Multiomic Analysis of Lung Tumors Defines Pathways Activated in Neuroendocrine Transformation

Lineage plasticity describes the capacity of cells to transition from one committed identity to that of a distinct developmental lineage. This phenotypic flexibility can promote survival of cancer cells under unfavorable conditions, such as hypoxia or selective pressure from oncogenic driver–targeted therapy (1–3). The histologic transformation of lung adenocarcinoma (LUAD) to an aggressive neuroendocrine (NE) derivative resembling small cell lung cancer (SCLC) is a signature example of lineage plasticity in cancer. Initially described in the prostate setting as a mechanism of resistance to androgen suppression (4–6), a similar process of NE transformation was identified in LUADs harboring EGFR mutations (7) and subsequently found to occur more broadly in lung cancers (8). Transformed SCLC (T-SCLC) is associated with a notably poor prognosis, similar to or worse than that of de novo SCLC (3). The increased practice of tumor rebiopsy upon disease progression has improved the ability to identify histologic transformation, which in EGFR-mutant LUAD may comprise up to 14% of cases of acquired resistance to osimertinib (9, 10).

Identification of the molecular mechanisms promoting lineage plasticity in clinical samples is key to identifying patients at high risk of transformation and may define strategies to prevent or treat this phenomenon. Little is known about the molecular alterations occurring during NE transformation in human tumors, including in lung cancer. Transcriptomic analyses of prostate cancer undergoing NE histologic transformation have been performed but only on relapsed and posttransformation samples (11, 12). A paucity of well-annotated paired pre- and posttransformation clinical samples has been a major hurdle in defining mechanisms of lineage plasticity in lung cancer. Previous genomic studies in small numbers of cases have suggested that concomitant inactivation of TP53 and RB1 is necessary but not sufficient and have identified few other recurrent genomic alterations (3, 9, 13).

On rare occasions, pathologic examination of resected cancers reveals more than one histology in single tumors. We hypothesized that such cases might represent lineage plasticity captured in temporal and spatial proximity to the occurrence of a histologic shift. Detailed molecular characterization of such cases could provide novel insight into key drivers of histologic transformation. Here we report the first comprehensive characterization of NE transformation, including genomic, transcriptomic, epigenomic, and protein analyses, in a cohort of mixed histology LUAD/SCLC samples. In addition to our primary analysis of mixed histology tumors with discrete areas of LUAD and SCLC, we include analyses in matched pre- and posttransformation cases, with reference to control “pure” LUAD and SCLC. Our strategy provides novel insights into molecular drivers and potential therapeutic vulnerabilities of NE transformation in lung cancer.

For in-depth characterization of NE transformation, we analyzed clinical specimens consisting of combined LUAD/SCLC histology exhibiting clear spatial separation (n = 11); pretransformation LUADs (n = 5) and posttransformation SCLCs (n = 3), including one matched case; never-transformed LUADs (n = 15); and de novo SCLCs (n = 18; Fig. 1A; Supplementary Tables S1–S4). Microdissection was performed for independent genomic, epigenomic, and transcriptomic analyses (Fig. 1B and C; Supplementary Fig. S1).

Our selection of combined histology samples for this analysis was predicated on the assumption that the LUAD and SCLC components were clonally related. Alternatively, it was possible that these represented “collision tumors” derived from two independent oncogenic events. To distinguish between these alternatives, we performed whole-exome sequencing (WES) of all LUAD and SCLC samples from combined histology specimens and the matched pre- and posttransformation pair (T12; Supplementary Tables S5 and S6). Sequencing revealed multiple shared mutations in all cases, confirming that matched LUAD and SCLC components were clonally related (Fig. 2A). For one case lacking WES data, we were able to confirm genetic relatedness in the RNA-sequencing (RNA-seq) data (Supplementary Fig. S2A). We therefore refer to these hereafter as T-LUAD and T-SCLC, with the T referring to histologic transformation, without presumption of directionality. Higher tumor purity in the T-SCLC component was observed, consistent with the low stromal content of SCLC relative to LUAD (refs. 14 and 15; Supplementary Fig. S2B). Suggestive of temporal proximity, we did not observe consistent differences in tumor ploidy, tumor mutation burden, or predicted neoantigen burden between T-LUAD and T-SCLC components (Supplementary Fig. S2C–S2E).

We next sought to define mutational processes that might contribute to lineage plasticity and histologic transformation through mutational signature analysis. Smoking signature was dominant in 7 of 11 cases but did not differ consistently between T-LUAD and T-SCLC (Supplementary Fig. S2F). APOBEC signature, previously proposed to be a predictor of SCLC transformation in triple EGFR/TP53/RB1-mutant tumors (9), was prominent in 5 of 11 of the T-LUAD samples (Supplementary Fig. S2F).

Analyses of the most prevalent mutations and copy-number alterations (CNA), including variants of both known and unknown significance (VUS), revealed almost universal TP53 loss in both T-LUAD and T-SCLC (93%; Fig. 2B), with only two T-LUADs (T-LUAD1 and T-LUAD8) showing wild-type TP53. RB1 mutations/deletions were less frequently detected (63% of samples), identified in 8 of 14 T-LUADs and in 8 of 11 T-SCLCs (Fig. 2B and C). However, IHC in samples for which tissue was available showed that Rb protein expression was lost in all but one T-LUAD with wild-type RB1 (T-LUAD1) and in all T-SCLC samples (Fig. 2C). These results show that loss of RB1 function can be independent of apparent genomic alterations, highlighting the importance of complementary genomic and IHC profiling for confirmation of RB1 activity. Additionally, to confirm whether the mutations in TP53 and RB1 occurring only in one of the two matched histologic components for a given case were truly private (T1 and T8), we performed targeted sequencing (Supplementary Fig. S2G) that confirmed TP53 and RB1 mutations in T1 are exclusive of the T-SCLC component. However, targeted sequencing showed that the TP53 mutation found in T-SCLC8 was not truly private, as it was detected at near subclonal levels also in T-LUAD8 [variant allele fraction (VAF) VAF = 0.06]. Oncogenic EGFR mutations were present in 33% of T-LUAD samples (Fig. 2B), further illustrating that NE transformation can occur outside the EGFR-mutant setting (8). Within matched pairs, we observed common mutations of both known and unknown significance, highlighting genetic relatedness. There were no recurrent mutational distinctions between paired T-LUAD and T-SCLC seen in more than two cases in this data set, suggesting that although a preexisting genetic context may facilitate plasticity, NE transformation itself may not be mutationally driven. Mutation profile of the T-SCLC samples was similar to that described for de novo SCLC (Supplementary Fig. S2H).

To better define the context that permits lineage plasticity, we focused on the most commonly altered genes in this sample set, present in both the T-LUAD and T-SCLC components (Fig. 2B). Notably, these include factors involved in WNT signaling (BCL9, LRP5), PI3K/AKT signaling (PTEN, PIK3CA, PIK3R1, etc.), Notch signaling (NOTCH4, IDH2), epigenetic regulation (KMT2A/C, CREBBP, FOXA1, etc.), and cell cycle/DNA repair (TRIP13 and TP53BP). The presence of these pathway alterations in T-LUAD samples implies that they may occur early in the NE transformation process and prime LUAD for lineage transition.

Next, we compared the frequency of mutations/CNAs identified in the T-LUADs in our cohort to those of The Cancer Genome Atlas (TCGA) LUADs (Fig. 2D and E; Supplementary Fig. S3A; Supplementary Table S7). We focused on the differentially mutated genes showing alterations in ≥20% of T-LUAD samples to filter for those more likely to have a role in transformation promotion. As expected, we found enrichment of TP53 (P = 0.008) and RB1 (P < 0.001) alterations in T-LUAD (3, 9). Consistent with previous reports of NE transformation in EGFR-mutant LUAD, we found enrichment in EGFR alterations (P = 0.012) in the T-LUAD cohort (9, 13). We noted decreased frequency of KRAS mutations in our T-LUAD (P = 0.008; Supplementary Fig. S3A); this may suggest that KRAS-mutant LUADs are less likely to undergo NE transformation, or, alternatively, may be attributable to the historical lack of potent targeted inhibitors of KRAS.

Novel observations in this analysis included mutations on NFE2L2 (P = 0.010), a transcription factor (TF) involved in response to oxidative stress (16); mutations on KMT2B (P = 0.014), an epigenetic regulator; and amplifications in EGFR (P = 0.002), TERT (P < 0.001), and TRIP13 (P < 0.001), a gene involved in cell division. These genes were rarely altered (<5%) in the TCGA LUADs (Fig. 2E; Supplementary Fig. S3A). Of note, none of these enrichments were significant when performing multiple hypothesis testing, most likely due to the small size of our cohort. Validation in a larger cohort would be required to support a role for these alterations as predictors of susceptibility to SCLC transformation. Interestingly, we observed recurrent loss of the 3p chromosome arm in 85% of our pretransformation LUAD cases, a significantly higher rate than observed in TCGA LUADs (P = 0.045; Fig. 2F).

The paired nature of the combined histology tumors provided an opportunity to explore serial events in branched evolution of the distinct histologic lineages. WES data were of sufficient quality to allow reconstruction of the clonal history for five of the cases under study (Fig. 3). For each of these, we identified exclusive or enriched mutations in the T-SCLC components, but no common driver mutations across cases were observed. The integer copy-number segmentation profile (Fig. 3, left) was suggestive of genomic instability consistent with TP53 inactivation, which was identified as an early event in most samples. Within these segments, at a gene level, we observed several recurrent focal non-VUS copy-number changes (Supplementary Table S8). Many of the oncogenic driver mutations were shared within each matched pair (Fig. 3, right), suggesting that they occurred before the histologic divergence. Interestingly, both pre- and posttransformation samples in all of these cases were genome doubled, implying that whole-genome doubling (WGD) occurred before transformation (Fig. 3, right).

Work from our group and others has highlighted the intertumoral heterogeneity of SCLCs (15). De novo SCLCs can be divided into discrete molecularly defined subtypes based on dominant expression of one of four transcriptional regulators: ASCL1 (SCLC-A), NEUROD1 (SCLC-N), POU2F3 (SCLC-P), and YAP1 (SCLC-Y). However, little is known about the molecular subtyping of T-SCLC tumors or whether these tumors consistently align with one of these four defined subtypes.

To study whether T-SCLCs were enriched in any subtype, we analyzed relative expression of these four transcriptional regulators (mRNA and protein by IHC), together with canonical NE markers (mRNA), in the T-LUAD and T-SCLC samples (Fig. 4A; Supplementary Fig. S3B; Supplementary Table S9). As expected, we observed increased expression of NE markers in T-SCLC versus T-LUAD consistent with NE transformation, which was further accentuated in de novo SCLC (Supplementary Fig. S3B). Expression of three TFs (ASCL1, NEUROD1, and POU2F3) was consistently low in the T-LUADs. However, expression of YAP1 was higher in all but one (T4) of the T-LUADs than in their matched T-SCLCs (Fig. 4A). YAP1 expression was higher in never-transformed LUADs than in T-LUAD (Supplementary Fig. S3C), consistent with the oncogenic role of this Hippo pathway effector in LUAD (17, 18) and with its incompatibility with NE features in lung cancer (19). We observed good concordance between IHC and RNA data. Where discrepant, we assigned the subtype based on relative RNA expression following current consensus (15).

Notably, we were able to detect all four SCLC subtypes among the T-SCLC samples, suggesting that lineage plasticity in LUAD can give rise to any of the four SCLC subtypes. Interestingly, two of the samples (T1 and T3) were categorized as SCLC-P, with high POU2F3 levels exclusive to the T-SCLC component, and no expression of any of the other TFs by IHC (Fig. 4A and B; Supplementary Table S9). Tuft cells, a rare population of lung cells, have been previously hypothesized to be the cell of origin for SCLC-P based on a very similar POU2F3-dependent gene-expression program exclusive in this cell type of the normal lung (20). However, no POU2F3 protein expression was observed in the matched T-LUAD components of these samples (Fig. 4A and B; Supplementary Table S9). Furthermore, mRNA levels of other tuft cell markers (20) were also not elevated in these T-LUADs relative to the rest of pretransformation or control LUADs (Supplementary Fig. S3D). This suggests that a tuft cell–like gene-expression program is induced in this T-SCLC subtype, independent of the cell of origin. These data demonstrate that T-SCLCs conform to all major subtypes of de novo SCLCs and suggest a tuft cell–independent origin of SCLC-P.

We performed transcriptomic (RNA-seq; Supplementary Table S10) and methylation (EPIC) analyses of T-LUADs, T-SCLCs, control LUADs, and de novo SCLCs (Fig. 1C; Supplementary Tables S1–S3 and S11). Principal component analyses (PCA) of RNA-seq data showed dissimilar expression patterns for control LUAD and de novo SCLC, as expected (Fig. 4C). T-LUADs clustered together in adjacency to control LUADs, and T-SCLCs in proximity to de novo SCLCs. T-LUAD and T-SCLC did appear to represent intermediate phenotypes and demonstrated substantial overlap in expression profile (Fig. 4C). This suggests that T-LUADs might be distinctly primed to transform, relative to other LUAD, and that T-SCLC retains some transcriptomic features of T-LUAD. PCA of methylation profiling by EPIC revealed that T-SCLCs exhibit distinct methylation profiles to those of de novo SCLCs and show proximity to the methylome of T- and control LUADs (Fig. 4D). This implies that tumors undergoing NE transformation retain broad-scale epigenomic features of the LUAD from which they derived.

To further analyze the transcriptional changes occurring during NE transformation, we performed differential gene expression and pathway enrichment analyses (gene set enrichment analysis, GSEA) of T-LUAD and T-SCLC samples (Fig. 4E). As expected, EGFR expression was downregulated upon transformation (Supplementary Fig. S4A). T-SCLC demonstrated increased expression of NE markers such as SYP, SYN1, and INSM1 and of genes associated with Notch signaling inhibition, including DLL3 and HES6 (15). Pathway enrichment analyses performed on differentially expressed genes (DEG) in T-LUAD versus T-SCLC samples (Fig. 4E and F) showed T-SCLC–specific upregulation of genes involved in (i) neural differentiation (including SEZ6, TAGLN3, and KCNC1); (ii) cell-cycle progression (including E2F2, CENPF, and FBXO5); (iii) DNA repair (including FANCB, EYA2 and RFC3); (iv) chromatin remodeling (including HDAC2); and (v) PRC2 complex (including HIST1H2BO, HIST1H2BL, and HISH1H4H; Fig. 4E and F). We further confirmed a consistent increase in the mRNA expression of EZH2, enconding the enzymatic subunit of the PRC2 complex (Supplementary Fig. S4B), previously strongly implicated in lineage plasticity and NE transformation in prostate cancer (4). GSEA also showed a gene-expression signature of induced WNT signaling in T-SCLC, with downregulation of the negative regulator of WNT signaling TCF7L2 and overexpression of WNT pathway activators such as WNK2, ASPM, and FZD3 (Fig. 4E and F). This was further supported by protein analysis results of T-LUAD and T-SCLC samples and patient-derived xenografts (PDX; Fig. 4G; Supplementary Fig. S4C and S4D; Supplementary Table S12). We observed increased expression of the major WNT signaling effector, β-catenin, and increased phosphorylation of PYK2, a protein involved in WNT signaling activation (21) in T-SCLC (Fig. 4G; Supplementary Fig. S4C; Supplementary Tables S11 and S12). Among other changes, NE transformation was also associated with global downregulation of receptor tyrosine kinase signaling, inhibition of apoptotic induction, suppression of antitumor immune activation, and induction of PI3K/AKT signaling (Fig. 4E and F), which was also confirmed at the protein level in protein array assays (Fig. 4G; Supplementary Fig. S4C) and Western blot (Supplementary Fig. S4D).

Integrative analyses of transcriptomic and epigenomic data showed that a substantial number of DEGs were also differentially methylated in T-SCLC relative to T-LUAD, consistent with epigenomic reprogramming upon NE transformation in the lung. We observed cell adhesion, neuron differentiation, cytokine signaling, and neutrophil degranulation pathways to be among the top pathways differentially affected by methylation (Fig. 5A; Supplementary Fig. S4E).

Methylation occurring in TF-binding motifs can inhibit TF engagement and affect regulation of target gene expression (22). Analysis of differential methylation of TF-binding motifs revealed hypomethylation of binding motifs for genes involved in (i) neuronal and NE differentiation (including ASCL1 and NEUROD1); (ii) WNT signaling activators (TCF4, EBF2); (iii) stemness (NANOG, BHLHA15); and (iv) epithelial–mesenchymal transition (EMT; SNAI1, TWIST1/2, ZEB1, among others) in T-SCLC relative to T-LUAD (Fig. 5B). We also found T-SCLC–specific hypermethylation of binding motifs for TFs involved in MAPK signaling (JUNB/D, AP-1, FOSL1/2) and WNT signaling suppression (SOX7/10/17; Fig. 5B). These data suggest that epigenomic reprogramming upon transformation leads to altered methylation of key TF-binding motifs, driving expression phenotypes observed during histologic transition (Fig. 4E and F).

Notably, three TFs—FOXN4 (β = 3.38, q-value = 0.031), ONECUT2 (β = 3.10, q-value = 0.014), and POU3F2 (β = 2.02, q-value = 0.083)—were among the top DEGs upregulated in T-SCLCs (Supplementary Fig. S5A). ONECUT2 and POU3F2 have previously been implicated in acquisition and maintenance of the NE phenotype in prostate cancer (23, 24). FOXN4 has previously been shown to interact with ASCL1 to modulate Notch signaling (25). To assess the role of these TFs as drivers of NE transformation, we overexpressed FOXN4, ONECUT2, and POU3F2 each independently in two EGFR-mutant LUAD cell lines (PC9 and HCC827; Supplementary Figs. S5B and S5C). Ectopic overexpression of these factors downregulated EGFR expression in both lines (Supplementary Fig. S5B), as also observed after ASCL1 or NEUROD1 overexpression (Supplementary Fig. S5D). ONECUT2 and POU3F2 overexpression increased osimertinib resistance in one of the cell lines under study (Supplementary Fig. S5E), and osimertinib exposure increased FOXN4 and ONECUT2 expression in both cell lines (Supplementary Fig. S5F and S5G). These results suggest that although these TFs might not individually induce NE transformation, they may contribute to transformation through downregulation of EGFR expression, a commonly observed phenotype in EGFR-mutant T-SCLC (3, 13). Additionally, these results highlight a link between anti-EGFR therapy and upregulation of TFs implicated in transformation.

Taken together, these data highlight that while epigenetic reprogramming in NE transformation results in induction of transcriptional changes affecting several key signaling pathways, some epigenomic features are maintained during NE transformation, differentiating these tumors from de novo SCLC. Transformation to an NE phenotype may be promoted by the PRC2 complex and other epigenetic modifiers and appears to be characterized by activation of the PI3K and WNT signaling pathways, acquisition of a mesenchymal phenotype, and suppression of antitumor immune response pathways.

To identify transcriptional changes that may predispose to NE transformation, we next compared the transcriptomic and methylomic profiles of T-LUAD and control (never-transformed) LUADs (Fig. 6A–C). In the T-LUAD samples, we observed relative downregulation of a variety of keratin genes (KRT7, KRT8, and KRT15, among others; Fig. 6C), consistent with a potential partial loss of the LUAD phenotype (26). As expected, we also observed multiple alterations in the RB pathway (Fig. 6A). RB1 mutations and Rb protein loss were found in 36% (4/11) and 86% (6/7), respectively, of T-LUADs. We also observed differential expression of members upstream of RB1 (possibly in compensation for RB1 functional deficiency; refs. 27, 28), including upregulation of CDKN2A associated with an increase in gene body methylation (Fig. 6A; Supplementary Fig. S6A), downregulation of CCND1 (Cyclin D1), and upregulation of CCNE1/2 (Cyclin E1/2). These results are consistent with prior observations that RB1 loss of function precedes NE transformation (3, 9).

We also identified DEGs representative of some of the same pathways identified when comparing T-LUAD and T-SCLC samples, suggesting progressive differential regulation of these pathways in NE transformation (Fig. 6B and C). These included upregulation of genes enriched in cell-cycle progression (TOP2A, CENPF, and FBXO5), DNA repair pathways (CLSPN, EXO1, and FANCB), and PI3K/AKT signaling (PIK3CA, PIK3R1, and AKT3), as well as downregulation of RTK signaling (DUSP6, ERBB2, and MAPK13), cell adhesion [CDH1 (E-cadherin), PCDHA11, PCDHA9], and antitumor immune response (multiple genes involved in neutrophil degranulation, TNF signaling, and antigen presentation). Consistent with the known role of Notch signaling in suppressing NE tumor growth (29), these analyses revealed early downregulation of genes involved in Notch signaling, including Notch receptors NOTCH1/2/3 and ligands JAG2 and DLL4 (Fig. 6B and C). Consistent with an overall retention of genome methylation patterns of LUAD, integrative analyses with transcriptomic and methylation data revealed that none of these pathways were likely being differentially regulated by gene-specific methylation (Supplementary Fig. S6B).

These results suggest that an intermediate phenotype is captured in T-LUAD specimens, which is further accentuated upon NE transformation to T-SCLC. This phenotype is characterized by partial loss of LUAD features and of dependence on RTK signaling, and by the upregulation of gene programs promoting AKT signaling, cell-cycle progression, and DNA repair, as well as downregulation of genes related to immune response and Notch signaling.

Finally, we sought to explore molecular differences between transformed and de novo SCLCs. Comparison of the transcriptome of T-SCLCs to that of our control de novo SCLCs revealed lower expression of genes involved in neuron differentiation (SALL3, DLX1, and NEURL1), Notch signaling (JAG2, DLL1/4, and NOTCH3), PI3K/AKT pathway (AKT1/2, BAD, and TSC2), and epigenetic regulation (HIST2H3D, SMARCA4, and ARID1B; Fig. 6D and E). We also observed higher expression of genes involved in stemness (such as CD44, NAMPT, or the aldehyde dehydrogenase ALDH1A2), IFN signaling (TLR2/3/7/8 and CLEC7A), lymphocyte chemotaxis (CXCL10/13/14, XCL, and CCL5), and TCR signaling (PAK2, UBE2D2, and NCK1) in T-SCLCs relative to de novo SCLCs. Integrative transcriptome/methylome analyses (Fig. 6F; Supplementary Fig. S6C) indicated that the suppressed neuronal phenotype in T-SCLCs was associated with a high number of differentially methylated genes in that pathway, suggesting epigenetic reprogramming (Fig. 6F). These results suggest that T-SCLC is characterized by decreased neuronal features, an accentuated stem-like/plastic phenotype, and an increased ability to induce an antitumor immune response relative to de novo SCLC. These data further support that inhibition of Notch signaling may be particularly key for SCLC transformation and persists after histologic transition.

The identification of novel therapeutic approaches to treat or prevent the emergence of T-SCLC is a major clinical need. Toward this end, we tested different targeted therapies on an EGFR-mutant PDX model (Fig. 7A–D; Supplementary Fig. S7A–S7D) exhibiting combined LUAD and NE histology with small cell features (T14-CH; Fig. 7A and C). This model was derived from a patient with LUAD whose tumor subsequently underwent SCLC transformation (T14; Supplementary Table S2). The resulting PDX from the original LUAD exhibited combined histology after serial passaging. Targeted sequencing confirmed the presence of the same EGFR mutation as in the pretransformation LUAD clinical sample (EGFR S768_D770dup).

We targeted the major signaling pathways upregulated upon NE transformation in our transcriptomic and protein analyses and prioritized targets with the highest potential for clinical translation. Inhibitors assessed included samotolisib (PI3K/KT/mTOR inhibitor; ref. 30), DS-3201-b (EZH1/2 inhibitor; ref. 31), and G007-LK (WNT inhibitor; ref. 32). The T14-CH model was treated with either vehicle or single-agent samotolisib, DS3201-b, or G007-LK without or with osimertinib. Osimertinib monotherapy significantly delayed tumor growth (Fig. 7B; T/C = 63.9%, P = 0.011 at the control group endpoint). Combinations of DS-3201b or G007-LK with osimertinib did not achieve greater tumor growth inhibition than osimertinib monotherapy alone (Supplementary Fig. S7A).

Interestingly, the AKT inhibitor samotolisib caused significant delay in tumor growth as a single agent (T/C = 54.2%, P < 0.001), and the combination of samotolisib with osimertinib further inhibited tumor growth in this model, with a T/C value of 33.54% at the control group endpoint (P < 0.001; Fig. 7B). The combination group tumors showed an average 55.8% size reduction relative to the osimertinib-treated group at experiment endpoint. Western blots for pAKT and pPRAS40 in these tumors confirmed AKT signaling inhibition by samotolisib (Supplementary Fig. S7B). Mouse body weight measurements suggested that the combination did not cause significantly greater toxicity than osimertinib monotherapy (Supplementary Fig. S7C).

Tumors were collected and analyzed for histology and for IHC expression of the LUAD markers TTF-1 and Napsin A, and of the NE markers ASCL1, NEUROD1, and chromogranin A (Fig. 7A and C). Control tumors demonstrated mixed histology, with areas of LUAD showing high TTF-1 and mild Napsin A staining with no expression of any of the NE markers tested, and areas of SCLC-like cells with lower expression of TTF-1 and no expression of Napsin A (Fig. 7A and C). The SCLC component showed high levels of NEUROD1 and chromogranin A, but no ASCL1 staining (Fig. 7A and C). Osimertinib monotherapy completely depleted the LUAD component of these tumors (Fig. 7D), suggesting that after transformation, osimertinib exerts a selective pressure enriching for the NE component.

Osimertinib-treated tumors showed areas of mutually exclusive staining for ASCL1 or NEUROD1, all of them positive for chromogranin A (Fig. 7C). Samotolisib-treated tumors showed enrichment in the LUAD component as compared with control tumors (Fig. 7C and D), suggesting that AKT inhibition may be exerting a selective pressure against the NE component in these tumors, which was consistent with increased pAKT and pRAS40 IHC staining in the NE component (Fig. 7A; Supplementary Fig. S7D). The addition of samotolisib to osimertinib attenuated the depletion of LUAD and completely suppressed the emergence of ASCL1-positive SCLC seen with osimertinib alone (Fig. 7B and D).

These results further support a role of AKT signaling in NE transformation in the lung and suggest that combined AKT and EGFR inhibition may constrain NE relapse in EGFR-mutant tumors at risk of transformation.

Cancer cell promiscuity in lineage commitment is a reflection of the exceptional heterogeneity of tumors and an important source of treatment failure. The advent of potent and specific targeted inhibitors for mutational drivers in LUAD, like the use of highly effective antiandrogenic agents in prostate cancer, has prompted increasing recognition of lineage plasticity as a primary barrier to successful management of cancer. Although frequently considered in the context of acquired therapeutic resistance, lineage plasticity in cancer is also evident independent of drug selection. In this study, we took advantage of the long-standing recognition of mixed histology lung cancers to gain insight into the molecular phenotypic landscapes underlying histologic transformation between LUAD and SCLC lineages. WES confirmed that the histologically distinct components of mixed tumors were clonally related, reflecting distinct lineage pathways derived from a shared tumorigenic founder. Tumor mutational burden (TMB) was similar in the LUAD and SCLC components in these samples, reflecting their presumed temporal proximity. By focusing primarily on a cohort of biphenotypic tumors in which the distinct lineages are in temporal and spatial proximity, we had the opportunity to identify consistent molecular changes that characterize this transformation. In this study, we provide the first comprehensive multiomic characterization of NE transformation in lung cancer, including genomic, transcriptomic, epigenomic, and protein analyses of matched samples. Primary limitations of this study include the use of formalin-fixed, paraffin-embedded (FFPE) samples, with reduced sequencing quality relative to fresh samples, and the limited number of samples available for analysis. These factors should be taken into consideration in interpreting this data set.

One conclusion that may be taken from our data concerns the degree to which activation of lineage plasticity can result in distinct cell fates. De novo SCLC has been classified into four distinct subtypes based on differential expression of master transcriptional regulators (3, 15, 33). Examining a cohort of mixed histology LUAD/SCLC tumors, we find that the T-SCLC derivatives do not consistently fall into one of these subtypes—rather we find all four subtypes clearly represented among just 11 cases. This observation underscores the degree to which plasticity in lung cancer can activate diverse transcriptional programs. Particularly surprising to us was the identification of mixed histology tumors in which the T-SCLC component expressed POU2F3, defining the subtype SCLC-P. LUAD is believed to derive from type II pneumocytes (34). Based on its expression profile, SCLC-P had been proposed to arise from transformation of tuft cells, a rare pulmonary cell that is the exclusive source of POU2F3 expression in lung (20). The identification of two independent cases of clonally linked T-LUAD and POU2F3-expressing T-SCLC calls into question the cell of origin of SCLC-P and highlights the capacity of lineage plasticity to allow cancer cells to transdifferentiate between clearly distinct biological lineages.

Several features of the analysis of mixed histology T-LUAD/T-SCLC tumors reflect prior observations made regarding NE transformation of LUAD, reinforcing the relevance of this approach. Consistent with previous publications (9, 13), we observe inactivation of TP53 and RB1 in essentially all T-SCLC, and in nearly all of the paired T-LUAD (Fig. 7E). Notably, inactivation of these two key tumor suppressors was not always evident in exome sequencing, highlighting the importance of determining TP53 and RB1 protein expression as a complement to genetic testing. Although NE transformation of LUAD was originally observed in EGFR-mutant LUAD under selective pressure of EGFR TKI treatment, we confirm here similar histologic transformation regardless of EGFR mutation, including in the treatment-naïve setting (3, 8). A novel finding here is the high frequency of 3p chromosome arm loss in T-LUADs (Fig. 7E; ref. 35). What genes resident on 3p singly or in combination could account for this observation is currently unclear, but 3p loss may represent a novel risk factor for NE transformation, which, combined with the determination of the RB1/TP53 genomic and expression status, may increase the sensitivity to predict this histologic transition.

The paucity of recurrent mutations across samples in our cohort suggests that NE transformation in the lung is not dependent on a common mutational driver but rather may be primarily dependent on epigenetic shifts in gene-expression programs. Transcriptional analysis of the T-LUAD and T-SCLC components of our mixed histology tumor set, relative to control (nontransformed) LUAD and de novo SCLC, suggests that T-LUADs and T-SCLCs occupy intermediate, transitional states—states that overlap with both with their apparent nontransforming histology and each other. Our data point to a number of signaling pathways that appear to shift in consistent patterns from T-LUAD to T-SCLC (Fig. 7E). These shifts include higher expression of genes in cell cycle and DNA repair, consistent with the highly proliferative capacity of SCLC tumors (36). Higher expression of NE and mesenchymal features in T-SCLC agrees with previous reports suggesting that NE transformation may occur through an intermediate EMT stem-like state (37, 38). Our data correlate this with putative methylation-induced repression of cell adhesion molecules, and induced expression of mesenchymal effectors such as CDH2 (N-cadherin) and NCAM1 associated with demethylation of binding motifs of key mediators of EMT, such as SNAI1 and TWIST1 in T-SCLC.

Our data implicate multiple pathways known to regulate stem and progenitor cell biology in lineage plasticity and NE transformation (Fig. 7E), notably including upregulation of PRC2 complex activity, induction of WNT signaling, and suppression of the Notch pathway. The induction of PRC2 activity is in keeping with its apparent role in NE transformation in prostate cancer (4, 39). WNT signaling too has previously been implicated in lineage plasticity (40) and in the maintenance of an NE phenotype in the prostate (41, 42). Given the previously defined role of Notch signaling in suppression of NE tumor growth, we believe that sustained inhibition of Notch signaling may be a prerequisite for NE transformation in lung.

We note that SCLCs are notoriously immune “cold” tumors relative to NSCLCs (14, 15, 43). Consistent with this, we see a progressive suppression in antitumor immune response pathways including cytokine signaling, T-cell immunity, and neutrophil degranulation from control LUAD to T-LUAD, from T-LUAD to T-SCLC, and from T-SCLC to de novo SCLC.

Finally, we also find consistent evidence of PI3K/AKT pathway activation in T-SCLC. Emerging data support a role for PI3K/AKT signaling in lineage plasticity and NE transformation (3, 44). Overactivation of this pathway may predict higher risk of transformation, as mutations on AKT pathway members are enriched in LUAD at high risk of transformation (9), and we consistently found upregulation of genes involved in this pathway in T-LUAD as compared with LUAD. AKT has also been identified as a driver of NE phenotypic shift in nontumoral prostate and lung cells (45). In line with these findings, our results suggest that T-SCLC may be sensitive to AKT inhibition, which delays NE relapse in an EGFR-mutant PDX model of NE transformation in combination with osimertinib. These results support a novel therapeutic approach to delay or revert NE transformation in lung.

NE transformation in lung cancer induces a highly lethal and recalcitrant tumor profile that currently lacks effective treatments. A better understanding of molecular drivers of NE transformation in lung cancer can nominate therapeutic targets to treat or prevent transformation. Through detailed analysis of transformation pairs, we provide a comprehensive molecular characterization of NE transformation in lung cancer, describing the signaling pathways and phenotypes altered during histologic transformation mediated by lineage plasticity, and defining a potential therapeutic approach to inhibit emergence of T-SCLC in targeted treatment of LUAD.

We identified 11 FFPE tumors with combined LUAD and SCLC histology, from which independent isolation of both histologic components was possible (N = 11; Supplementary Tables S1 and S2; Supplementary Fig. S1). As the components of these mixed histology tumors are not temporally ordered, we refer to the component parts of these mixed histology tumors as “T-LUAD” and “T-SCLC,” with the T referring to histologic transformation. We identified an additional five pretransformation LUAD and three posttransformation SCLC cases for which tissue material was available (Supplementary Tables S1 and S3). As controls, we included a group of never-transformed LUADs (N = 15) and a set of de novo SCLC samples (N = 18; Supplementary Tables S1 and S4). All study subjects had provided signed informed consent for biospecimen analyses under an institutional review board–approved protocol.

For microdissection, hematoxylin and eosin (H&E)–stained FFPE tumor slides of tumors with combined LUAD/SCLC were independently evaluated by two pathologists. Where possible, multiple FFPE blocks of each tumor were reviewed, with the aim of selecting areas containing exclusively the LUAD or the SCLC component. Where individual slides with pure components were not available, slides containing both histologic components with complete physical separation were selected. Between 10 and 20 unstained sections at 10 μm prepared on uncharged slides from corresponding FFPE blocks were used for microdissection of each case. Every 10 sections, an additional section was stained with H&E for confirmation of histology. The areas corresponding to each histologic component on the initial H&E were dissected using a clean blade and the tissue collected in 0.5 mL nuclease-free tubes for nucleic acid extraction. Alternatively, 1.0 to 1.5 mm core punches were made from LUAD and SCLC areas on the FFPE blocks and placed in 0.5 mL nuclease-free tubes for nucleic acid extraction, exclusively in cases where each histologic component was located in a different block and where no histologic cross-contamination was confirmed by pathologic review.

FFPE tissue was deparaffinized using heat treatment (90°C for 10 minutes in 480 μL PBS and 20 μL 10% Tween 20), centrifugation (10,000 × g for 15 minutes), and ice chill. Paraffin and supernatant were removed, and the pellet was washed with 1 mL 100% EtOH followed by an incubation overnight in 400 μL 1 mol/L NaSCN for rehydration and impurity removal. Tissues were subsequently digested with 40 μL proteinase K (600 mAU/mL) in 360 μL Buffer ATL at 55°C. DNA isolation proceeded with the DNeasy Blood and Tissue Kit (QIAGEN, #69504) according to the manufacturer's protocol modified by replacing AW2 buffer with 80% ethanol. DNA was eluted in 0.5X Buffer AE.

FFPE sections were deparaffinized in mineral oil. Briefly, 800 μL mineral oil (Fisher Scientific, #AC415080010) and 180 μL Buffer PKD were mixed with the sections, proteinase K was added for tissue digestion, and the sample was incubated at 56°C for 15 minutes. Phase separation was encouraged with centrifugation, and the aqueous phase was chilled 3 minutes to precipitate RNA. After centrifugation for 15 minutes at 20,000 × g, RNA-containing supernatant was removed for extraction, while DNA remained in the pellet. Nucleic acids were subsequently extracted using the AllPrep DNA/RNA Mini Kit (QIAGEN, #80204) according to the manufacturer's instructions. RNA was eluted in nuclease-free water and DNA in 0.5X Buffer ATE.

Frozen tissues were weighed and homogenized in RLT, and nucleic acids were extracted using the AllPrep DNA/RNA Mini Kit (QIAGEN, #80204) according to the manufacturer's instructions. RNA was eluted in nuclease-free water and DNA in 0.5X Buffer EB.

After PicoGreen quantification and quality control by Agilent BioAnalyzer, 100 to 500 ng of DNA was used to prepare libraries using the KAPA Hyper Prep Kit (Kapa Biosystems, KK8504) with 8 cycles of PCR. After sample barcoding, 100 ng of library was captured by hybridization using the xGen Exome Research Panel v1.0 (IDT) according to the manufacturer's protocol. PCR amplification of the postcapture libraries was carried out for 12 cycles. Samples were run on a HiSeq 4000 in a 100 bp/100 bp paired-end run, using the HiSeq 3000/4000 SBS Kit (Illumina).

After PicoGreen quantification and quality control by Agilent BioAnalyzer, 100 ng of library transferred from the DMP was captured by hybridization using the xGen Exome Research Panel v1.0 (IDT) according to the manufacturer's protocol. PCR amplification of the postcapture libraries was carried out for 8 cycles. Samples were run on a HiSeq 4000 in a 100 bp/100 bp paired-end run, using the HiSeq 3000/4000 SBS Kit (Illumina).

Tumor and normal specimens were sequenced to an average median coverage of 120× and 71×, respectively. The median duplication rate was 39%, and uniquely mapped reads were above 99% for both. In-depth sample-level coverage and alignment metrics as computed using Picard tools (http://broadinstitute.github.io/picard/) are provided in Supplementary Table S13. Tumor purity ranged from 25% to 89% tumor cell content as estimated from sequencing data except for case 10, where the algorithm was unable to predict purity, which is expected in low-purity samples.

We used a comprehensive in-house WES pipeline, TEMPO (Time-Efficient Mutational Profiling in Oncology; https://github.com/mskcc-cwl/tempo), that performs alignment using BWA-mem algorithm followed by mutation calling using Strekla2 and Mutect2 variant callers. The combined, annotated, and filtered variant calls were used for downstream analysis. Details of the variant call processing are described at https://ccstempo.netlify.com/variant-annotation-and-filtering.html#somatic-snvs-and-indels and have previously been described as well (46). Copy-number analysis was performed with FACETS (Fraction and Allele-Specific Copy Number Estimates from Tumor Sequencing; https://github.com/mskcc/facets), processed using facets-suite (https://github.com/mskcc/facets-suite), and manually reviewed and refitted using facets-preview (https://github.com/taylor-lab/facets-preview). To delineate mutational processes driving the acquisition of somatic alterations, mutational signatures were decomposed for all tumor samples that had a minimum of five single-nucleotide somatic mutations using the R package mutation-signatures (https://github.com/mskcc/mutation-signatures). Further, a given signature was considered to be “dominant” if the proportion of mutations contributing to the signature was at least 20% of all mutations detected in the sample.

Purity, ploidy, TMB, genome doubling, and cancer cell fractions for all mutations in all specimens were inferred from sequencing data. We estimated neoantigen load by taking the number of variants estimated to have strong class I MHC binding affinity by NetMHC 4.0 (47) and normalizing it by the TMB. We summarized the top occurring somatic variants located on cancer genes in an oncoprint using the R package ComplexHeatmaps version 2.0.0 (https://github.com/jokergoo/ComplexHeatmap; ref. 48). Cancer genes were genes defined as “OncoKB Annotated” on the Cancer Gene List downloaded in June 2020 (https://www.oncokb.org/cancerGenes). All other plots for this analysis were created using ggplot version 3.3.2 (https://github.com/tidyverse/ggplot2).

We compared somatic mutations and gene-level calls (CNA) in cancer genes in our T-LUAD samples to those in TCGA-LUAD cohort. The mutations for the TCGA-LUAD50 cohort were extracted using the R package TCGAmutations (https://github.com/PoisonAlien/TCGAmutations) and selecting cancer type as “LUAD” by using an in-built R function “TCGAmutations::tcga_load(study = “LUAD”).” For this set of TCGA samples and for our LUAD cohort, we predicted gene-level CNAs using the FACETS algorithm. A Fisher exact test was then performed via R function (https://rdrr.io/bioc/maftools/man/mafCompare.html) from maftools R package v.2.0.16 (https://github.com/PoisonAlien/maftools) to identify significantly altered mutations, whereas that for gene-level CNAs (amplifications and deletions) was performed separately using custom R code. For both mutations and CNAs, genes with P < 0.05 were considered significantly altered. The results were summarized in a volcano plot using the R packages ggplot and “EnhancedVolcano” version 1.7.4 (https://github.com/kevinblighe/EnhancedVolcano). FDRs and q-values were calculated for multiple hypothesis testing using the Benjamini–Hochberg method. All analysis was performed using the R environment for statistical computing.

Evolutionary relationships among samples were inferred using the union of somatic mutations called in any of the tumors for a given case/patient. The genomic evolution trees were manually constructed based on the most parsimonious sequence of events using mutations and copy-number events with the shared alterations represented by the trunk and the private alterations represented by the branches with the trunk and arm lengths being proportional to the relative number of shared and private alterations. Somatic variant sites included were those (i) covered at 20-fold or greater in all tumor and 10-fold or greater in matched normal specimens, (ii) supported by greater than three reads in the tumor, (iii) present in less than three reads in the matched normal, and (iv) with a VAF in any affected tumor of greater than 2% for hotspots and 5% for non-hotspots. In order to ensure that private alterations are truly private, (i) we genotyped BAM files from matched patient samples to find whether the said alteration is present, albeit at a subthreshold level; if the alteration was indeed genotyped, it was not considered as truly private; and (ii) we excluded mutation loci where the matched patient sample shows a copy-number loss in the region. Samples with sufficient purity (greater than 30%) were selected for this analysis. Genotyping was performed using GetBaseCountsMultiSample v.1.2.2 (https://github.com/mskcc/GetBaseCountsMultiSample). Mutational cancer cell fraction was inferred using annotate-maf-wrapper function from facets-suite (https://github.com/mskcc/facets-suite), which uses the VAF, locus-specific read coverage, and an analytical estimate of tumor purity, as previously described (46). WGD was also inferred using methods implemented in facets-suite utilizing FACETS (49) output, and mutational timing analysis in the context of WGD was performed as previously described (50).

After PicoGreen quantification (Thermo Fisher, #P11496) and quality control by Agilent BioAnalyzer, 170 tp 750 ng of genomic DNA was sheared using a LE220-plus Focused-ultrasonicator (Covaris, #500569). Samples were cleaned using Sample Purification Beads from the TruSeq Methyl Capture EPIC LT Library Prep Kit (Illumina, #FC-151-1002) according to the manufacturer's instructions with modifications. Briefly, samples were incubated for 5 minutes after the addition of SPB, 50 μL RSB was added for resuspension, and resuspended samples were incubated for 2 minutes. Sequencing libraries were prepared using the KAPA Hyper Prep Kit (Kapa Biosystems, KK8504) without PCR amplification. Postligation cleanup proceeded according to Illumina's instructions with 110 μL Sample Purification Mix. After purification, three to four samples were pooled, and equimolar and methylome regions were captured using EPIC oligos. Capture pools were bisulfite converted and amplified with 11 to 12 cycles of PCR. Pools were sequenced on a NovaSeq 6000 or HiSeq 4000 in a 150/150 bp or 100/100 bp paired-end run, using the NovaSeq 6000 S4 Reagent Kit (300 cycles) or HiSeq 3000/4000 SBS Kit (Illumina). The average number of read pairs per sample was 51 million.

The Bismark pipeline (51) was adopted to map bisulfite-treated EPIC sequencing reads and determine cytosine methylation states. Trim Galore v0.6.4 was used to remove raw reads with low quality (less than 20) and adapter sequences. The trimmed sequence reads were C(G) to T(A) converted and mapped to similarly converted reference human genome (hg19; ref. 52) using default Bowtie 2 (53) settings within Bismark. Uniquely aligned reads (52%–83%) were retained. Duplicated reads (30%–90%) were discarded. The remaining alignments were then used for cytosine methylation calling by Bismark methylation extractor. The average cytosine coverage ranges from 1.2× to 11× across 49 samples. The details of statistics on EPIC data can be found in Supplementary Table S14. Because CHG and CHH (H: one of A, T, C) show minimal presence (<1%), we focus on cytosines in CpG content. The number of CpGs was further filtered by C coverage ≥5 in all samples to yield 735,636 CpGs that will be used in the subsequent analyses.

Differentially methylated CpGs (DMC) were identified using DSS R package (54, 55) on the basis of dispersion shrinkage followed by Wald statistical test for beta-binomial distributions. Any CpGs with FDR < 0.05 and methylation percentage difference greater than 10% were considered significant DMCs. Differentially methylated regions (DMR) were subsequently called based on the DMCs. The called DMRs were required to satisfy the minimum length of 50 bp and minimum 3 CpGs in the region; two neighboring DMRs were merged if less than 50 bp apart; and significant CpGs were those that occupy at least 50% of all CpG population in the called DMRs as default in DSS package. Pairwise comparisons were conducted for pretransformation LUAD versus control LUAD, posttransformation SCLC versus de novo SCLC, and posttransformation SCLC versus pretransformation LUAD. The DMRs were mapped to gene regions at promoters and gene bodies, and differential methylation levels were subsequently associated with differential gene-expression values in selected pathways. In addition to pairwise comparisons, PCA and partial least square discriminant analysis (PLSDA) were performed to classify samples into groups and identify influential CpGs using mixOmics R package (55).

Differential methylation may influence TF binding. To identify overrepresented known TF motifs due to differential methylation for the posttransformation SCLC compared with pretransformation LUAD, “findMotifsGenome.pl” from HOMER (56) was applied to DMCs (±50 bp) overlapping with gene promoter regions. DMC regions with hyper- and hypomethylation in SCLC were explored separately to show the effects from different methylation status. The significantly enriched TFs were defined as those with P ≤ 0.05.

Approximately 500 ng of FFPE RNA or 100 ng of fresh-frozen RNA per sample was used for RNA library construction using the KAPA RNA Hyper library prep kit (Roche) per the manufacturer's instructions with minor modifications. Customized adapters with unique molecular indexes (UMI; Integrated DNA Technologies) and sample-specific dual-index primers (Integrated DNA Technologies) were added to each library. The quantity of libraries was measured with Qubit (Thermo Fisher Scientific), and quality was measured by the TapStation Genomic DNA Assay (Agilent Technologies). Equal amounts of each RNA library (around 500 ng) were pooled for hybridization capture with IDT Whole-Exome Panel V1 (Integrated DNA Technologies) using a customized capture protocol modified from the NimbleGen SeqCap Target Enrichment system (Roche). The captured DNA libraries were then sequenced on an Illumina HiSeq4000 with paired end reads (2Å ∼ 100 bp), at 50 million reads/sample.

In-line UMI sequences were trimmed from the sequencing reads with Marianas (https://github.com/mskcc/Marianas) and aligned to human GRCh37 genome using STAR 2.7.0 (https://github.com/alexdobin/STAR; ref. 57) with Ensembl v75 gene annotation. Hybrid selection–specific metrics and alignment metrics were calculated for the BAM files using CalculateHsMetrics and CollectRnaSeqMetrics, respectively, from Picard Toolkit (https://github.com/broadinstitute/picard) to determine the quality of the capture.

We quantified RNA-seq reads with Kallisto v.0.45.0 (58) to obtain transcript counts and abundances. Kallisto was run with 100 bootstrap samples, with sequence-based bias correction, and in strand-specific mode, which processed only the fragments where the first read in a pair is pseudoaligned to the reverse strand of a transcript. Differential gene-expression analysis, PCA, and transcript per million (TPM) normalization by size factors, were done from Kallisto output files using Sleuth v0.30.0 run in gene mode (59). DEG were identified using the Wald test. Genes were marked significant if the FDRs and q-values, calculated using the Benjamini–Hochberg menthod, were less than 0.05, and beta(Sleuth-based estimation of log₂ fold change)>1.25, which approximately correlated to a log₂ fold change of 2 in our data. The log of the normalized TPM values for selected significant genes was rescaled using a z-score transformation and plotted in a heatmap using the ComplexHeatmap Library in R. RNA-seq data quality control is provided in Supplementary Table S15.

To confirm genetic relatedness for the case lacking WES data, known tumor mutations from MSK-IMPACT were genotyped in both RNA samples using GetBaseCountsMultiSample v.1.2.2 (https://github.com/mskcc/GetBaseCountsMultiSample). For locations where transcripts were expressed, mutations with at least 10 variant alleles were annotated as present.

GSEA (60) was performed on full sets of gene-expression data across the previously mentioned three comparisons. Genes were ranked on P-value scores computed as −log₁₀(P value) × (sign of beta). Gene set annotations were taken from the Molecular Signatures Database (MSigDB v7.0.1; refs. 60, 61). Gene sets tagged by the Kyoto Encyclopedia of Genes and Genomes (KEGG; refs. 62, 63) and REACTOME (64) pathways were retained for further analysis. The significance level of enrichment was evaluated using permutation test, and the P value was adjusted by the Benjamini–Hochberg procedure. Any enriched gene sets with adjusted P ≤ 0.05 were regarded as significant. This analysis was conducted using ClusterProfiler R package (65). The enriched gene sets that are influenced by DMCs were selected and pathway annotations concatenated manually to remove redundancy and achieve high-level generality. When the pathway terms were merged, median enrichment score was taken as the new group enrichment score, P values were aggregated using the Fisher method from the Aggregation R package (66), and the core enrichment of genes was collapsed.

Protein samples were quantified with the Bradford method (#5000205, Bio-Rad), and 200 μg aliquots were used in the phospho-kinase array (#ARYC003C, R&D-Biotechne), which was performed using the manufacturer's instructions. Quantification of spots was performed using Image Studio software (Version 3.1, Li-Cor). Technical replicates (two per array) per sample were averaged. Two-tailed Student t test was performed on these values, comparing the T-LUAD and T-LUSC groups.

The PC9 cell line was purchased from Millipore Sigma (#90071810-VL), and the HCC827 cell line was purchased from ATCC (#CRL-2868). Both cell lines were authenticated by the STR method and regularly tested (last tested April 2021) for Mycoplasma with the MycoAlert kit (#LT07-218, Lonza) and maintained in RPMI-1640 10% FBS. Both cell lines were used after little passages after purchase (5–10). Lentiviruses were produced as previously described (67) with FOXN4 (#EX-I2262-Lv151, GeneCopoeia), POU3F2 (#EX-A3238-Lv151, GeneCopoeia), and ONECUT2 (#EX-Z4476-Lv151, GeneCopoeia) overexpression lentiviral plasmids, with an EGFP overexpression plasmid as control plasmid (#EX-EGFP-Lv151, Genecopoeia). Cell lines were transduced at a high multiplicity of infection as previously described (67) with overnight virus incubation.

Protein extraction and Western blot were performed as previously described (68). Antibodies for FOXN4 (#PA539174, Thermo Fisher), ONECUT2 (#ab28466, Abcam), POU3F2 (#12137, Cell Signaling Technology), EGFR (#4267, Cell Signaling Technology), ASCL1 (#556604, BD), NEUROD1 (#ab109224, Abcam), pAKT (S473, #4060S, Cell Signaling Technology), pPRAS40 (T246, #13175, Cell Signaling Technology), β-catenin (#8480, Cell Signaling Technology), Vinculin (#13901, Cell Signaling Technology), and actin (#3700, Cell Signaling Technology) were used.

RNA extraction, reverse transcription, and quantitative PCR were performed as previously described (69). FOXN4 expression was normalized to that of GAPDH. Fluorescent probes against FOXN4 (#4351372, Applied Biosystems) and GAPDH (#4331182, Applied Biosystems) were used.

For treatments, 5 to 10 NOD.Cg-Prkdc<scid> Il2rg<tm1Wjl>/SzJ (NSG) mice were engrafted per treatment arm and incubated until they reached 100 to 150 mm³. At that point, mice were randomized into groups and treated with either vehicle, osimertinib (25 mg/kg/day, by oral gavage), DS-3201b (50 mg/kg/day, by oral gavage), samotolisib (10 mg/kg/day, by oral gavage), G007-LK (20 mg/kg/day, intraperitoneally), or combinations of osimertinib with either of the other inhibitors using the same concentrations as in monotherapy, 5 days a week. For most treatment arms, mice were sacrificed when tumors reached approximately 1,500 mm³ and fixed in formalin 10% overnight for paraffin embedding. For the samotolisib and samotolisib plus osimertinib groups, mice were sacrificed in parallel to the osimertinib treatment group for direct comparison. Tumors and mouse body weight were measured twice a week. T/C values were calculated by normalizing average tumor size of the treatment arm of interest with the average tumor size of the treatment group, both at control arm endpoint (day 21). FFPE tumors were stained as previously described (33) using TTF-1 (#M3575, Dako), Napsin A (#NCL-NapsinA, Leica; Novocast), ASCL1 (#556604, BD Pharmingen), NEUROD1 (#ab205300, Abcam), and chromogranin A (#A0430, Dako). All animal experiments were approved by the Memorial Sloan Kettering Cancer Center Animal Care and Use Committee.

WES data of combined LUAD and SCLC tumor samplesare available under SRA accession ID PRJNA727969. RNA-seq data of combined LUAD and SCLC tumors to look at gene-expression changes are available under ArrayExpress accession ID E-MTAB-10399. Methylation capture EPIC sequencing data of combined LUAD and SCLC tumors to look at gene-expression changes are available under ArrayExpress accession ID E-MTAB-10617.

A. Quintanal-Villalonga reports other support from AstraZeneca outside the submitted work. M. Offin reports personal fees from PharmaMar, Targeted Oncology, Novartis, Jazz Pharmaceuticals, Merck Sharp & Dohme, Bristol Myers Squibb, and OncLive outside the submitted work. M.H. Roehrl reports personal fees from Trans-Hit, Universal DX, and Proscia outside the submitted work. T.J. Hollmann reports grants from Parker Institute for Cancer Immunotherapy, Bristol-Myers Squibb, and Calico Labs outside the submitted work. H.A. Yu reports personal fees and other support from AstraZeneca, Janssen, and Daichi; other support from Novartis, Cullinan, and Pfizer; and personal fees from Blueprint outside the submitted work. C.M. Rudin reports personal fees from AbbVie, Amgen, AstraZeneca, Epizyme, Genentech/Roche, Ipsen, Jazz, Lilly, Syros, Harpoon Therapeutics, Bridge Medicines, and Earli outside the submitted work. No disclosures were reported by the other authors.

A. Quintanal-Villalonga: Conceptualization, data curation, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. H. Taniguchi: Data curation, formal analysis, investigation, writing–review and editing. Y.A. Zhan: Formal analysis, methodology, writing–review and editing. M.M. Hasan: Formal analysis, validation, visualization, methodology, writing–review and editing. S.S. Chavan: Data curation, software, formal analysis, validation, investigation, methodology, writing–review and editing. F. Meng: Data curation, formal analysis, investigation, writing–review and editing. F. Uddin: Investigation, writing–review and editing. P. Manoj: Investigation, writing–review and editing. M.T.A. Donoghue: Data curation, software, formal analysis, writing–review and editing. H.H. Won: Formal analysis, writing–review and editing. J.M. Chan: Formal analysis, validation, methodology, writing–review and editing. M. Ciampricotti: Investigation, writing–review and editing. A. Chow: Resources, writing–review and editing. M. Offin: Resources, writing–review and editing. J.C. Chang: Formal analysis, validation, investigation, writing–review and editing. J. Ray-Kirton: Validation, investigation, writing–review and editing. S.E. Tischfield: Investigation, writing–review and editing. J. Egger: Data curation, writing–review and editing. U.K. Bhanot: Formal analysis, validation, investigation, writing–review and editing. I. Linkov: Formal analysis, validation, investigation, writing–review and editing. M. Asher: Investigation, writing–review and editing. S. Sinha: Investigation, writing–review and editing. J. Silber: Investigation, writing–review and editing. C.A. Iacobuzio-Donahue: Resources, writing–review and editing. M.H. Roehrl: Investigation, writing–review and editing. T.J. Hollmann: Investigation, writing–review and editing. H.A. Yu: Resources, writing–review and editing. J. Qiu: Investigation, methodology, writing–review and editing. E. de Stanchina: Investigation, writing–review and editing. M.K. Baine: Formal analysis, investigation, writing–review and editing. N. Rekhtman: Formal analysis, validation, investigation, visualization, methodology, writing–review and editing. J.T. Poirier: Resources, writing–review and editing. B. Loomis: Resources, data curation, software, formal analysis, investigation, methodology, writing–review and editing. R.P. Koche: Resources, data curation, software, formal analysis, investigation, methodology, writing–review and editing. C.M. Rudin: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. T. Sen: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, visualization, writing–original draft, project administration, writing–review and editing.

This study was supported by NCI R01 CA197936 and U24 CA213274 (to C.M. Rudin), the Druckenmiller Center for Lung Cancer Research (to C.M. Rudin, T. Sen, and A. Quintanal-Villalonga), a Parker Institute for Cancer Immunotherapy grant (to T. Sen), an International Association for the Study of Lung Cancer grant (to T. Sen), NIH K08 CA-248723 (to A. Chow), and the Van Andel Institute–Stand Up To Cancer Epigenetics Dream Team grant (to C.M. Rudin). Stand Up To Cancer is a division of the Entertainment Industry Foundation. Research grants are administered by the American Association for Cancer Research, the Scientific Partner of SU2C. The authors acknowledge the use of the Integrated Genomics Operation Core, funded by an NCI Cancer Center Support Grant (CCSG; P30 CA08748), Cycle for Survival, and the Marie-Josée and Henry R. Kravis Center for Molecular Oncology. they also acknowledge Maria Corazon Mariana and Emily Lin from the PPBC Biobank for their invaluable help. The PPBC Biobank and Pathology Core Facility are supported by the NCI CCSG; P30 CA08748.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.