A map of signaling responses in the human airway epithelium

SUMMARY

Receptor-mediated signaling plays a central role in tissue regeneration, and it is dysregulated in disease. Here, we build a signaling–response map for a model regenerative human tissue: the airway epithelium. We analyzed the effect of 17 receptor-mediated signaling pathways on organotypic cultures to determine changes in abundance and phenotype of epithelial cell types. This map recapitulates the gamut of known airway epithelial signaling responses to these pathways. It defines convergent states induced by multiple ligands and diverse, ligand-specific responses in basal-cell and secretory-cell metaplasia. We show that loss of canonical differentiation induced by multiple pathways is associated with cell cycle arrest, but that arrest is not sufficient to block differentiation. Using the signaling-response map, we show that a TGFB1-mediated response underlies specific aberrant cells found in multiple lung diseases and identify interferon responses in COVID-19 patient samples. Thus, we offer a framework enabling systematic evaluation of tissue signaling responses. A record of this paper’s Transparent Peer Review process is included in the Supplemental Information.

eTOC

Receptor-mediated signaling is critical to tissue homeostasis and regeneration. This study maps the response of the airway epithelium to diverse signaling pathways, providing a systematic framework for studying signaling responses in a regenerative tissue. McCauley et al. apply this approach to identify mechanisms involved in epithelial differentiation and lung disease.

Graphical Abstract

INTRODUCTION

The proper cell type composition of tissues is established through the action of extra-cellular signaling pathways, and changes in signaling occur ubiquitously in disease¹. Establishing how different pathways modulate cell-type composition, organization and behavior therefore represents a priority in the fields of developmental biology and tissue physiology.

The question of how a tissue responds to extra-cellular signals is exemplified in the airway epithelium, a regenerative tissue exposed to continuous environmental stimuli yet demonstrating substantial long-term stability against perturbations^2,3 The airway epithelium is composed of a basal stem cell pool that gives rise to six distinct mature cell types with roles in host defense and mucociliary clearance from the lung: secretory club cells, mucin-rich goblet cells, multi-ciliated cells, pulmonary neuroendocrine cells (PNECs), tuft cells, and pulmonary ionocytes^2,4–7. The relative abundance of each of these six cell types is actively regulated and responds to diverse environmental insults: infectious and allergic stimuli lead to increased goblet cell numbers, while injury leads to differentiation of squamous cells at the expense of mucociliary cells^2,8 and some cells undergo an epithelial-to-mesenchymal transition (EMT)^2,8. The persistence of these cell states is associated with lung diseases². It is a long-standing goal to identify signals that induce these changes in airway epithelial composition, and to better understand the effects of different ligands on the different epithelial cell types².

Multiple extracellular signaling pathways modulate the composition of the airway epithelium including Notch, Wnt, transforming growth factor beta-1 (TGFB1), epidermal growth factor (EGF), bone morphogenetic protein 4 (BMP4), fibroblast growth factors (FGFs), and interleukins (IL13 and IL17)^2,8–13 Other physiological cues, including mechanical strain^14,15 and epithelial jamming^16,17 can also alter tissue composition. Several pathways have been shown to drive the primary modes of abnormal differentiation: persistent signaling through Notch, IL13, and IL17 induce goblet cell hyperplasia; persistent EGF signaling induces squamous metaplasia; and TGFB1 has been implicated in EMT and epithelial senescence associated with pulmonary fibrosis^{2,3,10,18–22}.

Building on this extensive work, we set out here to construct a signaling–response map that could address the following open questions: (1) we still do not know how all but the best-studied pathways alter the abundance of all cell types, including rare cell types of the epithelium. The categorization of abundance changes into squamous- and secretory-cell metaplasia may lose information on how the epithelium re-organizes in response to different signals, and few signaling pathways have been evaluated for their potential to regulate the frequency of pulmonary ionocytes, tuft cells and PNECs despite their potential roles in disease^5,23–26. (2) We do not know how each cell type changes in gene expression in response to different signals, and how plastic are the phenotypes of cells. Different ligands could induce similar (i.e. canalized or convergent²⁷) phenotypic responses in a cell type, or not, while individual ligands could induce similar responses in different cell types, or act in a pleiotropic manner. And finally, (3) tissue atlases of human disease have recently identified disease-specific ‘aberrant’ cell states in tissues by single-cell RNA sequencing (scRNA-Seq)^28–31 and it would be useful to relate these states to signaling pathways that may be potential targets for therapeutic intervention. A signaling–response map could address these three questions by defining changes in cell type abundance and gene expression in an unbiased manner, identifying changes not evident in the canonical gene repertoire used to study the tissue. We here make use of scRNA-Seq to construct such a version of such a map, encompassing responses to 17 signaling pathways.

To facilitate analysis of human disease atlases, it is desirable to study signaling responses in primary human cells. We made use of air-liquid interface (ALI) organotypic cultures of primary human bronchial epithelial cells (hBECs), which form a mucociliary epithelium containing physiological cell types. We evaluated changes in cell type composition in response to different signaling stimuli, and transcriptional responses to signaling. Through these analyses we identified stereotyped axes of variation in cell type composition, evaluated change in rare cell type abundance, defined convergent and unique transcriptional signatures of different pathways, and identified the cell types transcriptionally responding to different stimuli. We then explored the use of this signaling map to infer signal activation in lung diseases from patient tissue atlases. Thus, this study provides a framework for quantitative characterization of signaling responses, and it serves as a resource for predicting tissue-specific signaling signatures in diseases of the airway epithelium.

RESULTS

Modeling human airway epithelial regeneration under stimulation of signaling pathways

To prioritize potential signaling pathways regulating the airway epithelium, we shortlisted receptors with cognate ligands³² and transcript expression enriched in hBECs relative to other human tissues (Fig. 1A). We did so by comparing RNA-Seq data from the GTEx Portal³³ with scRNA-Seq data from human airway epithelial cells⁷, to identify 97 candidate receptors. From these we focused on those involved in immune, developmental, and hormonal signaling, and shortlisted a subset of 16 ligands and one chemical agonist known to interact with 31 of the 97 receptors (Fig. 1A, S1A,B). The selected ligands have previously been studied in varying contexts of airway epithelial differentiation and disease as summarized in Table S1. Analyzing expression of the selected receptors across different cell types in airway epithelial cells (Fig. S1C) showed that some of them are expressed both in basal and luminal cells.

Figure 1. A single-cell map of receptor mediated signaling induction in human airway epithelial cells.

(A) Approach for selection of signaling ligands for this study (for analysis details see Fig. S1). Red arrows show cognate ligand-receptor interactions.

(B) Schematic of the organotypic regeneration assay for evaluating changes in hBEC differentiation under signaling stimulation.

(C,D,E) UMAPs of scRNA-Seq data (GEO #GSE246368, n=3 donors per condition, 2 donors for FGF2, and 1 donor for IFNG, OSM) from (C) untreated cells, colored by cell type annotations, (D) all cells colored by treatment condition, and (E) all cells with control cells highlighted.

(F) Expression of marker genes in annotated cell types in the control data. Color scale represents scaled counts per 10k total counts (CP10k).

(G) De-mixing of treated cell states from untreated cell states quantified by the observed/expected ratio of nearest-neighbor cell fraction coming from control condition (100 nearest neighbors used). Lower values indicate separation of treated from untreated cells. Here and in all figures, CHIR = CHIR99021.

To test the effect of the selected signaling molecules on airway epithelial composition, hBECs were treated as shown in Fig. 1B: the cells were first differentiated without treatment into a pseudo-stratified epithelium that recapitulates the physiological cell types of the airway, after which the luminal cells were stripped through calcium depletion, leaving the remaining basal cells to regenerate the tissue. This technique was selected to bypass differences in outgrowth at initial plating, and to recapitulate injury-like conditions. We verified by scRNA-Seq at 6 hours post-stripping and by immunofluorescence immediately post-stripping that the remaining cells were depleted of mature cell types (Fig S1D, E), as previously reported^34,35. The remaining basal cells upregulate squamous gene programs including KRT17, KRT6A, IL1RN, and S100A1/S100A2, similar to the response seen in vivo after loss of luminal cells⁷.

Cells were allowed to regenerate post-injury in the presence of a signaling agonist added to each well at saturating dosage (concentrations in Fig. S1B). For 16 pathways we applied purified ligands, and for one – the canonical Wnt pathway – we applied a small molecule agonist, a GSK3 inhibitor (CHIR99021) (Fig. 1A). After 2 weeks of differentiation, the final composition of the tissue was analyzed by scRNA-Seq and imaging.

As a technical control for the efficacy of the ligands, we tested whether the expression of induced transcriptional targets (1–3 per pathway, Table S2) was significantly increased (Fig. S1F). For 14 out of 17 pathways, induced transcriptional targets increased after stimulation (family-wise error rate<0.05). Previously reported targets of three treatments (ActA, FGF2 and IGF1) did not change with statistical significance after multiple hypothesis correction, but they still showed an increase in average expression. We cannot rule out the possibility that these ligands did not activate their respective pathways.

Signaling responses extend the transcriptional landscape of the airway epithelium

After filtering for low quality cells, we obtained 77,568 cell transcriptomes. Of the 18 conditions (17 treatments, and one control) all treatments were represented by 3 donors with exception of FGF2, ActA, Leptin (2 donors) and IFNG, OSM (initially one donor). For IFNG and OSM, we carried out a second experiment contributing three additional donors and matched controls (23,983 additional cell transcriptomes). To obtain a first view of the data, we carried out donor-demultiplexing based on SNPs (Fig. S2A–C), performed batch correction between donors^36,37, and then generated UMAP embeddings for the control data (Fig. 1C) and for the full data set (Fig. 1D–E, S2A,B,D). In untreated controls, we observed clusters representing all major airway cell types (Fig. 1C), indicating that the ALI cultures after luminal stripping fully regenerate the mucociliary epithelium. The clusters expressed canonical markers as expected (Fig. 1F, Table S3) – KRT5+ basal cells, MUC5B+ secretory cells, FOXJ1+ multiciliated cells, FOXI1+/CFTR+ ionocytes, ASCL1+/SST+ PNECs and POU2F3+/ASCL2+ tuft cells (Fig. S2E). The untreated cells also defined two additional KRT5+ (basal-like) states: a KRT13^hi/TP63^lo state that expressed intermediate levels of basal and luminal keratins (KRT5 and KRT8 respectively) and variably expressed KRT4 (Fig. S2F); and a separate, rare group of cells (138/7564) expressing low levels of both basal and luminal keratins (KRT5/8) and enriched for KRT17 (Fig. 1C). The KRT13+ state does not relate to any classical airway epithelial state but appears homologous to a cell state found in mouse and human primary tissue samples^5,38 and may represent parabasal/suprabasal cells³⁹ that share a similar pattern of low TP63 expression and low/intermediate levels of KRT8^11–13,40. The KRT17^hi/KRT5^lo cell state did not resemble any cell state of the mucociliary epithelium sampled from healthy mice or humans by scRNA-Seq^5,38, but KRT17^hi cells have also recently been reported in scRNA-Seq analyses of bronchial samples from patients diagnosed with pulmonary fibrosis^28,31 and we have previously identified these states in HBEC ALI cultures⁷. KRT17 has served as a marker of injury and malignancy in epithelial tissues^41–43 and is associated with expression of antimicrobial small proline-rich proteins (SPRR genes)^44,45, which are indeed expressed by these cells (Fig. S2G). In total, the transcriptomes of the cell types emerging in the regenerating cultures, and their abundances, establish that the regeneration assay recapitulates physiological cell types and sets a baseline against which to interpret responses to signaling pathway stimulation.

The UMAP visualization of the full data set (Fig. 1D–E, S2D) offers a simplified first view of the many changes in response to the 17 signaling ligands. These plots suggest that several treatments gave rise to cells that were transcriptionally similar to untreated cells, while others (IFNA, IFNG and TGFB1) gave rise to cells that were not. However, the usage of UMAPs to construct 2D embeddings distort or even hide variation in complex tissues that are heterogeneous in differentiation state, cell cycle state, and signaling response. Hence, we looked at the differences between signaling responses by calculating the average density of untreated cell transcriptomes in high-dimensional gene expression space (Fig. 1G). This analysis confirmed and quantified the trends seen in the UMAP plots: the smallest and largest deviations from untreated states were respectively associated with those treatments showing the least and greatest separation from untreated states on the UMAP. For the treatments that showed the smallest effects (ActA, Leptin, HGF, IGF1, and Adiponectin), we re-treated ALI cultures at 3 different concentrations of the ligands and examined whether the ligands induced a dose-dependent response in canonical target genes by qPCR (Fig. S2H). Our results suggest that most of these pathways may show no response due to a lack of receptor stimulation, as no significant change in target gene expression was detected. Leptin was an exception, showing a response of a canonical target (FOS), despite leading to no observed change in the UMAP embedding (Fig. S2H). For those pathways that do show changes in UMAP embedding, in the following sections we dissect these changes as they manifest (1) in cell type abundance, and (2) in altering transcription in response to signaling pathway stimulation.

Mapping changes in cell type abundance upon signal pathway stimulation

To evaluate how pathway stimulation altered the abundance of epithelial cell types, we assigned the cell transcriptomes from treated conditions into annotated cell types using a classifier trained on the untreated cells (Fig. 2A). This classification was refined by k-nearest-neighbor voting, and a final filtering step based on a requirement that cells classified to each cell type show enriched expression for associated marker genes (Figs. 2B, S3A). A fraction of the cells appeared transitional between basal and secretory (Fig. 2B, S3B–D), and between secretory and multiciliated (Fig. 2B, S3E–G). These transitional cells expressed intermediate levels of marker genes associated with their respective mature cell types (Figs. S3D,G), and their transcriptomes were embedded in transitional regions of the UMAP plot (Figs. S3C,F). Transitional multiciliated cells also expressed genes associated with ciliary assembly and centrosome organization including CDC20B, CCNO and FOXN4 (Table S3), with a rare subset of the cells forming a discrete state high in these markers as seen previously⁷. These cells did not express elevated total transcript counts and thus do not appear to be doublets (Fig. S3H). We grouped these cells with mature secretory or multiciliated cells, and confirmed that doing so preserved the fold-changes in cell type abundances upon treatment that are analyzed below (Fig. S3I). With these annotations, we calculated the frequency of cell types in each condition (Fig. 2C).

Figure 2. Distinct modes of basal and secretory cell metaplasia in response to different signals.

(A) UMAP of all scRNA-Seq data from Fig. 1D (this and panel B use data GEO #GSE246368, n=3 donors per condition, 2 donors for FGF2, and 1 donor for IFNG, OSM), colored by canonical cell-type annotations learnt from untreated controls.

(B) Gene expression heatmap showing that the classified cells across treatments preserve expression of marker genes for their respective cell types. Each row represents a single meta-cell showing average expression of 10 nearest neighbors; classified cell types on left.

(C) Frequency of cell types after perturbation (logarithmic scale). Top: dynamic range of signaling-induced changes. This and panels (D-F, H-J,L) incorporate data (GEO # GSE246441, 3 additional donor replicates for OSM, IFNG, and control analyzed by 10X). Red=maximum; Blue=minimum; Black=untreated baseline. Bottom: heatmap of donor-averaged cell type frequencies in all conditions.

(D,E) First two Principal Components of the cell type frequency matrix, after per-donor normalization, showing (D) values for each treatment, and (E) cell type loadings. PC1 corresponds to basal cell metaplasia, and PC2 corresponds to goblet cell hyperplasia.

(F) Fold change in cell type frequencies for four conditions with highest PC1 values corresponding to loss of canonical differentiation. ND = Not detected

(G) Representative immunofluorescence images of cross-sections of differentiated HBEC cultures (15–50 individual cells represented per image) treated with indicated cytokines and stained for KRT5 (white), MUC5B (green), and acetylated alpha-tubulin (red). Scale bars, 25 μm.

(H) Log-Fold change goblet cell abundance for the three conditions with >2-fold increase in goblet cell frequency. Points = donors; bar = mean.

(I) Comparison of changes in frequency of goblet and club cells. Conditions inducing goblet cell hyperplasia are highlighted (black); remaining conditions shown in gray. See also Fig. S4B. Both axes in (I,J) are in logarithmic scale.

(J) Comparison of changes in frequency of goblet and multiciliated cells in the context of goblet cell hyperplasia. Colors as in (I). See also Fig. S4C.

(K) Gene expression heatmap showing differences in goblet cell states induced by IL13 and IL17. On the left we have single goblet cells where each column is a single meta-cell as in panel B. On the right we have plotted average expression across all goblet cells in control, IL17A and IL13 conditions. CP10k = Counts per total 10k counts.

(L) Comparison of changes in frequency of ionocytes and PNEC+Tuft cells across all conditions, indicating tandem variation in the frequency of rare cell types. See also Fig. S4D.

The cell type frequencies in Fig. 2C vary across multiple treatments as compared to their control values, and when considered independently multiple changes are found to be statistically significant (85 null hypotheses rejected by Fisher’s Exact test at 5% FDR, donors p-values integrated by Fisher’s method; Fig. S4A, Table S4). To identify patterns in these many changes, we carried out a principal component (PC) analysis of the normalized cell type frequency matrix (Fig. 2C, Table S4). The first two PCs account for 79% of donor-normalized, log-fold-change variation in cell type frequencies (Fig. 2D) and they spontaneously recapitulate two well-known axes of airway epithelial metaplasia: PC1 (58% of the variation) corresponds to expansion of basal-like cells at the expense of luminal cells, and PC2 (23% of the variation) describes goblet cell hyperplasia (Fig. 2E). The full observations, however, reveal differences within each of these two conditions, as we describe here.

PC1 shows an expansion of basal cells and associated suppression of normal luminal differentiation with increasing severity, IFNG → CHIR → BMP4 → TGFB1 (Fig. 2D) consistent with prior studies carried out individually for each of these pathways (Table S1), but with differences in the degree of loss of each cell type (Fig. 2F). CHIR-treated cells permitted club cell differentiation but repressed multiciliated and goblet cells; BMP4 led to a modest suppression of mucociliary cells; IFNG repressed secretory and rare cell differentiation while TGFB1 led to a near-total loss of all differentiated luminal cells. The effects of CHIR in this system recapitulated previous reports of loss of mucociliary differentiation of hBECs^46,47.

Whether these pathways only suppress the differentiation of secretory and multiciliated cells, or also lead to loss of rare cell types (ionocyte, PNEC and tuft cells), was until now not known. We found that all the above-mentioned conditions led to depletion of rare luminal cell types, and with CHIR and BMP4 depleting the rare cell types to a much higher extent than secretory and multiciliated cells [6% (CHIR) and 30% (BMP4) reduction in total mucociliary cell fraction, compared to 91% (CHIR) and 100% (BMP4) loss of total ionocytes, PNECs and tuft cell fraction] (Fig. 2F, S4A, Table S4). We confirmed the loss of luminal differentiation after regeneration with immunostaining of ALI cultures for luminal and basal markers for IFNG, TGFB1, CHIR and BMP4 (Fig. 2G).

PC2 identified several conditions leading to increased goblet cell abundance. As expected, the largest increase in goblet cell frequency was observed following IL13 treatment (36-fold expansion; FDR < 0.001), and we also observed expansion in response to IL17A¹⁰, and IFNA¹⁰ (Fig. 2H). The effect of IFNA is consistent with prior work suggesting it can stimulate secretory cell differentiation¹⁰, and contrasts with other reports showing reduced secretory differentiation in murine airway cells exposed to IFNA⁴⁸. It is possible that inconsistencies in prior work could be explained by differences in how goblet cells expanded: IL13-mediated goblet cell expansion occurred at the expense of club cells (Fig. 2I, S4B), but not multiciliated cells (Fig. 2J, S4C). By contrast, IL17A increased both club and goblet cell frequency while the frequency of multiciliated cells was reduced, while IFNA increased goblet cell frequency while both club and multiciliated cell numbers were reduced (Fig. 2I,J). Thus, these ligands may act at different stages of mucociliary differentiation, and their effect may be missed by staining for pan-secretory markers. Moreover, the goblet cells produced in each of these conditions were transcriptionally distinct and differed in expression of canonical goblet cell markers MUC5AC, SPDEF and MUC5B, as well as other genes (Fig. 2K). Together, these results suggest that goblet cell hyperplasia is not a monolithic phenotype: it encompasses multiple states of tissue composition and secretory cell phenotypes.

The response we observed to IL13 differed from previous reports^49,50, wherein IL13 repressed multiciliated cell differentiation. Differences in response may reflect our use of luminal cell-stripping, rather than directly culturing basal cells in the presence of IL13 as done in previous studies. To gain confidence in our observation we repeated IL13 treatment followed by immunostaining for MUC5AC and FOXJ1. MUC5AC staining increased after treatment but we did not observe a significant reduction in the fraction of FOXJ1+ cells (Fig. S4E), thus supporting the results seen by scRNA-Seq.

We also examined changes in the frequency of ionocytes, tuft cells, and PNECs across treatments. Multiple conditions led to a loss of these rare cell types (Figs. 2C, F) but none of the conditions we examined led to statistically significant increases in any of these cell types (Fig. S4A). Changes in the rare cell types could be hard to evaluate because of their low frequency (~1%, Fig. 2L), which reduces statistical power in identifying changes in their abundance in this study; however, we noticed that the ratio of ionocyte to tuft and PNECs remained roughly uniform across conditions (Fig. 2L, Pearson correlation R=0.85, S4D), consistent with their frequencies not being modulated independently of each other by any of the pathways studied here.

In summary, the analysis of canonical cell type abundances supports that (1) activation of several signaling pathways leads to loss of normal differentiation and expansion of basal-like cells with signaling-specific differences in the loss of different luminal cell types; that (2) goblet cells expand in several conditions that induce major differences in goblet cell gene expression, as well as differences in club and multiciliated cell frequencies; and that (3) none of the stimuli studied expanded rare cell types.

Convergent and unique transcriptional responses to signaling

We next sought to evaluate the transcriptional response to each signaling stimulus. Discovery of differentially expressed genes (DEGs) between treated and untreated cells for each cell type (rank-sum testing, 5% FDR; Table S5) revealed that the cell state consistently showing the largest changes after stimulation were those classified as basal (Fig. 3A, S5A). This bias in transcriptional response across all treatments may reflect the plasticity of undifferentiated basal cells to undergo alternative modes of differentiation. Among all conditions, the largest responses were seen upon treatment with IFNG, IFNA, IL13, CHIR and TGFB1 (> 100 genes, FDR < 0.05). IFNA stood out as inducing a response in all cell types, whereas five of the 17 ligands (ActA, Adiponectin, IGF1, HGF, FGF10) induced few differentially-expressed genes in any cell type.

Figure 3. Evidence of convergent and pathway-specific transcriptional responses including loss of canonical cell identity.

(A) Number of genes showing >2-fold differential expression in basal, secretory (club, goblet), multiciliated and rare (ionocyte, tuft, PNEC) cells following each treatment. Empty boxes = no genes; N/A = no cells present.

(B) Schematic for gene program analysis of signaling responses.

(C) Mean usage of eleven treatment-induced gene programs across all cells from each treatment condition the presence of convergent (Shared-1–3) and pathway-specific programs (remaining programs). A further nine control programs are shown in Fig. S5B. The heatmap is first column-normalized (sum=1) and then row-normalized (max= 1).

(D) Transcriptional responses vary across cell types as seen from the mean usage of signaling programs. For shared programs, usage is averaged across all treatments; for perturbation-specific programs, usage is calculated for cells from one perturbation.

(E) Loadings of top 10 genes for each of the signaling induced program (left) and of epithelial KRT genes (right) across all signaling programs; full table for gene loadings is provided in Table S6.

(F) Putative loss of canonical basal cell identity, but not mucociliary cell identities, is observed by the near-complete replacement of control transcriptomic programs (grey) in basal cells by induced programs (red) in response to CHIR, IFNG, and TGFB1. Other treatments are shown for contrast.

(G) Representative immunofluorescence images (300–400 individual cells per image) of whole-mount differentiated HBEC cultures treated with indicated cytokines and stained for F-actin (white), acetylated-alpha-tubulin (red), and DNA (Hoechst; blue). Scale bars, 50 μm. Bottom row conditions, corresponding to loss of basal cell identity in panel F, show cytoskeletal disorganization.

(H) Quantification of epithelial permeability measured by transit of lucifer yellow dye across the epithelial surface of cultures treated with indicated cytokines. Bars represent mean ± SEM, n = 3 HBEC donors (IFNG: n=2; IFNA: n=1) biological replicates shown as points, normalized to untreated fluorescence = 1. * = p-value ≤0.05 by Wilcoxon rank sum test.

To gain insight into the nature of these transcriptional responses, we factorized the full scRNA-Seq data matrix into programs with variable usage across cells by consensus non-negative matrix factorization (cNMF)⁵¹ (Fig. 3B). Unlike DEG identification, matrix factorization is agnostic to both perturbations and cell types and so it offers a tool to define gene programs that recur across multiple cell types or stimuli, without the typical loss of sensitivity seen in DEG analysis for rare cell types. We defined 20 cNMF programs from our dataset, of which 9 were associated with unperturbed mucociliary epithelium (Fig. S5B–E) and 11 programs induced by signaling perturbations (Figs. 3C–E, Table S6). We do not discuss the control programs further, and instead focus on the transcriptional programs induced by the signaling perturbations.

The 11 perturbation programs collectively describe transcriptional changes in response to all signaling stimuli. Their usage pattern across conditions reveals a logic that is simple: of these programs, three were induced by multiple perturbations (Fig. 3C, shared programs 1–3). The remaining programs appeared only in response to either one or two signaling conditions. We named these programs by the signaling condition that induces them (Fig. 3C).

The cNMF analysis also defines which cell types induced each of the different response programs (Fig. 3D, S5C), and the genes that define them (Fig. 3E, Table S7). The three shared programs were most enriched in basal cells (Fig. 3D). One of these programs (shared-3) was defined by increased expression of laminin genes, metallotheinin (MT2A), and inhibitors of TGF family members (Follistatin and Inhibin A), potentially indicating changes in extracellular matrix (ECM) properties and signaling (Fig. 3E). The two other convergent programs (shared-1 and shared-2) were enriched in genes observed in squamous epithelial tissues and which have been used as markers of squamous metaplasia in the airway: the Small Proline Rich Protein (SPRR) family in shared-1, and SPRR, DSP and KRT6A in shared-2 (Fig. 3E, Table S7). These shared programs may thus represent distinct forms or progressive stages of basal cell differentiation into a squamous-like epithelium, and indeed they were induced by ligands that led to loss of luminal cell types and emergence of squamous-like morphologies (BMP4, IFNG, TGFB1) (Fig. 2G, 3C).

For the ligand-specific programs, the response across cell types was more varied (Fig. 3D): the two programs induced by TGFB1 (TGFB1–1, TGFB1–2), the program induced by CHIR (CHIR), and one of the programs induced by IFNG (IFNG-2) were most strongly induced in basal cells. However, the IFNG programs were also induced in the few luminal cells still present after IFNG treatment, with one (IFNG-3) showing maximal expression in secretory cells. For IL13, the transcriptional response was maximal in goblet cells but the same program was also induced in club cells, multiciliated cells and basal cells. Thus, the IL13 transcriptional response is not directly a measurement of increased goblet cell numbers; it includes a large number of genes including the lipoxygenase ALOX15, whose activity promotes goblet cell differentiation in human airway⁵². For IFNA, the specific response was induced across all cell types, as expected from the DEG analysis (Fig. 3A). Of the genes upregulated by these programs (Fig. 3E, Table S7), we highlight that program TGFB1–2 identifies cells that induced canonical markers of epithelial-to-mesenchymal transition (SPARC, CDH2, FN1), whereas TGFB1–1 did not and corresponds to induction of diverse genes including cell cycle inhibitors (CDNK2B, GADD45A). Thus, these programs clarify the complexity in the responses to the different ligands, and they decouple convergent squamous-like responses from TGFB1-induced EMT-associated phenotypes and other pathway-specific phenotypes. The programs also clarify the pattern of expression of common markers used in tissue staining. Both TGFB1 programs, for example, induced expression of the cytokeratin KRT17 (Fig. 3E), but the map revealed KRT17 to also be induced by several other programs including the convergent program shared-2 (Fig. 3E, Table S7).

Loss of basal cell identity correlates with loss of epithelial barrier integrity

A question that can be asked from a systematic analysis of signaling responses is whether some ligands induce programs that qualitatively change cell identity. Definitions of canonical cell types historically depended on cell and tissue morphology, and on expression of marker genes including lineage-specifying transcription factors or unique structural proteins such as keratins. With access to whole-transcription information, we wondered whether the extent of remodeling of the cell transcriptome could offer an alternative and unbiased way of defining departures from canonical cell types. To formalize this idea, one can examine the expression of ‘control’ programs – those that specify the transcriptional state of cells in absence of treatment. Without treatment, control cNMF programs (defined in Fig. S3A and Table S6) composed >90% of the median transcriptome of luminal cells, and >80% of the median basal cell (Fig. 3F), while after treatment, control programs in basal cells exposed to TGFB1 and IFNG was almost entirely lost (median usage 13% and 11% respectively). BMP4, IFNA also led to somewhat reduced basal cell control program usage, but to a lesser extent, and CHIR represented an intermediate case. The residual luminal cells in all conditions continued to express control programs (>50% median usage in multiple conditions) (Fig. 3F) suggesting that luminal cells tend towards a more canalized identity in the face of perturbations as compared to basal cells. The near-complete loss of untreated transcriptional programs in basal cells suggests that TGFB1 and IFNG lead to a qualitative change in cell identity.

We expected that these large changes in whole-transcriptome state in response to TGFB1 and IFNG, together with the loss of luminal cells in these conditions, might be readily evident in the morphology and epithelial barrier function of the tissue after treatment with these ligands, but not after BMP4 or IFNA treatment. Staining the treated tissues for F-actin indeed revealed disorganization of the epithelial tissue in response to IFNG and TGFB1 (Fig. 3G). Further, dye-transport assays revealed a loss of epithelial integrity in response to these two treatments, while BMP4- and IFNA-treated cells had intact barrier function (Fig. 3H). As tight junctions form at the apical surface of polarized cells, this is consistent with BMP4/IFNA permitting differentiation of sufficient polarized luminal cells⁵³. CHIR showed loss of epithelial integrity comparable to that seen for IFNG and TGFB1 (Fig. 3G). The loss of basal cell identity as seen from whole-transcriptome analysis together with alterations in luminal cell abundance thus correspond to changes in epithelial cell organization.

Loss of differentiation induced by multiple pathways is accompanied by cell cycle arrest

Commonalities in the responses to perturbations can be used to identify shared mechanisms of action⁵⁴. Here, we noticed that the conditions that led to loss of luminal differentiation (TGFB1, IFNG, BMP4 and CHIR) showed a downregulation in a control cNMF program associated with cell cycle (Fig. 4A, S5B), and an upregulation of CDK inhibitors (Fig. 4B). One hypothesis is therefore that cell cycle arrest is associated with loss of mucociliary differentiation. Indeed, it has previously been observed that cell proliferation in airway epithelial cells is suppressed by BMP4⁵⁵, TGFB1^56,57, and IFNG⁴⁸. We tested here whether the decrease in cell cycle gene expression and increase in CDK inhibitor expression during regeneration after luminal stripping was indeed concurrent with cell cycle arrest by assaying cell number and EdU incorporation in these four conditions. Within 48 hours of epithelial stripping, we observed decreased epithelial cell density and loss of EdU incorporation (Figs. 4C–E) upon treatment with TGFB1, IFNG, BMP4 but not CHIR. Given this association between loss of normal differentiation and cell cycle arrest in three out of four conditions, we next asked whether cell cycle arrest is sufficient to alter differentiation. We arrested cell cycle in hBEC cultures using three different approaches: at the G1 phase via CDK4/6 inhibition (PD0332991); at G1/S and S phase using thymidine block for 48 hours; and again at G1/S and S phase using aphidicolin, an inhibitor of DNA synthesis⁵⁸. We then carried out Ca2+ depletion to strip the luminal epithelium as above, while maintaining cell cycle inhibition. All compounds inhibited cell division, as measured by EdU incorporation 48 hours following luminal cell stripping (Fig. 4F). We observed differentiation of secretory cells in all conditions measured by immunostaining for MUC5B and by qPCR (Fig. 4G,H) and saw no upregulation of markers of the convergent programs associated with squamous epithelia (Fig. 4H). Aphidicolin-mediated inhibition resulted in a loss of multiciliated cells measured by immunostaining and qPCR (Fig. 4G,H), but as thymidine block did not, we may conclude that the loss of mucociliary differentiation is unlikely to be cell cycle dependent, and that proliferation is not specifically required for normal luminal cell differentiation in ALI cultures regenerating after luminal cell loss. The sensitivity of multiciliated cells to aphidocolin is curious and could warrant further exploration.

Figure 4. Cell cycle arrests during signaling-induced loss of airway epithelial mucociliary differentiation, but is not required for differentiation.

(A) Fold change in cell cycle transcriptional program usage (defined in Fig. S5A) predicts a reduction in cell cycle in response to several signaling conditions.

(B) Expression of cell cycle inhibitor genes in control, BMP4, CHIR, IFNG and TGFB1 induced cells.

(C) Representative immunofluorescence images of whole-mount undifferentiated hBEC cultures (300–400 individual cells per image) treated with indicated cytokines and stained for DNA (Hoechst; white) indicate reduced cell densities in conditions showing reduced cell cycle programs and increased cell cycle inhibitor gene expression. Scale bars, 50 μm.

(D) Quantification of areal cell density from panel C. Bars represent mean ± SEM, n = 3 HBEC donors (BMP4: n=2), calculated from 1 mm stitched images. *p≤0.05 by Wilcoxon rank-sum test.

(E,F) Fraction of cells incorporating EdU after 48 hours of continuous EdU incubation (E) 48-hours following epithelial stripping and treatment with the indicated cytokines, or (F) 48-hours following epithelial stripping and treatment with aphidicolin (2 μg/mL), PD0332991 (PD033, 100 nM), or thymidine (thy.) block (2 mM). Bars represent mean ± SEM, n=3 HBEC donors.

(G) Top: Representative images for immunofluorescence of whole mount differentiated hBEC cultures (approximately 1000 individual cells per image) treated with aphidicolin, thymidine block, and PD0332991 and stained for MUC5B (red) and FOXJ1 (green). Bottom: Quantification of images to calculate percent MUC5B and FOXJ1 cells. Scale bars, 100 μm.

(H) Fold change of mRNA expression in differentiated HBEC cultures treated with aphidicolin/PD0332991 over untreated differentiated HBECs. Axis in logarithmic scale. Bars represent mean ± SEM, n=3 HBEC donors.

Predicting signatures of signaling in human disease

Systematic maps of signaling in human tissues could help identify pathways that induce tissue disorganization in disease. This can be done by comparing the features of primary patient tissues to unique signatures associated with each pathway. Such features could be based on imaging, but transcription has many advantages in being scalable and offering multiple dimensions through which to identify signaling responses. Conventionally, transcriptional changes observed in disease have been studied by gene set enrichment analysis, but universal gene sets^59–61 do not account for tissue-specific differences in signaling responses. We explored the use of transcriptional responses to signaling to generate hypotheses for signaling pathway activity in disease.

We developed a strategy to identify enrichment of the signaling-associated programs learnt from our data by gene signature enrichment⁶², here adapted to compare matched cell types from disease samples to their counterparts in control samples (Fig. 5A). We applied this strategy to published datasets of three lung diseases covering a total of 124 control or patient tissue samples from idiopathic pulmonary fibrosis (IPF), chronic obstructive pulmonary disease (COPD), and Covid-19 (Fig. 5B)^28–31.

Figure 5: Evidence of direct signaling-induced states in lung disease atlases.

(A) Schematic to infer signaling signatures in disease data by comparing matched cell states between disease and control samples. A signaling score is calculated using gene loadings from Fig. 3E for each cell state, and then compared by rank-sum testing with random downsampling of genes to ensure no single gene dominates the score. See Methods for detailed approach.

(B) List of disease datasets used for the analyses. Each dataset is assigned a color that is used to represent these datasets in the subsequent figure panels.

(C) Fold change increase in signaling program usage in disease versus control states across all diseases. Colors indicate dataset of origin for each cell state, as in B.

(D) Fold change expression in immune signaling programs in Covid-19 cells. Each point represents a cell state. The cell state annotations used are from the respective papers mentioned in B.

(E) Induction of TGFB1 signaling programs in cells identified as aberrant in the original papers, compared to other cells. Aberrant states include Krt17+/5−, aberrant basaloid, aberrant basaloid and ECM-high states in the 4 disease datasets respectively.

(F) Heatmap shows expression of top 20 genes of TGFB1–1 and TGFB2–1 programs in different states and datasets.

This strategy revealed a statistically significant enrichment of all eleven signaling-induced programs, relative to matched control samples, in a range of cell types across the four data sets (Fig. 5C). We highlight two observations that build confidence in the analysis. First, in the COVID-19 samples in particular, multiple cell types showed upregulation of inflammatory programs as expected during a viral response. However, in this disease not all immune programs were expressed: the IFNA and IFNG-1/2 programs were upregulated across multiple cell types after COVID-19 infection, but IL13 and IFNG-3 programs were not (Fig. 5D, S6). These observations suggest that our map indeed captures specific epithelial responses to viral infection and predicts which inflammatory pathways are most active. Second, the IPF and COPD disease data sets studied here have reported the presence of an aberrant cell state high in KRT17 expression. These cells have been suggested to locally activate TGF-beta through their expression of integrin α_vβ₆ subunits and through their localization directly lining myofibroblast foci^31,63,64. In addition, analyses of these data suggested an enrichment of TGFB1 signaling responses across the entire data set based on Gene Ontology enrichment⁵⁹, consistent with a central role for TGFB1 in fibrotic disease progression. We found here that the while TGFB1 programs were indeed broadly induced across airway cell types in IPF (Fig. 5C), the aberrant cells in these data sets expressed the two TGFB1-induced programs much more strongly than canonical cell types (Fig. 5C,E), and almost entirely recapitulated the program induced by TGFB1 in vitro (Fig. 5F). Notably, the gamut of genes enriched in this aberrant state including fibronectin (FN1), collagen 1A1 (COL1A1) and TGF-beta induced (TGFBI) arise uniquely from TGFB1 stimulation out of the pathways that we evaluated (Fig. 3E, 5F). This analysis also shows enrichment for other signaling pathway signatures in these aberrant cell state, with very similar patterns between the different IPF samples and COPD. By contrast, in COVID patient samples, an aberrant ‘ECM-high’ state has been reported. This state was enriched for only one of the two TGFB1 programs as compared to healthy lungs, and showed strong enrichment for an IFNG response that was absent in IPF and COPD. Thus, mapping signaling response in airway epithelium may provide a gateway to identifying direct cellular signaling responses in disease. The approaches used in this study can be expanded beyond airway epithelium to understand how signaling acts on other complex tissues, and drives the changes induced in disease.

DISCUSSION

In this study, we constructed a map of changes in cell type composition and cell type-specific gene expression in a regenerating culture of the human airway epithelium, in response to stimulation of 17 signaling pathways. We observed that the two principal axes of variation in cell type abundance after treatment recapitulate the primary forms of tissue metaplasia seen in diseases of the airway (Fig. 2): the loss of luminal differentiation, and goblet cell hyperplasia. However, the detailed changes in cell type frequencies and the gene expression programs induced by perturbation (Fig. 3) revealed far more granular phenotypes induced by signaling, including both convergent responses and unique signatures of several pathways evaluated here. In some cases, a single stimulus induced variable responses in different cells, as seen in the case of IFNG that induced three independent programs characterized by peak expression of either IL1RL1, or CXCL11, or B2M; and in the case of TGFB1 that induced an EMT response (FN1-hi) as well as a second program (CDKN2B-hi), and CHIR that induced both a convergent squamous-like program (Shared-1, high in SPRR genes) and a specific KRT6C-hi response (Fig. 3, Table S7), consistent with previous reports^46,47. We showed that the near-total loss of control transcriptional programs in IFNG and TGFB1 is associated with changes in epithelial organization seen by loss of epithelial integrity. We further demonstrated that multiple perturbation programs associated with loss of normal differentiation induced cell cycle inhibition, but that the latter is not sufficient to arrest differentiation (Fig. 4).

These maps also offer an opportunity to investigate the regulation of rare cell types including the FOXI1+ pulmonary ionocyte, which have not so far been evaluated in almost any signaling context. Previously we showed a requirement for Notch signaling in ionocyte differentiation⁷, however, little is known about the role of other signaling pathways in maintenance and differentiation of this or other rare cell types. Here, we found that no signals clearly led to increased differentiation into ionocytes, PNECs or tuft cells, and indeed the proportion of these cells was maintained across multiple signaling conditions, suggesting that their differentiation may be under shared control not explored here. However multiple signals led to loss of this population alongside a global repression of luminal cell differentiation. The approaches utilized here could help to identify signaling conditions that modulate rare cell type abundance in the lung, with potential therapeutic relevance.

While this study offers a strategy to define signaling responses during stem cell differentiation and identify these responses in disease, it has two types of limitations. First, it studies the epithelium in isolation from its native environment, and lacking interactions with immune cells, fibroblasts, smooth muscle cells and extracellular matrix. For human tissues, the possibility to assay complete physiological responses is practically limited, but some of the responses identified here could be evaluated in animal models, but most likely at lower throughput due to differences in ligand pharmacokinetics or use of specialized genetic models for each pathway. Second, the number of pathways evaluated is not exhaustive, and we have not evaluated the response of the ligands over a range of concentrations or durations of exposure. We have also not evaluated their combinatorial effects, or the effects of inhibiting signaling pathways, or the contribution of secondary stimulation of one pathway by another by simultaneously activating and inhibiting pairs of pathways. As single cell analytical tools have advanced in the last few years, and allow for systematic sample multiplexing, one can now consider extending the map in these directions to define the range of signaling responses occurring in human tissue maintenance, regeneration, and disease.

In spite of these limitations, these results offer a birds-eye view of the major axes by which the airway epithelium can be remodeled. They also provide a platform by which to interpret changes observed in scRNA-Seq atlases of human tissues. In the context of lung disease, our map established that a recently identified disease-specific (‘aberrant’) cell state directly recapitulates the expression response of airway basal cells to TGFB1. These cells have previously been suggested to locally activate TGFB1 through their expression of α_vβ₆ integrin subunits. Our results support the view that these cells not only have a potential to activate TGFB ligands, but that they are directly induced by these ligands, a signature of a positive feedback loop in TGFB signaling^63,64. This supports the hypothesis that lung fibrosis is sustained by a fibrotic cascade where abnormal signaling is perpetuated by the altered niche environment, and that epithelial cells may remodel their own niche^63,64. We expect that this map will be useful to interpret further studies of lung disease, and to generate hypotheses for how signaling pathways maintain aberrant cells states in these diseases.

STAR Methods

Resource Availability

Lead Contact:

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Allon Klein (allon_klein@hms.harvard.edu).

Materials Availability:

This study did not generate new materials.

Data and Code Availability:

• Source data statement: Processed next-generation sequencing data and associated metadata have been deposited at GEO and are publicly available. Accession numbers are listed in the Key Resources Table. Signaling perturbation data can also be explored on SPRING browser:

  • All conditions:

    • https://kleintools.hms.harvard.edu/tools/springViewer_1_6_dev.html?datasets/kalki/airway_signaling3/all_cells_umap
  • Unperturbed data:

    • https://kleintools.hms.harvard.edu/tools/springViewer_1_6_dev.html?datasets/kalki/airway_only_control/umap

• Code statement: All functions used for analyses are described below in Methods. Original code has been deposited at github.com/AllonKleinLab/paper-data/ (doi: 10.5281/zenodo.10602177).

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Mouse monoclonal to acetylated alpha-tubulin (clone 6–11B-1)	Millipore Sigma	T6793
Mouse monoclonal to MUC5AC (clone 45M1)	ThermoFisher	MS-145-P0
Purified anti-keratin 5 polyclonal chicken antibody	BioLegend	905903
Mouse monoclonal to MUC5B (clone 5B19–2E)	ThermoFisher	37–7400
Mouse monoclonal to FOXJ1 (IgG1)	ThermoFisher	14–9965-82
Rabbit polyclonal to FOXJ1	Millipore Sigma	HPA005714
Goat anti-Mouse IgG2b Cross-Adsorbed Secondary Antibody, Alexa Fluor 647	ThermoFisher	A-21242
Goat anti-Mouse IgG1 Cross-Adsorbed Secondary Antibody, Alexa Fluor 546	ThermoFisher	A-21123
Goat anti-Mouse IgG2b Cross-Adsorbed Secondary Antibody, Alexa Fluor 546	ThermoFisher	A-21144
Goat anti-Mouse IgG1 Cross-Adsorbed Secondary Antibody, Alexa Fluor 488	ThermoFisher	A-21121
Goat anti-Chicken IgY,	ThermoFisher	A-21449
Alexa Fluor 647 Phalloidin	ThermoFisher	A22287
Bacterial and virus strains
Biological samples
Human bronchial epithelial cells; see Methods Table 1 for detailed donor information)	Lonza	CC-2540
Chemicals, peptides, and recombinant proteins
Hoechst	ThermoFisher	H3570
Recombinant human Activin A	R&D Systems	338-AC-010
Recombinant human BMP4	Peprotech	120–05
Recombinant human EGF	ThermoFisher	PHG0311
Recombinant human FGF10	Sino Biological	10573-HNAE
Recombinant human FGF2	ThermoFisher	PHC9534
CHIR99021	Stem Cell Technologies	72052
Recombinant human HGF	ThermoFisher	PHG0254
Human IFN-Alpha (alpha 2A)	PBL Assay Science	11100
Recombinant human IFN-Gamma	Peprotech	300–02
Recombinant human IL-13	Peprotech	200–13
Recombinant human IL-17A	ThermoFisher	PHC9714
Recombinant human Oncostatin M	Peprotech	300–10
Recombinant human TNF-alpha	Peprotech	300–01A
Recombinant human Adiponectin	Peprotech	450–24
Recombinant human Leptin	R&D Systems	398-LP
Trypan Blue Solution	ThermoFisher	15250061
Triton X-100	Millipore Sigma	X-100
Tween-20	Millipore Sigma	P2287
16% Paraformaldehyde	Fisher Scientific	50–980-487
Lucifer Yellow CH dipotassium salt	Millipore Sigma	L0144
Aphidicolin from Nigrospora sphaerica	Millipore Sigma	A0781
Palbociclib (PD0332991)	Millipore Sigma	PZ0383
ProLong Diamond AntiFade Mountant	ThermoFisher	P36965
OptiPrep Density Gradient Medium	Millipore Sigma	D1556
Formalin solution, neutral buffered	Millipore Sigma	HT501128
Thymidine	Millipore Sigma	T9250
Normal goat serum	ThermoFisher	38172
Critical commercial assays
Click-iT EdU Pacific Blue Flow Cytometry Assay Kit	ThermoFisher	C10418
High-Capacity RNA-to-cDNA Kit	ThermoFisher	4388950
Taqman assay: FOXJ1	ThermoFisher	Cat. # 4331182, assay ID Hs00230964_m1
Taqman assay: MUC5B	ThermoFisher	Cat. # 4331182, assay ID Hs00861595_m1
Taqman assay: MUC5AC	ThermoFisher	Cat. # 4331182, assay ID Hs01365616_m1
Taqman assay: SCGB1A1	ThermoFisher	Cat. # 4331182, assay ID Hs00171092_m1
Taqman assay: SPRR1A	ThermoFisher	Cat. # 4331182, assay ID Hs00954595_s1
Taqman assay: SPRR3	ThermoFisher	Cat. # 4331182, assay ID Hs01878180_s1
Taqman assay: KRT6B	ThermoFisher	Cat. # 4331182, assay ID Hs00749101_s1
Taqman assay: EGR1	ThermoFisher	Cat. # 4331182, assay ID Hs00152928_m1
Taqman assay: FOS	ThermoFisher	Cat. # 4331182, assay ID Hs04194186_s1
Taqman assay: SOCS3	ThermoFisher	Cat. # 4331182, assay ID Hs02330328_s1
Taqman assay: FST	ThermoFisher	Cat. # 4331182, assay ID Hs01121165_g1
Taqman assay: HES1	ThermoFisher	Cat. # 4331182, assay ID Hs00172878_m1
Taqman assay: MET	ThermoFisher	Cat. # 4331182, assay ID Hs01565584_m1
Taqman assay: SOX2	ThermoFisher	Cat. # 4331182, assay ID Hs04234836_s1
Taqman assay: PPARG	ThermoFisher	Cat. # 4331182, assay ID Hs01115513_m1
TaqMan assay: 18S ribosomal RNA	ThermoFisher	Cat. # 4448481, assay ID Hs03928985_g1
TaqMan Fast Universal PCR Master Mix (2X), no AmpErase UNG	ThermFisher	4366072
RNEasy Plus Mini Kit	Qiagen	74134
Deposited data
Single cell RNA sequencing: ALI with 17 signaling conditions	This paper	GSE246368
Single cell RNA sequencing: Additional ALI data with control, OSM and IFNG	This paper	GSE246441
Single cell RNA sequencing: Timecourse of calcium depletion assay	This paper	GSE247613
Experimental models: Cell lines
Experimental models: Organisms/strains
Oligonucleotides
Recombinant DNA
Software and algorithms
ImageJ	National Institutes of Health	https://Imagej.nih.gov/ij/
Indrop.py pipeline	Zilionis et al. (2017)	https://github.com/swolock/indrops
ImageStudio (version 5.2)	LI-COR	https://www.licor.com/bio/image-studio/
Zen Blue	Carl Zeiss	https://www.zeiss.com/microscopy/en/products/software/zeiss-zen.html
scSplit	Xu et al. (2019)58	https://github.com/jon-xu/scSplit
Scanpy	Wolf et al. (2018)59	https://scanpy.readthedocs.io/en/stable/index.html
cNMF	Kotliar et al. (2019)38	https://github.com/dylkot/cNMF
Python 3.7 or above	Anaconda	https://www.python.org/
Custom code to generate analyses for this paper	This paper	http://github.com/AllonKleinLab/paper-data
Other
BEGM Bronchial Epithelial Cell Growth Medium BulletKit	Lonza	CC-3171
Bronchial Epithelial SingleQuots Kit	Lonza	CC-4175
96-well F-bottom black assay plate	Greiner	655076
12 mm Transwell with 0.4 μm Pore Polyester Membrane Insert, Sterile	Corning	3460
75cm2 U-Shaped Canted Neck Cell Culture Flask with Vent Cap	Corning	430641U
Micro Cover Glasses, Rectangular, no. 1	VWR International	48393–026
Frosted Micro Slides	VWR International	48312–004
Trypsin-EDTA (0.25%)	ThermoFisher	25200056
Phosphate buffered saline, pH 7.4	ThermoFisher	10010072
Fetal bovine serum	ThermoFisher	16000044
Hank’s Balanced Salt Solution	ThermoFisher	14025092
Bovine serum albumin	ThermoFisher	B14
Trypsin Neutralizer Solution	ThermoFisher	R002100
MEM	ThermoFisher	10370088
DMEM, high glucose	ThermoFisher	11965175

Experimental Models and Subject Details

Selection of receptors and pathways

Data sources and pre-processing.

A list of genes encoding receptors with at least one known ligand supported by published data were obtained from an annotated ligand-receptor database (Table S4 of Ramilowski et al.³²). Genes were considered if their receptor was paired with ligands and annotated as “known” and “literature-supported” in column “Pair.Evidence” of the database. The resulting list of receptor gene symbols is denoted R. Selection of receptors from this list was performed on scRNA-Seq data obtained from Plasschaert et al.⁷, GEO accession GSE102580. Cell-by-gene counts matrices were obtained from files GSE102580_filtered_normalized_counts_human.tsv and GSE102580_filtered_normalized_counts_human_viral_transduction.tsv, and cell type annotations were obtained from files GSE102580_meta_GSE102580_filtered_counts_human.tsv and GSE102580_meta_filtered_counts_human_viral_transduction.tsv. Cells with annotation “Basal” were pooled from both data sets, averaged across all cells, and then normalized to units of transcripts per million (TPM) to give the mean expression of gene $k$ in basal cells, $x_k^{(\text{basal)}}$. For comparison to other tissues, data were downloaded from the GTex Portal Bulk RNASeq database (gtexportal.org/home/datasets; file GTEx_Analysis_2017–06-05_v8_RNASeQCv1.1.9_gene_median_tpm.gct.gz). Data was normalized to TPM, and the average $m_k$ and standard deviation ${\sigma }_k$ expression for each gene $k$ across all tissues was calculated. A standardized expression of each gene $k$ in basal cells relative to GTEx tissues was then calculated as $z_k^{(\mathit{basal})}=(x_k^{(\mathit{basal}\text{)}}-m_k)/{\sigma }_k$.

Receptor-ligand selection and prioritization.

Genes were then filtered according to three criteria: (1) the gene is in the list $R$; (2) it has minimal expression $x_k^{\text{(}\mathit{basal}\text{)}}>10$ TPM; and (3) it is expressed at average or higher levels relative to other tissues $z_k^{(\mathit{basal}\text{)}}>0$. Receptors that passed these three filtering steps were then categorized based on known literature function into categories “Developmental signaling,” “Immune signaling,” “Hormone receptors,” “Matrix interactions,” and “Other,” with the latter including gene products implicated in cell-cell adhesion, cell motility, and proliferation as well as decoy receptors and death receptors. “Developmental signaling, ” “Immune signaling, ” and “Hormone receptors” were selected. Ligands cognate to the selected 31 receptors were selected from Table S4 of Ramilowski et al.³² and published literature. Where multiple ligands were available, those binding to multiple receptors were prioritized.

hBEC Culture and Cytokine Treatments

Primary human bronchial epithelial cells (HBECs) from normal donors were obtained from Lonza (Cat. # CC-2540) and were expanded for two passages with Growth Medium (BEGM supplemented with one SingleQuots kit, Lonza Cat. # CC-3171) in uncoated T75 flasks (Corning Cat. # 430641U). Donor information is cataloged in Methods Table 1. After expansion, cells were seeded on uncoated 12-well TransWell (Corning Cat. # 3460) plates at a density of 1e6 cells per plate and expanded for 7 days in Differentiation Medium (50:50 BEGM (Lonza Cat. # CC-3171)/high-glucose DMEM (ThermoFisher Cat. # 11965175) supplemented with one SingleQuots kit with no T3 and final retinoic acid concentration of 50 nM, Lonza # CC-4175). After 7 days, media was removed from the apical surface and cells were differentiated at air-liquid interface for an additional 30–45 days with media replaced every 48–72 hours. Cells were then washed 3X with warm 1XS PBS and treated with MEM (ThermoFisher Cat. # 10370088, 500 μL apical, 1 mL basal) for 1 hour at 37°C to remove differentiated luminal cells. Following incubation, luminal cells were removed by pipetting gently up and down 3–5X before removing apical media and washing with 500 μL MEM in apical chamber³⁵. Cytokines were diluted in Differentiation Medium at concentrations typically 10–100X higher than reported IC50 doses, as indicated in Supplemental Figure 1B; as we chose to only use one dose per ligand due to the technical demands of this approach, the intention was to saturate the signaling pathway to ensure receptor activation as best as possible. Cells were cultured for an additional 14 days, replacing media every 48–72 hours.

Methods Table 1.

Donor ID numbers and demographic information for human bronchial epithelial cell (hBEC) primary cells used for experiments (Lonza, CC-2540).

Donor ID	Age (years)	Sex	Lung disease status
429581	9	F	Healthy
221175	42	M	Healthy
323353	7	M	Healthy
134626	43	F	Healthy
646466	38	M	Healthy
627466	24	F	Healthy
033435	73	F	Healthy
347553	43	F	Healthy
073038	69	F	Healthy

Single-cell dissociation and capture for single cell sequencing

HBECs were collected 14 days post-injury by treating with 0.05% Trypsin-EDTA (prepared by dilution of 0.25% Trypsin-EDTA (ThermoFisher Cat. # 25200056) with calcium/magnesium-free PBS), for 20 minutes at 37 °C. Trypsin dissociation was halted with Trypsin Neutralizer Solution (ThermoFisher Cat. # R002100) at 1:1 dilution and cells were filtered through a 40 μm strainer and pelleted by centrifugation at 300× g for 5 minutes. Cells were resuspended in phosphate buffered saline (ThermoFisher Cat. # 10010072) with 0.1% bovine serum albumin (ThermoFisher Cat. # B14) and counted on a Countess Cell Counter (ThermoFisher) to determine cell number and viability. For cell counting, 10 μL of cell suspension was mixed with 10 μL of Trypan Blue Solution (ThermoFisher Cat. # 15250061) and 10 μL of that 1:1 mix was loaded into Countess chip and read with default program. Viable cell numbers were used to determine cell concentration. Optiprep Density Gradient Medium (Sigma Cat. # D1556) was added to achieve a final concentration of 15% Optiprep and 120–150,000 cells/ml. Cells were run in two separate experiments: donor #429581 was collected individually; donors #221175 and #323353 were pooled and captured simultaneously. We collected additional data from three new donors (#21TL033435, #21TL347553, #22TL073038) for validation of IFNG and OSM induced responses. Cells from these donors were also pooled for sequencing and demultiplexed later computationally.

The scRNA-Seq experiments for characterizing calcium depletion assay and testing the 17 signaling conditions were performed using inDrops⁶⁵, following previously described protocols⁷. Briefly, single cells and hydrogel beads tagged with barcoding primers for reverse transcription (RT) were encapsulated together in 3–4 nL droplets. RT was carried out in droplets in an emulsion. The emulsion was broken for library production consisting of (i) second strand synthesis, (ii) in vitro transcription for linear amplification, (iii) RNA fragmentation, (iv) RT, and (v) amplification by polymerase chain reaction (PCR). The resulting libraries were sequenced on NovaSeq and NextSeq Illumina platforms in paired-end mode at length of 2×76 base pairs and converted into FASTQ files using standard Illumina pipelines. Data was processed into BAM files using published pipelines⁷.

The additional scRNA-seq experiment for validation of response for OSM and IFNG was performed using 10X Chromium kit (Cat#1000269 and Cat#1000127). Resulting libraries were sequencing on NovSeq Illumina platform in paired-end mode as described in the 10X protocols (https://www.10xgenomics.com/support/single-cell-gene-expression/documentation). Analyses of this data is described in the methods section - Analyses of additional data collected for validation of OSM and IFNG response.

Single-cell RNA sequencing read processing, data filtering and normalization

Gene expression counts matrix was generated from raw FASTQ files using indrop.py pipeline⁶⁵ (https://github.com/indrops) using human genome assembly GRCh38/91 (genome assembly/ENSEMBL release). For donor lots #221175 and #323353 where cells were pooled prior to droplet capture (see above), the cell barcodes were then assigned to each donor respectively by their SNP profiles using the scSplit package⁶⁶. Out of all the pooled cells, we had 10.2% inferred doublets and 0.02% unassigned cells. To build confidence in our donor assignments, we also performed donor demultiplexing using Vireo package⁶⁷ and observed near complete overlap (> 96% in all conditions) in the assignments (Fig. S2C). To retain high quality transcriptomes, total count and mitochondrial count filters were applied. Transcriptomes with more than N total UMI counts and less than 40% of counts coming from mitochondrial genes were retained. The UMI threshold N was chosen separately for each library within the range 500–800 after manual inspection of the UMI/cell distribution per library. The total UMI counts per cell in each sample can be found in the metadata files _metadata.csv.gz. on GEO. For mitochondrial gene count calculations, genes with a gene symbol starting with “MT-” defined the list of mitochondrial genes. Subsequently, scRNA-Seq data were processed in Python version 3.7+ using the scanpy (sc) package (version 1.6+)⁶⁸. For normalization, genes expressed in less than 3 cells and with less than 6 total counts across all cells were filtered out using the function and parameters – sc.pp.filter_genes(min_cells=3, min_counts=6) and then cell counts were normalized to total 10,000 counts for each cell using sc.pp.normalize_per_cell(counts_per_cell_after=1e4) to define $x_{ij}$ as the expression of gene $j$ in cell $i$ in units of counts per 10,000 (CP10K). Counts were log10-normalized using the function sc.pp.log and default parameters.

Dimensionality reduction and donor integration of single cell data

Dimensionality reduction was carried out twice – one for untreated cells only (Fig. 1C,E), and one for the complete data set (Fig. 1D and all subsequent analyses). In each of the two cases, to reduce dimensionality, the top 4000 highly variable genes were identified using sc.pp.highly_variable_genes(flavor=‘cell_range’, n_top_genes=4000), and counts were z-scaled using sc.pp.scale to give $z_{ij}(\mathrm{\Omega })=\frac{{\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left(x_{ij}+1\right)-E_{\mathrm{\Omega }}\left({\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left(x_{ij}+1\right)\right)}{\sqrt{{\mathrm{V}\mathrm{a}\mathrm{r}}_{\mathrm{\Omega }}\left({\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left(x_{ij}+1\right)\right)}}$, where $\mathrm{\Omega }$ is the set of cells being analyzed (untreated cells or all cells), and $E_{\mathrm{\Omega }}(\cdot )$, ${\mathrm{V}\mathrm{a}\mathrm{r}}_{\mathrm{\Omega }}(\cdot )$ are respectively the mean and variance over cells $i\in \mathrm{\Omega }$. For untreated cells, this step was followed by Principal Component (PC) analysis with number of PCs = 50, and a k-nearest graph was constructed with correction for donor variation via the bbknn algorithm (batch balanced k-nearest neighbors)³⁶ (version = 1.31, function: bbknn.bbknn(batch =‘assigned_donor’, neighbors_within_batch = 10)). For the full data set, we tested several methods for donor batch correction, and found upon visual inspection that Harmony (version 0.05) adequately preserved within-donor variation. We constructed a batch corrected embedding with Harmony using harmony.run_harmony(vars_use=‘experiment’)function with PC coordinates as input. We then constructed a k-nearest neighbor graph with k=20 nearest neighbors in the donor-corrected embedding. In both cases, the embedded data was visualized into two dimensions using a UMAP projection using function: sc.tl.umap(random_state = 0) (Figs. 1C,D). Note: We used bbknn for untreated cells instead of using harmony since harmony resulted in artifactual clusters that were not observed when we visualized data without batch correction or visualized donors separately.

Annotation of untreated hBEC sample scRNA-Seq data

After performing dimensionality reduction and k-nearest neighbor graph construction (above), the untreated data was clustered using Leiden clustering⁶⁹ (function: sc.tl.leiden, default parameters). The clusters were manually annotated as: basal, KRT13^hi/TP63^lo, KRT17^hi/KRT5^lo, secretory, multiciliated or a cluster of “rare cells” including PNECs, tuft cells and ionocytes. Annotation was determined according to the mean expression of cell type-specific marker genes shown in Fig. 1F. The “rare cells” cluster was then subclustered and cells in the resulting clusters were annotated as ionocytes, tuft and PNEC cells after inspection for marker genes shown in Fig. 1F.

Composite gene scores

The subsequent sections make use of composite gene scores as defined here. For a set of cells $\mathrm{\Omega }$, we define the $\mathrm{z}$-standardized gene expression for gene $j$ in cell $i$ as $z_{ij}(\mathrm{\Omega })$ (see above for z-standardization). We then define the knn-smoothed value of $z_{ij}(\mathrm{\Omega })$ to be ${\bar{z}}_{ij}(\mathrm{\Omega },\mathrm{k})=\frac{1}{k}{\sum }_{m\in \left\{N_i(k-1),i\right\}}{Z}_{mj}(\mathrm{\Omega })$, where $N_i(k)$ is the set of $k$ nearest-neighbors of cell $i$ in the 50-dimensional PC space (see above). For a gene set $S=\left\{g_1,g_2,\dots ,g_N\right\}$, we define a composite gene score for cell $i$ as the average of the smoothed $\mathrm{z}$-standardized expression of the genes in the set, ${{\chi }_i}^{\left(S\right)}(\mathrm{\Omega },\mathrm{k})=\frac{1}{N}{\sum }_{j\in S}{\bar{z}}_{ij}(\mathrm{\Omega },\mathrm{k})$.

Marker gene identification

Marker genes (Table S3) were defined as previously⁶⁵. We reproduce the definition here for convenience and to define the specific method parameters: pseudocount $pc=0.1$, threshold fraction $f_{\mathit{min}}=0.01$, target FDR $\alpha =5\mathrm{\%}$, threshold max-to-second-max ratio $r=1.5$. Let $S$ be the set of all cells considered, partitioned into $n$ disjoint subsets $s_1,\dots ,s_n$. Define ${\bar{g}}_{ij}$ as the average expression (CP10k) of gene $j$ in subset $i$, and ${\bar{o}}_i^{(j)}$ as the values of ${\bar{g}}_{ij}$ for gene $j$ sorted in descending order. Define ${stepratio}_j^{(k)}(\bar{g}):=({\bar{o}}_k^{(j)}+pc)/({\bar{o}}_{k+1}^{(j)}+pc)$ as the ratio of the k-th highest average subset expression of gene $j$ to the $(k+1)$-th expression value, after adding a pseudo count of $pc=0.1$ (CP10k) to all average gene expression values.

A gene $j$ is a marker gene for cell subset $s_i$ if:

1. Gene $j$ is detected in fraction $f_{min}$ of cells in $S$.

2. Gene $j$ is statistically significantly higher in expression in subset $s_i$ compared to the complement set (all cells $S$ not in $s_i$). To establish significance, we used a two-tailed Mann-Whitney U test with multiple hypothesis correction, Benjamini Hochberg $\mathrm{F}\mathrm{D}\mathrm{R}<\alpha$.

3. Gene $j$ has maximal average expression in subset $i$, i.e. ${\bar{g}}_{ij}={\bar{o}}_1^{(j)}$.

4. Gene $j$ satisfies ${stepratio}_j^{(1)}(\bar{g})>r$.

Cell type classification in treated samples

To classify perturbed cells, a logistic regression classifier was trained on annotations of untreated data. Python’s machine learning package sklearn (version: 0.24.2+) was used with function and parameters: sklearn.linear_model.LogisticRegression(c_param=1, n_iterations = 1000). The 50 PC values learnt from untreated cells (see above) were used as features for classification. The annotations “basal”, “KRT13^hi/TP63^lo” and “KRT17^hi /KRT5^lo” were merged into one class (“basal”). The gene expression values for treated cells were standardized to the control samples $z_{ij}(\mathrm{\Omega }=\mathrm{u}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d}\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{s})$ and then projected onto the PCs of untreated cells. The trained classifier was applied to label each treated cell. Three subsequent steps were then carried out to (1) relabel the cells classified as “secretory” as either “goblet” or “club”, (2) Update annotations based on nearest neighbor labels and (3) identify cells with low-confidence assignment and relabel these as “indeterminate”. These steps are as follows:

Classifying goblet vs club cells.

A composite “goblet gene score” ${{\chi }_i}^{(S)}(\mathrm{\Omega }=\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{s},\mathrm{k}=10)$ was calculated with $S=\{MUC5AC,SPEDF\}$. Cells classified as secretory were sub-labeled as “goblet” if ${{\chi }_i}^{(S)}>1$, and “club” otherwise.

Update annotations based on nearest neighbor labels.

We performed knn-smoothing, that is, for every cell, we found 10 nearest neighbors in its 50-dimensional Harmony space and that cell was reassigned the state of maximum number of its neighbors. Python sklearn function sklearn. neighbors was used to find nearest neighbors of a cell.

Declassifying indeterminate cells.

For multiciliated cells, ionocytes, tuft cells and PNECs, a composite gene score ${{\chi }_i}^{(S)}(\mathrm{\Omega }=\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{s},\mathrm{k}=10)$ was calculated for each cell using the following gene sets:

Multiciliated: $S_{\mathit{multiciliated}}=$ FOXJ1, C9orf24, RSPH1, TPPP3 and SNTN

lonocyte: $S_{ionocyte}=$ FOXI1, TMPRSS11E, STAP1, CFTR and PDE1C

Tuft: $S_{\mathit{tuft}}=$ POU2F3, LRMP, NREP, CRYM and HOTAIRM1

PNEC: $S_{PNEC}=$ ASCL1, SST, GRP, CALCA and CHGA

Let $l_i$ be the annotated cell type label assigned to cell $i$ using the logistic regression classifier. A cell was re-labeled as “indeterminate” if it expressed its label-associated marker gene score below a threshold value ${{\chi }_i}^{\left(S_i\right)}<{\zeta }_{l_i}$. The threshold value ${\zeta }_{l_i}$ was manually determined for each annotation by comparing the distribution of ${{\chi }_i}^{\left(S_l\right)}$ between untreated cells with annotation $l_i=l$ to all cells with other annotations $l_i\ne l$. The two distributions for each cell type are shown in Fig S3A. Cells labelled as “indeterminate” were excluded from downstream cell type specific analyses in Figs. 2–3.

Fig. S3A also shows comparable gene scores for basal and secretory cells, using the gene scores:

Basal: $S_{\mathit{basal}}=$ KRT5, KRT13, S100A2, KRT14 and CSTA

Secretory: $S_{sec}=$ MUC5B, SCGB3A1, VMO1, BPIFA1 and BPIFB1

We further defined secretory cells as transitional secretory or mature secretory cells (Fig. 2B) based on the level of expression of secretory genes. Secretory cells are defined as transitional if ${{\chi }_i}^{\left(S_{\mathit{sec}}\right)}<1$ for $i\in$ secretory cells. Multiciliated cells were defined as transitional multiciliated cells if they expressed secretory marker genes. Thus, transitional multiciliated cells included cells ${{\chi }_i}^{\left(S_{\mathit{sec}}\right)}>1$ where $i\in \mathrm{m}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{l}\mathrm{i}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d}\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{s}$.

Fraction control neighbor analyses (Fig. 1G)

For each donor separately, let $C_0$ be the set of untreated cell transcriptomes in the data set post-filter, and let $C_t$ be the set of cells from treatment $t$. For each cell $i$, the set of 100 nearest neighbor transcriptomes $N_i(k=100)$ was determined by Euclidean distance on the Harmony-corrected principal component space (see section “Dimensionality reduction and with donor integration of single cell data”). Then, $M_{t,0}={\sum }_{i\in C_t}\left|N_i(k=100)\cap C_0\right|$ is the total number of control cells among the 100-nearest-cell neighborhood of cells in treatment $t$, and $M_{t,t}={\sum }_{i\in C_t}\left|N_i(k=100)\cap C_t\right|$ is the respective number of cells from treatment $t$. Defining the control fraction $f_t=M_{0,t}/\left(M_{t,t}+M_{0,t}\right)$, a null expectation for $f_t$ is ${\hat{f}}_t=N_0/\left(N_0+N_t\right)$ where $N_{\text{.}}=\left|C_{\text{.}}\right|$ are the total number of cell transcriptomes in each set respectively. This null represents the assumption that treatment does not alter cell states relative to untreated controls. The observed/expected ratio is $O/E=f_t/{\hat{f}}_t$. This quantity was calculated for each donor and treatment condition. The mean and SEM across replicate donors are plotted in Fig. 1G.

Gene expression heatmaps

For all gene expression heatmaps in the paper, gene expression was quantified as $z_{ij}(\mathrm{\Omega })$ [z-score of log₁₀(CP10k+1)], or as $\mathrm{z}$-score of CP10k as specified in figure captions. For reporting gene expression of single cells rather than cluster means (Fig. 2B, 2K), we plot graph-smoothed gene expression ${\bar{z}}_{ij}(\mathrm{\Omega },\mathrm{k}=10)$ as defined above. For Fig. 2B, the cells have been ordered first according to their cell type annotations and then within cell type, they have been ordered in decreasing order of mean expression of cell type specific marker genes. For Fig. 2K, goblet cells are first separated by condition and then ordered by decreasing order of MUC5AC expression.

Analyses of additional data collected for OSM and IFNG (Fig. S2B)

For IFNG and OSM, a second experiment was carried out contributing three additional donors (donor IDs: #21TL033435, #21TL347553, #22TL073038) and matched controls (23,983 additional cell transcriptomes after filtering). The data was processed using cell ranger pipelines (https://support.10xgenomics.com/single-cell-gene-expression/software/overview/welcome): cellranger mkfastq to make demultiplexed FASTQ files for the three libraries containing control, IFNG and OSM samples, and cellranger count with the argument --include-introns=false to generate count matrices from the FASTQ files. Donor demultiplexing was performed using Vireo v1.2.3 (https://github.com/single-cell-genetics/cellsnp-lite). Data filtering, normalization and dimensionality reduction was performed similar to original data as described in the methods sections – Single-cell RNA sequencing read processing, data filtering and normalization and Dimensionality reduction and donor integration of single cell data. Since, we only had 3 conditions, we manually annotated the data by clustering and assigning labels to each cluster as described in the section Annotation of untreated hBEC sample scRNA-Seq data. This was followed by further annotation of secretory cells as either goblet or club cells using goblet marker genes and updating annotations based on nearest neighbor labels as described Classifying goblet vs club cells and Update annotations based on nearest neighbor labels sections above.

In this dataset, it was not possible to distinguish ionocyte versus tuft cells versus PNEC cells. Hence, we annotated the rare cells as a single cluster called rare.

We did not use batch correction for any of the analyses described above. We used batch correction using bbknn (version = 1.31, function: b.bknn.b.bknn (batch = ‘assigned_donor’, neighbors_within_batch = 10)) for data visualization in Fig. S2B.

Testing and visualizing changes in cell type abundance (Fig. S4A, Table S4)

For p-values in Table S4, Fisher’s exact test was used to evaluate statistical significance in the changes in cell type frequency between treated and control samples separately for each donor. The resulting p-values across donors were combined by Fisher’s chi-squared method and controlled for false discovery rate by the Benjamini-Hochberg method. For Fig. S4A, we plot ${\mathrm{l}\mathrm{o}\mathrm{g}}_2\left(\frac{f_{tl}}{f_{0l}}\right)$ for each donor, where $f_{tl}$ corresponds to the frequency of cell type $l$ in the $t$-th treatment condition, and $f_{0l}$ is the respective cell type frequency for the untreated control. Where $f_{tl}=0$, we note “n.d.” (not detected). Line averages in Fig. S4A and bar chart values in Figs. 2F,H plot ${\mathrm{l}\mathrm{o}\mathrm{g}}_2\left(\frac{E_d\left[f_{tl}\right]}{E_d\left[f_{0l}\right]}\right)$ where $E_d[\cdot ]$ is an average over donors for which each treatment was carried out. For principal component (PC) analysis of changes in cell type abundance (Figs. 2C,D), we calculated $w_{tl}=E_d\left[{\mathrm{l}\mathrm{o}\mathrm{g}}_2\left(\frac{f_{tl}+e}{f_{0l}+e}\right)\right]$, where e=0.1% is a pseudocount. PC analysis was then carried out on the scaled values $\frac{w_{tl}-E_t\left[w_{tl}\right]}{\sqrt{{Var}_t\left[w_{tl}\right]}}$ using Python sklearn. decomposition. PCA (sklearn package version 0.24.2+), where $E_t$ and ${Var}_t$ are the mean and variance of each cell type / over all treatments $t$.

For the additional data collected for OSM and IFNG, since we could not separate rare cells into ionocyte, tuft and PNEC cells, we assigned equal number of cells in the three categories from the total number of rare cells.

Differential gene expression per cell type

Tests for differential gene expression were performed between cells from each treatment and untreated sample, for cells with matched cell type labels – basal, secretory (club + goblet) and rare cells (ionocytes + tuft + PNEC cells). Normalized filtered gene expression counts $x_{ij}$ (see above) were used for testing. Statistical significance was calculated by the Wilcoxon rank sum test python scipy (version 1.9.3), with multiple hypothesis correction controlling for false discovery rate using Benjamini and Hochberg’s method (python statsmodels package, version 0.13+). Gene fold-changes in expression per cell type, reported in Table S5, were calculated with a pseudo-count correction ${\mathrm{l}\mathrm{o}\mathrm{g}}_{2}FC={\mathrm{l}\mathrm{o}\mathrm{g}}_2\frac{E[x_{i\in C_{l,t},j}]+1}{E[x_{i\in C_{l,0},j}]+1}$ where $C_{l,t}$ is the set of cells with label $l$ and treatment $t$. The number of genes with ${\mathrm{l}\mathrm{o}\mathrm{g}}_{2}FC>1$ and FDR < 0.05 are reported in Fig. 3A, S5A.

Treatment program factorization using cNMF

Gene expression programs were learnt by consensus non-negative matrix factorization (cNMF)^51,70 using the Python package (cNMF, version 1.1) with parameters: number of components k = 20, percentage of replicates used as nearest neighbors for outlier detection = 30%, and local density threshold for defining outliers = 0.2. Usage and program matrices are provided in Tables S5, S6. The resulting program usages, $u_{ji}$, are normalized such that ${\sum }_j{u}_{ji}=1$ for programs $j$ in each cell $i$. The gene weights $w_{jk}$ are normalized such that ${\sum }_k{w}_{jk}=1$ for program $j$ over all genes $g_k$ (shown in Fig. 3E, Table S7).

Visualization of cNMF program usages (Figs. 3C,D,F S5B,C)

For visualization of treatment-specific usages in Figs. 3C, S5B, the program usages were averaged across all cells and then all donors for each condition: for every donor $i$, program $j$ and treatment condition $t$, program usage $U_{ijt}={\sum }_{\mathrm{i}}\frac{u_{ji\in C_{it}}}{N_{it}}$ where $C_{it}$ and $N_{it}$ are respectively the set of cells from treatment $t$ and donor $i$, and $N_{it}=\left|C_{it}\right|$. The average usages $U_{ijt}$ were then averaged across donors, $U_{jt}=E_{\text{donor}\mathrm{s}i}\left[U_{ijt}\right]$, and scaled for plotting ${\tilde{U}}_{jk}=\frac{U_{jk}-\mathrm{m}\mathrm{a}\mathrm{x}\left(U_j\right)}{\mathrm{m}\mathrm{a}\mathrm{x}\left(U_j\right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(U_j\right)}$ in Figs. 3C, S5B (top). For Fig. S5B (bottom), the fold-changes in usage shown for each treatment and each program are ${\theta }_{jt}={\mathrm{l}\mathrm{o}\mathrm{g}}_2{E}_{\text{donor}\mathrm{s}\text{i}}\left[\frac{U_{ijt}}{U_{ij0}}\right]$, were $t=0$ corresponds to the untreated controls. For visualization of cell type-specific usages of control programs (Figs. 3D, S5C), mean usages were similarly calculated, $U_{jl}={\sum }_{\mathrm{i}}\frac{u_{ji\in {C_l}^{\mathrm{*}}}}{\left|{C_l}^{\mathrm{*}}\right|}$, but now using the following cell sets ${C_l}^{\mathrm{*}}$. For S5C, we plot $U_{jl}$ only for cells from the untreated control samples, ${C_l}^{\mathrm{*}}=C_l\cap C_0$. For shared perturbation programs (Fig. 3D first panel), we plot $U_{jl}$ only for cells from the treated samples, ${C_l}^{\mathrm{*}}=C_l-C_0$. And for perturbation programs unique to specific signals (Fig. 3D, except first panel), we plot mean usage of only the cells from the respective signaling perturbation, ${C_l}^{\mathrm{*}}=C_l\cap C_t$.

Aggregate control program usage

To plot usage of control programs in specific conditions for different cell types (Fig. 3F), usages of all control programs was summed for each cell, $u_{ctrl,i}={\sum }_{j\in (\text{ctrl}\text{prog}.)}{u}_{ji}$. The median value of $u_{ctrl,i}$ across all cells of the indicated cell types (basal, secretory, multiciliated and rare cells) in the indicated treatment conditions was plotted. BMP4 and IFNG have ≤ 1 rare cells and hence are not included in the rare cell plot (Fig. 3F, last panel). Aggregate perturbed program usage in each cell is equal to $1-u_{ctrl,i}$.

Enrichment of signaling programs in disease (Fig. 5)

Human ScRNA-Seq data sets consist of an expression matrix $x_{ik}$ (units of CP10K) for cell $i$ and gene $k$, with annotation (state label) $l_i$ for the $i$-th cell. To define signaling responses, we considered changes in expression of gene sets $G_j(M)$ each consisting of $M$ genes. The genes comprising each set are the $M$ genes with the largest cNMF weights $w_{jk}$ for each program $j$ (defined above). To calculate a composition gene expression score for $G_j(M)$, we define a sparsified weight matrix ${\widehat{w}}_{jk}(M)=\left\{{\alpha }_M{w}_{jk}\right.\mathrm{f}\mathrm{o}\mathrm{r}k\in G_j(M);0\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\}$, with normalization factor ${\alpha }_M$ chosen such that ${\sum }_k{\widehat{w}}_{jk}(M)=1$. We used this score to report expression of signaling programs in disease datasets as in Fig. 5E. To calculate fold changes in the expression of signaling programs in disease and control donors (Fig. 5C, D), we further define a (randomly) downsampled sparsified weight matrix ${{\widehat{w}}^{\mathrm{*}}}_{jk}\left(M,{\mathrm{\Theta }}_m\right)=\left\{{\alpha }_M{w}_{jk}\right.\mathrm{f}\mathrm{o}\mathrm{r}k\in G_j(M)\left[{\mathrm{\Theta }}_m\right];0\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\}$ with ${\mathrm{\Theta }}_m$ being a random sample of $m$ integers between 1 and $M$ without repetition. Hereafter we use $M=20$ and $\mathrm{m}=7$, and for brevity omit argument $M$ for ${\widehat{w}}_{jk},{{\widehat{w}}^{\mathrm{*}}}_{jk}$. The randomly-downsampled signature score for program $j$ in a single cell $i$ is then calculated as a weighted sum ${s^{\mathrm{*}}}_{ij}\left({\mathrm{\Theta }}_m\right)={\sum }_k{{\widehat{w}}^{\mathrm{*}}}_{jk}\left({\mathrm{\Theta }}_m\right)\left(0.01+x_{ik}\right)$. To avoid any single gene dominating the signature scores, we bootstrap a distribution of ${s^{\mathrm{*}}}_{ij}$ for each cell by 100 instantiations of ${\mathrm{\Theta }}_m$. For each bootstrap, the magnitude of signaling response was calculated for each labeled state $l$ as $\delta \left({\mathrm{\Theta }}_m\right)={\mathrm{l}\mathrm{o}\mathrm{g}}_2\left(E_{i\in C_{l,d}}\left[{s^{\mathrm{*}}}_{ij}\left({\mathrm{\Theta }}_m\right)\right]/E_{i\in C_{l,h}}\left[{s^{\mathrm{*}}}_{ij}\left({\mathrm{\Theta }}_m\right)\right]\right)$ where $C_{l,d}$ represent the set of cells with matched state label $l$ in the disease $(d)$ and healthy $(h)$ donor samples respectively. Let ${\widehat{\mathrm{\Theta }}}_m$ be the median bootstrap, $\delta ({\widehat{\mathrm{\Theta }}}_m)=\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}[\delta ]$. The value of $\delta ({\widehat{\mathrm{\Theta }}}_m)$ response is plotted for each cNMF program $j$ (Fig. 5C). A one-tailed Wilcoxon rank-sum test was then carried out between the ${s^{\mathrm{*}}}_{ij}({\widehat{\mathrm{\Theta }}}_m)$ values for cells annotated with the same label $l$ between disease and healthy donor samples to find programs that are significantly upregulated in diseased cells. As an exception, for the IPF and COPD “aberrant basaloid” state and “KRT17+IKRT5-” state there are less than 0.15% corresponding healthy cells; for these states we instead carried out comparison to cells annotated as “basal” in healthy donor samples. The resulting p-values were then controlled by the Benjamini-Hochberg procedure and significance was determined at 5% FDR.

Analysis of calcium depletion by single cell sequencing

Cells from n=2 hBEC donors (#323353, #646466) were differentiated at air-liquid interface for 7 days and stripped by calcium depletion as described above. Cells were harvested at Day 7 pre-stripping, 6 hours following replacement of Differentiation Media post-stripping, and after 14 days of post-stripping outgrowth as single cells, as described above. Single cells were captured by inDrops, and raw data were processed to filtered, normalized counts tables as described above. Data from each donor were analyzed separately. After performing dimensionality reduction and k-nearest neighbor graph construction (as described above), the untreated data was clustered using Leiden clustering (function: sc.tl.leiden, resolution=0.4). Top genes for each cluster were identified (function: sc.tl.rank_genes_groups, default parameters) and clusters were manually annotated based on this gene expression as “basal,” “multiciliated,” and “secretory.” Multiciliated and secretory clusters were labeled as “luminal” and percentage of total cell number of basal vs. luminal cells was calculated for each donor. Data from both donors were compiled (function: sc.AnnData.concatenate, default parameters) and renormalized, and violin plots of top genes by rank_genes_groups at the 6-hour timepoint were plotted (function: sc.pl.stacked_violin, default parameters).

Immunostaining of ALI cultures

For immunofluorescence on cross-sections of TransWell cultures (Fig. 2), membranes were fixed in 10% neutral-buffered formalin overnight at room temperature (12–24 hours), then transferred to 70% ethanol solution, excised by scalpel and embedded in paraffin blocks using standard dehydration and processing approaches. 5 μm sections were cut on a tissue microtome, mounted on 1.0 mm glass slides, dried and de-paraffinized following standard protocols.

Whole mount cells were fixed in fresh 4% paraformaldehyde (PFA, diluted from 16% stock in 1XPBS) for 1 hour at room temperature, washed 3 × 20 minutes with PBS at room temperature, and stored at 4°C for no more than 14 days until staining.

Both sections and whole-mount cultures were incubated for 1 hour at room temperature in immunofluorescence buffer (IF Buffer, 130 mM NaCl, 7 mM Na₂HPO₄, 3.5 mM NaH₂PO₄, 7.7 mM NaN₃, 0.1% bovine serum albumin, 0.2% Triton X-100, and 0.05% Tween- 20) supplemented with 10% normal goat serum. Primary antibody was incubated overnight at 4°C, washed for 3 × 20 minutes with IF buffer, and secondary antibody plus 1:5000 Hoechst 33342 was incubated for 1 hour at room temperature. Cells were washed 3 × 20 minutes at room temperature in 1X PBS then mounted with coverslips with ProLong Diamond Antifade Reagent (ThermoFisher). F-Actin staining was performed by incubating with phalloidin dye (ThermoFisher) diluted 1:300 in 1XPBS for 30 minutes at room temperature. The stained samples were imaged on a confocal microscope: Axiovert 200 microscope (Carl Zeiss) with Yokogawa CSU-X1 spinning disc head and an Evolve 512 electron-multiplying charge-coupled device camera (Photometrics). Raw images were processed by adjusting brightness and contrast, merging channels, adding scale bars, and adding false-color, using ImageJ. Antibodies used for immunofluorescence are detailed in Key Resources Table.

Quantification of immunofluorescence by high-content imaging (Fig. 4G, S4E)

Immunofluorescence was quantified by imaging using a Yokogawa CV8000 high-content imager. ImageStudio version 5.2 (LI-COR) was used to quantify images. 10 Z-stack images were captured per well and projected to a single plane to the highest intensity regions; low-quality frames were manually removed. Nuclei were identified for each image by Hoechst 33342 staining and were set as each cell center. Intensity of each marker stain was evaluated and per-cell localization was called based on proximity to identified nucleus. Intensity and per-cell expression of each marker gene was normalized across 10 images per replicate stain and relative expression was calculated by subtracting the percentage of cells called for each marker gene from the control condition.

Permeability assay

Differentiated HBEC cultures were prepared from n=1–3 donors as above (Lonza, donors 134626, 646466, 627466). Following 1 week of differentiation, cultures were stripped as described previously and differentiated for an additional two weeks in media supplemented with CHIR99021, rhTGFB1, rhBMP4, IFNG, and IFNA, at concentrations listed Fig. S1B.

Cells were washed 3 x with HBSS and incubated for 1 hour at 37°C in HBSS (basal chamber) and HBSS + 0.1 mg/mL Lucifer Yellow (apical chamber, Sigma). 150 μL of media was transferred from the basal chamber to a 96-well assay plate (Greiner) and measured in a spectrofluorometer (Clariostar, BMG LabTech; excitation=485 nm, emission=535 nm). Data were normalized to a blank well containing HBSS and the resulting value from the control well was set to 1.

EdU measurement in cytokine-treated cultures

Differentiated HBEC cultures were prepared from n=3 donors as above (Lonza donor #s 134626, 646466, 627466). Following 1 week of differentiation, cultures were stripped as described previously and cultured for 48 hours in media supplemented with CHIR99021, rhTGFB1, rhBMP4, or IFNG, at concentrations listed in Fig. S1B. After 48 hours, cells were treated with 2 μM ClickIt EdU labeling reagents (ThermoFisher) diluted in culture media for an additional 48 hours at 37°C. EdU incorporation was measured by flow cytometry on a Cytoflex cytometer (Beckman).

Nuclear quantification in cytokine-treated cultures

Differentiated HBEC cultures were prepared from n=3 donors as above (Lonza donors 134626, 646466, 627466). Following 1 week of differentiation, cultures were stripped as described previously and cultured for 14 days in media supplemented with CHIR99021, rhTGFB1, rhBMP4, or IFNG, at concentrations listed in Fig. S1B. Cells were fixed with 4% PFA and nuclei were stained with Hoechst dye (1:5000 dilution, ThermoFisher) at room temperature for 1 hour. Membranes were excised using 0.8 cm biopsy punch and mounted on 1.0 mm glass slides with glass coverslips (#1, 30×22 mm at 13 mm thickness) in ProLong Diamond Mounting reagent. One BMP4-treated sample was lost during mounting, so BMP4 data is presented as n=2. For each condition, nuclei were imaged on a confocal microscope using automated built-in stitching software with starting point at center of mounted membrane and 10% overlap between images (Zen Blue Microscopy software; Carl Zeiss) to generate 3260 × 3260 μM 20x images. Raw images were processed and analyzed using ImageJ.

1. Loaded in full stitched image and ran auto-adjust brightness/contrast (built-in ImageJ function)

2. Automatically applied built-in ImageJ functions for smoothing, thresholding (Huang fuzzy thresholding method⁷¹), and watershedding. This approach is used to automatically distinguish closely-grouped nuclei to allow counting of individual cells.

3. Cropped image to 1 mm² to eliminate edge effect variability across samples

4. Calculated centroid measurements of minimum size 1 pixel² (1 centroid = 1 cell)

Effect of cell cycle inhibitors on differentiation

Differentiated HBEC cultures were prepared from n=3 donors as above. Cells were pre-treated for 48 hours with 2 μg/mL aphidicolin (Sigma) or 1 μM PD0332991 (Sigma), then stripped as described above. Media was replaced every 48 hours to ensure sustained cell cycle inhibition over the course of differentiation. 11–14 days post-stripping, cells were treated with 2 μM ClickIt EdU labeling reagents (ThermoFisher) for 48 hours, and EdU incorporation was measured by flow cytometry on a Cytoflex cytometer (Beckman). Cells were harvested for immunofluorescence by fixation with 4% PFA and stained as above with primary antibodies against MUC5AC, FOXJ1, and acetylated alpha-tubulin. Antibody information can be found in Key Resource Table.

Reverse Transcriptase Quantitative RT Polymerase Chain Reaction (qRT-PCR)

Cells were lysed with RLT Buffer (RNEasy kit, Qiagen), RNA was extracted following RNA was extracted following manufacturer’s protocol, and 1 ug of RNA was transcribed to cDNA using reverse transcription reagents (High-Capacity RNA-to-cDNA kit, Thermo). cDNA was diluted between 1:4 and 1:6 and 4 μL of cDNA was added to each 10 μL TaqMan Fast Universal PCR Master Mix (ThermoFisher) qPCR reaction. Gene expression was detected using TaqMan (ThermoFisher) probes (gene-specific primer information in Methods Table). Relative gene expression for each sample was calculated using the 2^(−ΔΔCT) method by normalizing the cycle number (Ct) for each sample to an 18S control and to a control sample, where baseline of control was defined as fold change = 1. TaqMan assay probes used are detailed in Key Resource Table.

Supplementary Material

3

Table S3. List of cell type specific marker genes, related to Figure 1.

Worksheets:

Sheet S3–1 (Mean expression): Mean gene expression per annotated cell state for unperturbed cells. Expression is defined as counts per total 10k counts (CP10K). The raw data and associated metadata is uploaded to GEO (#GSE246368). We include only genes expressed in >=3 cells and that have >=6 total counts (before normalization).

Sheet S3–2 (Marker genes): Scoring marker genes in unperturbed cells. These represent genes that are uniquely expressed in that state. Columns: Log2 fold change: The log fold-change of expression in the specified cell state vs the cell state that has second maximum expression (max-to-second-max ratio). P-value: calculated using rank sum test and then corrected to control for FDR using the Benjamini-Hochberg method (see “Marker genes” methods section for details on calculation).

4

Table S4. Fold change abundances and frequencies of cell types, related to Figure 2.

Worksheets:

Sheet S4–1 (Fold changes): Log₂ fold change of relative abundance of cell types between signaling condition and untreated sample averaged between donors. Donor specific p-value was calculated using Fisher’s Exact test and integrated by Fisher’s method. FDR calculated using Benjamini Hochberg method.

S4–2 (Frequencies donor 1): Frequencies of cells in all annotated cell type for donor 1

S4–3 (Frequencies donor 2): Frequencies of cells in all annotated cell type for donor 2

S4–4 (Frequencies donor 3): Frequencies of cells in all annotated cell type for donor 3

S4–5 (Frequencies donor 4): Frequencies of cells in all annotated cell type for donor 6

S4–6 (Frequencies donor 5): Frequencies of cells in all annotated cell type for donor 7

S4–7 (Frequencies donor 6): Frequencies of cells in all annotated cell type for donor 8

Donors 4,5 and 6 correspond to additional data collected for IFNG and OSM

5

Table S5. Differentially expressed genes upon stimulation of 17 signaling pathways, analyzed separately for each canonical cell type, related to Figure 3.

These include number of genes showing >2-fold differential expression and FDR < 0.05 (rank sum test, Benjamini Hochberg correction) in basal, secretory (club, goblet), multiciliated and rare (ionocyte, tuft, PNEC) cells following each treatment. See attached Excel spreadsheet.

6

Table S6. Usages of cNMF programs in our dataset, related to Figure 3.

Worksheets:

S6–1 (Averaged usage): Usage averaged per condition and annotated cell type.

S6–2 (Usage per cell): Usage per cell for all cells in the data

7

Table S7. Gene loadings for cNMF programs for 3000 highly variable genes, related to Figure 3.

Worksheets:

S7–1 (Top 20 genes): Top 20 genes for each program

S7–2 (All gene loadings): Gene loadings for all 3000 genes

HIGHLIGHTS.

• Single-cell “signaling-response maps” link ligand exposure to changes in cell state

• We constructed a signaling-response map for 18 ligands acting on the airway epithelium.

• Ligands induce shared and unique responses, modifying differentiation and cell cycle.

• The map predicts ligands elevated in lung diseases including IPF and COVID-19.