CCR20) – lignin biosynthesis. Only cells from the well‐watered control are used. (h) Plant maps showing the pseudobulked, normalized expression of Response to Abscisic Acid 16A for different cell types in all growth conditions tested. (i) Plant maps showing differential expression between each stressor and control; only cell types in which the gene is differentially expressed (false discovery rate (FDR) < 0.01, |Log₂ Fold Change| > 0.5) are colored.

We annotated cell types within each tissue system by subsetting, reprojecting, and clustering the transcriptomes and examining the expression of known marker genes (Fig. S8). Cell‐type marker genes were extracted from the literature (Matsukura et al., 2000; Fitter et al., 2002; Nishimura et al., 2002; Yamaji et al., 2008; Wei et al., 2011; Mao et al., 2012; Pettkó‐Szandtner et al., 2015; Lu et al., 2019; Wang et al., 2021; Yan et al., 2022; He et al., 2024; Swift et al., 2024; Wang et al., 2025) and from the scPlantDB resource (He et al., 2024; Table S4). For cell types for which marker gene availability was limited, particularly the vasculature, we incorporated GO term enrichment of cluster markers to infer cell types from known functions (Tables S5, S6). For example, the cluster we annotated as fiber cells was enriched for GO terms associated with secondary cell wall biosynthesis and lignin deposition (Ambavaram et al., 2011), and bundle sheath cells were enriched for trehalose metabolism (Tonetti et al., 2025). Among the nuclei from the vasculature, we were able to annotate nuclei as procambium, bundle sheath, mestome sheath, xylem tracheary elements, xylem parenchyma, phloem sieve elements, phloem companion cells, and fiber (Fig. 2c,d). We applied the same annotation strategy for the other tissue systems and annotated precursor and mature cell types (Figs 2e, S9). To confirm the cell‐type annotations, we selected three transcripts with preferential expression in various cell types to validate using RNA ISH, including mesophyll marker RBCS4 (Os12g0292400), fiber marker LAC5 (Os01g0842500), and procambium marker OSH1 (Os03g0727000; Fig. S10). We were unable to confidently identify nuclei derived from mature epidermal cell types, such as guard cells and bulliform cells, and stone cells which may stem from a lack of resolution in our data or a need for more reliably expressed marker genes for these cell types. Despite this, our dataset is sufficiently annotated to address the transcriptional responses of cells early in leaf development.

III..4. 4. Data visualization tool development

To facilitate data exploration and hypothesis development, we created a pseudobulk visualization tool using the ggPlantmap R package (Jo & Kajala, 2024). This package allowed us to create intuitive and quantitative heatmaps of the snRNA‐seq data visualized as a cross‐section of a rice leaf. We manually mapped borders of all cells in a leaf cross‐section micrograph (Fig. S11), and then annotated each cell based on detailed definitions and micrographs from Zeng et al. (2016) (Fig. 2f). Pseudobulk expression for fully differentiated cell types can be used to map the expression of markers for a cell type or a biological function of interest (Fig. 2g). Here, we show the expression of APL (Os02g0174000), a phloem marker gene (Yan et al., 2022), and CCR20 (Os08g0441500), which is involved in lignin biosynthesis (Gong et al., 2024), both of whose expression is localized to the expected cell types. Additionally, pseudobulk expression of any gene can be visualized as normalized expression values (Fig. 2h) or as log₂ fold‐change values between each stressor and well‐watered control (Fig. 2i); for the log₂ fold‐change values, in each cell type, only genes that are differentially expressed (FDR < 0.01, |LFC| > 0.5) are assigned colors. Expression of RAB16A (Os11g0454300), a common marker for osmotic stress (Ganguly et al., 2012; Chakraborty & Roychoudhury, 2022), demonstrates differential expression between stress type, severity, and cell specificity. The data and images are also available on the BAR eFP browser so that the community can easily explore the data set reported here (https://bar.utoronto.ca/cellxgene/view/OsLeafStressIntegrated.h5ad/).

III..5. 5. Transcriptional stress responses are characteristic of cell type, stress type and stress extent

We anticipated that the extent of stress‐induced perturbation would vary across cell types, stress types, and stress extents. To determine whether transcriptome changes were unique to a particular treatment, stressor or extent of stress, or common to all stress conditions, we calculated DEGs (FDR < 0.01, |LFC| > 0.5) between each stressor and the well‐watered control for each annotated cell type (Fig. 3a; Table S7). Strikingly, we found that over half of the DEGs in each cell type were uniquely differentially expressed in response to a single experimental condition (ranging from 50% of DEGs in the meristem to 64% of DEGs in vascular cell types). Within each stressor, c. 45% of DEGs were specific to a single cell type, with an additional c. 17% and c. 10% shared between two and three cell types, respectively (Fig. S12). The next largest set of DEGs were common to two or three of the more severe stressors. We note that the largest number of DEGs were found in the meristem and differentiating tissues (i.e. precursors and procambium) than mature cell states, similar to other findings in single‐cell studies of plant stress responses (Hill et al., 2026). We performed GO term enrichment analysis of all gene sets (Table S8). Exploring the epidermal precursor cells' stress responses, we found that DEGs that were common across all stressors were enriched for GO terms associated with general stress, including responses to heat and hydrogen peroxide, while DEGs particular to salinity responses were enriched for ‘photosynthesis’ and other sugar and light related GO terms (Fig. 3b). This confirmed that some gene functions are perturbed in stress‐specific manner.

Fig. 3.

Transcriptional stress responses in Asian rice (Oryza sativa) are influenced by stress type, stress degree, and developmental stage . (a) Number of differentially expressed genes (DEGs) between each stressor and well‐watered control (false discovery rate (FDR) < 0.05, |Log₂ Fold Change| > 0.5) in representative expression patterns in all cell types. Darker pink color indicates larger numbers of DEGs. (b) Top two enriched Gene Ontology (GO) terms of each DEG expression pattern in the epidermal precursor cells (P < 0.05). P‐values indicated on the right. (c, d) Plant maps highlighting absolute expression values of Heat Shock Protein 22A (HSP22A) (c) and basic leucine zipper protein 18 (bZIP18) (d) relative to well‐watered control (top panel). In situ hybridization of each transcript stained with Western Blue is shown, with darker purple indicating higher transcript abundance. Bars, 50 μm. (e) Pseudobulked expression of select genes within mature cell types from the single‐nucleus RNA‐sequenced transcriptomes (n = 2). Bars indicate the mean expression of replicates for each growth condition relative to the well‐watered control with SD indicated with vertical bars. (f) Expression of the same genes in bulk leaf samples, quantified by quantitative reverse transcriptase polymerase chain reaction (qRT‐PCR; n = 3). Bars indicate the mean expression of replicates for each growth condition relative to the well‐watered control with SD indicated with vertical bars. In (e, f), the bars correspond to samples in the following order, from left to right: well‐watered, mild drought, moderate drought, mild salinity, moderate salinity.

To confirm that quantitative changes in transcript abundance calculated from snRNA‐seq data could be recapitulated using an orthogonal qualitative method, we selected two genes that responded to stress according to snRNA‐seq for confirmation by ISH. HSP22A was higher in both moderate salinity and moderate drought treated in many cell types, with the greatest change in the mesophyll cells in both snRNA‐seq and ISH (Fig. 3c). Similarly, bZIP18 was preferentially expressed in the bundle sheath cells and, in response to the mild salinity treatment, in a broader variety of cell types by both methods (Fig. 3d). We used qRT‐PCR analysis of several genes to confirm the quantitative differences in transcript abundance between treatments, including HSP22A (Os04g0445100), bZIP18 (Os02g0203000), LHCA3 (Os02g0197600), PHI1 (Os02g0757100), and RNS4 (Os09g0537700). These transcripts were selected because they had distinct stress‐responsive abundance patterns. The qPCR data, which were generated from whole leaves that were developmentally matched to those from which the snRNA‐seq analysis was performed, closely paralleled pseudobulked expression patterns from the snRNA‐seq data (Fig. 3f). These results show that snRNA‐seq can be used to profile qualitative and quantitative stress‐responsive transcriptome changes in plants.

III..6. 6. An ensemble‐based Perturbation Index quantifies transcriptional changes

We recognized a need to develop a tool to characterize stress perturbation that made use of the unique characteristics of snRNA‐seq data, namely, the sparsity of the data for each cell and the high replication of each cell type. To identify which cells of the developing leaves were the most perturbed by each stressor, we developed a Perturbation Index that incorporated four orthogonal strategies for measuring stress‐responsive transcriptome changes. DESeq2 (Love et al., 2014) uses the Wald test to quantify the number of DEGs between treatments (FDR‐adjust P < 0.05); Augur (Skinnider et al., 2021) uses a machine learning approach to estimate differences between treatments in high‐dimensional space; and, scDist (Nicol et al., 2024) uses a mixed‐effect model to estimate distances between cell populations in Euclidean space. For these three methods, the data had to be preclustered before perturbation could be estimated. Because cluster size is known to influence the outputs of downstream analyses (Zhang et al., 2019; Grabski et al., 2023), we included four clustering resolutions (resulting in 9, 14, 24, and 33 clusters) for each tool in our index (Figs 4a, S13–S15; Table S9). The fourth method, Mixscape (Papalexi et al., 2021), estimates perturbation between neighboring nuclei, which produces a unique perturbation score for each individual nucleus (Fig. S16). For each method and clustering resolution, the perturbation metric (i.e. DEG count, area under the curve (AUC), distance) for each cell was scaled from 0 to 1 across all four stressors. These 13 scores were then averaged to generate a single Perturbation Index for each nucleus, which can be used to compare the degree of perturbation between stressors as well as for each cell in response to each stressor (Fig. 4b). While meristematic and mesophyll cells had consistent index values across all tools, the large parenchyma cells had a high Perturbation Index according to scDist, and a low index according to the other tools (Figs S13–S15). It was only by using multiple tools that we were able to discover this divergence, thus, highlighting the need to incorporate multiple tools to assess transcriptome perturbation.

Fig. 4.

Ensemble‐based Perturbation Index (PI) quantifies transcriptional changes in individual Asian rice (Oryza sativa) cells. (a) Uniform Manifold Approximation and Projection (UMAP) are shown for each tool and clustering resolution, only results for moderate drought are shown here. Tools used are in bold and include DESeq2, Augur, and scDist; the text beneath each tool name is the metric produced by the tool (number of differentially expressed genes (DEGs), area under the curve (AUC) of precision/recall, Euclidean distance (distance)). Four clustering resolutions were used, resulting in 9, 14, 24, and 33 clusters, respectively. Lower PIs are indicated with darker colors. (b) UMAP of PI for each cell in all osmotic stress conditions. (c) Violin plots of the PI in each cell, grouped by condition and tissue system. Each dot represents the PI for a single cell. The width of the violin plot is proportional to the number of cells with a given PI score. Osmotic stressors perturb regulatory programs underlying cell division and differentiation.

Stress‐induced transcriptome changes varied across cell types, stress types, and stress extents (Fig. 4b,c). As anticipated, the mild drought stress treatment led to the lowest extent of perturbation in all cell types relative to the other stress treatments. Notably, in response to mild drought and salinity treatments, cells in the meristem were less perturbed, or less distinguishable from control cells, than cells in the more differentiated tissue systems. By contrast, both the moderate drought and salinity treatments resulted in a high extent of perturbation across all cell types, with the highest Perturbation Index in meristematic cells. This analysis emphasizes the diversity of responses across cell types, so we next wanted to perform a more targeted analysis on early differentiating cells.

III..7. 7. The fates of early differentiating cells can be predicted from the transcriptomes of nuclei in the meristem

As the previous analyses identified the meristem as a highly stress‐perturbed region of the developing leaf (Fig. 4), we next investigated the impact of stress on the regulatory programs of the meristem, precursor, and procambium transcriptomes. Within the meristematic cells, distinct cell populations at different stages of the cell cycle were identified by the expression of cell cycle markers, including 15 markers of the G1/S‐phase and 11 markers of the G2/M‐phase (Fig. 5a; Table S10) (Abe et al., 2008; Caillaud et al., 2009; Ogawa et al., 2011; Pettkó‐Szandtner et al., 2015; Lin et al., 2022; Chandran et al., 2025). Actively dividing G2/mitotic nuclei clustered together on the UMAP while nuclei in the G1/S‐phase were seen in three separate populations leading to each tissue system. We subsetted and clustered the meristematic nuclei and identified five distinct populations: nuclei derived from initial cells differentiating into each of the vascular, mesophyll, and epidermal tissue systems, mitotic nuclei, and a population of nuclei whose dominant transcriptome attribute was stress response (Fig. 5b). The vascular, mesophyll, and epidermal initial populations were annotated based on the proximity to the precursor populations.

Fig. 5.

Osmotic stressors perturb regulatory programs underlying cell division and differentiation in Asian rice (Oryza sativa). (a) Module scores of cell cycle markers for G1/S‐phase and G2/M‐phase on a Uniform Manifold Approximation and Projection (UMAP) of meristem nuclei. (b) Annotated clusters of nuclei within the meristem, that differentiate into each tissue system. Arrows indicate the developmental trajectories. (c) Proportion of nuclei from the stress‐responsive nuclei cluster in each of the growth conditions. (d) Total unique molecular identifier (UMI) counts of all nuclei in each meristem population (gene count has been log₁₀ transformed). Each data point represents the UMI count for a single nucleus; the width of the violin reflects the frequency at which a given UMI count is reported. (e) Predicted regulons from Mini‐EX. Color indicates the number of target genes regulated by transcription factors within each condition and nuclei population. (f) Enriched transcription factor binding motifs (TFBMs) (false discovery rate (FDR) < 0.05) in the promoter region of all target genes within each nuclei population are represented as sequence LOGOs. Family names of the transcription factors (TFs) that bind to a given TFBM are listed beside the LOGO. (g) Regulatory network of ethylene response factors (ERFs; blue) and target genes associated with abscisic acid (ABA; red) and gibberellic acid (GA; green) signaling within the stress‐responsive nuclei population. Arrowheads indicate directed regulation from regulon TFs to target genes.

The stress‐responsive nuclei had transcriptomes that were intermediate to those of the mesophyll and epidermal populations. To ascribe function to these nuclei, we identified marker genes for each meristem cell population (FDR < 0.01, LFC > 1; Table S11) and noted that several of the top markers for the stress‐responsive population were genes that regulate cell division and progression through the cell cycle under abiotic stress (Takahashi et al., 2003; Xu & Cai, 2014; Shen et al., 2023) and genes involved in oxidative damage mitigation (Garg et al., 2010; Zhang et al., 2011; Li et al., 2013; Fig. S17). The proportion of nuclei within this cluster differed between conditions, representing 1.3% of meristematic cells from the well‐watered condition, 1.5% of mild drought, 3.0% of mild salinity, 3.3% of moderate drought, and 3.6% of moderate salinity (Fig. 5c). Notably, treatments with the highest proportion of nuclei in this cluster were those with the shortest division zone length (Fig. 1d). Additionally, nuclei in this cluster had a larger number of UMI counts relative to other populations of nuclei (Fig. 5d), consistent with transcriptome sizes of cells in the G2 phase undergoing DNA replication before division (Schwabe et al., 2020). Taken together, this may indicate that this cell population represents stressed cells that are inhibited from further dividing.

III..8. 8. Transcription factors regulating cell division are active in stress responses in early differentiating cells

To model how stress responses in early cell states are regulated, we used Motif‐Informed Network Inference based on single‐cell EXpression data (MINI‐EX) to predict cell type‐ and condition‐specific regulons consisting of TFs and their target genes (Ferrari et al., 2022). The putative target genes must be localized to the same cell type and have the cognate TF‐binding motif (TFBM) in their promoter to be included in the predicted network. With MINI‐EX, we generated a list of TF‐target interactions in each cell type for each condition; to retrieve interactions that were perturbed by stress, we removed interactions that were common between the well‐watered control and any of the stressors. This filtered network consisted of 594 regulons, with 101 TFs and 4489 target genes, with regulons containing a median of 213 targets (Table S12). Of the 101 TFs, 45 were active in only one cell population, indicating a high degree of specificity of stress regulation in differentiating cells (Fig. S18). For clarity in visualization, we included only TFs predicted to have > 500 targets across all cell types and conditions, and hierarchically clustered the regulons to identify regulatory patterns across stressors (Fig. 5e). The regulons include TFs involved in known developmental pathways and abiotic stress responses. For example, we found several Rice Outermost Cell‐Specific (ROC) family TFs active in the epidermal initials, which are involved in formation and development of epidermal cells (Xu et al., 2021), and two Golden2‐like (GLK) TFs active in the mesophyll initials, which initiate chloroplast development in photosynthetic tissues (Fitter et al., 2002). Among TF families known to regulate abiotic stress responses, our network included five basic helix–loop–helix (bHLH) TFs, which are commonly involved in drought responses (Wei & Chen, 2018), and seven ERF TFs, which regulate transcriptional responses to osmotic and temperature stressors and modulate plant growth (Mizoi et al., 2012).

Based on the size of the regulons active within each cell type, we found that moderate drought induced a larger regulatory shift in epidermal and mesophyll initials than did other stressors, but moderate drought and moderate salinity had a similar number of targets with the same regulators in vascular initials and mitotic cells. Most regulons in the stress‐responsive meristematic nuclei were of similar size for all stressors, indicating shared regulatory programs that were active regardless of stress type or extent. To further characterize the regulatory relationships between genes in this network, we calculated enriched TFBMs within the proximal promoter region of target genes in each cell population. We identified TFBMs within a narrower region of −453 to +137 bps from the Transcription Start Site (TSS), as this has been shown to be the most active region of TF‐binding activity in rice (Wilkins et al., 2016). We found that each annotated group of nuclei was enriched for a distinct combination of TFBMs associated with different TF families (Fig. 5f; Table S13; FDR < 0.05). The epidermal initials were most enriched for TFs of the myeloblastosis (MYB) and heat shock factor (HSF) families, mesophyll initials were enriched for NAM, ATAF, and CUC (NAC) and basic leucine zipper (bZIP) families, vascular initials were enriched for DNA binding with one finger (DOF) and Cysteine2‐Histidine2 (C2H2) families, and mitotic cells were enriched for MYB and bZIP families. Notably, the targets in the stress‐responsive nuclei population were highly enriched for TFs from the APETALA2/ERF (AP2/ERF) family, including both the classical AP2/ERF DNA‐binding domain (GCCGCC) and the dehydration response element domain (CCGAC; Nole‐Wilson et al., 2000). This suggests that combinatorial regulation by multiple TF families mediates developmental responses to stress in distinct tissue systems.

Given the enrichment of AP2/ERF TFBMs in the promoters of target genes within the stress‐responsive nuclei population, as well as the expression of seven ERF TFs in these nuclei, we next looked at targets within the ERF regulons. We identified genes related to stress phytohormone biosynthesis, signaling, and responses, particularly related to gibberellin (GA; 4 genes) and abscisic acid (ABA; 12 genes; Table S12), all of which are upregulated under stress. We plotted a GRN of the seven ERF TFs and 21 targets (Fig. 5g), which identified a hierarchy of ERF‐mediated regulation, with higher‐order TFs, such as ERF48 (Os08g0408500), ERF3 (Os01g0797600), and ERF64 (Os03g0183300) regulating other ERFs as well as common target genes. Examples include RCAR10 (Os10g0573400) and PP2C9 (Os01g0846300), both involved in ABA signaling and growth modulation under drought in rice (Verma et al., 2019; Miao et al., 2020), as well as GID1 (Os05g0407500) and GASA3 (Os04g0465300), involved in GA reception and signaling, respectively, under drought and salinity (Du et al., 2015; Muhammad et al., 2019). As antagonistic ABA and GA signaling interactions have a strong influence on plant growth under osmotic stress conditions (Du et al., 2015; Liao et al., 2023), this further supports the hypothesis that cells in the stress‐responsive nuclei population are inhibited from further dividing, driven by phytohormone signaling. Taken together, we find that osmotic stress responses are integrated with developmental programs that shape leaf development, and these responses are distinct within nuclei populations differentiating into each tissue system.

III..9. 9. Mild salinity perturbs mesophyll development and function

To explore the heterogeneity of stress responses along a developmental trajectory, we explored the stress responses in the photosynthetic mesophyll nuclei. We selected these nuclei for further characterization because they represent a single cell type in different developmental states, they represent the largest portion of all cell types in the dataset (32.9%; 41 433 nuclei), and they exhibit a high degree of perturbation in response to all stressors (Fig. 4). We reprojected the nuclei in UMAP space and applied a pseudotime analysis to identify a developmental trajectory of cell states (Fig. 6a). Nuclei were assigned to one of eight pseudotime bins, each with an equal number of nuclei (1598; Bin 1 most meristematic–Bin 8 most differentiated). To first explore how this developmental progression is affected by stress, we averaged the Perturbation Index of nuclei in each bin for each stressor (Fig. 6b). Interestingly, the Perturbation Index scores across development had a similar profile for both drought stressors, with the highest index value in Bin 4 and a decrease toward the mature cell states. On the contrary, moderate salinity peaked in Bin 6 before a decrease toward maturation, while mild salinity continued to increase along the entire developmental progression. This indicates that drought mainly perturbs cell states undergoing maturation, while salinity continues to disrupt mesophyll cells late into development.

Fig. 6.

Mild salinity perturbs mesophyll development and function in Asian rice (Oryza sativa). (a) Uniform Manifold Approximation and Projection (UMAP) of mesophyll precursors and mature mesophyll cells, colored according to pseudotime score from early to late development. (b) Averaged Perturbation Index scores in each pseudotime bin. (c) Upset plot of mesophyll differentially expressed genes (DEGs) (false discovery rate (FDR) < 0.01, |Log₂ Fold Change| > 0.5) between each stressor and well‐watered control, which includes genes that are differentially expressed in any of the pseudotime bins. (d) Select expression patterns of mesophyll DEGs across pseudotime bins. Pseudobulked expression values were scaled with z‐scores. Summarized enriched Gene Ontology (GO) terms (P < 0.05) are shown. (e) Log₂ fold change of DEGs annotated with the ‘photosynthesis’ GO term in mesophyll, epidermis, large parenchyma, and bundle sheath cell types. (f) Plant maps of Photosystem II subunit P (PSBP), methionine sulfoxide reductase B1(MSRB1), and basic leucine zipper protein 18 (bZIP18).

To identify which genes and functions are affected by each stressor along this developmental progression, differential expression analysis was conducted by comparing each stressor to the control for each pseudotime bin. In this way, 1714 unique DEGs were identified across the stressors, with the largest number in mild salinity (Table S14). We used an Upset plot to visualize patterns of shared mesophyll DEGs between the stressors (Fig. 6c). The most common pattern was DEGs common to the moderate drought, mild salinity, and moderate salinity responses (295 DEGs). The next four most common patterns were genes differentially expressed in a single stressor: mild salinity (291 DEGs), moderate drought (242 DEGs), mild drought (153 DEGs), and moderate salinity (110 DEGs). This indicates that there are both shared and stressor‐specific perturbations in the mesophyll.

The 1714 mesophyll DEGs were then hierarchically clustered based on their expression across all treatments, and we identified seven stress response profiles among the DEGs (Fig. S19). Of these, three profiles were characterized by strong upregulation of genes across mesophyll development in response to stress (Fig. 6d). The expression of genes in Cluster 1 (260 DEGs) did not change across the developmental trajectory, but were uniformly increased in mild and moderate salinity, and especially moderate drought. These genes were enriched for GO terms related to protein folding and refolding, as well as responses to heat, hydrogen peroxide, salt stress, water deprivation, and alcohol, indicating an elevation of broad stress responses. Genes in Cluster 2 (426 DEGs) increased along the developmental trajectory in all conditions, with mild salinity in particular having the highest expression, and were enriched for photosynthesis‐related functions, such as sugar metabolism, Chl and carotenoid biosynthesis, and photosystem assembly and repair. Genes in Cluster 3 (80 DEGs) were expressed most highly in the middle of the trajectory, but were only increased in the mild salinity treatment, and were enriched for protein folding, refolding, and stabilization, as well as water transport and water deprivation. There were several notable groups of genes in this cluster with common functions suggesting cell expansion: water transporters (PIP1A, PIP1;2, and TIP2;2), and cell wall modifiers (PME3, AGP2, AGP26, AGP27, GLP1, FLA2, FLA19, EXPA7, EXPB11, RALF4, RALF8, RALF31, and RALF42). Taken together, these results indicate that moderate drought and moderate salinity have shared classical abiotic stress responses induced across mesophyll development, while mild salinity has a more significant response by upregulating genes related to photosynthesis and cell expansion.

While exploring the expression of these genes across all cell types, we noted a striking pattern of photosynthesis‐related genes that were also differentially expressed in the epidermis, large parenchyma, and bundle sheath under mild and moderate salinity. While typically considered photosynthetically inactive, these cell types contain chloroplasts and exhibit low expression of some, but not all, photosynthesis genes (Tsutsumi et al., 2006; Sage & Sage, 2009; Cackett et al., 2025). We took all genes with a GO annotation of ‘photosynthesis’ that were detected in our dataset (92 genes) and checked whether they were differentially expressed in these cell types under mild or moderate salinity (Fig. 6e); of these, 59 were differentially expressed, 54 of which were upregulated. Of the upregulated photosynthesis‐related genes: 51 and 40 were differentially expressed in the mesophyll of mild and moderate salinity, respectively; 49 and 40 were differentially expressed in the epidermis, respectively; 44 and 26 were differentially expressed in the large parenchyma, respectively; and 16 and 5 were differentially expressed in the bundle sheath, respectively. These photosynthesis‐annotated genes were overwhelmingly involved with light capture, including many photosystem I (PSA‐) and II (PSB‐) subunits, light‐harvesting complex I (LHCA‐) and II (LHCB‐) subunits, and Chl‐binding proteins (CP‐). The number of DEGs was far fewer in moderate and mild drought or other cell types in the salinity treatments (Fig. S20), indicating that mesophyll in salinity conditions induces a strong upregulation of light‐capture functions which carry over into neighboring cell types. Recently, a single‐cell study on photomorphogenesis in rice and sorghum leaves found that bundle sheaths in rice have a significantly reduced response to light relative to sorghum bundle sheaths (Swift et al., 2024), suggesting that the light responsiveness detected here may only occur under stress.

Finally, we visualized genes that may be related to this process with plant maps (Fig. 6f). An example from (Fig. 6d): PSBP (Os07g0141400) demonstrates a light‐capture gene that is unchanged in drought but upregulated in multiple tissues in salinity, notably in the bundle sheath of mild salinity. Notably, using CoroNa green, we show that plants grown in saline soil accumulated sodium in the bundle sheath (Fig. 1h). We also noted the presence of several DEGs in the bundle sheath related to oxidative stress, including MSRB1 (Os06g0472000) which was highly upregulated in the large parenchyma and bundle sheath in mild salinity, indicating that this increased light capture may lead to oxidative stress. Finally, bZIP18 (Os02g0203000) was upregulated in the mesophyll and large parenchyma in moderate drought and mild and moderate salinity, but additionally highly upregulated in the bundle sheath under mild salinity. These are cell types in which sodium ions accumulate. bZIP18 has been found to be a key regulator of light responses under abiotic stress (Duan et al., 2025), suggesting that mild salinity induces regulatory changes that increase light capture in cell types with typically low contributions to photosynthetic activity.

IV. Discussion

Osmotic stress is a major constraint on rice productivity, yet its effects on leaf development, a critical predictor of yield, is poorly understood. By generating a single‐nucleus transcriptome atlas of rice leaves developing under mild and moderate drought and salinity stresses, we have gained insight into how stress type and extent alter transcriptional and regulatory responses in differentiating cell types. We find that although drought and salinity produce similar leaf‐scale phenotypes (i.e. reduced growth and carbon assimilation), the underlying transcriptional programs invoked are stress‐specific and vary between cells of different lineages and developmental stages. This is consistent with findings in Arabidopsis that found that salinity and osmotic stressors induced more stress‐specific differentially expressed proteins than shared between stressors (Rodriguez Gallo et al., 2023). It has also been shown that transcriptional responses to drought highly differ along leaf developmental zones in Brachypodium (Verelst et al., 2013), maize (Nelissen et al., 2018), and rice (Robertson et al., 2025), and differ between cell types in Arabidopsis (Berrío et al., 2025; Illouz‐Eliaz et al., 2025) and sorghum (Stata et al., 2025). These findings emphasize the heterogeneity of stress responses across developing cell populations and highlight the need to examine individual components of environmental stressors.

Another important finding of this study is that mild and moderate stressors did not simply induce scaled versions of the same responses. While the increasing stress severity led to decreasing leaf growth, division zone length, and photosynthetic capacity, we showed that increasing stress led to qualitatively different transcriptome changes, beyond the quantitative changes we had anticipated. First, we determined that mild stressors mostly affected the transcriptomes of the mature cell states, while the moderate stress treatments had profound effects on the transcriptomes of meristematic cells. The differences between the responses to the mild and moderate stresses are consistent with the finding that increasing severity of drought stress in rice induced nonlinear reductions of leaf growth through different soil matric potential thresholds for limiting cell expansion and cell division (Todaka et al., 2017). Second, we found that in response to the mild stressors, fully differentiated cell types responded with distinct transcriptome changes, reflecting perturbation to the primary functions of the cell types; whereas, in response to the moderate stress, meristematic cells responded with changes in the expression of genes involved in damage mitigation, including chaperone proteins, ROS scavengers, and cell division. These findings may reflect distinct strategies undertaken by the developing leaves: moderate osmotic stress triggers a stress‐avoidance strategy in which cell division and expansion is inhibited to conserve energy and prevent further damage in mature cells, whereas mild osmotic stress triggers a stress‐tolerance strategy where mature cells impacted by low water availability or ion accumulation react to the stress as it arises.

The response to mild salinity was characterized by increasing perturbations along the mesophyll developmental axis. During leaf expansion, plants take up water to support the rapid growth and expansion of cells. In mild salinity conditions, the growth rate is not slowed, but the final leaf length is reduced compared with well‐watered leaves; cell growth and size in the division and elongation zones are unchanged or may even increase due to the reduced osmotic potential from sodium ions (Fricke & Peters, 2002; Hu & Schmidhalter, 2008). As water is taken up, sodium ions are pulled into expanding tissues, including the mesophyll as shown by the high sodium ion accumulation in the mild salinity. We hypothesize that the elevated sodium levels in the mesophyll cells result from the rate of ion uptake outpacing the expanding leaf's ability to sequester the ions within cells in the vascular system. As the ions accumulate, photosynthesis rates decrease as sodium toxicity damages photosystems (Demetriou et al., 2007) and disrupts chloroplast proton‐motive forces (Bose et al., 2017). To maintain mesophyll development under sodium toxicity, water transporters and cell wall modifiers are upregulated to allow cell expansion and may also lower sodium concentration (Munns & Tester, 2008). To prevent further damage to the mesophyll, light capture is offloaded to the large parenchyma and bundle sheath cells. This is consistent with hypotheses that the bundle sheath and parenchyma serve as stress‐protective buffers between the mesophyll and vasculature to mitigate rapid changes in water availability (Sage, 2001). Moderate salinity does not exhibit this response; instead, the reduced expansion rate induced by osmotic stress slows the uptake of water, allowing additional time for vascular cells to sequester and compartmentalize sodium ions into less sensitive cell types, such as fiber or bundle sheath cells. This response is more consistent with the typical onset of salinity stress, where plant growth and function is initially inhibited by osmotic stress, followed later by ionic stress (Munns & Tester, 2008).

Stress responses occur at the level of cells, but studying the responses of individual cells to stress using conventional methods will almost always underestimate the degree of perturbation. As single‐cell RNA‐sequencing data are sparse with small amplitude changes in transcript abundance, assessments of whole‐transcriptome perturbations are needed. This is especially true when studying long‐term and/or mild stress: while the onset of stress is typically characterized by a large number of DEGs with high fold changes, sustained stress returns close to baseline levels with smaller adjustments to transcript levels (Wilkins et al., 2016; Illouz‐Eliaz et al., 2025). Thus, using DEG numbers as a metric to quantify stress only accounts for a small subset of genes with high magnitude changes, and which excludes biologically relevant genes with small changes (Khang & Lau, 2015) and is heavily influenced by selected significance thresholds (Stupnikov et al., 2021). To study stress perturbation, assessment of the whole transcriptome, rather than DEGs, is needed. Many approaches for quantifying perturbation in single‐cell transcriptome data have been developed, including simple models based on frequentist statistics, mixed machine learning models, and deep learning or generative models (Gavriilidis et al., 2024). Here, we selected and combined tools that did not require any prior knowledge, and that quantified different metrics of the data, such as gene expression and position in high‐dimensional space. By combining the output of these models in an ensemble perturbation index, we took advantage of the different strengths of each model while mitigating the weaknesses. Ensemble approaches increase accuracy, reduce bias, and are more robust to noise than single models (Webb & Zheng, 2004; Saraswat et al., 2026). Our approach identified perturbed tissue systems and developmental stages that would not have been detected with any single method.

Overall, these results indicate that leaf development occurring in salt and drought stress conditions is the product of a complex interplay of developmental, environmental, and cell‐type contexts. Moreover, these data demonstrate that stressors in the environment lead to heterogeneous stress conditions between tissues and cell types. Perturbation of leaf development by stress is tractable to study using single‐cell technologies, which can help guide the engineering of climate change‐resilient crops by identifying the cell types and developmental processes that are most perturbed by environmental challenges. New methods of analysis and visualization tools are needed to expand the scope of biological questions and biological systems that can be explored using single‐cell sequencing technologies.

Competing interests

None declared.

Author contributions

SMR and OW contributed to the conceptualization and writing – original draft preparation. SMR, AP and NJP contributed to the data curation and software. SMR and CD contributed to the formal analysis, validation, and visualization. SMR, CD, CS and OM contributed to the investigation and methodology. NJP and OW contributed to the funding acquisition, project administration, resources, and supervision. SMR, CD, CS, NJP and OW contributed to the writing – review and editing.

Disclaimer

The New Phytologist Foundation remains neutral with regard to jurisdictional claims in maps and in any institutional affiliations.

Supporting information

Fig. S1 Days after planting until third leaf initiation and emergence.

Fig. S2 Lengths of cells in the leaf division zone.

Fig. S3 Chla + b content of third leaves developing in each condition.

Fig. S4 Ratio of Na⁺ : K⁺ content in tissues of seedlings developed in well‐watered and salinity conditions.

Fig. S5 Cross‐sections of third leaves developed in well‐watered and salinity conditions.

Fig. S6 Tissue system annotation and proportions.

Fig. S7 UMAPs of each replicate of each snRNA‐seq library.

Fig. S8 Marker gene expression within annotated cell types.

Fig. S9 UMAPs of Mesophyll and Epidermis tissue systems.

Fig. S10 Validated cell‐type markers, showing transcript abundance on UMAP, and in situ hybridization.

Fig. S11 Rice leaf cross‐section micrograph used as a reference image to trace cell borders for ggPlantMap.

Fig. S12 Number of stress‐induced DEGs shared across cell types within each stressor.

Fig. S13 Perturbation Index component: DEGs from DESeq2.

Fig. S14 Perturbation Index component: area under the curve (AUC) from Augur.

Fig. S15 Perturbation Index component: Distance from scDist.

Fig. S16 Perturbation Index component: perturbation score from MixScape.

Fig. S17 Expression of select stress‐responsive genes in meristem nuclei populations.

Fig. S18 Stress‐responsive regulons of all 101 TFs in the meristem regulatory network.

Fig. S19 Expression patterns of mesophyll DEGs in pseudotime bins between each stressor and control.

Fig. S20 Log₂ fold change of stress‐induced DEGs annotated with the ‘photosynthesis’ GO term in all cell types.

Methods S1 Extended description of methods used.

Table S1 Sequencing and alignment metrics for all snRNA‐seq libraries.

Table S2 Primers used for in situ hybridization and qRT‐PCR.

Table S3 Tissue system marker genes from other snRNA‐seq datasets.

Table S4 Cell‐type marker genes from the literature.

Table S5 Top 200 marker genes for each vascular cell type.

Table S6 Enriched GO terms for each vascular cell type, calculated from the top 200 markers.

Table S7 Differentially expressed genes in each cell type between each stressor and well‐watered control.

Table S8 Enriched GO terms for DEGs between each stressor and well‐watered control within each cell type.

Table S9 Perturbation metrics for each clustering resolution and condition used to calculate the Perturbation Index.

Table S10 Marker genes for cell cycle phases from the literature.

Table S11 Meristem nuclei population markers.

Table S12 All regulons predicted with MINI‐EX.

Table S13 Enriched transcription factor binding motifs in the promoters of all regulon targets in each meristem population.

Table S14 Differentially expressed genes in each mesophyll pseudotime bin between each stressor and well‐watered control.

Please note: Wiley is not responsible for the content or functionality of any Supporting Information supplied by the authors. Any queries (other than missing material) should be directed to the New Phytologist Central Office.

Data availability

Raw reads and processed count matrices are freely available as accession no.: GSE313726 on NCBI's Gene Expression Omnibus (Edgar et al., 2002). The plant map generated for this project is available for use in the ggPlantMap R package (https://github.com/leonardojo/ggPlantmap; Jo & Kajala, 2024). The code for generating multi‐panel plant maps with absolute or differential expression is available on GitHub so that users can explore any relevant data set (https://github.com/Wilkins‐Lab).