In humans, moderate to severe OSA is associated with interstitial lung disease with remodeling of the extracellular matrix (Kim et al., 2017). Lung is the primary organ that senses episodes of hypoxia and is therefore exposed to large fluctuations in the oxygen concentrations compared to other tissues throughout the body. For these reasons, we sought to identify initial changes in gene expression pathways in the lung in response to IH.

Mice were initially entrained to the same light:dark schedule to synchronize active and inactive phases. After 14 days of entrainment in the 12 hr:12 hr light:dark cycle, mice were exposed to IH or room air (normoxia) for the entire 12 hr inactive phase for 9 days (Figure 1A). Hematoxylin and eosin (H and E) staining of whole lungs did not show comprehensive changes in architecture or inflammatory remodeling after exposure to IH (Figure 1—figure supplement 1). Bulk RNA sequencing (Bulk RNA-seq) was performed to explore transcriptomic effects of IH on lung tissue at the organ level. There were 374 genes (Figure 1—figure supplement 2A) upregulated and 149 downregulated in mouse lung after exposure to IH (Benjamini-Hochberg q value (BHQ) < 0.05 and fold change > 1.5). Not surprisingly, the top upregulated genes included well-known HIF-1 target genes (e.g. Edn1, Bnip3, and Ankrd37; Figure 1—figure supplement 2B).

We performed DAVID (Huang et al., 2009) enrichment analysis to identify biological processes associated with the top 200 up- and downregulated genes from mice exposed to IH. Pathways induced in response to hypoxia included circadian rhythm, angiogenesis, and extracellular matrix organization (Figure 1B,C). As previously reported, there is a tight interaction between clock genes and HIFs (Edgar et al., 2012; Gu et al., 2000; Hogenesch et al., 1998; McIntosh et al., 2010; Taylor and Zhulin, 1999). Several circadian clock repressors (e.g. Nr1d1, Nr1d2, Bhlhe40, and Per3) were significantly upregulated in the IH group (Figure 1C). Increased expression of RNAs associated with angiogenesis, such as vascular endothelial growth factor (VEGF), was observed after IH, consistent with findings in a mouse model of prolonged exposure to IH (Reinke et al., 2011a) and after 72 hr of IH exposure to endothelial cells in vitro (Wohlrab et al., 2018). There is a tight relationship between angiogenesis and pulmonary hypertension (Tuder and Voelkel, 2002), a clinical consequence associated with OSA. Unexpectedly, RNAs associated with immune responses and cell cycle were significantly downregulated after 9 days of IH (Figure 1B,C). Present findings contrast with the general concept that HIFs are important regulators of inflammation and immune responses (Eltzschig and Carmeliet, 2011; Scholz and Taylor, 2013; Taylor et al., 2016). For example, activation of neutrophils by HIFs is largely considered proinflammatory (Walmsley et al., 2005; Peyssonnaux et al., 2005; Taylor et al., 2016). The cell cycle and cell division downregulation may be associated with HIFs induced in cell cycle arrest in response to IH (Koshiji et al., 2004).

We detected significant differences related to a number of biological functions at the tissue level. We then applied single-cell transcriptomics to identify cell type-specific effects of IH. We performed three biological replicates in each group to improve the statistical power in differential gene expression analysis at the single cell level. In total, we sequenced 12,324 and 16,125 pulmonary cells from IH and control mice, with 3542–5641 cells per biological replicate. Unsupervised analysis identified 25 transcriptionally distinct cell clusters, corresponding to 19 distinct cell types (Figure 2A,B) based on the expression of established marker genes (see Materials and methods), including stromal, epithelial, endothelial, immune, and small numbers of other cell types. In brief, AltAnalyze identified 40–60 marker genes for each cell cluster. We annotated each cluster to a cell type using enrichment analysis between these marker genes and a comprehensive reference marker gene list, which is collected from public databases and published scRNA-seq studies performed in mouse or human lungs. As shown in Figure 2C, the cell-type assignment was also validated with multiple well-known cell-type markers. The proportion of endothelial, alveolar epithelial type II (AT2), and fibroblast/myofibroblast cells were modestly increased, but the proportion of immune cells (e.g. B and T cells) was decreased in the IH-exposed mice (Figure 2—figure supplement 1). Overall, the proportional variation of lung cell types was small (BHQ > 0.05). We further performed confocal immunofluorescence to evaluate variation in lung structure and to quantify the differences in cell number of major cell types in mice exposed to IH compared to controls.

Similar to H and E histology, confocal immunofluorescence microscopy did not demonstrate comprehensive inflammatory remodeling. However, some modest changes were noted with specific cell populations. Immunostaining for endothelial markers FOXF1 and LYVE1 showed modest increases in the number of endothelial cells in mice exposed to IH (Figure 3C,D) compared to control mice (Figure 3A,B). Also important, staining for MKI67 did not show statistically significant changes in cell proliferation (Figure 3E–H, Figure 3—figure supplement 1). Expression levels of HOPX and SFTPC in alveolar type I and II cells, respectively, were not different for IH (Figure 3K,L) versus control (Figure 3I,J) mice. The trend for small increases in AT2 cells, although not significant, was also seen from the scRNA-seq data (Figure 2—figure supplement 1, Figure 3—figure supplement 1). Immunostaining for the progenitor marker SOX9 (Figure 3I–L) or the extracellular matrix marker POSTN (Figure 3M–P) did not demonstrate any significant changes after IH exposure. Overall, there was a modest increase (11.5%; BHQ < 0.05) of endothelial cells in mice exposed to IH compared to controls (Figure 3—figure supplement 1); however, this does not necessarily reflect a biologic difference. This trend for an increasing number of endothelial cells in the histologic samples was similar to the changes in cell numbers seen from the scRNA-seq data (Figure 2—figure supplement 1, Figure 3—figure supplement 1). Other cell percentages, including AT2 cells and proliferating cells, were not significantly different between IH and control groups. We did not see comprehensive changes in alveolar area and alveolar wall thickness based on morphometric quantification for mice exposed to IH compared to controls (Figure 3—figure supplement 2A–C). The average MaxFeret90 measurements and alveolar areas were 21.0 µM and 419 µM² in mice exposed to IH, while the average MaxFeret90 measurements and alveolar areas were 23.5 µM and 564 µM² in control mice. The mean alveolar wall thickness was 4.81 µM for mice exposed to IH and controls.

We further explored the early cell type-specific response to IH in mouse lung by aggregating single-cell data into ‘pseudo-bulk’ data to compare biological replicates for each identified cell type (see Materials and methods for details). For each biological replicate, we summed the reads for each gene in the same cell type. Total read counts between replicates and the two conditions were normalized by size factors in DESeq2 to reduce the impact of cell number differences. Using DESeq2 (Love et al., 2014), the number of up- or downregulated genes in different lung cell types in response to IH were not equal at the same statistical cutoff (Figure 4—figure supplement 1). For example, there were 607 and 550 genes up- and downregulated in endothelial cells, while there were 186 and 139 genes up- and downregulated in B cells at p<0.05. We selected the top 200 up- and downregulated genes (ranking by the p-value) from each cell type in the pathway enrichment analysis to balance the input gene number difference. From the DAVID enrichment analysis, diverse biological processes were up- and downregulated in different cell types in response to IH (Figure 4A–C, Figure 4—figure supplement 2A,B). For example, hypoxia-responsive and circadian pathways were enriched in those upregulated genes in response to IH in endothelial cells, myofibroblasts, and AT2 cells. Surprisingly, circadian pathways were highly enriched in multiple cell populations, not just epithelial cells, a population that is important for circadian rhythmicity in the lung (Gibbs et al., 2009). The co-upregulation of genes in these two pathways suggest that the genome-wide co-regulation of hypoxia and the circadian pathway is tighter in lung endothelial, myofibroblast, and AT2 cells than other cell types.

Our data demonstrate that cell adhesion genes are upregulated in endothelial cells, myofibroblasts, fibroblasts, and pericytes after exposure to IH. These results suggest that lung endothelial and stromal cells may contribute to the increased circulating cell adhesion molecules in OSA patients. Interestingly, cell adhesion molecules that control leukocyte trafficking are increased in OSA patients (Ohga et al., 1999) and reduced after continuous positive airway pressure (CPAP) treatment (Chin et al., 2000; Pak et al., 2015). Immune response-related and antigen processing and presentation were enriched among those downregulated genes in monocytes, macrophages-dendritic cells, NK cells, and erythroblasts. The decrease among immune related genes in these cell types is consistent with the bulk RNA-seq data. It is known that a hypoxic tumor microenvironment will induce immunosuppression (Sitkovsky et al., 2008). Hypoxia may also cause immunosuppression in patients with OSA and contribute to the increased incidence of lung cancer in this population (Campos-Rodriguez et al., 2013).

As expected from specific genes in each biological process, the response level of the same genes were different in multiple cell types (Figure 4B,C). For example, the circadian repressor gene, Nr1d1, was more responsive to IH in endothelial and AT2 cells than lymphatic endothelial cells and AT1 cells. We also noted cell-specific responses for the downregulated genes. For example, immune response genes (e.g. Iglc3, S100a8, and Oas3) decreased more in monocytes than macrophages in response to IH. These results suggest that a cell type-specific response to short exposures of IH exists at a single-gene level, findings that may be related to the progression of OSA-associated chronic pulmonary diseases. For example, bleomycin-induced fibrosis is induced in mice by an Nr1d1 mutation in specific lung cells (Cunningham et al., 2020).

Recent studies show distinct vascular endothelial cell subpopulations in mouse and human lung (Gillich et al., 2020). Our vascular endothelial populations were annotated to endothelial artery, vein, capillary aerocytes (Cap-a), and general capillary (Cap-g) cells (Figure 5A, Figure 5—figure supplement 1). Interestingly, we found that endothelial cells demonstrated profound changes in gene expression profiles in response to IH. The endothelial capillary cells were more responsive to IH compared to endothelial artery and vein cells (Figure 5B). For example, at BHQ < 0.2, more than 100 genes were significantly upregulated in aerocytes and general capillary cells. However, only one gene in the arterial endothelial cells and 57 genes in the venular endothelial cells were significantly upregulated at the same cutoff. This trend persisted at other BHQ cutoffs (Figure 5—figure supplement 2). Given the location of lung capillary cells around the alveoli and positioned at the air–blood barrier, these results indicate that the capillary cells may be among the first lung endothelial cell groups that respond early after initial exposure to IH. Among the capillary cells, general capillary cells are more responsive to hypoxia than aerocytes. More genes were significantly up and downregulated in general capillary cells than aerocytes (Figure 5B). Hypoxia-responsive genes (e.g. Pdk1, Vegfc, Slc2a1, Pkm, and Flt1) showed higher levels of expression variation (fold change and significance) in general capillary cells than aerocytes (Figure 5C), demonstrating variation at the subpopulation level. Aerocytes are specialized for gas exchange, and general capillary cells function in capillary homeostasis and aerocyte production during repair (Gillich et al., 2020). Our observed response from general capillary cells may be related to its role in capillary homeostasis. Alternatively, proximity and interaction with other cell types, such as fibroblasts or immune cells, may also help to explain these findings. For example, general capillary cells contact stromal cells while aerocytes lie in close proximity to AT1 cells (Gillich et al., 2020).

We further compared the response of general capillary cells and aerocytes to IH at the pathway level. Glycolytic process was upregulated and cell migration was downregulated in both aerocytes and general capillary cells in response to IH (Figure 5D,E). Given the association between glycolysis, cytoskeletal remodeling, and cell migration in other cell types (Shiraishi et al., 2015), the similar enrichment trends for these pathways is not surprising. Additionally, without vascular growth associated with chronic IH, glycolysis may be used to meet metabolic demand. Interestingly, changes in the glycolytic process were specific to endothelial cell types (Figure 4—figure supplement 2). On the other hand, aerocytes and general capillary cells demonstrated more differences in enrichment pathways (Figure 5D,E). For example, regulation of cell proliferation and regulation of angiogenesis were only enriched for those up- and downregulated genes in general capillary cells in response to IH. These pathways are tightly correlated with its specific role in capillary homeostasis as a capillary progenitor cell and in regulation of vasomotor tone (Gillich et al., 2020). While phagocytosis, immune response, and regulation of cell shape were only enriched for those downregulated genes in aerocytes (Figure 5D,E), this may reflect aerocyte’s specific role in leukocyte trafficking and gas exchange, both highly influenced by exposure to IH.

We used SINCERA (Guo et al., 2015) to predict key transcription factors (TFs) in the inferred transcriptional regulatory network in general capillary and aerocytes (Figure 5—figure supplement 3A,B; Materials and methods). Hypoxia-associated TFs (e.g. Epas1, Jun, and Ahr), Foxf1 (canonical endothelial cell marker), and Ybx1 (glycolysis related) were among the top-ranked TFs in both aerocytes and general capillary cells. Tbx2, Tbx3, and Meox1 were among the top-ranked TFs only in aerocytes, which are considered aerocyte-specific TFs (Gillich et al., 2020). Interestingly, three immune response-related TFs (Hlx, Rarg, and Smarca2) were only top-ranked in aerocytes. Additionally,Tcf4 and Sox17 were among the top-ranked TFs only in general capillary cells, which are associated with endothelial cell regeneration or stem cell self-renewal (van Es et al., 2012; Liu et al., 2019). Elk3 and Casz1 were among the top-ranked TFs in general capillary cells and are related to angiogenesis (Heo and Cho, 2014; Charpentier et al., 2013). The pathway-level differences between aerocytes and general capillary cells may be driven by these predicted top-ranked TFs, which need further experimental validation.

Evaluation of the expression pathways at the single-cell resolution demonstrated significant changes in multiple cell types. With this information, we then wanted to identify potential candidates for therapeutic intervention in response to IH.

OSA is associated with an array of pulmonary diseases, such as interstitial lung disease (ILD) (Kim et al., 2017), idiopathic pulmonary fibrosis (IPF) (Lancaster et al., 2009), and pulmonary hypertension (PH) (Sajkov and McEvoy, 2009; Chaouat et al., 1996). IH led to significantly more upregulated than downregulated pulmonary disease-associated genes (Figure 6A, Figure 6—figure supplement 1). IH-induced expression in myofibroblasts and fibroblasts demonstrated enrichment of genes associated with pulmonary fibrosis (PF). However, the disease genes were not equally expressed or upregulated in these cell types (Figure 6B). For example, Ptgis is a PH-associated gene highly expressed in myofibroblasts and fibroblasts compared to other cell types (Figure 6—figure supplement 2). Ptgis is also a target gene for epoprostenol – a drug used for treating PH (Sitbon and Vonk Noordegraaf, 2017). The IPF-associated gene, Thbs1 (Idell et al., 1989; Kuhn and Mason, 1995; Yehualaeshet et al., 2000), was highly expressed and more responsive to IH in myofibroblasts compared to other cell types. THBS1 is upregulated in lung stromal cells of patients with chronic interstitial lung diseases (e.g. IPF) compared to healthy controls (Gene Explorer under Banovich/Kropski in http://www.ipfcellatlas.com/; Habermann et al., 2020). Msr1 (Silverman et al., 2002; Hersh et al., 2006) is a chronic obstructive pulmonary disease (COPD)-associated gene, which was highly expressed in macrophage-DCs, basophils, and monocytes, consistent with scRNA-seq data from patients with IPF (Gene Explorer under Banovich/Kropski in http://www.ipfcellatlas.com/; Habermann et al., 2020). These data highlight the similarity of the IH-associated signatures with cell type-specific responses in an array of pulmonary diseases.

The top 200 up- or downregulated genes in each cell type exposed to IH were linked with drug target genes extracted from DrugBank (Wishart et al., 2008). We obtained 42 drugs used as activators or inhibitors of 23 target genes to treat pulmonary hypertension, pulmonary fibrosis, COPD, and asthma (Figure 6C). Interestingly, nearly one third of these drug target genes were more responsive to IH in endothelial cells compared to other cell types (Figure 6C, Figure 6—figure supplement 1). We further evaluated changes in the interactions of these drug-targeted disease genes in each cell type using CellPhoneDB (Vento-Tormo et al., 2018). There were 32 ligand–receptor interactions found in the database using these drug-targeted disease genes. Overall, myofibroblasts were involved in 98 of 176 gains of interaction (BHQ = 4.28e-50), demonstrating the significance of this cell type in early IH-associated responses (Figure 6D, Figure 6—figure supplement 3). Here, we showed the importance of myofibroblasts in activating the fibroblast growth factor (FGF) signaling pathway (Figure 6—figure supplement 3). We then performed further analysis on myofibroblasts. The GSEA showed upregulation of multiple extracellular matrices (ECM)-associated pathways (e.g. ECM proteoglycans, non-integrin membrane ECM interactions, and collagen chain trimerization) in myofibroblasts (Figure 6—figure supplement 4) after exposure to IH. FGF signaling pathway has a tight interaction with ECM (Taipale and Keski-Oja, 1997; Ornitz and Itoh, 2015), which may be associated with activation of the FGF signaling pathway in myofibroblasts in response to IH. In the inferred transcriptional regulatory network, the majority of the 25 top-ranked TFs were shared between the control and IH groups (Figure 6—figure supplement 5). Interestingly, only Dbp and Hlf were top-ranked TFs in mice exposed to IH (Figure 6—figure supplement 5). These results suggest that IH-induced circadian dysregulation may drive the transcriptional changes in myofibroblasts.

Use of animals and all procedures were approved by the Institutional Animal Care and Use Committee at Cincinnati Children’s Hospital Medical Center and complied with the National Institutes of Health guidelines. Male C57BL/6J wild-type mice (RRID: IMSR_JAX:026133) aged 6 weeks were purchased from The Jackson Laboratory (Bar Harbor, ME) and entrained to a 12 hr:12 hr light:dark cycle for 2 weeks prior to exposure.

All studies were conducted in 8–10 weeks old male mice. Mice were housed in light boxes and entrained to a 12:12 light:dark cycle for 2 weeks prior to initiation of IH. Mice were randomly assigned to IH or room air exposures. For the experimental group, mice were maintained in a commercially designed gas control delivery system (Model A84XOV, BioSpherix, Parish, NY) during the inactive (light) phase from (ZT 0–12). Mice were provided with food and water ad libitum. For each episode of IH, the fractional inhaled oxygen (O₂) was reduced from 20.9% to 6% over a 50 s exposure period, followed by an immediate 50 s recovery period to 20.9%. The fractional oxygen was maintained at 20.9% for approximately 15 s before the cycle was repeated, allowing for ~30 hypoxic events per hour. Ambient temperature in the hypoxia chamber was maintained between 22 and 24°C to match room air. Mice in the experimental group were maintained at room air during the active phase, ZT 13–24. Mice in the control group were maintained at room air throughout the circadian cycle, ZT 0–24. Experimental and control mice were exposed to IH vs room air for 9 days, followed by sacrifice at ZT 3 on day 10 of exposure. This was immediately followed by organ harvest and preparation for bulk RNA-seq, scRNA-seq, or histopathology.

Animal models of IH are performed by exposing mice to cycling levels of room air (fraction of inspired oxygen [FiO₂] – 21%) to different trough/nadir values (most commonly from 5 to 15%), producing dose–response effects (Nagai et al., 2014; Lim et al., 2016; Gallego-Martin et al., 2017; Farré et al., 2018). These fluctuations in FiO₂ translate to measured oxygen saturation (SaO₂) nadir values of approximately 50–70% in mice (Jun et al., 2010; Lim et al., 2015; Reinke et al., 2011b). These cycling SaO₂ nadirs in mice closely correspond to partial pressure of oxygen (PaO₂) values seen in humans with OSA (Farré et al., 2018). PaO₂ levels in humans are lower than for mice for any given SaO₂ value (Farré et al., 2018). Hence, SaO₂ nadir values that commonly occur in patients with OSA (70–90%) correspond to SaO₂ nadir values (40–70%) in mice exposed to our cycling FiO₂ trough levels. Based on the diagnostic criteria for severe OSA in humans (Kapur et al., 2017), our protocol for 30 hypoxic events an hour with cycling FiO₂ troughs of 6% in mice represent clinically severe OSA in humans.

In total, six experimental and six control mice were used for bulk RNA-seq. For bulk RNA-seq, the lung was quickly harvested and snap-frozen in liquid nitrogen. Organs were later homogenized in TRIzol reagent (Invitrogen) and processed using a bead mill homogenizer (Qiagen Tissuelyser). RNA was then isolated from lung homogenates by phase separation using chloroform and phase separation columns. The aqueous phase was then applied to an RNeasy column following the manufacturer’s protocol (Qiagen) to extract and purify the RNA.

RNA from the lungs of control and experimental mice were sent for bulk sequencing separately. Approximately 0.4 µg of total RNA was used for library preparation. mRNA enrichment and library preparation were performed using the Polyadenylated (PolyA+) mRNA Magnetic Isolation Module (New England Biolabs) and NEBNext Ultra II RNA Library Prep Kit for Illumina (New England Biolabs), following the manufacturer’s protocol. All 12 samples were then pooled together and sequenced in one lane using Illumina Novaseq 6000 platform with paired-end 150 bp (Supplementary file 1a). The raw fastq files from RNA-seq were mapped to GRCm38 mouse genome reference using STAR (version 2.5) with default parameters. More than 90% (Supplementary file 1a) of sequenced paired-end reads (above 50M reads for each library) were mapped to the mouse genome by STAR (Dobin et al., 2013; RRID:SCR_015899). HTSeq (Anders et al., 2015; RRID:SCR_005514) (version 0.6.0) was used to quantify gene expression, with Ensembl GRCm38.96 as a reference. DESeq2 (RRID:SCR_015687; version 1.24.0) was used to perform the differential expression analysis on the HTSeq quantified count per gene. The top 200 up- and downregulated genes (Supplementary file 1b), ranked by p-value from low to high with fold change above 1.5 (or log2[fold change] > 0.58), were used for biological process enrichment analysis in the DAVID database. Biological process terms with at least five differentially expressed genes and BHQ < 0.15 were selected. For aggregating redundant biological processes, GOSemSim (Yu et al., 2010) (version 2.10.0) was used to calculate the semantic similarity (‘Jiang’ method from GOSemSim) between significant biological processes. Redundant biological processes were manually merged into biological process categories (Supplementary file 1c).

For scRNA-seq, a total of three biological replicates (three IH and three controls in the first experiment and two of each in the other replicates). On the last day of exposure, the mice were sacrificed, lung harvested, and tissue immediately placed in ice-cold phosphate-buffered saline (PBS). Dissociation of the pooled whole mouse lung for each group (IH vs control) was performed as previously described (Guo et al., 2019). Briefly, minced lung was placed in collagenase/elastase/dispase digestion buffer (Sigma–Aldrich, St. Louis, MO; Worthington Biochemical, Lakewood, NJ). After mixing on ice for approximately 3 min, the lung was minced again. After resting the suspension, the supernatant was passed through a 30 µM filter. A Bacillus licheniformis mix (Sigma–Aldrich) was added to the cell suspension, mixed on ice for approximately 10 min, and passed through a 30 µM filter. The suspension was spun at 500 g for 5 min at 4°C. The pellet was rinsed with a red blood cell lysis buffer. This was again passed through a 30 µM filter and then spun at 500 g for 5 min at 4°C. The cell suspension was resuspended in PBS/bovine serum albumin and manually counted with a hemocytometer. The volume was adjusted to obtain a final concentration of approximately 1000 cells/µl to be loaded to the 10× Chromium platform.

The single-cell suspension was applied to the 10× Genomics Chromium platform (San Francisco, CA) to capture and barcode cells, as described in the manufacturer’s protocol. Libraries were constructed using the Single Cell 3′ Reagent Kit (v2 Chemistry). The completed libraries were then sequenced using HiSeq 2500 (Illumina, San Diego, CA) running in Rapid Mode. Each sample was loaded onto two lanes of a Rapid v2 flow cell.

To identify differentially expressed genes in each lung cell type between the control and IH groups with multiple biologic replicates, all cells assigned to the same cell type were aggregated into a ‘pseudo-bulk’ data library by library. For each library, the sum of the reads per gene from cells assigned to the same cell type were used to represent the cell-type-specific gene expression profiles. Percentage of cells expressed per gene were calculated as a fraction of cells with ≥1 read(s) for the gene in each cell type. Count per one million UMI (CPM) for each cell type was calculated as the (sum of reads per gene/sum of reads) * 1,000,000 for each library. Differential expression analysis was performed with DESeq2 for each cell type, using the sum of reads per gene as input, with three replicates in each of the control and IH groups. Ranked by p-value from low to high, the top 200 up- and downregulated genes (Supplementary file 2d) with fold change above 1.2 (or log2[fold change] > 0.26) were used for biological process enrichment analysis in the DAVID database. Selecting and aggregating biological processes (Supplementary file 2e) were performed as described in the Materials and methods section labeled, ‘Bulk RNA sequencing and analysis’. We further extracted those genes enriched in circadian rhythm and immune response and selected well-established genes in each biological process to demonstrate their expression variation under IH exposure in each cell type based on literature searches.

To improve the accuracy for classifying endothelial cells, we reran the AltAnalyze-ICGS2 clustering algorithm using only the 5579 cells annotated to pulmonary vascular endothelial cells from IH-exposure groups, followed by a realignment of all vascular endothelial cells from both control and IH groups to these clusters using cellHarmony. AltAnalyze-ICGS2 produced six clusters with 314 marker genes (Supplementary file 3a). We matched the ICGS2-selected marker genes with vascular endothelial subpopulation marker genes presented in a study from Travaglini et al., 2019. This annotated the clusters into four subpopulations of endothelial artery, vein, capillary aerocytes, and general capillary cells (Supplementary file 3b). We further aggregated all cells of the same vascular endothelial subpopulations into ‘pseudo-bulk’ data library by library. DESeq2 was used to detect differentially expressed genes for each subpopulation between the control and IH-exposure groups. To select differentially expressed genes in each endothelial subpopulation, the cutoff was set as BHQ < 0.2 and fold change > 1.2 (or log2[fold change] > 0.26). Those differential expression genes (Supplementary file 3c) in endothelial capillary cells were used for BP enrichment analysis in the DAVID database. Selecting and aggregating biological processes (Supplementary file 3d) were performed as described in the Materials and methods section labeled, ‘Bulk RNA sequencing and analysis’. Transcription factor analysis for capillary aerocytes and general capillary cells were run using SINCERA (Guo et al., 2015; RRID:SCR_016563). The analysis for control vs IH-exposure groups were run separately. All cells from the three biological replicates were used for a combined analysis. Cell annotations from AltAnalyze were used to select cells annotated to aerocytes and general capillary cells, respectively. For each cell cluster, we selected for marker genes using the FindMarkers function in Seurat (Satija et al., 2015) and filtered with parameters pct.1 ≥ 0.2, avg_log2FC ≥ log(1.5), p_val_adj < 0.1. The filtered marker genes were used as the target genes for driving force analysis in SINCERA. The predicted key TF list for aerocytes and general capillary cells for control and IH-exposure groups are shown in Supplementary file 3e, f. A network regulatory figure is plotted using igraph (Csardi and Nepusz, 2006) package in R. The top 25 predicted TFs from control and IH-exposure groups were selected.

Gene–disease association information was downloaded from the DisGeNET (Piñero et al., 2017; RRID:SCR_006178) database (curated gene–disease associations). The downloaded file was filtered with keywords to specifically select genes linked to pulmonary diseases, which includes ‘allergic asthma’, ‘asthma’, ‘chronic obstructive airway disease’, ‘pulmonary hypertension’, ‘chronic thromboembolic pulmonary hypertension’, ‘idiopathic pulmonary arterial hypertension’, ‘familial primary pulmonary hypertension’, ‘idiopathic pulmonary hypertension’, ‘interstitial lung fibrosis’, ‘pulmonary fibrosis’, and ‘idiopathic pulmonary fibrosis’. One-tailed Fisher’s exact test was used to determine whether the top 200 up- or downregulated genes in each cell type exposed to IH significantly overlapped with these pulmonary disease-associated genes (Supplementary file 4a). The significant cutoff was set with a BH adjusted p-value from a Fisher’s exact test of <0.05 and at least five overlapped genes with any pulmonary disease gene set. For the association analysis between IH-responsive genes and pulmonary drug targets, the xml file was downloaded from the DrugBank (Wishart et al., 2008; RRID:SCR_002700) (version 5.1.4). The drugbankR package (https://github.com/yduan004/drugbankR; Duan, 2019) was used to parse the xml file to get each drug and its target genes. The parsed drug table was linked with the top 200 up- or downregulated genes in each cell type exposed to IH by drug target genes. The drug table was further filtered with respiratory tissue and disease-associated key words (e.g. asthma, lung, bronchus, airway, etc.) to keep candidate drugs used to treat pulmonary diseases. The filtered table was manually curated to select drugs mainly indicated to treat pulmonary diseases. In the association analysis, the ‘homologene’ package (https://github.com/oganm/homologene; Mancarci, 2019) was used to find the human-unique homolog of mouse genes. CellPhoneDB (RRID:SCR_017054) is used to predict ligand–receptor interactions between cell types. The pulmonary disease-relevant drug-targeted gene list (Supplementary file 4b) is used to select ligand–receptor interaction pairs. Genes with unknown drug effects were filtered out from the analysis. Human homolog gene is used for this analysis. Gain of interactions and loss of interactions are calculated by summing up the number of ligand–receptor interactions that were only significantly present in either hypoxia or control (p-value<0.05 and mean>0.1), respectively, for each cell pair comparison. Two-tailed Fisher’s exact test was used to calculate significance changes of each cell type. Hive plot is made using the HiveR package (https://github.com/bryanhanson/HiveR; Hanson, 2017). The gene set enrichment analysis (GSEA; Subramanian et al., 2005) was performed for myofibroblasts (Supplementary file 4c). The expressed genes in myofibroblasts were ranked by the fold change calculated from pseudo-bulk expression values in control and IH-exposure samples. The ‘homologene’ package was used to find the human-unique homolog of correlated mouse genes. The canonical pathway list was downloaded from MSigDB (http://www.gsea-msigdb.org/gsea/msigdb/collections.jsp). The fgsea R package (Korotkevich et al., 2016) was used to run GSEA on the ranked genes in myofibroblasts. The gene sets were restricted to REACTOME subsets with at least 10 genes but less than 500 genes. The ‘nperm’ was set to 100,000. The significantly enriched gene sets were selected by padj<0.15 and absolute NES values above 1.7. With this cutoff value, no gene sets with downregulated leading edge genes after exposure to IH are significantly enriched. The predicted key TFs for myofibroblasts (Supplementary file 4d) were performed in a similar way as with endothelial subpopulations. To highlight the IH-driven key TFs, the top 25 TFs predicted from lung myofibroblasts after IH exposure were compared to all significant TFs predicted from lung myofibroblasts in control samples.

For histological staining, mice were sacrificed, and lung inflation fixation was immediately performed. After exposure of the trachea and lungs, the trachea was cannulated with a scalp vein cannula (EXELINT, Redondo Beach, CA), and 10% neutral buffered formalin was gravity perfused into the lung at a height of 25 cm. After infusion, the lung was harvested and placed in formalin for 24 hr. Whole lung was then dehydrated in 70% ethanol and embedded in paraffin. For morphologic evaluation, 5 µm thick sections were cut from the paraffin blocks and stained with hematoxylin and eosin.

Immunofluorescence staining on 10% formalin-fixed mouse lung was performed on 5 µm thick paraffin-embedded tissue sections. Tissue slides were melted at 60°C for 2 hr, following rehydration through xylene and alcohol, and finally in PBS. Antigen retrieval was performed in 0.1 M citrate buffer (pH 6.0) by microwaving. Slides were blocked for 2 hr at room temperature using 4% normal donkey serum in PBS containing 0.2% Triton X-100 and then incubated with primary antibodies diluted in blocking buffer for approximately 16 hr at 4°C. Primary antibodies included ACTA2 (1:2000, Sigma–Aldrich; RRID:AB_262054), EMCN (1:200, R and D Systems; RRID:AB_2100035), FOXFI (1:100, R and D Systems; RRID:AB_2105588), HOPX (1:100, Santa Cruz Biotechnology; RRID:AB_2687966), MKI67 (1:100, BD Biosciences; RRID:AB_396287), POSTN (1:100, ABCAM), Pro-SFTPC (1:1000, Seven Hills Bioreagents; RRID:AB_2335890), SOX9 (1:100, Millipore; RRID:AB_2239761), and LYVE1 (1:100, ABCAM; RRID:AB_301509). Secondary antibodies conjugated to Alexa Fluor 488, Alexa Fluor 568, or Alexa Fluor 633 were used at a dilution of 1:200 in blocking buffer for 1 hr at room temperature. Nuclei were counterstained with DAPI (1 μg/ml) (ThermoFisher). Sections were mounted using ProLong Gold (ThermoFisher) mounting medium and coverslipped. Tissue sections were then imaged on an inverted Nikon A1R confocal microscope using a NA 1.27 objective using a 1.2 AU pinhole. Maximum-intensity projections of multi-labeled Z-stack images were generated using Nikon NIS-Elements software (RRID:SCR_014329).

Heterogeneous cell populations from paraffin-embedded immunofluorescent confocal images (60× magnification) were characterized in Imaris (Bitplane; RRID:SCR_007370), version 9.6. An average of 185 cells from control mice and 301 cells from IH-treated mice (n = 4 mice, six images per mouse) were counted specifically for proliferating cells (MKI67), endothelial cells (FOXF1), and alveolar type II cells (SFTPC) using spot detection or manual counting of cells. The cell percentage endothelial cells was calculated as the ratio between cells stained with FOXF1 and totally counted nucleus stained with DAPI in each image. Similarly, we got the cell percentages of AT2 and proliferating cells. All the remaining cells without FOXF1, SFTPC, and MKI67 staining or a cell stained with both MIK67 and SFTPC were taken as one composition group in the composition difference analysis. For testing the composition difference of endothelial, AT2, and proliferating cells in mouse lung between experimental and control groups, a centered log ratio transformation was performed on the percentage of each cell type before applying the t-test (two tailed). The statistical p-value from the t-test was adjusted with the BH method.

Alveolar wall thickness was quantified in Imaris (Bitplane), version 9.6, using the measurement tool for autofluorescence to define the alveolar walls. A total of 20 alveolar wall thickness measurements were taken for each 60× magnification confocal image (n = 2 mice per treatment, six images per mouse). Alveolar area and MaxFeret90 area were quantified in Nikon Elements, version 4.5. Maxferet90 is the distance measured orthogonally to the maximum feret chord. The larger Maxferet90 value indicates more openness in the alveolus. General analysis was run on maximum-intensity projection confocal images (20× magnification), n = 2 mice for control- and IH-treated mice, with three images per mouse. Autofluorescence with the TRITC channel was used to define alveolar areas. Thresholding was turned on, and binary processing was created for inverting, removing objects touching borders, and filtering an object area. Airways, arteries, veins, and alveolar ducts were filtered out using a minimum value of 50 and a maximum value of 4000 using the filter on object area binary. Feature area and MaxFeret90 values were exported using Microsoft Excel.