Lineage-specific intolerance to oncogenic drivers restricts histological transformation

Abstract

Lung adenocarcinoma (LUAD) and small cell lung cancer (SCLC) are thought to originate from different epithelial cell types in the lung. Intriguingly, LUAD can histologically transform into SCLC following treatment with targeted therapies. Here we designed models to follow the conversion of LUAD to SCLC and found the barrier to histological transformation converges on tolerance to Myc, which we implicate as a lineage-specific driver of the pulmonary neuroendocrine cell. Histological transformations are frequently accompanied by activation of the Akt pathway. Manipulating this pathway permitted tolerance to Myc as an oncogenic driver, producing rare, stem-like cells, transcriptionally resembling the pulmonary basal lineage. These findings suggest histological transformation may require the plasticity inherent to the basal stem cell, enabling tolerance to previously incompatible oncogenic driver programs.

One-Sentence Summary:

By modeling histological transformation of lung cancer, we uncover neuroendocrine-specific tolerance to Myc as an oncogenic driver.

Histological transformation (abbreviated hereafter as “HT”) is a poorly understood process whereby a cancer’s initial histology is altered and presents as a new histologic type of cancer. These changes are presumed to be under selective pressure of an oncogene-targeted therapy. Best described in the context of EGFR inhibition in lung adenocarcinoma (LUAD) (1–3) and androgen receptor inhibition of prostate adenocarcinoma (4, 5), these transformations most commonly lead to neuroendocrine and/or squamous differentiation (6). Considered “off-target” forms of acquired resistance, the cancer’s proliferative character is no longer dependent on the original oncogenic driver pathway. In the earliest reports of HT, it was unclear whether the recurrent tumors were independent, primary small cell lung cancer (SCLC), sub-clones selected for under the pressure of targeted therapy, or products of the direct conversion of LUAD into SCLC. Supporting evidence for direct conversion relies on the continued presence of a LUAD driver oncogene within a cancer that is histologically SCLC. Oncogenic mutations in the genes that commonly drive LUAD have rarely been encountered in genome sequencing studies of primary, treatment-naïve SCLC (7, 8). Critically, some post-transformation samples have shared, clonally related mutations present in the LUAD prior to HT; however, protein expression of the LUAD driver is conspicuously absent post-transformation (9). Furthermore, loss of the RB tumor suppressor is observed in cases of HT in both lung and prostate cancers (4, 10) and patients with EGFR-driven LUADs harboring inactivation of TP53 and RB1 are at higher risk for developing SCLC following targeted therapy (1, 11).

Primary LUAD and SCLC are thought to develop from distinct cell types in the lung: the alveolar type II (AT2) cell and the pulmonary neuroendocrine cell (PNEC), respectively. Much is known about surfactant-producing AT2 cells as a cell of origin for LUAD, for which the epidermal growth factor (EGF) signaling pathway is an established mitogenic program (12–14). Hence, activating mutations and amplifications of genes that encode proteins in the mitogen activated protein kinase (MAPK) pathway, including RAS and EGFR are common in LUAD (15). In contrast, SCLC is thought to arise predominantly from pulmonary neuroendocrine cells (PNECs) (8, 16, 17). PNECs are rare, found near anatomic branchpoints of the large airway and function as sentinels for inhaled pathogens and environmental changes – signaling to other cells through both electrochemical innervation and secretory function (17–19). Due to their scarcity, much less is known about PNEC biology, including what signaling events drive their proliferation.

Combining newly-developed models of genetically engineered lung tumorigenesis, lineage tracing and chronologic single cell RNA-sequencing, we interrogate the dynamic interplay between cell of origin and oncogenic driver programs and their contributions to HT. Through this, we address several fundamental unknowns: (i) how an adenocarcinoma transforms to a high-grade neuroendocrine cancer, (ii) what the intermediate steps in HT are, and (iii) what the oncogenic driver program of the pulmonary neuroendocrine lineage is.

Results

Histologically distinct lung tumors in an isogenic mouse model

To chronicle HT, we generated a new genetically engineered mouse model that combines the conditional expression of Myc, rtTA3 and tdTomato (lox-stop-lox alleles) with the loss of tumor suppressors Rb1 and Trp53 (floxed alleles) and a doxycycline (DOX)-inducible, oncogenic EGFR transgene (20). The model (herein abbreviated “ERPMT”) allows for two classes of manipulations: (i) control of the cell of origin, through lineage-restricted expression of Cre recombinase following intratracheal infection using adenoviral vectors with cell type-specific promoters (21), and (ii) oncogenic EGFR^L858R expression, controlled in a DOX-dependent manner (Fig. 1A). Tumors initiated in the AT2 or PNEC lineages in ERPMT mice produced histologically distinct lung tumors with opposing dependencies on the presence or absence of DOX. If expression of Cre was initiated in the AT2 lineage and the mice received DOX, the ERPMT model developed an aggressive LUAD. In contrast, no mice succumbed to disease in this timeframe if off DOX. In mice receiving DOX, we observed multifocal, glandular lesions consistent with LUAD throughout all lobes of the airway, demonstrating homogenous expression of tdTomato (tdTom) and lack of the neuroendocrine marker synaptophysin (Fig. 1B). Conversely, if expression of Cre was initiated in the PNEC lineage, then the cohort off DOX developed an aggressive, neuroendocrine SCLC and mice on DOX appeared healthy at a time when 100% of the cohort off DOX was moribund (Fig. 1C).

Figure 1. A GEMM model to generate distinct histologic subtypes of lung cancer.

A) Nomenclature used for the ERPMT model with abbreviated alleles bolded/underlined and general strategy to use promoter-restricted adenovirus to initiate tumorigenesis in specific airway cells; B) Survival and histologic appearance of the ERPMT model initiated in alveolar type 2 cells (AT2; red cartoon cell) or C) pulmonary neuroendocrine cells (PNEC; blue cartoon cell) when mice are on doxycycline (DOX) containing (dashed line; n = 10 per group) or control diet (solid line; n = 10 per group). Representative sagittal H&E and tdTomato (“tdTom”) immunohistochemistry (IHC) lung sections alongside high powered (100um scalebar) H&E and synaptophysin (Syp) IHC from the LUAD model initiated from AT2 cells on DOX (above) or the SCLC model initiated from PNEC cells off DOX (below). D) Mean imputed expression of AT2 and PNEC lineage markers (table S1) (77) in single cells isolated from LUAD (red; n = 5,394 cells) or SCLC (blue; n = 4,371 cells) ERPMT models with overlaid kernel density estimates reflecting cell density; tdTom+ tumor cells sorted and pooled from n = 3 mice at 8wks post-infection.

LUAD and SCLC tumorigenesis developed with similar penetrance and latency in the ERPMT model. To compare these models transcriptionally, we isolated tdTom+ cells in the on DOX group from the AT2 lineage (labeled hereafter as LUAD; red) and the off DOX group from the PNEC lineage (labeled hereafter as SCLC; blue) at 8wks on study and performed single cell RNA-sequencing. As expected, cells from these tumors were transcriptionally distinct (fig. S1A) and closely resembled their precursors (Fig. 1D and fig. S1, A to B). LUAD cells exhibited mixed AT2 and AT1 character, consistent with prior observations in the regenerating airway (22) and in murine (12) and human tumors (23). Compared to LUAD, SCLC had lower expression of components of the major histocompatibility (MHC) class II antigen pathway (fig. S1C); consistent with MHC class II deficiency in primary SCLC (24, 25) and AT2-specific expression of MHC Class II molecules (26). Instead, most SCLC tumor cells expressed factors that suppress Notch signaling, including the inhibitory ligand Dll3 and transcription factor Hes6 (27). Conversely, the LUAD model was enriched for Notch stimulating factors, consistent with this model presenting as a non-neuroendocrine, alveolar-derived LUAD (fig. S1D). Additionally, bulk ATAC-sequencing revealed regions of chromatin that mapped to Notch2, Hes1, and Sftpc were differentially accessible in LUAD compared to SCLC (28, 29). In contrast, the SCLC model had greater accessibility of neuroendocrine genes, including Insm1, Chga, and Ascl1 (fig. S1E). Furthermore, differentially accessible peaks in the MAPK-driven LUAD were significantly enriched for AP-1 motifs like Jun and Fos whereas several bHLH regulatory elements involved in neurogenesis (Tcf21, Tcf4, and Ascl1) enriched in the SCLC model (fig. S1F) (30). Hence, this mouse model can form two histologically distinct lung cancers that are united by the process of HT – prompting us to ask whether the LUAD tumors in these mice could be encouraged to transform to SCLC.

EGFR removal and the emergence of neuroendocrine character.

We speculated that the ERPMT model could be used to understand conversion of LUAD to SCLC, provided adequate selective pressure was applied against the LUAD driver. Following the development of late-stage LUAD, we randomized cohorts to one of three arms: remain on DOX (ERPMT on), come off DOX for 1mo and then re-start DOX (ERPMT on > off > on), or come off DOX and remain off for the duration of the study (ERPMT on > off). Terminal lung cancers developed at statistically different rates across these perturbations (Fig. 2A). Critically, if DOX was permanently removed, the resulting lung tumors were consistent with SCLC, but if DOX was re-started (ERPMT on > off > on), tumors were consistent with LUAD, albeit with fewer, larger lesions than in mice continually on DOX (Fig. 2B).

Figure 2. Tracing the origins and transitions between LUAD and SCLC in vivo.

A) Survival of AT2-derived ERPMT model on DOX (red), after DOX removal at 8wks (teal), and after restarting DOX diet one month following initial DOX removal (yellow; n = 10 mice per group); ***p<0.001, ****p<0.0001. B) Representative sagittal lung H&E sections from moribund mice in each group in A. C) tdTom+ burden in the airway after one month of EGFR inhibition via DOX removal (pink; n=9) or daily treatment with osimertinib (10mg/kg; PO d1–5 of 7; white; n = 9) as compared to 8wks timepoint before DOX removal (red; n = 12). DOX removal or osimertinib treatment [on DOX] were initiated at 8wks following the development of extensive ERPMT-derived LUAD tumorigenesis. For clarity, comparisons are effectively an 8wk pre-treatment cohort to ~12wk post-treatment cohorts (1mo of treatment); ***p<0.001, ****p<0.0001. D) Outline of single cell samples sequenced at distinct timepoints along the transition between LUAD and SCLC following DOX perturbations (as in A) E) Force-directed layout of cell states captured along the transition from at AT2 cell to ERPMT LUAD and finally towards a neuroendocrine fate as compared to de novo SCLC tumorigenesis (transparent blue); colored by sample (annotated in D; n = 15,828 cells pooled from 3 mice per sample). F) PCA projection of the lung epithelial lineage probability space (see Methods; same data as in E). Individual cells are colored by their max lineage probability and archetypes (see Methods) are overlaid as colored nodes (fig. S4, C to E). G) Flow plot showing relative abundance of cells assigned to their nearest lineage archetype, ordered by sampling time. Bar height is normalized by sample size (log-scale). H) Heatmap of scaled imputed and min-max normalized highly variable transition genes (see Methods), top ranked macrostate genes (n = 5, see Methods), and lineage markers along terminal cell state probabilities computed using CellRank (see Methods) for all cells in E. For each gene, expression was smoothed along the terminal state probability using a generalized additive model (GAM) as described in CellRank (34). The top 20 HVGs correlated with the bottleneck macrostate (table S2) and select lineage-specific markers are labeled on the right. A stacked KDE reflecting sample abundance (lower) and a KDE reflecting total cell frequency (upper, grey) are shown above the ranked heatmap. Above this, scaled imputed gene expression trends for the oncogenic drivers Myc and EGFR, modeled using a GAM along the terminal probability.

To better understand the proliferative nature of the residual disease we randomized ERPMT mice developing LUAD to come off or stay on DOX following three daily pulses of the S-phase label EdU. Three weeks later we labeled again, now with a different S-phase analog (BrdU), thereby labeling cells with a proliferative history that continued to cycle following tumor regression (fig. S2, A to B). Tumors that remained on DOX were EdU+/BrdU+, whereas off DOX, residual tdTom+ cells were present as single cells and labeled EdU+/BrdU−. This suggested that residual cells did not continue to cycle following DOX removal (fig. S2, C to D).

Manipulating transcription of EGFR with DOX is a convenient and effective experimental approach, but it does not precisely recapitulate the clinical scenario in which targeted therapies against the EGFR kinase domain are the primary therapeutic strategy and usual precursor to HT. To address this difference, we compared inhibiting the kinase activity of EGFR using osimertinib to the effect of removing DOX, thereby gradually extinguishing EGFR expression. After eight weeks of LUAD development, the frequency of tdTom+ cells in the airway of ERPMT mice ranged from ~5–35%. However, this was reduced to <1% in all animals following 1mo of treatment with osimertinib (“osi”) (31) or the removal of DOX (Fig. 2C). We isolated tdTom+ cells from these groups and found genetic – as opposed to pharmacologic – suppression of EGFR led to a greater increase in Ascl1 expressing cells, the definitive lineage marker of the PNEC (32) (fig. S3, A to B). Critically, cells expressing both Ascl1 and EGFR were extremely rare, suggesting that a dual-positive state is either short-lived or unviable. Furthermore, independent of Ascl1 expression, residual tumor cells retained expression of AT2 lineage markers and lacked high expression of a proliferative program characteristic of de novo SCLC (fig. S3C). These data provided the rationale to extinguish oncogenic EGFR transcriptionally by withdrawing DOX as we explore HT in the ERPMT model.

An undifferentiated, stem-like state emerges during HT

To generate conditions likely to favor HT, we removed DOX from ERPMT mice with late-stage LUAD and transcriptionally profiled single tdTom+ cells from pools of mice before, during and after EGFR removal (Fig. 2D; see Methods). We observed a continuous cell state transition from a normal AT2 cell, through EGFR-driven LUAD, and finally towards transformed SCLC at later timepoints. Notably, a minority of tumor cells isolated from samples off DOX formed a bottleneck adjacent to the de novo LUAD model prior to breaking out towards de novo SCLC (Fig. 2E). Five extreme phenotypic states (“archetypes”; see Methods) (33) were identified that corresponded to mature lung epithelial lineages, including AT1, AT2, PNEC, secretory and basal cell types (Fig. 2F and fig. S4B); however, one archetype exhibited features of a highly undifferentiated state (Fig. 2F and fig. S4C) and conspicuously mapped to the bottleneck between de novo LUAD and transformed SCLC (Fig. 2, F to G and fig. S4, B to D). This undifferentiated archetype was not associated with any lung epithelial lineage, but instead expressed modest levels of basal stem cell programs, as well as Myc and Sox2 target genes (fig. S4C). Notably, it retained features of the original AT2 lineage, but had yet to express neuroendocrine markers, including Ascl1 (fig. S4E). Tumor cells belonging to this undifferentiated state were a minority in primary LUAD, expanded in the residual LUAD (1mo off DOX), and diminished in the transforming (2mo off DOX) and transformed (>3mo off DOX) populations, which were composed largely of neuroendocrine tumor cells (Fig. 2G).

To model the dynamic transition from an AT2 origin to an EGFR-driven LUAD, and finally a transformed neuroendocrine state, we applied CellRank (34) for single-cell fate mapping. Four stable cellular phenotypes (“macrostates”) were identified along this trajectory capturing the normal AT2, de novo LUAD, the HT bottleneck, and finally a transformed SCLC population, respectively (Fig. 2H, below heatmap). As the LUAD oncogenic driver was removed, MAPK pathway activity (35) decreased along this trajectory; conversely, we observed a time-dependent increase in Myc transcriptional output (table S2). During the course of this trajectory, a highly specific bottleneck to transformation emerges that is stem-like (Tm4sf1) (36, 37) and highly proliferative (38), showing features of neuronal differentiation (Creb) (39–41) (Fig. 2H) and Myc downstream signaling (fig. S4F). Transcription factor regulatory modules (“regulons”) that characterize this bottleneck are likewise associated with neuronal plasticity (Creb5), airway stemness (Sox9) and basal cell function (Trp63) (fig. S4G, see Methods). Thus, on the path of HT, as levels of EGFR transcript wane, there may be selection for a basal stem-like state most fit to be driven by high levels of Myc (Fig. 2H and fig. S4F) and cells that break through this bottleneck may be rapidly transformed by Myc towards a neuroendocrine fate (Fig. 2H).

AT2 and PNEC lineage driver differences

To directly compare SCLC transformation efficiencies between AT2 and PNEC cells, we infected RPMT mice (which differ from ERPMT mice by lack of the EGFR transgene) with equal titers of Ad5.Spc-Cre (for AT2) or Ad5.Cgrp-Cre (for PNEC). Initially the tdTom+ frequency was greater in the airway of mice infected with Ad5.Spc-Cre, consistent with a higher baseline frequency of AT2 cells in the lung. However, this was short-lived, followed by exponential expansion of the Ad5.Cgrp-Cre group (Fig. 3A). At 8wks post-infection, macroscopic disease was clearly visible in the PNEC-derived RPMT model, but not the comparator (Fig. 3B).

Figure 3. Cell of origin and oncogenic driver incompatibility.

A) Frequency of tdTom+ cells in the airway of RPMT mice following infection with equivalent titers of adenovirus (~10⁶ pfu per mouse) using Ad5.Cgrp-Cre (blue) or Ad5.Spc-Cre (red) over a period of 8 weeks; n = 4 mice per timepoint); *p<0.01. B) Comparative histology of RPMT (no Tg.TetO-EGFR^L858R allele) mice 8wks post-infection following infection with Ad5.Cgrp-Cre (blue outline) or Ad5.Spc-Cre (red outline). C) Lineage tracing oncogenic Myc^T58A (down-pointing triangles) or EGFR^L858R (circles) on AT2 (Spc^CreERT2; red) or PNEC (Ascl1^CreERT2; blue) cells in the airway over time; n = 3 mice per timepoint. Control traces (tdTomato only; “wildtype”) are shown as grey squares; **p<0.001, ****p<0.0001. D) Long-term survival for cohorts shown in B; Ascl1^CreERT2 > EGFR^L858R (n = 13) or Myc^T58A (n = 15) and below, Spc^CreERT2 > EGFR^L858R (n = 14) or Myc^T58A (n = 12). Mice having a single copy of Rosa26^LSL-tdTom and Rosa26^LSL-rtTA3 were maintained on DOX chow throughout studies investigating lineage trace allele-mediated expression of Tg.TetO-EGFR^L858R.

To confirm whether our results with RPMT mice reflect the differential tolerance of cell lineages to the two oncogenic drivers, Myc and EGFR, we performed experiments in which production of a tamoxifen-activated Cre is governed by cell-specific promoters active in neuroendocrine (Ascl1^CreERT2) or AT2 (Spc^CreERT2) lineages, in the absence of Cre-susceptible loci for Rb1 and Trp53. Following tamoxifen administration to activate Cre^ERT2, we found that Myc expanded the airway Ascl1+ population and EGFR led to an eventual decline. Conversely, EGFR expanded the AT2 lineage, where Myc was detrimental as compared to wildtype controls (Fig. 3C). In a larger cohort of mice, we observed that Myc expression alone from the Ascl1+ lineage was sufficient to produce a lethal, fully penetrant phenotype, whereas EGFR was not. In contrast, EGFR expression alone was sufficient to transform the AT2 lineage, but Myc was not (Fig. 3D). These data strongly support cell lineage-specific differences in the tolerance of oncogenic drivers, Myc and EGFR, in the lung.

The pulmonary neuroendocrine cell is refractory to transformation by oncogenic EGFR

The phrase “terminal neuroendocrine” has been used to describe the eventual histology that the ERPMT model transitions towards following EGFR withdrawal; however, it was unclear whether this was a terminal state, or if we could study HT in reverse – converting a SCLC tumor towards a LUAD state. Like earlier experiments, we initiated tumorigenesis from PNECs in the ERPMT model and followed cohorts of mice that were on or off DOX. There was a significant delay in the lethality of the model on DOX (fig. S5A); however, in lungs of mice on DOX where EGFR protein should be produced, we observed low to absent EGFR by immunohistochemistry (IHC) in tdTom+ tumor regions, histologically consistent with SCLC (fig. S5B). Immunofluorescence for EGFR^L858R and Ascl1 showed patchy regions of EGFR-positivity that were excluded from larger areas of Ascl1-positivity (fig. S5C), with the Ascl1-positive/EGFR-negative SCLC component of these tumors having the greatest burden in all mice examined. Taken together, these data support the conclusions that cells in the PNEC lineage resist transformation towards an EGFR-driven LUAD state, just as cells in the AT2 lineage cannot be easily transformed to SCLC, even though the latter form of HT can occasionally occur under certain conditions.

Oncogenic EGFR led to a gradual elimination of PNECs over months, without evidence of acute intoxication (Fig. 3C). It was unclear whether elevated signaling through the MAPK pathway (via EGFR) was a disfavored situation for the PNEC, or if EGFR incompatibility arose through some other mechanism. If excessive MAPK signaling suppressed PNEC proliferation, then inhibition of this pathway should increase Ascl1+ cells. We traced Ascl1+ cells and randomized mice to receive a diet formulated with or without the MEK inhibitor (MEKi) trametinib to suppress Mek>Erk signaling (fig. S6A). We terminated the study after 3mo on MEKi due to toxicity, with adult mice experiencing weight loss approaching our protocol limits (fig. S6B). However, no significant differences were observed in the abundance of tdTom+ cells between the two cohorts (fig. S6C). Moreover, despite achieving chronic suppression of phosphorylated Erk (fig. S6D) we did not observe any change in the location or proliferative status of tdTom+ cells in mice treated with MEKi diet (fig. S6E), suggesting that physiologic Mek>Erk signaling was not suppressing proliferation of PNECs.

Myc is sufficient to transform the PNEC

Single oncogene lineage tracing demonstrated a clear difference in the sensitivity of the AT2 and PNEC cell types to oncogenic transformation by Myc (Fig. 3); however, a hallmark of SCLC is inactivation of the RB1 and TP53 tumor suppressors, in addition to heightened expression of a Myc family member and its transcriptional targets (7). Inspecting the bronchioles of mice expressing Myc driven by the Ascl1^CreERT2 lineage trace (abbreviated “Ascl1>Myc”) 1mo post-labeling revealed clusters of proliferative, tdTom+ cells that had not yet invaded surrounding tissues, consistent with carcinoma in situ (fig. S7A). Instead, all Ascl1>Myc mice were dying of cancer localized to the thyroid (fig. S7B). To date, we have been unsuccessful in activating the Ascl1^CreERT2 allele specifically in the lungs, while sparing the trachea and thyroid. We therefore isolated tdTom+ cells from the airways of four distinct genotypes of mice combining Ascl1 lineage-traced Myc with loss of Rb1, Trp53 or both tumor suppressor genes. We expanded cells ex vivo using organotypic culture conditions and engrafted equivalent cell numbers into the flanks of athymic mice. Combined loss of Rb1 and Trp53 accelerated the growth of these Myc-driven tumors, but all genotypes were sufficient to form transplantable cancers. Moreover, all tumors had a similar histologic appearance consistent with high-grade neuroendocrine cancer (fig. S7C). These data suggest that PNECs could be transformed by Myc [alone] if expanded ex vivo, but this experiment did not demonstrate that Myc was required for tumor maintenance.

To test whether PNEC-derived tumors are dependent on Myc, we sorted tdTom+, rtTA3-expressing PNECs from the airway and infected cells with lentiviruses containing tetracycline-promoter driven Myc constructs. Coupling rtTA3 expression to the lineage trace (Ascl1^CreERT2) removed the likelihood of infecting lineage-negative cells present as contaminants. While these cells were sparse (fig. S8), we could consistently generate organoid cultures from as few as 50 cells when DOX was present in the culture to drive Myc expression. Removal of DOX resulted in near complete growth suppression of organoids expressing Myc^WT as compared to Myc^T58A (fig. S9, A to C), a long-lived Myc variant (42) (fig. S9C). Engrafting PNECs expressing inducible Myc^WT into immunocompromised mice demonstrated DOX-dependent tumor growth. Removing DOX from tumor-bearing mice dramatically reduced tumor volumes over the course of several weeks (fig. S9D), with residual, fibrotic tumor tissue comprised of non-cycling cells (fig. S9E). Consistent with their rapid proliferation, Myc-driven PNEC organoids were sensitive to compounds that exacerbated replication stress such as topoisomerase inhibitors (etoposide), as well as inhibitors of enzymes required for cell cycle progression, including Cdk4/6 (palbociclib) and Wee1 (adavosertib) (fig. S9, F to G). Additionally, direct inhibition of the Myc-Max interface using the small molecule MYCi975 (43) provided similar reduction in organoid growth as compared to DOX removal. (fig. S9F).

AT2 cells are refractory to transformation by Myc

While Myc alone may be sufficient to drive transformation and expansion of PNECs, our earlier results suggested it was insufficient to transform the AT2 lineage (Fig. 3C). To further investigate what underlies this bottleneck, we established AT2 organoid cultures using recently published methods (44) from lineage-traced, wildtype or Myc^T58A expressing cells. These organoids expanded rapidly as compared to their wildtype counterparts but were unsustainable beyond 3 passages (fig. S10A). In early timepoints following the expression of Myc^T58A in AT2 cells in vivo, we noted incorporation of EdU in tdTom+ cells. However, a year following the initiation of the trace, tdTom+ cells in the airway failed to incorporate EdU, suggesting they were no longer proliferative or were eliminated (fig. S10B). Ex vivo, AT2 organoids expressing Myc^T58A demonstrated increased DNA damage sensing, replication stress and markers of programmed cell death as compared to wildtype (fig. S10C). It is unlikely this resulted from an excess of Myc protein, as levels were significantly lower in AT2 cells compared to levels tolerated in PNEC organoids (45) (fig. S10D). Finally, consistent with the observation that Ras signaling through PI(3)K relieves Myc-induced apoptosis (46), we likewise observed that Myc significantly accelerated oncogenic EGFR-driven LUAD. This implies that enhanced signaling via the EGFR>Ras>Mek>Erk pathway can relieve intolerance to Myc in the AT2 lineage (fig. S10, E to F).

While our data suggest the different oncogenes that drive AT2 and PNEC lineages are in stark contrast, it remained unclear whether other epithelial airway cells can transform in response to Myc alone. To address this, we performed a generalized, conditional trace using an Nkx2.1^CreERT2 allele that will drive Cre-mediated recombination in a broad range of cell types derived from the anterior foregut endoderm, including the trachea, pituitary, thyroid and most of the lung (47). At early timepoints, lineage-labeled tissues within the thyroid (both Ascl1+ and Ascl1-) expanded following Myc expression compared to control mice (fig. S11A); however, at later timepoints there was an outgrowth of tdTom+/Ascl1+ cells in the bronchioles not observed in wildtype controls (fig. S11B). Together these data suggest the Ascl1+ PNEC is unique in its tolerance to Myc, but we had not yet explained how intolerance to Myc could be overcome in the AT2 lineage during HT.

Deletion of Pten removes a barrier to Myc transformation

Genes upregulated in tumor cells as they escape the bottleneck to HT in our ERPMT model (termed “breakout”; fig. S12A) were notably associated with PI(3)K signaling (fig. S12, B to C) as compared to cells stuck in the bottleneck and unable to adopt neuroendocrine fate (48). Thus, we asked whether increasing PI(3)K-dependent Akt signaling through deletion of Pten would enable Myc^T58A-driven transformation in an AT2 lineage trace. Strikingly, we observed fully penetrant Myc-driven transformation in an AT2 cell (using Spc^CreERT2/+) when Pten^Fl/wt was inactivated (abbreviated “Spc>Pten;Myc”; Fig. 4A). At the median period where animals in the Spc>Pten;Myc cohort were moribund, lungs from Spc>Myc mice (Pten^wt/wt) showed no evidence of macroscopic disease (Fig. 4B). Similar observations were made when combining Myc expression with a conditionally active, mutant PI(3)K allele (E545K) or inactivating both copies of Pten (Pten^Fl/Fl; Fig. 4B). Examining Spc>Pten;Myc animals 3mo post-recombination reinforced the observation that lesions were variable in their frequency, size and histologic appearance (fig. S13). However, they were Ascl1-negative and glandular in appearance, suggesting they were adenocarcinoma-like and excluding the likelihood of SCLC, squamous or mixed histology (fig. S13). AT2 lineage markers, including pro-surfactant protein C (pro-SPC) and MHC Class II, were low or absent, whereas basal epithelial markers like Sox2 and keratin 18 were variably expressed (fig. S13). Using single cell RNA-sequencing, we also observed a notable increase in the basal-like stem state following combined loss of Pten and expression of Myc in AT2 cells – not seen with either genetic perturbation alone (Fig. 4C). Notably, this basal-like state was associated with greater Myc transcriptional output (Fig. 4D).

Figure 4. Loss of Pten in the AT2 lineage removes the barrier to Myc transformation.

A) Survival of mice where Myc (n = 12), Pten^Fl/wt (n = 5) or the combination of these alleles (n = 23) are initiated in AT2 cells using a single copy of Spc^CreERT2. Data were censored between 250–325d in the non-lethal arms. B) Comparative histology of whole lungs from representative Spc>Myc, Spc>Pten^Fl/+;Myc), Spc>Pten^Fl/Fl;Myc or Spc>PI(3)K^LSL-E545K;Myc mice at ~3mo post-labeling. Higher magnification regions (boxed) are provided at right; 200um scalebar. C) Bar plot showing fraction of each epithelial lineage archetype detected per sample as in (Fig. 2G). D) Bar plot of Hallmark gene sets significantly enriched (FDR < 0.01) within the undifferentiated cell state of the Spc>Pten^Fl/+;Myc sample pool, as in (fig. S4F). E) Effect of Pten deletion combined with deletion of p53, Rb1, and expression of Myc and tdTomato (RPPtenMT or RPMT) in neuroendocrine (blue) or AT2 (red) cells; n = 10 per arm with x-axis split for clarity, ***p<0.001. F) Clustered heatmap of imputed and z-normalized expression of AT2 and PNEC signature genes, model oncogenic drivers (Myc and Tg.EGFR), and SCLC subtype identifiers (NeuroD1, Pou2f3, and Yap1) for all tumor-epithelial cells from the RPPtenMT model (green) and the de novo LUAD (red) and SCLC (blue) models described (Fig. 1). Genes and cells are clustered using the Euclidean distance method.

Pulmonary basal cells efficiently generate SCLC

The basal stem-like transcriptional program associated with AT2 cells capable of tolerating Myc supported the possibility that an intermediate state during HT may be “basal-like.” More generally, this raised the possibility that the basal cell may serve as a cell of origin for SCLC, as previously speculated (49). Targeting pulmonary basal cells in mouse models is limited by their anatomic location, noted to be refractory to viral infections delivering Cre (50). Indeed, we observed an absence of tdTom+ cells in the lungs of mice one-month post-labeling when using a Krt5^CreERT2 allele to target basal cells (fig. S14A). However, following regeneration in the proximal lung induced by naphthalene damage of secretory cells (51, 52), tdTom+ cells were detected within the lungs of mice (fig. S14B). As basal cells are known to serve as multi-potent progenitors (fig. S14C) (51–53), we then generated mice to delete Rb1 and/or Trp53 from the basal lineage and found Rb1 loss alone was sufficient to skew tdTom+ cells towards a neuroendocrine fate (fig. S14D).

Consistent with this, we observed fully penetrant neuroendocrine tumorigenesis from the basal lineage in mice that have lost both Rb1 and Trp53 with (Krt5>RPMT) or without (Krt5>RPT) Myc expression (fig. S14E). However, a significantly shorter tumor latency was observed in mice producing oncogenic Myc. Expression of the conditional tdTomato reporter allele was easily observed throughout the keratinized epithelium of mice following recombination, but interestingly, tumorigenesis was restricted to the proximal airway (fig. S14, F and H). Histologically, both models produced SCLC-like tumors; however, we noted that most tumors were adjacent to or within the thymus and did not invade the lungs, unless we damaged lungs with naphthalene (fig. S14, G to I). These data suggest basal cells may serve as cells of origin for SCLC, and that their progeny and anatomic location can be influenced by genotype and injury.

An efficient model of AT2-derived SCLC following the loss of Pten

Deletion of Pten was sufficient to lower the barrier to transformation by Myc in the AT2 lineage; however, the resulting tumors were not neuroendocrine (fig. S13). We suspected Rb1 loss was required (10) – in addition to adaptation to Myc – for neuroendocrine transformation. Thus, to recapitulate bona fide transformation from an AT2 cell to SCLC efficiently, we generated a model in which we could delete Rb1, Trp53, a copy of Pten, and express Myc and tdTom (abbreviated “RPPtenMT”). If tumorigenesis was initiated in PNECs, there was no difference in latency between the RPMT and RPPtenMT models, but a significant difference emerged if the model was initiated in AT2 cells (Fig. 4E). AT2-derived RPPtenMT tumor cells displayed neuroendocrine expression profiles most like de novo SCLC (Fig. 4F and fig. S15, A to B). Notably, two subpopulations of neuroendocrine cells stratified by expression of Ascl1 (fig. S15B). The Ascl1^low group was differentially enriched with neuronal genes including, Stnm2, Nfix and Mapt. Whereas, the Ascl1^low group also expressed high levels of NeuroD1 – consistent with prior work showing that Myc can drive subtype plasticity from an Ascl1^high to Ascl1^low / NeuroD1^high transcriptional profile in multiple models of SCLC (54–58) (fig. S15C). Pathology review revealed extensive heterogeneity with mixing of classic SCLC and large cell neuroendocrine (LCNEC) tumor features, marked by variable expression of neuronal and neuroendocrine markers (fig. S15D).

Rb1 loss is necessary but insufficient for HT

The ERPMT model provided an efficient system to study HT, but we had not addressed a core requirement for Myc in this process. Thus, we generated another model in the absence of transgenic over-expression of Myc (“ERPT”) and found recurrent disease following the removal of DOX (ERPT on > off) was not fully penetrant (Fig. 5A). Moreover, while the de novo ERPT LUAD showed an absence of Ascl1 and Myc protein, only some recurrent tumors appeared to histologically resembled SCLC (Fig. 5, B to D). Importantly, the ERPT tumors recurring as SCLC did show an increase in Myc protein (Fig. 4C). Both the ERPT and ERPMT models transcriptionally maintained AT2 identity, whereas Myc and embryonic stem cell target genes - only moderately elevated in the ERPT model – expectedly increased in the ERPMT model (Fig. 5E). True paired cases of human LUAD pre-/post-HT are limited, but some studies have compared unrelated de novo LUAD to transformed (T-SCLC) in an attempt to understand this phenomenon (1, 3, 59, 60). Re-analysis of a publicly available human data revealed that Myc target genes were indeed differentially up-regulated in T-SCLC as compared to LUAD, suggesting Myc may play a role in facilitating HT in the clinical setting (Fig. 5F).

Figure 5. Rb1 loss cooperates with Myc expression to facilitate HT.

A) Survival of mice with the ERPT genotype (no Myc transgene) initiated with Ad5.Spc-Cre on DOX (ERPT on; dark green, n = 23), or on and then off DOX once mice developed advanced disease (ERPT on > off; light green, n = 20). As before, DOX diet was removed when individual mice exhibited signs of labored breathing and/or significant weight loss with a hunched appearance. For each model, cohorts of mice were followed until ~3X the median latency elapsed, at which point lungs were collected and the study was ended. B) Histologic appearance of “ERPT on” lungs by H&E and IHC staining for Ascl1 and Myc (scalebar 100 um). C) Similar to B, now with mice on then off DOX representative of a SCLC-like tumor. D) As in B, now with mice on then off DOX representative of a LUAD-like tumor. Pathologic interpretation provided below each representative example. E) Dot plot (left) showing frequency of expressing cells (node size) and log-transformed expression (node color) of AT1, AT2, and basal cell marker genes in normal AT2, ERPT LUAD, and ERPMT LUAD models. Genes shown are expressed in at least 20% of cells within at least one condition. Kernel density estimate (KDE, right) plots showing mean log-transformed expression by condition for select gene signatures. F) Volcano plot showing differentially expressed genes from bulk RNAseq of human transformed SCLC vs LUAD. Genes from the Hallmark MYC Targets V1 signature are colored according to conditional enrichment, top genes (abs(log2FC) > 1 and FDR < 1e-5) are labeled, and the pathway NES and FDR from GSEA are inset (bottom right). G) Tumorigenesis initiated using an adenoviral (n = 10; Ad5.Spc-Cre) or AT2 lineage trace allele (n = 19; Spc^CreERT2) in the EPMT model (Rb1^+/+) produces LUAD that does not relapse following DOX removal (n = 5 per group). H) Representative sagittal lung H&E sections from each group in G on DOX at point of moribund disease or one month following the removal of DOX from an otherwise moribund animal. I) Clustered heatmap of gene signatures differentially enriched between the tumor-epithelial cells of the EPMT and ERPMT models before and after DOX removal. All pathways are significantly enriched (NES > 0 and FDR < 1e-5) in at least one condition. Less significant signatures (FDR < 0.01) are transparent and signatures not meeting this threshold are blank. Rows and columns are clustered using the complete Manhattan distance method and metric. J) J) Bar plot showing fraction of each epithelial lineage archetype detected per sample (as in Fig. 2G) for the EPMT (left) and ERPMT (right) models before and after removal of DOX.

Rb1 loss cooperates with expression of Myc to facilitate neuroendocrine transformation

To evaluate the relative contributions of Rb1 loss and Myc over-expression for HT, we generated a model in which Rb1 was wildtype (abbreviated herein as “EPMT”) and initiated tumorigenesis in AT2 cells using Ad5.Spc-Cre or a lineage trace (abbreviated herein as Spc^CreERT2>EPMT). If on DOX, then both models produced LUAD with differences in latency likely reflecting the broad initiation achievable using a lineage trace allele as compared to sparse infection using inhaled adenovirus. Critically, however – LUAD did not recur in EPMT mice taken off DOX (Fig. 5, G to H). Instead, we only observe up-regulation of transcriptional programs associated with HT (i.e. Myc target genes, embryonic stem, and neuroendocrine gene programs) in off-DOX residual tumor cells in the context of Rb1 loss (Fig. 5I). Indeed, our results suggest that Rb1 loss is required for emergence of the residual stem-like state capable of full neuroendocrine transformation (Fig. 5J). Thus, these events likely cooperate as neither Rb1 loss nor Myc over-expression alone efficiently facilitates HT (Fig. 5G, 4A and fig. S13).

Taken together, we conclude that the AT2 cell is highly refractory to transformation by oncogenic Myc, as are many other cell types within the lung – the noted exception being the PNEC. The intolerance of AT2 cells to Myc can be relieved through activation of the Akt signaling pathway, such as the deletion of the tumor suppressor Pten, to generate a permissive, stem-like state and full conversion to a Myc-driven, high grade neuroendocrine cancer requires the additional loss of Rb1.

Discussion

Carcinogenesis and related processes, like tumor progression, therapy resistance and HT remain incompletely understood. To better understand how HT occurs in lung cancer, we have developed new mouse models in which lung tumorigenesis can be initiated in different cell lineages to follow the transformation of EGFR-driven LUAD to Myc-driven SCLC. In doing so, we demonstrate that HT can be simply understood as a change in oncogenic driver compatibility as tumor cells transition between alveolar and neuroendocrine fates. The driver oncogene in the PNEC lineage is Myc and the bottleneck in HT can be relieved by mechanisms that allow an AT2 cell to become a stem-like cell that is driven to an oncogenic state by high levels of Myc.

Understanding events that facilitate HT clinically has been limited by the lack of samples obtained from the same patient before, during and after HT. Comparisons of de novo SCLC to transformed SCLC have highlighted pathways activated following neuroendocrine differentiation (59). A more recent study relied on “putative” HT samples in which a LUAD oncogenic driver was detected in a histologically-confirmed SCLC. Interestingly, when compared to de novo SCLC, these cases were significantly enriched for mutations that activate the PI(3)K signaling pathway (61). Consistent with this, we found that loss of Pten or activation of Pik3ca were sufficient to break a barrier in the AT2 lineage to transformation by Myc, but insufficient to lead to neuroendocrine transformation.

Prior work has established a combination of genetic events capable of reprogramming human cell types to neuroendocrine cancers (62); however, the intermediate steps in this process remain unclear. We build upon this work by developing intact genetically engineered animals and sampling tumor cells throughout the lifetime of HT. Moreover, we find that while Rb1 loss is necessary for HT to a neuroendocrine cancer, it is the extinction of the driver oncogene transcriptionally that is required prior to the emergence of neuroendocrine features.

Beyond the genetic manipulations and changes in gene expression described here, we speculate that there are likely other mechanisms that allow for such lineage conversion, cellular plasticity or trans-differentiation. Lineage tracing has demonstrated that basal progenitors can give rise to PNECs (63, 64). More recently, tracheal basal cells have been shown to differentiate towards a PNEC fate under hypoxic conditions, but the intermediate transcriptional steps in this complex process are uncharted (65). In our attempts to describe the transition state between LUAD and SCLC, we found that these tumor cells appear “basal-like,” based on their transcriptional profile, but lacked definitive basal lineage markers. Instead, these cells are more accurately described as a “lineage negative,” stem-like progenitors described to arise in the mouse lung following major airway injury (66). Thus, the airway cell most capable of such plasticity may be the pulmonary basal cell or a stem-like cell that has yet to be fully described.

Finally, it is unclear whether therapeutic targeting of pathways facilitating HT in patients poised to undergo HT will have efficacy. Our results suggest that non-invasive monitoring for activation of the Akt signaling pathway [such as the appearance of PI(3)K mutations in circulating tumor DNA] may serve to alert physicians to the likelihood of HT, prior to the emergence of an aggressive, transformed SCLC. It is not yet known whether direct targeting of Myc in such Myc-driven scenarios (including HT or de novo SCLC tumorigenesis) will be fruitful; however, strategies to inhibit the transcriptional activity of Myc proteins have advanced substantially over the last decade and might ultimately have clinical utility in multiple contexts (43, 67, 68).

Limitations of the study

There are several key limitations to this study: First, we are attempting to understand HT by modeling a human phenomenon in the mouse where the complexity of cell types and microenvironments are not identical. Recent descriptions of the human distal airway have found a diversity of cell types in human lungs that are not observed in the mouse, including specialized regenerative cells with hybrid alveolar-secretory character and multiple subtypes of basal cells (19, 69, 70). For example, basal cells are not present in the mouse lung, unless the airway is damaged, which we demonstrate using naphthalene. Furthermore, most laboratory mice are housed under pathogen-free conditions. We know that lifetime exposure to various carcinogens and particulates fundamentally alters the microenvironment of the lung, including the likelihood that a cancer will develop (71, 72). Second, there are processes in human cells not found in the mouse that may be critical for HT, including the role for APOBEC-mediated hypermutation. Studies of EGFR-mutated LUAD have found APOBEC mutagenesis signatures following treatment with EGFR-targeted therapies (11, 73, 74) and this mutational signature is enriched post-HT (1, 61). It is unclear whether APOBEC mutations are directly responsible for activating the PI(3)K signaling pathway, thereby [hypothetically] relieving intolerance of the AT2 cell to Myc (1, 59, 61); however, other studies would suggest this is possible (75, 76). Finally, we were unable to initiate tumorigenesis in the ERPMT model using the Spc^CreERT2 lineage trace allele. This proved impossible because, in the mouse genome, the Rb1 locus is ~3Mb from the Sftpc locus (GRCm38/mm10 build). Such proximity precludes generation of the desired genotype with both homozygous floxed copies of Rb1 and the Spc^CreERT2 knock-in allele.

Data and Materials Availability

Lead contact: Requests for resources should be directed to and will be fulfilled by Eric E. Gardner (eeg2001@med.cornell.edu) or Ashley M. Laughney (asl4003@med.cornell.edu).

Materials: All mouse models, organoids derived from mice and plasmids described in this work will be made available to investigators through an institutional or third-party Material Transfer Agreement (MTA), upon reasonable request. Select plasmids will be submitted to Addgene upon manuscript acceptance. Not all mouse strains are currently active – please contact Eric E. Gardner (eeg2001@med.cornell.edu) for specific information and availability.

Data and code: The processed single cell data and relevant code, including Docker environments with Jupyter notebooks demonstrating key analyses, will be available on the Gene Expression Omnibus (GEO) and Laughney Lab GitHub upon manuscript acceptance.