Multi-omics analysis reveals distinct spatial compartmentalization of lung repair niches in pediatric ARDS

Abstract

Background

Pediatric acute respiratory distress syndrome (PARDS), often triggered by viral infections, is a life-threatening condition. Despite its severity, children demonstrate significantly better survival rates and superior lung repair compared to adults. However, the mechanisms underlying this age-specific advantage remain incompletely understood.

Patients and methods

We conducted a pilot multi-omics study of influenza-associated PARDS integrating single-cell RNA sequencing (scRNA-seq) of pediatric lung tissue and bronchoalveolar lavage fluid (BALF), spatial transcriptomics, and plasma proteomics. Analyses were harmonized with the Human Lung Cell Atlas (HLCA) reference, reanalysis of public pediatric PARDS airway scRNA-seq, and contextual comparisons to adult lethal COVID-19 lung.

Results

Tissue scRNA-seq and spatial data indicated outcome-linked divergence in PARDS. Survivor showed spatially restricted repair with preserved alveolar type II (AT2) cells, AT2-to-alveolar type I (AT1) differentiation signatures, and higher KRT17, whereas fatal case and adults exhibited diffuse immune activation with pro-fibrotic and pro-apoptotic signaling. In BALF, KRT17-positive airway stress–repair epithelial cells (hillock-like) increased from the acute to recovery phase, and plasma proteomics showed higher circulating KRT17 in survivors. HLCA-based label transfer strengthened cell-type definitions and enabled pediatric–adult comparisons suggesting biological and developmental differences; the adult lethal COVID-19 atlas provided a benchmark with attenuated epithelial repair and prominent collagen CTHRC1-pathologic fibroblasts. Fibroblast programs were regionally compartmentalized, with injury-enriched CTHRC1⁺ states versus alveolar fibroblasts in preserved areas, and showed stronger injury–homeostasis anti-correlation in fatalities. Myeloid remodeling included BALF transitions from FCN1-high inflammatory states toward FABP4-positive resident-like states, consistent with public pediatric datasets showing reduced inflammatory and interferon-stimulated gene (ISG) modules and severity-linked increases in aged neutrophils.

Conclusions

This pilot multi-omics case series outlines putative pediatric lung repair niches in influenza-associated PARDS. KRT17-positive transitional epithelium, preserved AT2 differentiation, and restoration of resident-like macrophages may align with recovery, whereas diffuse immune activation and CTHRC1-enriched fibroblast programs may accompany worse outcomes. HLCA-guided annotations and adult benchmarks indicate possible age-related differences, warranting validation in larger multi-center cohorts.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-025-07588-8.

Introduction

Viral infections are a leading cause of pediatric acute respiratory distress syndrome (PARDS) and lung injury [1]. Viruses directly damage lung tissue, causing alveolar epithelial apoptosis and dysfunction, which disrupt the integrity of the alveolar-capillary barrier and impairs gas exchange [2]. The host immune response further amplifies injury, as pro-inflammatory cytokines trigger immune cell infiltration [1, 3]. Persistent inflammation can further induce fibrotic remodeling that compromises lung elasticity and function, creating a cycle of injury and impaired repair [4–6]. Despite these challenges, the lung possesses a remarkable capacity for repair, particularly in pediatric cases [7]. During recovery, stem and progenitor cells are activated to regenerate damaged tissue. Alveolar type 2 (AT2) cells, in particular, play a pivotal role in this process by self-renewing and differentiating into alveolar type 1 (AT1) cells, which are indispensable for restoring gas exchange [8, 9]. The extracellular matrix (ECM) provides structural and biochemical support for this regeneration by facilitating cell migration, proliferation, and differentiation [10]. Additionally, signaling pathways, including Wnt, Notch, and Hippo, influence cellular processes to promote tissue repair and restore homeostasis [11].

Notably, pediatric patients exhibit distinct outcomes compared to adults with ARDS, with higher survival and greater regenerative capacity [12]. These differences underscore the importance of understanding age-specific mechanisms of lung regeneration. Pediatric lungs may harbor distinct cellular and molecular features that confer enhanced repair potential, including differences in immune response, stem cell activity, and ECM composition [13]. However, the precise mechanisms underlying successful repair in pediatric cases remain poorly understood.

To fill this gap, we conducted an exploratory, pilot multi-omics case series focused on influenza-associated PARDS. We integrated single-cell and spatial transcriptomics of lung tissues from both survivors and fatal cases with confirmatory plasma proteomics and BALF single-cell analyses. To enhance interpretability and external context, we benchmarked cell states against the Human Lung Cell Atlas (HLCA) and reanalyze public pediatric PARDS airway single-cell datasets, and we included comparisons to adult lethal COVID-19 lung as a contextual reference. This work aims to generate testable hypotheses regarding the immune and reparative dynamics of epithelial, fibroblast, and macrophage lineages, and interaction networks that may be leveraged for precision interventions to improve PARDS outcomes.

Materials and methods

Human subjects and tissue acquisition

Lung tissues were obtained from two pediatric ARDS cases (10-year-old survivor [LIVE], 4-year-old fatality [DEATH]), six pediatric controls, and one adult ARDS case. The LIVE case improved after 50 days of ventilation and 14 days of extracorporeal membrane oxygenation (ECMO); the DEATH case, refractory to ECMO, died of shock. Biopsies underwent pathology, scRNA-seq, and 10x Visium. Controls were histologically normal tissues from six children undergoing elective thoracic surgery; all were lesion-free intraoperatively. Adult tissue was sampled from the right lower lobe margin of a previously healthy 70-year-old with viral ARDS requiring ECMO. For plasma proteomics, blood was collected from 13 elective pediatric controls and 20 PARDS patients (10 survivors, 10 non-survivors). BALF was collected at ECMO initiation (T1, n = 4) and weaning (T2, n = 4) in survivors. These data are not intended to represent fatal trajectories. Please refer to Supplementary Methods 1 for detailed clinical data.

Single-cell RNA-sequencing (scRNA-seq)

Fresh lung tissues were washed with 1× PBS and minced into approximately 0.2–0.6 mm³ fragments. The tissue fragments were then digested in a 10 mL enzyme cocktail containing Dispase (50 caseinolytic U/mL; Corning, Cat#: 354,235), collagenase (2 mg/mL), elastase (1 mg/mL), and DNase (30 μg/mL). Digestion was carried out at 37 °C for 1 hour with agitation every 10 min. The enzymatic reaction was quenched by adding 20 mL of PBS supplemented with 10% fetal calf serum (FCS). The resulting cell suspension was filtered through a 70-μm strainer and centrifuged at 300 × g for 10 minutes. The cell pellet was then treated with 5 mL of red blood cell lysis buffer for 2 min at room temperature to remove erythrocytes. The lysis was terminated by adding 20 mL of PBS with 10% FCS. After a final centrifugation step (300 × g for 10 min) and filtration through a 70-μm strainer, the cell pellet was resuspended in 1 mL of ice-cold PBS. Cell viability was assessed at > 75%.

BALF samples were filtered, centrifuged, and resuspended; viability was ensured above 85%. Cells were prepared as single-cell suspensions by gentle pipetting and fine-mesh filtering to prevent clumping. The final concentration of lung and BALF cells was adjusted to 7–12 × 10^5 cells/mL. Libraries were prepared using the 10x Genomics Chromium Next GEM Single Cell 5′ Kit v2 (PN-1000263) according to the manufacturer’s protocol [14].

scRNA-seq data analysis

Raw sequencing data were processed using Cell Ranger (v9.0.0) to align reads to the human reference genome (GRCh38) and generate gene-by-cell count matrices. Potential doublets were identified and removed using the DoubletFinder R package (v2.0.3) [15]. Subsequent analysis was conducted using the Seurat R package (v5.2.0, https://satijalab.org/seurat). For quality control, cells were retained if they expressed between 200 and 6,000 genes and had a mitochondrial gene content below 10%. Data were then normalized, and the top 2,000 highly variable genes were identified for dimensionality reduction. After scaling the data, principal component analysis (PCA) was performed. To correct for batch effects between samples, we applied the Harmony algorithm (v0.1.0) [16] to the PCA-reduced data. Shared nearest neighbor (SNN) graphs were built, clusters identified, and UMAP visualized. Major cell types were annotated by canonical markers, followed by iterative sub-clustering for subtype resolution.

Cell–cell communication analysis for scRNA-seq data

CellChat was employed to infer and analyze intercellular communication networks based on its ligand-receptor interaction database (http://www.cellchat.org/) )[17]. For each experimental group, the communication probability between cell types was calculated based on the expression of cognate ligand-receptor pairs. To identify significant interactions, we focused on ligands and receptors that were differentially expressed in each cell population (p < 0.05). To compare signaling across different conditions, the inferred communication networks were merged and contrasted to identify differentially regulated pathways and interaction strengths. Finally, network centrality analysis was employed to identify and rank the dominant signaling pathways. This allowed us to pinpoint the key sending and receiving cell populations that orchestrate communication within the tissue microenvironment.

Single-cell pseudotime analysis

Alveolar epithelial cells were converted to a Monocle2 CellDataSet; genes expressed in ≥ 3 cells were kept, and size factors/dispersions ersions estimated. Ordering genes combined highly variable epithelial genes with those differentially expressed among the three clusters (q < 0.01). DDRTree (max_components = 2) generated the trajectory, and orderCells rooted it in the surfactant-high AT2 cluster. Marker gene dynamics (AT2, AT2-to-AT1, AT1) were visualized across pseudotime.

Public reference mapping

We benchmarked our single-cell data to the Azimuth Human Lung Cell Atlas reference “human_lung_v2” [18]. Query objects were QC-filtered and harmonized to the reference using SCTransform with the provided reference SCT model. We computed transfer anchors, transferred labels at ann_level_3 and ann_finest_level, and obtained MappingScore and per-class prediction scores. Query embeddings were integrated to the reference, neighbors recalibrated (NNTransform), and cells projected onto the reference UMAP. We defined hlca_label as predicted.ann_finest_level and quantified uncertainty as 1 − max class probability; optional entropy was computed from full prediction vectors. For biological summaries, we aggregated per-sample/per-group proportions of HLCA states of interest and compared groups with two-sided Wilcoxon tests. Robustness was assessed via confusion matrices/ARI/NMI against internal labels and by module-score/marker heatmaps for curated gene panels. Please refer to Supplementary Methods 2 for detailed experimental procedures.

Public BALF scRNA-seq cohort, processing, subtyping, and module scoring

We analyzed 18 pediatric BALF scRNA-seq samples (14 PARDS timepoints from survivors and 4 controls) from the PARDS study by Flerlage et al. [19]. Grouping followed sample_metadata: ARDS Acute (P002_0, P003_0, P004_0, P005_0 [Moderate]; P008_0, P009_0 [Severe]; P012_0 [Moderate]); ARDS Recovery (P002_2, P003_2, P004_2, P008_2, P009_2 [Mild]; P005_2, P012_2 [No PARDS]); Controls: C001/C002/C003/C005. Matrices were merged into Seurat, normalized (vst 2,000 HVGs), scaled, PCA, and Harmony-integrated by sample; clustering (resolution 0.5) and UMAP used Harmony embeddings. Major lineages were assigned via module scores, then re-clustered: epi (Secretory, Ciliated/Deuterosomal, Hillock-like, SOX5+), MON_MAC (FCN1+ Inflammatory, SPP1_repair, FABP4+ Resident-like, ISG_high, Repair_like), and NEU (mature CXCR2hi, aged CXCR4hi, intermediate). ISG and mononuclear phagocytes (MNP) modules were scored. Proportions were summarized per sample and compared by group/severity (Wilcoxon/Kruskal–Wallis+Dunn; Spearman trends). Please refer to Supplementary Methods 3 for detailed experimental procedures.

Cross-dataset alignment to adult fatal COVID-19 lung

We reanalyzed the adult fatal COVID-19 lung atlas (Nature 2021; GSE171524) using the authors’ original fine cell-type definitions as reference [19]. Raw counts (GSM5226574–GSM5226600) and metadata were imported, barcodes harmonized, Seurat objects created with standard QC, log-normalized, and 3,000 HVGs selected. Without altering the reference labels, we performed label transfer (FindTransferAnchors/TransferData, dims 1–30) to map our pediatric PARDS data onto the adult label hierarchy, recording prediction scores and conducting reverse mapping for consistency. Analyses were restricted to shared genes. We visualized each dataset in its own UMAP, and used marker heatmaps/dot plots plus module scores for damage-associated transition progenitor (DATP)/pre-alveolar type-1 transitional cell state (PATS)-like epithelium (or variously called ADI or KRT8+ cells [20]) and CTHRC1+ fibroblast programs, with within-cohort nonparametric tests only. Please refer to Supplementary Methods 4 for detailed experimental procedures.

Spatial transcriptomics data collection pre-treatment

Formalin-fixed, paraffin-embedded (FFPE) tissue sections were placed on Visium Spatial Gene Expression slides. The sections first underwent hematoxylin and eosin (H&E) staining to generate high-resolution histological images, which served as a morphological reference map. Following imaging, the tissue was permeabilized to facilitate the release of mRNA molecules. These molecules hybridized to capture probes on the slide, each containing a unique spatial barcode. On-slide reverse transcription was then performed to synthesize spatially barcoded cDNA. The resulting cDNA was collected, amplified, and fragmented to construct Illumina-compatible sequencing libraries. These libraries were sequenced using a 150-bp paired-end configuration, with a target depth of at least 50,000 reads per tissue-covered spot. The raw sequencing data were processed using the Space Ranger software pipeline (v2.0, 10x Genomics). This involved aligning reads to the GRCh38 reference genome and subsequently mapping the transcriptomic data back to their original spatial coordinates on the H&E image. This process generated a matrix that links gene expression profiles to distinct locations within the intact tissue architecture. The final output enabled the visualization and quantitative analysis of whole-transcriptome data in its native morphological context, with each data point corresponding to a tissue spot of approximately 55 µm in diameter.

Spatial transcriptomic data analysis

We processed Visium gene-by-spot matrices in Seurat v5. Spots with more than 200 detected genes were kept, while low-abundance genes (fewer than 10 counts or present in fewer than three spots) were removed. Data were normalized with SCTransform to reduce depth-related artifacts. We summarized spot counts and basic QC metrics, then aligned H&E images using the 10x Visium workflow with manual verification. No notable batch effects were seen based on UMAP and concordance with morphology. To profile spatial biology, we computed enrichment scores for IFN-I, IFN-II, apoptotic, and inflammatory pathways using AddModuleScore with MSigDB sets and visualized them with SpatialFeaturePlot. We ran PCA, constructed an SNN graph, and clustered at resolution 0.8; clusters were reviewed by a pathologist and mapped to 10 lung-injury regions, then displayed by UMAP. We extracted KRT17, SFTPC, and AGER, applied global Min-Max normalization, and derived combined scores. Region-level ALV_normal and ALV_injury signatures from normalized cell subtype proportions were averaged by region and group and visualized as a heatmap. Spatial transcriptomics workflow establishing SCT or Spatial assay, min–max normalizing WT1 and SPP1 to compute WT1_SPP1_joint, extracting SPP1 macrophage proportions from celltypeprops, aligning spot coordinates, estimating Hedges g and Cliff’s delta with 2000 bootstrap BCa CIs, testing Moran’s I via kNN weights, and visualizing forest-style effect sizes. Please refer to Supplementary Methods 5 for detailed experimental procedures.

Spatial deconvolution using RcppML

We integrated single-cell RNA sequencing (scRNA-seq) data with spatial transcriptomics (ST) data using the RcppML package to perform non-negative least squares (NNLS) deconvolution (https://github.com/zdebruine/RcppML). First, we normalized and identified variable features in both datasets, followed by dimensionality reduction (PCA and UMAP) for the scRNA-seq data. Using the RunNNLS function, we deconvoluted the ST data by leveraging scRNA-seq-derived cell type expression profiles, specifying the cell type labels and a minimum cell count threshold. The resulting cell type proportions were stored in a new assay (celltypeprops) and visualized using SpatialFeaturePlot to map the spatial distribution of selected cell types. The results were saved for further downstream analysis.

RCTD deconvolution and spatial analyses

We performed cell type deconvolution with spacexr (RCTD v2.2.0) [21]. Spatial data was aligned to tissue images and coordinates; the scRNA-seq reference provided celltype2 labels, retaining cell types with ≥50 cells. Counts from total_sce built the RCTD Reference. For each sample, spots with both expression and coordinates were used; Spatial counts were filtered for all-zero genes and aligned (x, y). RCTD ran with doublet_mode=“doublet” and max_cores = 1–4. From weights, we computed entropy, max weight, and Simpson index; hard labels required max weight ≥0.2. SpatialFeaturePlot visualized key genes. Cohen’s d with 1000 bootstraps compared LIVE vs DEATH. ROI scores, including Epithelial_Repair (KRT17, SFTPC, AGER), Inflammatory (SPP1, CHI3L1, MMP9), Regenerative (FABP4, MARCO), and Fibrotic (CTHRC1, COL1A1, FN1), were contrasted via KS tests. NNLS-derived proportions were correlated with RCTD weights to assess concordance. Please refer to Supplementary Methods 6 for detailed experimental procedures.

Cell–cell communication analysis for spatial transcriptomics data

Cell-cell communication analysis was performed using CellChat v2.1.0 to investigate spatially resolved interactions in lung tissures of LIVE and DEATH [22]. Cell type labels were assigned to each spot by selecting the cell type with the highest predicted probability from Seurat object metadata. Normalized gene expression matrices were extracted from the ‘SCT’ assay, and common genes across both samples were retained. Sample-specific prefixes were added to column names prior to merging expression matrices. Metadata tables were created containing cell type labels and sample identifiers, with both factors explicitly defined to ensure consistent grouping. Spatial coordinates were obtained from full-resolution tissue positions and converted to micrometers using scale factors derived from the ratio of the theoretical spot size (65 μm) to the spot diameter in pixels. A unified CellChat object was initialized using the merged expression data, metadata, spatial coordinates, and scale factors, specifying the analysis as spatial transcriptomics. The Secreted Signaling subset of CellChatDB.human was selected for ligand-receptor interaction analysis. After subsetting the data to signaling genes, over-expressed ligands/receptors and their interactions were identified using parallel computing. Cell-cell communication probabilities were computed with spatial constraints, modeling diffusible signals within a 250 μm range and contact-dependent signaling within 100 μm, while applying a truncated mean (trim = 0.1) to mitigate outlier effects. Interactions involving cell groups with fewer than 10 cells were filtered out to ensure robustness.

Mass SpectrometryAssay for data independent acquisition (DIA)

Plasma (2 μL) was diluted in 50 mM ammonium bicarbonate, heat-inactivated (95 °C, 3 min), and digested overnight with trypsin (1:25, 37°C). Peptides were dried, reconstituted in 0.1% formic acid, trapped (2 cm, 100 μm ID, 3 μm C18) and separated (15 cm, 150 μm ID, 1.9 μm C18) on an EASY-nLC 1200 coupled to a Q Exactive HF-X. DIA was acquired with MS1 300–1400 m/z (60k resolution, AGC 4e5, 50 ms) and 30 DIA windows (15k, Automatic gain control 1e6/Auto), Higher-energy collisional dissociation Normalized collision energy 27%, profile mode, default MS2 charge 3. Raw abundance matrices were processed in R (numeric coercion; non-numeric to missing), samples identified by S/L-prefixed IDs, and grouped as DEATH, LIVE, or Control (CC). Three-group differences used Kruskal–Wallis with Dunn–Holm; visualization used log10 groupwise medians, ranking, and generalized additive model smoothing. Prognostic analyses focused on DEATH vs LIVE. For C1QB, C1QC, CFD, IFNGR1, ORM1, CRP, TNFRSF10B, TNFRSF1A, IL6, IL1R2, TNFSF10, LTF, KRT17, SFTPC, SFTPD, AGER, CAV1, KRT5, MUC5AC, covariates (age, ECMO, SOFA) were merged and per-gene linear effects removed via regression residuals; groups were compared by Welch’s t-test with FDR across genes, with ANCOVA as robustness. Discrimination was evaluated using ROC analysis, reporting AUC with 95% confidence intervals by DeLong’s method and two-sided tests against an AUC of 0.5. Multivariable logistic regression adjusted for age, ECMO, and SOFA was used to estimate odds ratios with 95% confidence intervals from profile likelihood; p values were obtained from Wald tests. ECMO was modeled as a binary indicator. Please refer to Supplementary Methods 7 for detailed experimental procedures.

Immunofluorescence

Paraffin lung sections were fixed in 4% PFA, deparaffinized, and subjected to heat-induced epitope retrieval (10 mM sodium citrate, pH 6.0, 95 °C, 15 min). After blocking/permeabilization, sections were incubated overnight at 4 °C with anti-WT1, followed by Alexa Fluor secondary (1:500) and DAPI, then mounted and imaged by confocal microscopy with identical settings. Multiplex TSA immunofluorescence was performed on FFPE sections from LIVE, DEATH PARDS, and adult ARDS cases using SFTPC, KRT17, KRT5, and AGER primaries. HRP-conjugated secondary antibody and tyramide deposition enabled sequential rounds with distinct fluorophores, including peroxidase quenching, blocking, retrieval as needed, and DAPI counterstaining. Please refer to Supplementary Method 8 for detailed experimental procedures.

Enzyme-linked immunosorbent assay

Human plasma concentrations of KRT17 were measured using a commercially available ELISA kit (Jianglai Biotech, catalog# JL15259) according to the manufacturer’s instructions. Plasma samples were centrifuged to remove particulates and diluted appropriately to fall within the detection range. After incubation with sample/standards, plates were washed and sequentially incubated with biotinylated detection antibody and streptavidin-horseradish peroxidase (HRP). Color development was achieved by adding tetramethylbenzidine (TMB) substrate, and the reaction was stopped with sulfuric acid. Absorbance was read at 450 nm using a microplate reader, and KRT17 concentrations were calculated by interpolating sample optical densities against the standard curve.

Statistical analyses

All analyses were conducted in R (v4.4.0). Unless otherwise specified, tests were two-sided and p values were adjusted for multiple comparisons using the Benjamini–Hochberg method. Continuous variables were summarized as mean ± SEM or median (IQR) as appropriate. Group comparisons used Wilcoxon rank-sum or Kruskal–Wallis tests for continuous data and Fisher’s exact test for categorical data; paired data (e.g., T1 vs T2 within survivors) used Wilcoxon signed-rank tests.

Results

Spatial and cellular signatures of divergent lung repair outcomes in PARDS

We integrated scRNA-seq from six healthy pediatric lungs, one surviving PARDS case and one fatal PARDS case, together with spatial transcriptomics from the two PARDS cases (Fig. 1A), to delineate molecular and spatial programs that either promote repair or intensify injury. Visium data showed robust quality in both samples, with 4992 spots in LIVE and 4927 in DEATH, alongside strong per-spot median gene and UMI counts and exclusion of spots with fewer than 500 detected genes (Supplementary Table 1, Fig. 1B). Differential expression analysis revealed that the DEATH case was characterized by upregulated viral and LPS response pathways, heightened IFN-I regulation, and enhanced cytokine production (Fig. 1C). These features were consistent with widespread immune activation and inflammatory responses, which likely contributed to diffuse tissue damage and impaired repair. In contrast, the LIVE case exhibited transcriptional programs favoring vasculature development, hormone response, and TGFβ signaling, which are critical for tissue regeneration and homeostasis (Fig. 1C). Dimensionality reduction further revealed ten spatial regions with distinct pathological and molecular features (Fig. 1D–F), highlighting spatial compartmentalization of repair signatures. Notably, apoptosis, inflammation, and interferon signaling scores were diffusely elevated across all spatial regions in the DEATH case, indicating a failure to localize repair programs (Fig. 1G). Conversely, the LIVE case demonstrated localized regenerative signatures, including enhanced epithelial repair, mesothelium proliferation, and vasculature integrity. The LIVE case also exhibited higher proportions of diffuse damage regions, while the DEATH case was dominated by fibroblast-expanding regions (Fig. 1H). These observations were further corroborated by spatial analysis of apoptosis, IFN, and inflammation scores, which showed consistently elevated levels across all regions in the DEATH case (Fig. 1I). Together, these findings emphasized the importance of spatially restricted repair programs and cellular interactions in determining the trajectory of lung recovery or progression to fibrosis and organ failure in PARDS.

Fig. 1.

Multi-omics spatial lung atlas reveals pathological regions and molecular features in LIVE and DEATH ARDS cases (A) Experimental design of the multi‑omics spatial lung atlas, including fresh/frozen samplings for single‑cell RNA sequencing (scRNA‑seq; six controls, one live case, one death case) and Visium spatial transcriptomics (st; one live case, one death case). (B) Violin plots showing higher genes‑per‑spot and UMI counts in LIVE versus DEATH; spatial feature maps of nFeature_spatial highlight signal‑dense regions in LIVE and DEATH. (C) Gene Ontology (GO) enrichment analysis of differentially expressed genes (DEGs) in ST comparing LIVE and DEATH across entire tissue sections. (D) Sequential clustering of ST data identifies 10 pathological regions visualized on UMAP. (E) Spatial projection of cell types onto hematoxylin and eosin (H&E) images of lung tissues. (F) Top five contributing DEGs per region assisting pathological region identification; bar heights indicate relative contribution. (G) Spatial feature plots of signature scores for IFNG, IFNA, apoptosis, and inflammation derived from curated gene sets. (H) Proportional composition of each pathological region in LIVE and DEATH. (I) Region‑level scores for apoptosis, IFNA, IFNG, and inflammation across LIVE and DEATH

Single-cell transcriptomics-assisted analysis of regional cellular composition in spatial transcriptomics

Using single-cell RNA sequencing (scRNA-seq) of lung tissues from controls and PARDS cases, we identified 39 distinct cell subtypes (Fig. 2A–B, Figure S1A). We mapped our scRNA-seq dataset to the Azimuth HLCA human_lung_v2 reference and projected cells onto the reference UMAP (Figure S1B). Transferred labels showed strong concordance with our curated annotations, supported by confusion matrices and external agreement metrics (Fig. 2C). Label certainty was high for dominant epithelial, fibroblast, macrophage, B and T cell lineages (Figure S1C-E). FABP4+ alveolar-resident macrophages aligned to “Alveolar macrophages”, SPP1+ macrophages aligned to “Monocyte-derived macrophages”, lipofibroblasts to “Myofibroblasts/Mesothelium”, and AT2-to-AT1 aligned to “AT2/AT1/pre-TB secretoy” (Fig. 2C). Importantly, the key states used in our downstream analyses—captured by the module scores for “CTHRC1_Fib”, “Alveolar_Fib”, “FABP4_Macrophage”, “KRT17+stress-repair_Epi” and “SPP1_Macrophage”—were faithfully recapitulated in the HLCA space (Fig. 2D; Figure S1F–G). Module-score genes heatmaps for curated marker panels further confirmed expected separation of these states (Figure S1H). HLCA mapping provides an external reference that supports the validity of our cell definitions used for spatial deconvolution and pathway analyses. CellChat analysis revealed that DEATH exhibited upregulated inflammatory and apoptotic signaling pathways, including TWEAK, TNF, FASLG, COMPLEMENT, VEGF, HGF, and CXCL, which likely exacerbated tissue damage and impaired repair. In contrast, LIVE demonstrated activation of immunomodulatory and anti-inflammatory pathways, such as CD40, BTLA, and IL10, which are essential for promoting tissue regeneration and resolving inflammation (Fig. 2E). Integration with spatial transcriptomics further revealed the heterogeneity of repair programs in PARDS. In the LIVE case, the diffuse damage zones contained higher proportions of AT2 cells, while AT1/AT2_AT1 cells were localized to bleeding areas, and bronchial regions were enriched with club, ciliated, and goblet cells. Conversely, in the DEATH case, the diffuse damage regions were enriched with various T cell, neutrophil, and monocyte subtypes, indicating a robust local immune response. In the fibrotic-expanding regions of survivor lung, an enrichment of classical monocytes, B cells, GZMB+ CD8 T cells, and FOLR2+ macrophages was observed. The macrophage-enriched areas in LIVE showed an abundance of both FABP4+ and FCN1+ macrophages, whereas these regions in DEATH were predominantly populated by FCN1+ macrophages. Furthermore, a unique air-leak region observed only in LIVE was enriched with mesothelial cells, lipofibroblasts, and SPP1+ macrophages (Fig. 2F). Concordance analysis demonstrated strong Pearson correlations between RCTD- and NNLS-derived cell-type proportions across shared spots, supporting robustness of the deconvolution-based annotations (Figure S2A-B). RCTD deconvolution of the two spatial sections revealed marked compartmentalization of epithelial program. Effect-size analysis (Cohen’s d, 1000 bootstraps) supported these trends: KRT17 and SFTPC were elevated in LIVE (Figure S2C). Program-level ROI scores confirmed spatially restricted Epithelial_Repair/Regenerative signals in LIVE (Figure S2D).

Fig. 2.

Single-cell transcriptomics and cellular communication analysis reveal cell type composition and interactions in LIVE and DEATH ARDS cases (A) UMAP embedding of single cells from eight samples (six controls, one live, one death), annotated into 39 cell types by canonical markers. (B) Bar plots showing cell‑type compositions per group. (C) Confusion heatmap (row‑normalized) comparing study‑defined cell types (rows) to Human lung cell atlas (HLCA) predictions (columns); darker tiles indicate stronger correspondence. (D) UMAP feature plots of module scores for five gene sets—CTHRC1 fibroblasts (Score_CTHRC1_Fib), alveolar fibroblasts (Score_alveolar_fib), FABP4 alveolar macrophages (Score_FABP4_AM), KRT17+ stress‑repair epithelium (Score_KRT17_stress‑repair_epi), and monocyte‑derived SPP1 macrophages (Score_MDM_SPP1)—colored low (blue) to high (red). (E) Comparison of relative information flows between LIVE and DEATH inferred by CellChat in scRNA‑seq. (F) Heatmap of predicted cell‑type proportions per pathological region by non‑negative least squares (NNLS) deconvolution

Survivor lung tissue exhibits an active epithelial reparative niche

Single-cell pseudotime and spatial analyses revealed that AT2 cells differentiate toward AT1 via a transitional AT2-to-AT1 state marked by a sharp, transient surge of KRT17. Along this trajectory, AT2 surfactant genes declined, the narrowly peaked and higher-amplitude KRT17 signal precedes sustained induction of AT1 barrier/structural genes (AGER, CAV1), delineating stage-specific transcriptional programs (Figure S3A-D). In spatial transcriptomics, the LIVE case showed elevated proportions of AT1, AT2, and AT2-to-AT1 transitional cells, which were confirmed by higher epithelial marker levels (Fig. 3A; Figure S3E; Figure S4A-B). KRT17 co-expressed with SFTPC in LIVE (Fig. 3B–C). Label transfer from an adult COVID-19 lung atlas aligned the major compartments in our data, with strong diagonal agreement between pediatric annotations and adult fine labels, enabling harmonized comparisons without hard integration (Figure S5A–C). Pediatric AT2-to-AT1 cells defined here mapped within the reference-defined AT2 and AT1 compartments. We found that atlas-defined alveolar macrophages broadly overlapped with FABP4, FCN1, FOLR2 macrophages and monocytes in our PARDS dataset, suggesting monocyte-derived macrophages constitute the major component of alveolar macrophages in adults (Figure S5C). In adults, DATP/PATS-like scores were highest in AT2, consistent with KRT8/CLDN4/CDKN1A states (Figure S5D–F). After mapping, pediatric PARDS exhibited a similar pattern of enrichment. In children, KRT17 localized to AT2 cells rather than basal cells, supporting a role in pediatric alveolar repair. By contrast, in adults KRT17 was predominantly expressed in basal and other airway epithelial cells, with minimal expression in AT1 and AT2 (Figure S5G-H). Using four-color immunofluorescence co-localization for KRT5, KRT17, SFTPC, and AGER on lung tissues from pediatric survivors, pediatric non-survivors with ARDS, and adults with ARDS, we confirmed that KRT17 is highly expressed in AT2 cells in children. In adults, KRT17 was highly expressed in KRT5+ bronchial epithelial cells, indicating aberrant repair within the alveolar region (Fig. 3C). Spatial CellChat analysis revealed DEATH exhibited enhanced cross-talks between club cells areas and other epithelial, lipofibroblast, and myofibroblast regions, suggesting a dysregulated repair environment. Conversely, LIVE displayed CLUBTOAT2 and AT2-dominated communication niches, which are indicative of localized repair programs favoring epithelial regeneration (Figure S3F-G). Together, these findings underscore the critical roles of cellular heterogeneity, spatial compartmentalization, and intercellular communication in determining repair outcomes in PARDS, providing valuable insights into therapeutic strategies for promoting lung regeneration.

Fig. 3.

Spatial mapping of key epithelial cell types and genes in PARDS (A) Spatially deconvolved proportion maps (Visium ST) for alveolar type I (AT1), alveolar type ii (AT2), transitional AT2-to-AT1 cells (putative AT2→AT1 intermediates) and DATP/PATS score. (B) Composite colocalization map showing spatial coexpression of SFTPC, AGER, and KRT17, highlighting transitional/remodeling niches Data are displayed as mean±SEM. ***p < 0.001 (C) Multiplex TSA immunofluorescence to colocalize KRT17 with SFTPC, AGER and KRT5 in lung tissues from PARDS LIVE, PARDS DEATH and adult ARDS case. Scale bars, 200 and 25 μm (magnified image)

KRT17 as a candidate prognostic biomarker differentiating survival outcomes in PARDS

To further verify whether biomarkers of lung epithelial injury could be monitored in peripheral blood and differentiate between LIVE and DEATH PARDS groups, we conducted proteomic analyses on peripheral plasma samples collected from 13 children in the control group, 10 surviving PARDS patients, and 10 deceased PARDS patients (Fig. 4A). There were statistically significant differences in platelet and LDH levels between surviving and deceased children during the acute phase of ARDS (Table 1; Supplementary Table 2). PCA demonstrated significant discrimination among all groups (Fig. 4B). We quantified an average of 3725 (Control), 4057 (PARDS-LIVE) and 4123 (PARDS-DEATH) proteins per plasma sample (Figure S6A). The heatmap of differentially expressed genes revealed that the DEATH group exhibited a greater number of upregulated proteins compared to the LIVE group (Figure S6B). Across the three groups, nonparametric comparisons indicated distributional differences for a subset of proteins, and ranked, log10-transformed median abundance profiles with GAM smoothing illustrated modest shifts in the DEATH group over specific rank ranges (Figure S6C). Gene Ontology (GO) analysis showed that the upregulated proteins in the DEATH group were primarily enriched in biological processes such as apoptosis regulation, immune response, and cytokine production. In contrast, the proteins upregulated in the LIVE group were mainly associated with pathways involved in epithelial differentiation, WNT signaling, lymphocyte immune response, and progenitor cell differentiation (Fig. 4C). Focusing on prognosis, we compared LIVE and DEATH after merging clinical covariates and linearly adjusting for age, ECMO, and SOFA score. We found higher complement activation (C1QB/C, CFD), acute inflammatory proteins (ORM1, CRP), TNF families (TNFRSF10B, TNFRSF1A, TNFSF10) and cytokine signaling (IFNGR1, IL6, IL1R2) in DEATH (Figure S6D), and covariate-adjusted residuals with boxplots and ROC supported consistent elevations (Figure S6E), indicating inflammation/complement dysregulation linked to mortality. Faceted boxplots of adjusted expression showed that most epithelium-repair related proteins were not significant while KRT17 remained significant, consistent with Welch t-tests with FDR correction across seven prespecified markers including KRT17, SFTPC, SFTPD, AGER, CAV1, KRT5, and MUC5AC (Fig. 4D). Discrimination analyses showed modest to moderate AUCs overall with KRT17 performing best, with DeLong 95% confidence intervals (Fig. 4E, Supplementary Table 3). In multivariable logistic regression adjusted for age, ECMO, and SOFA, KRT17 retained an independent association with outcome with odds ratios whose confidence intervals did not cross one, whereas most other proteins were not significant (Figure S6F). We further confirmed a significant reduction of KRT17 in the plasma of deceased PARDS patients by ELISA detection (Figure S7).

Fig. 4.

Plasma proteomics revealed that KRT17 levels were significantly higher in surviving PARDS patients than deceased cases (A) Flowchart of peripheral blood plasma proteomic profiling in control children and children with pediatric ARDS (PARDS). (B) Principal component analysis (PCA) of quantified proteins across all plasma samples; points color‑coded by group to visualize separation and batch effects. (C) GO annotations for upregulated proteins in survival versus deceased PARDS (p < 0.05). (D) Adjusted expression of epithelial markers across outcome groups; box‑and‑jitter plots show residuals from linear models adjusted for age, extracorporeal membrane oxygenation (ECMO), and Sequential Organ failure Assessment (sofa) score at ICU admission for AGER, CAV1, KRT17, KRT5, MUC5AC, SFTPC, and SFTPD. (E) Receiver operating characteristic (ROC) curves comparing LIVE versus DEATH for seven epithelial markers. Data are displayed as mean±SEM. **p < 0.01

Table 1.

Comparison of baseline information between deceased and surviving PARDS patients

		Group
		Live	Death
n		10	10	p
Sex (%)	F	5 (50.0)	2 (20.0)	0.35
	M	5 (50.0)	8 (80.0)
Pathogens_recovered (%)	CMV	1 (10.0)	1 (10.0)	0.853
	Flu.A	5 (50.0)	4 (40.0)
	Flu.A/Flu.B	1 (10.0)	0 (0.0)
	Flu.B	1 (10.0)	0 (0.0)
	RSV	1 (10.0)	3 (30.0)
	SARS-CoV-2	1 (10.0)	2 (20.0)
ECMO_(yes/no) (%)	no	5 (50.0)	6 (60.0)	1
	yes	5 (50.0)	4 (40.0)
Age (median [IQR])		6.50 [4.00, 8.00]	4.88 [4.00, 8.25]	0.79
BMI (median [IQR])		14.70 [14.27, 15.83]	14.25 [13.62, 16.52]	0.733
SOFA (median [IQR])		14.50 [13.00, 17.50]	15.50 [14.25, 16.75]	0.592
Leukocytes_(10^3/ul) (median [IQR])		13.40 [11.83, 17.75]	18.42 [11.95, 34.06]	0.406
Lymphocytes_(%)		8.9 [7.0–14.0]	14.0 [6.5–24.4]	0.51974
Granulocytes_(%)		87.8 [82.1–89.7]	77.8 [71.9–87.2]	0.18555
Platelets_(10^3/ul) (median [IQR])		287.10 [231.75, 353.25]	122.70 [102.50, 247.10]	0.016
Procalcitonin_(ng/ml) (median [IQR])		0.46 [0.11, 2.08]	0.30 [0.11, 0.46]	0.344
CRP_(mg/l) (median [IQR])		69.90 [41.17, 150.25]	100.90 [39.64, 127.79]	0.821
LDH_(U/l) (median [IQR])		439.00 [353.00, 514.92]	596.00 [510.50, 726.25]	0.049
IL_6_(ng/l) (median [IQR])		84.95 [62.01, 136.70]	196.00 [92.30, 260.62]	0.112

Abbreviations: F, female; M, male; CMV, cytomegalovirus; Flu.A, influenza A virus; Flu.B, influenza B virus; RSV, respiratory syncytial virus; ECMO, extracorporeal membrane oxygenation; IQR, interquartile range; BMI, body mass index; SOFA, Sequential Organ Failure Assessment; CRP, C-reactive protein; LDH, lactate dehydrogenase; IL-6, interleukin-6

Fibroblast heterogeneity and spatial dynamics dictate divergent repair outcomes in PARDS

Our findings underscored the pivotal role of fibroblasts in determining repair outcomes in PARDS. Fibrosis-related genes exhibited distinct expression patterns across spatially defined regions. In the DEATH case, fibrosis-associated genes such as COL1A1, COL3A1, CTHRC1, TNC and TIMP1 were upregulated in regions characterized by alveolar enlargement (Alv_enlarged), bleeding, diffuse damage, fibroblast expansion (Fib_expanding), and macrophage-enriched zones (Mac_enriched) (Fig. 5A). In contrast, the LIVE case demonstrated upregulation of genes promoting matrix remodeling and tissue repair, including MMP2, MMP14, and POSTN, especially in vascular-associated regions. Notably, SFRP1 and SFRP2 were predominantly expressed in vascular areas, including smooth muscle cell (SMC) and vessel zones, suggesting their role in maintaining vascular homeostasis (Fig. 5A). Single-cell clustering further identified six distinct fibroblast subtypes, revealing functional and spatial heterogeneity (Fig. 5B–C). Spatial transcriptomics demonstrated that the LIVE case was enriched with alveolar fibroblasts (alv_Fib) in bleeding, SMC, Alv_enlarged, and vessel zones, indicative of localized repair programs supporting epithelial and vascular regeneration. Conversely, the DEATH case accumulated lipofibroblasts (lip_Fib) in these regions as well as macrophage-enriched zones, reflecting a dysregulated fibrotic response that likely contributed to impaired repair and tissue remodeling (Fig. 5D–E). Interestingly, pleural leak sites in the LIVE case showed mesothelium accumulation with elevated expression of MMP14, TIMP1, COL3A1, and APOE, highlighting the role of mesothelial cells in sealing pleural leaks and promoting localized repair (Fig. 5F). RCTD deconvolution aligned with these observations, revealing diffuse lipofibroblast signals in the DEATH case (Figure S2B). Effect-size analysis supported SPP1, and CTHRC1 elevation in DEATH (Figure S2C). Program-level ROI scores confirmed spatially restricted Fibrotic/Inflammatory programs in DEATH, with KS tests indicating significant distributional shifts (Figure S2D, Supplementary Table 4). Cross-cohort analyses in adult fatal COVID-19 identified a robust CTHRC1-positive pathological fibroblast program marked by high-expression of CTHRC1 (Figure S8A-C). Pediatric mapped fibroblasts showed similar CTHRC1+ increases in pathological states versus alveolar/adventitial types, with our dataset’s self-defined lipofibroblasts predominantly residing within the pathological fibroblast compartment. (Figure S5C, S8D-F). CTHRC1+ pathological fibroblast scores increased while alveolar fibroblast score decreased in PARDS (Figure S8G-I). Pathological and intermediate pathological fibroblasts were prominent in adults and showed parallel but smaller increases in pediatric intermediate or pathologic fibroblast states (Figure S8J-K). Cross-cohort conclusions are descriptive due to age/etiology differences, yet directionally concordant, providing insights into fibroblast-mediated lung recovery or fibrosis in PARDS.

Fig. 5.

Fibroblast heterogeneity and spatial distribution in lung repair versus fibrotic remodeling in PARDS. (A) Heatmap of fibroblast‑related lung repair and fibrosis genes across pathological regions. (B) UMAP embedding of fibroblast‑lineage cells resolved into six transcriptional subtypes. (C) Expression of typical marker genes per fibroblast subtype. (D) Box plots comparing predicted proportions of the six fibroblast phenotypes between LIVE and DEATH across regions. (E) Spatial deconvolution maps showing projected proportions of the six fibroblast subtypes across sections. (F) Spatial feature maps of extracellular matrix (ECM), remodeling, proliferative, and lipid‑associated fibroblast genes (POSTN, SFRP2, TNC, MMP2, CTHRC1, COL3A1, MMP14, MMP9, TIMP1, MKI67, COL1A1, APOE), highlighting localized fibrotic/reparative niches

Macrophage subtypes and spatial interactions define divergent immune responses in PARDS

Macrophages, the dominant cell type identified in our scRNA-seq analysis, were classified into five distinct subtypes, which highlighted their functional diversity (Fig. 6A–B). Spatial analysis revealed notable differences in macrophage subtype distribution between DEATH and LIVE cases. Deconvolution of LIVE showed enrichment for FABP4+ macrophages, and DEATH exhibited diffuse FCN1 and SPP1+ macrophage signals in macrophage-enriched zones. Effect-size analysis (Cohen’s d, 1000 bootstraps) supported FABP4 elevation in LIVE, while SPP1 favored DEATH (Figure S2B-D, Supplementary Table 4), as were proved by AM-resident scores and MDM-fibrotic scores (Fig. 6C–D). Interestingly, in LIVE, SPP1_MAC specifically localized to pleural injury sites, which aligned with the observed mesothelial repair responses in these regions (Fig. 6E–F). Spatial CellChat analysis provided further insights into the intercellular communication dynamics. In DEATH, lipofibroblasts exhibited more active interactions with epithelial cell regions, potentially contributing to a dysregulated repair environment. However, in both DEATH and LIVE cases, monocytes and macrophages maintained cell communication strictly within their own niches, reflecting a conserved spatial organization of immune responses (Fig. 6G).

Fig. 6.

Macrophage subtype dynamics, spatial distribution, and cell-cell communication networks in LIVE and DEATH ARDS cases (A) UMAP distribution of five macrophage subtypes. (B) Expression of typical marker genes per macrophage subtype. (C) Box plots comparing alveolar‑resident macrophage scores and monocyte‑derived fibrotic macrophage scores by group. (D) Box‑and‑jitter plots showing region‑level expression scores of macrophage subtypes (FABP4‑MAC, FCN1‑MAC, FOLR2‑MAC, Prolif.MAC, SPP1‑MAC) across regions, comparing death (red) and live (teal). (E) Projected proportions of predicted macrophage subtypes in ST. (F) Spatial feature plots of SPP1, FOLR2, CD14, and FABP4 expression in tissue sections. (G–H) Circle plots of cell–cell communication networks in death (G) and live (H); right‑side legends list pathway names

Dynamic shift of BALF immune and epithelial subtypes in PARDS

To elucidate the dynamic changes in cell subtypes during the acute and repair phases of PARDS, we performed scRNA-seq on BALF samples collected at T1 and T2 time points from surviving pediatric patients. Based on classical markers, we classified the cells into six major groups: neutrophils, macrophages, T cells, dendritic cells (DCs), epithelial cells, and B cells (Fig. 7A, Figure S9A). At T2, the proportion of neutrophils significantly decreased, whereas the proportions of macrophages and T cells markedly increased (Fig. 7B). Focusing on the epithelial composition in BALF, epithelial cells were further subdivided into three subtypes: ciliated cells, KRT17_epi cells, and mucus cells (Fig. 7C, Figure S9B-C). At T1, mucus cells constituted the majority of epithelial cells, while at T2, the proportion of KRT17_epi cells increased significantly (Fig. 7D). GO enrichment analysis of differentially expressed genes in KRT17_epi cells revealed that these cells were in a highly dynamic stress state, characterized by autophagy activation and tissue repair processes (Fig. 7E). This finding highlighted the critical role of KRT17_epi cells during the repair phase. Macrophages were then classified into three subtypes: FABP4_MAC, FCN1_MAC, and SPP1_MAC (Fig. 7F, Figure S9D-E). At T1, the macrophages in BALF were predominantly FCN1_MAC, which were highly active in inflammatory responses (Figure S9F). By T2, FABP4_MAC replaced FCN1_MAC as the dominant subtype (Fig. 7G, Figure S9G). SPP1_MAC was present in relatively low abundance in BALF, and FOLR2_MAC was not detected, suggesting the absence of interstitial macrophages in BALF. To validate our findings, we reanalyzed the study by Flerlage et al., confirming consistency across datasets. Harmony-integrated scRNA-seq of BALF identified an immune-dominant composition (NEU, MNP, NK/T, B, DC) with a minor epithelial fraction (Figure S10A). Epithelial cells included secretory, ciliated, deuterosomal, hillock-like, and SOX5+ transitional types. KRT17 expression was predominantly localized to hillock-like cells rather than other epithelial lineages (Figure S10B)—emphasizing the transcriptional similarity to our KRT17_epi gene signature (Figure S9B)—and was higher within hillock-like cells from patients in the recovery phase or with no/mild ARDS (Figure S10C–D), as was reflected by KRT17/PATS score (Figure S10E). Consistently, the proportion of hillock-like cells showed a modest increase during the recovery phase overall (Figure S10F-G). Across severity strata, hillock-like cells were relatively enriched in No PARDS and mild ARDS, declined in moderate ARDS, and were lowest in severe ARDS (Figure S10F), indicating that this stress–repair epithelial state is progressively blunted with increasing disease severity. Myeloid subtyping revealed resident-like MNP, SPP1+ MNP, repair-like MNP, inflammatory MNP, and ISG_high MNP (Figure S11A). As in Flerlage et al., inflammatory MNP proportion and module score dropped from acute to recovery and with lower severity, while SPP1+ MNP proportion and score rose. ISGhigh MNP frequency and ISG module scores decreased in the recovery phase (Figure S11D) and were also lower in severe disease (Figure S11B-E). Mature and intermediate neutrophils increased during acute ARDS and decreased in recovery (Figure S11F). By severity, Aged and Intermediate increased with severity, while mature fall (Figure S11G). This is consistent with prior observations of an accumulation of the neutrophil “aged” phenotype in severe PARDS.

Fig. 7.

Dynamic shift of BALF immune and epithelial subtypes in PARDS (A) Schematic workflow of bronchoalveolar lavage fluid (BALF) collection and processing at Timepoint 1 (acute) and Timepoint 2 (recovery). (B) UMAP embedding of single cells from eight paired BALF samples (Timepoint 1 → Timepoint 2) annotated into six major lineages by canonical markers; adjacent stacked bars show lineage composition per time point. (C) UMAP of BALF epithelium annotated into three cell types by canonical markers. (D) Bar plots showing epithelial subtype proportion changes from Timepoint 1 to Timepoint 2. (E) GO enrichment analysis of DEGs in KRT17+ epithelium. (F) UMAP of BALF macrophages annotated into three cell types by canonical markers. (G) Bar plots showing macrophage subtype proportion changes from Timepoint 1 to Timepoint 2

Niche-specific signaling pathways orchestrate alveolar repair versus injury in PARDS survival

To further delineate alveolar injury and repair dynamics, we developed normalized alveolar niche scores (ALV_normal, ALV_injury), which were correlated with distinct immune, inflammatory, and apoptotic patterns. ALV_injury scores were elevated in DEATH, while ALV_normal scores were higher in LIVE, indicating a more balanced and regenerative microenvironment in survivors (Figure S12A). Regional scoring revealed spatial specificity in repair mechanisms. In DEATH, injury signals predominated in almost all areas, conversely, LIVE exhibited stronger ALV_normal signatures in fibroblast expanding area, reflecting active repair and regeneration (Figure S12B). Spatial mapping showed CTHRC1+ pathologic fibroblast score enriched in injury regions (Figure S12C-D). Adipocyte-like programs concentrated in macrophage-enriched region (Figure S12E). Region-level correlations confirmed inverse axes (injury/pathologic/adipocyte-like vs. ALV_normal/alveolar), strengthened in DEATH (Figure S12F-G).

At pleural leak sites in LIVE, co-upregulation of SPP1 and WT1 matched the accumulation of mesothelium and SPP1_MAC, underscoring the coordinated role of these cells in sealing pleural injuries and promoting localized repair (Figure S13A-D). Moreover, spatial CellChat analysis revealed that IL6_IL6R_IL6ST and SPP1_CD44 interaction pathways were highly active in LIVE, supporting immune modulation and tissue regeneration. However, in DEATH, these pathways exhibited extensive upregulation, potentially amplifying inflammatory and apoptotic signals, thereby exacerbating tissue damage (Figure S13E-F). In LIVE, WT1/SPP1 joint expression and SPP1+ macrophage signals showed a striking, localized elevation at pleural air‑leak sites, whereas in DEATH, SPP1+ macrophage signals were globally elevated but diffusely distributed, yielding only small-to-moderate within-section contrasts for macrophage-enriched ROIs (Figure S13G).

Discussion

To our knowledge, this pilot multi-omics case series integrating single-cell and spatial transcriptomics with plasma proteomics provides an initial view of lung tissue repair features in pediatric ARDS and suggests differences between survivors and fatal cases that may highlight the role of spatially organized repair programs. These data indicate that fatal cases may show more diffuse pro-inflammatory and apoptotic signals across lung tissues, whereas survivors may exhibit more localized regenerative programs with comparatively preserved alveolar epithelial repair capacity. In survivors, maintenance of AT2 populations and the presence of AT2-to-AT1 transition markers are consistent with effective epithelial regeneration mechanisms, aligning with prior observations that alveolar repair capacity may relate to outcomes in pediatric respiratory failure [11, 23–25].

Our study is the first, to our knowledge, to validate that plasma KRT17 levels significantly discriminate between surviving and deceased pediatric ARDS patients. Anchoring to the adult lethal COVID‑19 atlas (Nature 2021) [19], DATP/PATS scores (KRT8/CLDN4/CDKN1A) were highest in adult AT2 and low in AT1/airway states; mapping our pediatric epithelium to this reference reproduced the same subtype pattern. However, pediatric KRT17 localized to AT2‑lineage cells rather than airway/basal cells, indicating a pediatric, AT2‑linked stress‑repair trajectory distinct from adult patterns. Prior work has shown that during AT2-to-AT1 differentiation in lung injury model systems, a transitional KRT17+KRT5− epithelial state arises [26]. In human adults, KRT17+/KRT5− transitional cells are enriched in end-stage fibrotic lung tissue [27] and can also localize to histologically normal-appearing alveolar regions in pulmonary fibrosis patients [26]. Nevertheless, age-related cohort differences, distinct immune features of the pediatric alveolar niche, the inherent difficulty of inferring causality from clinical specimens, and the self-limiting nature of the bleomycin mouse fibrosis model [28, 29] complicate interpretation; thus, in adult PF contexts, the persistence of these cells has been interpreted as a fibrosis-linked transitional state with maladaptive potential. Importantly, our survey of adult public lung datasets did not reveal a consistent, significant upregulation of KRT17 within alveolar epithelial subtypes, whereas our pediatric lung tissue analyses by multiplex IF consistently demonstrate high KRT17 expression in AT2-lineage cells (this study). Extending beyond prior adult fibrosis–focused studies that emphasized KRT17 changes within KRT5+ basal-like programs, our data show that KRT17 is highly expressed within the AT2 compartment in pediatric ARDS, indicating a pediatric- and phase-specific epithelial repair trajectory that may not be generalizable across disease phase and age from adult fibrotic paradigms; notably, we do not equate pediatric AT2-associated KRT17+ cells with the pathologic KRT17+/KRT5− population described in end-stage adult fibrosis. This distinction is clinically meaningful because pediatric ARDS exhibits substantially lower mortality and a greater capacity for complete lung recovery compared with adults, and, critically, lung tissue data from children with ARDS have been scarce—our study provides a rare and informative dataset that refines current understanding of KRT17 biology in lung injury. Our data indicate that the presence of KRT17+ cells is associated with effective lung injury repair and survival in pediatric ARDS, representing an active reparative program within the alveolar epithelium based on image- and fluid-based evidence. This is corroborated by detectable KRT17 in both plasma and BALF from surviving children, with concordant directional changes across specimen types. Thus, plasma KRT17 quantification in pediatric ARDS may constitute a pediatric-specific, clinically actionable biomarker for prognostic stratification, pending validation of cutoffs and assay standardization.

Alveoli function through coordinated interactions among epithelial, fibroblast, endothelial, and immune cells, but this delicate balance can be disrupted by inflammation or aging, leading to impaired repair and lung dysfunction [30]. The study reveals critical fibroblast heterogeneity, with fatal cases dominated by fibrotic lipofibroblasts and CTHRC1 signature while survivors maintained alveolar fibroblast populations, mirroring findings in adult ARDS about divergent remodeling pathways [29, 31, 32]. Spatial analysis demonstrates that successful repair depends not only on molecular signatures but also on their precise anatomical compartmentalization, with survivors containing injury to specific zones while preserving near-normal alveolar niches elsewhere. This spatial organization of repair programs aligns with emerging concepts about the importance of regional heterogeneity in lung injury resolution [4]. Across adult and pediatric lung datasets, HLCA and prior lung atlases consistently report the emergence of CTHRC1-positive pathologic (myofibroblast-like) fibroblasts during severe injury and fibrotic remodeling, alongside alveolar and adipocyte-like fibroblast states [33]. In adult lethal COVID-19 lung tissue, expansion of CTHRC1+ fibroblasts co-occurs with hypoxic/inflammatory signaling and diffuse epithelial damage, providing a benchmark for disease-associated remodeling [33]. In our PARDS cohort, injury-enriched regions (ALV_injury) showed higher scores for CTHRC1, COL1A1, SERPINE1, and HIF1A modules, whereas preserved regions (ALV_normal) retained alveolar fibroblast features and lower pathologic scores (this study). Importantly, HLCA-based label transfer indicated high certainty for alveolar fibroblast/macrophage labels—consistent with disease-affected states—supporting a model in which pathologic fibroblast programs commingle with adipocyte-like fibroblasts under inflammatory stress, while survivors preferentially maintain alveolar fibroblast niches [19, 33]. These public-dataset concordances reinforce that spatially localized repair in pediatric ARDS is associated with containment of CTHRC1-driven remodeling, in contrast to the more diffuse expansion observed in adult fatal disease [19].

Macrophage polarization patterns differ markedly between outcomes [4], with fatal cases displaying widespread SPP1+ macrophages linked to tissue fibrosis [34] whereas survivors retained protective FABP4+ macrophages in defined regions, supporting previous biomarker studies suggesting macrophage phenotype influences PARDS mortality [18]. Orthogonal plasma proteomics and clinical data likewise indicate a cytokine-skewed milieu in fatalities. Extending these observations, integration of tissue and BALF datasets indicated a trajectory from FCN1-high inflammatory monocyte-like states in the acute phase toward FABP4-positive resident-like alveolar macrophages during recovery, accompanied by dampening of interferon-stimulated gene modules and reduced neutrophil aging signatures, in line with public pediatric PARDS airway profiles [4, 18, 33, 34]. These data motivate targeted cytokine assays to refine inflammatory endotyping in pediatric ARDS. The pleural repair response in survivors, marked by mesothelial enrichment and SPP1+ macrophage accumulation, suggests a unique mechanism for resolving air leaks that may be particularly relevant in pediatric patients given developmental differences in pleural biology. Together, these findings reinforce a model in which restoration of resident-like FABP4+ alveolar macrophages aligns with epithelial repair and favorable outcomes, whereas expansion of SPP1+ and FCN1-high inflammatory programs associates with fibrosis-prone trajectories and mortality.

These findings provide foundational insights into age-specific lung repair mechanisms after viral infection, extending previous work on PARDS sub-phenotypes [35] by revealing the spatial architecture of successful recovery. Given the small tissue cohort and pathogen homogeneity with influenza, our conclusions should be interpreted as hypothesis generating rather than definitive across all PARDS etiologies. Because the BALF arm included survivors only, we do not extrapolate BALF findings to fatal outcomes. Instead, we used BALF to contextualize airway-facing programs that overlap with, but are distinct from, tissue signatures observed in DEATH. Despite this limitation, this study has identified, potential therapeutic targets including fibroblast subpopulation modulation and macrophage polarization strategies that could promote regeneration in PARDS, building on prior observations about personalized ventilation approaches [36]. Although limited by sample size and event counts that can yield unstable adjusted odds ratios in multivariable models, the consistently high AUC for plasma KRT17 supports strong standalone discrimination for survival which addresses a different question from adjusted independent effects, and concordant signals across spatial transcriptomics, tissue and BALF single cell RNA sequencing, and plasma proteomics enhance biological plausibility. When harmonized to HLCA, elevated label transfer uncertainty in ARDS cell types likely reflects true disease related phenotypic divergence rather than technical noise [33], and while age specific features of the pediatric alveolar niche and the rarity of pediatric ARDS lung tissue constrain external generalization, these same factors underscore the novelty and clinical relevance of our dataset which nominates KRT17 as a measurable biomarker that warrants prospective validation and supports the need for spatially targeted therapies in pediatric lung injury.

supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

ABC

Ammonium bicarbonate

ADI

Alveolar differentiation intermediate

AGER

Advanced glycosylation end product specific receptor

AGC

Automatic gain control

ALV

Alveolar niche

ALV_normal

Preserved alveolar niche score

ALV_injury

Injury‑enriched alveolar niche score

ANCOVA

Analysis of covariance

ARDS

Acute respiratory distress syndrome

ARI

Adjusted Rand Index

AT1

Alveolar type 1 cell

AT2

Alveolar type 2 cell

AT2-to-AT1

Transitional AT2→AT1 epithelial state

AUC

Area under the curve

BALF

Bronchoalveolar lavage fluid

BCa

Bias‑Corrected and Accelerated bootstrap

BMC

BioMed Central

BTLA

B and T lymphocyte attenuator

CC

Healthy control group in plasma cohort

Cliff’s delta

Cliff’s Δ effect size

CTHRC1

Collagen triple helix repeat containing 1

CXCL

CXC motif chemokine ligand

DAPI

4′,6-diamidino-2-phenylindole

DC

Dendritic cell

DATP

Damage-associated transient progenitor

DEATH

Fatal pediatric ARDS case

DEG

Differentially expressed gene

DIA

Data-independent acquisition

DNase

Deoxyribonuclease

ECM

Extracellular matrix

ECMO

Extracorporeal membrane oxygenation

ELISA

Enzyme-linked immunosorbent assay

FACS

Fluorescence-activated cell sorting

FABP4

Fatty acid binding protein 4

FASLG

Fas ligand

FC

Fold change

FCN1

Ficolin 1

FCS

Fetal calf serum

FDR

False discovery rate

FFPE

Formalin-fixed, paraffin-embedded

FN1

Fibronectin 1

GAM

Generalized additive model

GO

Gene Ontology

GRCh38

Genome Reference Consortium Human Build 38

H&E

Hematoxylin and eosin

HCD

Higher-energy collisional dissociation

Hedges g

Hedges’ g effect size

HGF

Hepatocyte growth factor

HIF1A

Hypoxia inducible factor 1 alpha

HLCA

Human Lung Cell Atlas

HRP

Horseradish peroxidase

IFN

Interferon

IL

Interleukin

ISG

Interferon-stimulated gene

KRT5

Keratin 5

KRT8

Keratin 8

KRT17

Keratin 17

KS

Kolmogorov–Smirnov

LDH

Lactate dehydrogenase

LIVE

Surviving pediatric ARDS case

LIMMA

Linear models for microarray analysis

LPS

Lipopolysaccharide

MAC

Macrophage

MAST

Model-based analysis of single-cell transcriptomics

MDM

Monocyte-derived macrophage

Moran’s I

Moran’s I spatial autocorrelation

MMP

Matrix metalloproteinase

MNP

Mononuclear phagocytes

MS

Mass spectrometry

MSigDB

Molecular Signatures Database

NEU

Neutrophil

NMI

Normalized Mutual Information

NNLS

Non-negative least squares

NNTransform

Nearest neighbor transform

PARDS

Pediatric acute respiratory distress syndrome

PATS

Pre-alveolar type 1 transitional cell state

PCA

Principal component analysis

PBS

Phosphate-buffered saline

PDGFRA

Platelet-derived growth factor receptor alpha

PFA

Paraformaldehyde

PICU

Pediatric intensive care unit

POSTN

Periostin

RcppML

package for non-negative matrix factorization/machine learning

RNA-seq

RNA sequencing

ROC

Receiver operating characteristic

ROI

Region of interest

scRNA-seq

Single-cell RNA sequencing

SCT

Single-cell transform normalization

SEM

Standard error of the mean

SMC

Smooth muscle cell

SNN

Shared nearest neighbor

SOFA

Sequential Organ Failure Assessment

SPP1

Secreted phosphoprotein 1 (osteopontin)

ST

Spatial transcriptomics

T1

Timepoint 1 at ECMO initiation

T2

Timepoint 2 at ECMO weaning

TGFβ

Transforming growth factor beta

TMB

Tetramethylbenzidine

TNF

Tumor necrosis factor

TWEAK

TNF-like weak inducer of apoptosis

UMAP

Uniform manifold approximation and projection

VEGF

Vascular endothelial growth factor

WT1

Wilms tumor 1

Author contributions

LX conceived and supervised this project. HC, XH, and LS. designed experiments. LS. conducted data analysis and designed the bioinformatics pipelines. YY and RS performed experiments with the help of YL, DS, and WG. LS wrote the manuscript with help from the other authors. LX, XH, and HC offered suggestions for revision.

Funding

The work was supported by the National Natural Science Foundation of China (82172109) (L.X.).

Data availability

Data sequenced from human lung tissues in this project were deposited in National Center for Biotechnology Information (NCBI). scRNA-seq data (GSE223793) was deposited in Gene Expression Omnibus (GEO). Spatial transcriptomics data were deposited in China National Center for Bioinformation (CNCB). The raw sequences were access controlled, PRJCA042363. All analysis scripts and reproducible workflows used in this study are openly available at GitHub under the repository Lung-Repair-Niches-in-Pediatric-ARDS at https://github.com/songlicheng666/Lung-Repair-Niches-in-Pediatric-ARDS.

Declarations

Ethics approval and consent to participate

Our study was conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines and regulations. Human lung tissues, blood, and BALF samples were obtained with the approval of the Ethics Committee of the Chinese PLA General Hospital (approval No. 2,022,052,701,006), and the protocol was approved by the Institutional Review Boards of the 8th Medical Center of the Chinese PLA General Hospital. Written informed consent was obtained from participants or their legal guardians, and all procedures complied with applicable national and institutional standards for research involving human subjects.

Consent for publication

Not applicable.

Competing interests

The authors declare no conflicts of interest.

Statement

The manuscript has been read and approved by all the authors. The requirements for authorship as stated earlier in this document have been met. Each author believes that the manuscript represents honest work.