In vivo functional screens reveal KEAP1 loss as a driver of chemoresistance in small cell lung cancer

Abstract

Exquisitely chemosensitive initially, small cell lung cancer (SCLC) exhibits dismal outcomes owing to rapid transition to chemoresistance. Elucidating the genetic underpinnings has been challenging owing to limitations with cellular models. As SCLC patient-derived xenograft (PDX) models mimic therapeutic responses, we perform genetic screens in chemosensitive PDX models to identify drivers of chemoresistance. cDNA overexpression screens identify MYC, MYCN, and MYCL, while CRISPR deletion screens identify KEAP1 loss as driving chemoresistance. Deletion of KEAP1 switched a chemosensitive SCLC PDX model to become chemoresistant and resulted in sensitivity to inhibition of glutamine metabolism. Data from the IMpower133 clinical trial revealed ~6% of patients with extensive-stage SCLC exhibit KEAP1 genetic alterations, with activation of a KEAP1/NRF2 transcriptional signature associated with reduced survival upon chemotherapy treatment. While roles for KEAP1/NRF2 have been unappreciated in SCLC, our genetic screens revealed KEAP1 loss as a driver of chemoresistance, while patient genomic analyses demonstrate clinical importance.

In vivo genetic screens and patient genomic data analyses reveal that KEAP1 pathway alterations drive chemoresistance in SCLC.

INTRODUCTION

Small cell lung cancer (SCLC) is highly aggressive and lethal with a 5-year survival rate of less than 6% (1). Approximately two-thirds of patients exhibit a strong response to initial chemotherapy, but relapse occurs within months. The addition of immune checkpoint blockade to the standard of care in 2019 only modestly improved treatment responses (2, 3). Furthermore, we have a poor understanding of the genes and pathways that confer chemoresistance in SCLC, in part owing to limitations with the model systems available. Study of SCLC cell lines has not provided major insights into chemosensitivity and resistance. Comparison of cell lines derived from chemo-naïve SCLC patients versus chemotherapy-treated patients showed no difference in sensitivity to a panel of chemotherapeutic agents (4). Also, genetically engineered models of SCLC harboring Rb/p53 mutation do not show the exquisite chemosensitivity seen in most patients with SCLC, although they do exhibit cytostatic responses (5). In contrast, patient-derived xenograft (PDX) models of SCLC maintain similar response to chemotherapy as seen in the patients from which the models were derived, suggesting that PDX models may be ideal tools to study the biology of chemotherapy response and resistance in SCLC (6–10). By treating chemosensitive PDX models with cycles of chemotherapy and then comparing parental models to more chemoresistant variants, the Rudin lab demonstrated that epigenetic suppression of SLFN11 expression promotes SCLC chemoresistance (11). Recent work from our lab used in vivo lentiviral perturbation studies to demonstrate that MYCN or MYCL overexpression switches chemosensitive PDX models of SCLC to become resistant to cisplatin-etoposide (cis-eto) (5), a finding corroborated by recent analyses of patient samples and serially derived PDX models (10). Here, we sought to leverage the fidelity of PDX models of SCLC to test, in a multiplexed manner, hundreds to thousands of perturbations for ability to make chemosensitive PDX models chemoresistant. cDNA overexpression and CRISPR knockout functional screens confirmed known pathways and revealed previously unknown pathways underlying SCLC chemoresistance, including roles for KEAP1. To evaluate clinical importance, we queried genomic data from patients in the placebo + chemotherapy arm of the IMpower133 clinical trial (2), revealing that the KEAP1/NFE2L2 pathway is genetically altered in SCLC and pathway activation correlates with impaired response to chemotherapy in patients with SCLC.

RESULTS

In vivo cDNA screening in a chemosensitive SCLC PDX model that identifies MYC family members as drivers of tumor cell growth and chemoresistance

Our prior work using single-gene lentiviral perturbation of chemosensitive SCLC PDX models in vivo demonstrated that MYCN or MYCL overexpression could switch these models to become chemoresistant (5). To determine whether we could query many genes in a multiplexed manner for ability to promote in vivo tumor growth under chemotherapeutic pressure, we modified a barcoded open reading frame (ORF) library (12) to add MYCN and MYCL and conducted an in vivo screen (Fig. 1A). We used a chemosensitive SCLC PDX model (SC151) that is Achaete-Scute Family BHLH Transcription Factor 1 (ASCL1) high (fig. S1A) and was previously only propagated in vivo. We briefly explanted tumors ex vivo for lentiviral transduction, minimizing the time the cells were in culture (<18 hours) before implanting 1 million cells per flank into a cohort of immunocompromised nonobese diabetic scid gamma (NSG) mice. To establish a population zero (P0) reference point, we maintained a subset of the cells in culture for 72 hours following lentivirus exposure before collection. In the P0 population, 353 of the 363 cDNAs in the library passed library representation filters. Once flank tumors reached 300 to 350 mm, the mice were treated with either saline or 2 to 3 weekly cycles of cis-eto chemotherapy, which led to ablation of tumors followed by regrowth, with tumor collection an average of 58 days following the end of cis-eto treatment. Tumors from the saline-treated group (n = 13) and those that returned following cis-eto chemotherapy (n = 45) had genomic DNA (gDNA) extracted and sequencing libraries prepared. We performed edgeR (13) analysis to identify cDNAs enriched (relative to the P0 reference point) in either the saline-treated or cis-eto–treated groups. Plotting the average log₂ fold change (log₂FC) versus −log₁₀ false discovery rate (−log₁₀FDR) for each ORF in the saline and cis-eto treatment groups reveals MYC, MYCN, and MYCL as top hits in both treatment arms (Fig. 1B and also see table S1). Waterfall plots depicting cDNA representation in individual tumors show strong and consistent enrichment of MYC, MYCN, and MYCL cDNAs across individual tumors in both treatment groups (Fig. 1C). We also compared cDNA representation in the cis-eto versus saline conditions, showing that MYCN had a clearly stronger effect in the cis-eto group versus the saline group (FDR < 0.05) (fig. S1B). Other hits identified were also enriched in both treatment arms, including CCND1, encoding CYCLIN D1 and TMEM241. CCND1 was an interesting hit, as this gene is amplified in a subset of RB1 wild-type SCLC (14), although we have not validated this result, and Western blot analysis shows the absence of pRB protein in the SC151 PDX model (fig. S1A). As our previous work demonstrated that lentiviral single-gene overexpression of MYCN or MYCL in chemosensitive PDX models results in a switch to chemoresistance (5), a finding supported by recent independent studies (10), these data now show that in vivo multiplexed positive selection screens in this system reflect a powerful approach to identify drivers of chemoresistance in a multiplexed manner.

Fig. 1. In vivo cDNA screening in a chemosensitive SCLC PDX model identifies drivers of chemoresistance.

(A) Schematic of in vivo cDNA screen design in the SCLC PDX model SC151. (B) cDNA-level volcano plots of edgeR analysis results showing average log₂ fold change (log₂FC) enrichment versus P0 and −log₁₀ false discovery rate (−log₁₀FDR) for each cDNA across tumors for the saline (SAL) (n = 13) and cis-eto (n = 45) treatment (trx) arms. Each dot depicts one cDNA. MYC family members are displayed in orange, while additional top hits are displayed in blue (log₂FC > 2 and FDR < 0.05 in the cis-eto versus P0 comparison). cDNAs that were also enriched in cis-eto versus saline are denoted by * (FDR < 0.2) or ** (FDR < 0.05). The cis-eto versus saline comparisons are shown in more detail in fig. S1B. (C) Waterfall plots depicting top six hits (log₂FC > 2 and FDR < 0.05 in the cis-eto versus P0 comparison) at the individual tumor level for both the saline and cis-eto versus P0. Each vertical bar depicts one tumor.

In vivo CRISPR-Cas9 deletion screens identify drivers of chemoresistance in SCLC PDX models

SCLC is dominated by inactivation of tumor suppressor genes rather than activation of oncogenes (15–17). With demonstration that in vivo cDNA overexpression screens in SCLC PDX models are feasible, we next performed a CRISPR-Cas9 deletion screen. Moving to the CRISPR approach allowed us to rapidly generate libraries targeting genes relevant to SCLC. We generated a focused library, targeting 400 high priority genes with six guide RNAs (gRNAs) per gene and included 120 nontargeting gRNAs as controls for a total of 2520 gRNAs (see table S2, which shows the focused library composition). We mined genomic data from SCLC to identify genes in regions of deletion or that exhibited truncating mutations (15, 16, 18). The focused library also targeted genes such as Schlafen 11 (SLFN11) and WNT pathway members, previously implicated in SCLC chemoresistance (11, 18), as well as additional genes of interest that emerged from a pilot genome scale screen. To generate the SCLC focused library, single guide RNA (sgRNA) sequences from the Brunello Human CRISPR Knockout Pooled Library were cloned into the lentiCRISPRv2 backbone (19). We propagated the chemosensitive ASCL1-expressing PDX model FHSC14 (20) in vivo, explanted tumors into cell culture for addition of lentivirus, and, ~18 hours later, injected transduced cells subcutaneously into the flanks of NSG mice at 1 × 10⁶ cells per flank (Fig. 2A). We also cultured a subset of cells for 72 hours before gDNA isolation to establish a P0 reference population. We screened the FHSC14 model with this focused gRNA library just as described for the cDNA screen above and collected 27 saline and 48 cis-eto return tumors for sequencing, with the kinetics of tumor response to cis-eto and later regrowth off therapy shown in Fig. 2B. A challenge with in vivo screens is bottlenecking results in only a subset of implanted cells contributing to the tumor, substantially reducing library representation (21, 22). We developed a strategy to randomly assign independent tumors into pooled sets and computationally pool read counts, normalizing per tumor read depth to increase library representation in the pooled tumor sets. While individual tumors in the saline group had a median of <50% of all gRNAs present, library representation was improved by computational pooling, with six tumors per pool increasing library representation to >95% (Fig. 2C). Applying the Model-based Analysis of Genome-wide CRISPR/Cas9 Knockout (MAGeCK) algorithm (23) to the computationally pooled tumors (six tumors per pool), in the saline condition (with n = 4 pools analyzed), there were three genes for which at least three of their corresponding sgRNAs were enriched relative to P0 (Fig. 2D and also see table S3, which shows gene-level analysis and results for the focused CRISPR knockout screen, and table S4, which shows sgRNA-level results). These genes included MBOAT4, OTOP3, as well as CREBBP, which is a gene mutated in ~9% of SCLC (15, 16, 24, 25) that our group showed functions as a tumor suppressor in genetically engineered mouse models of SCLC (26). In the cis-eto condition (with n = 8 pools analyzed), six genes had at least three of their corresponding sgRNAs enriched relative to P0 (Fig. 2E and table S4). Five of six of the most highly enriched sgRNA-targeted genes in our screen were members of the SAGA (Spt-Ada-Gcn5 acetyltransferase) complex, including USP22, TAF5L, TADA2B, TAF6L, and TADA1. Roles for SAGA complex members in SCLC chemoresistance are under exploration in our laboratory and will be reported independently. Also included among these six genes with highly enriched sgRNAs was KEAP1, a master regulator of the nuclear factor erythroid 2-related factor 2 (NRF2) oxidative stress response pathway (Fig. 2, E and F). Previous in vivo work demonstrated that KEAP1 deletion can drive non–small cell lung cancer (NSCLC) tumorigenesis and therapeutic resistance by stabilizing NRF2, increasing proliferation, and helping to resist oxidative stress (27, 28), but there has been limited study of the KEAP1/NFR2 pathway and little functional evidence for pathway importance in SCLC.

Fig. 2. In vivo CRISPR-Cas9 deletion screens in an SCLC PDX model identifies drivers of chemoresistance.

(A) Schematic of in vivo sgRNA screen. NT, nontargeting. (B) FHSC14 tumor volume kinetics with 3 to 4 weekly cycles of cis-eto. (C) Computational pooling of tumors and percentage of library sgRNAs per pool. Each point depicts one pool generated by randomly selecting the number of tumors specified on the x axis for pooling. For each number of tumors per pool, tumors were randomly pooled 10 times to generate distributions shown. Six tumors per pool was selected for subsequent analyses based on high sgRNA coverage in saline group. Boxes indicate 25th percentile (bottom), median (horizontal line), and 75th percentile (top) of individual points. Vertical lines attached to boxes extend to minimum or maximum values or up to 1.5 times the interquartile range (IQR) from box edges. (D) Gene-level volcano plots of MAGeCK analyses results. Gene-level log₂FC was calculated as median of sgRNA-level log₂FC values. SAGA complex members displayed in blue. FDR < 0.05 for genes with right-side up triangles, 0.05 to 0.20 for genes with upside down triangles, and >0.2 for genes with filled circles (E) Waterfall plots depicting log₂FC enrichment versus P0 of sgRNAs targeting CREBBP and KEAP1, showing top three sgRNAs per gene for all FHSC14 tumor pools. Each vertical bar depicts one sgRNA from one pooled tumor. (F) Relative abundance of KEAP1 sgRNAs in each pooled tumor in cis-eto and saline groups, expressed as log₂FC versus mean abundance in P0. Counts of top three sgRNAs for KEAP1 were summed. Relative abundance of KEAP1 sgRNAs was significantly higher in the cis-eto group [log₂FC = 4.1, FDR < 0.05, based on the MAGeCK analysis from (D)]. Horizontal lines indicate median abundance, while box indicates the IQR. Vertical lines indicate data range, up to 1.5× IQR from the upper or lower quartile.

KEAP1 loss leads to chemoresistance in the SCLC PDX model

To validate hits from our loss-of-function genetic screens in vivo, we generate isogenic PDX models with CRISPR-Cas9 single-gene deletions. We focused on KEAP1, given its known roles in chemotherapy response in NSCLC but minimal characterization in SCLC. In normal conditions, KEAP1 binds to the CUL3-RBX1 ubiquitin ligase complex and the transcription factor NRF2, leading to proteasomal degradation of NRF2. Interaction of reactive oxygen species or other molecular stressors with KEAP1 sensor cysteine residues triggers release of NRF2, allowing its accumulation and enabling transcription of genes that promote stress responses (29–31). We cloned the top enriched sgRNA from the FHSC14 CRISPR deletion screen as well as a nontargeting control sgRNA into lentiCRISPRv2-ZsGreen (see table S5, which shows the gRNA sequences used). Dissociated FHSC14 tumors were transduced in vitro with lentivirus containing these vectors for up to 18 hours and re-engrafted in NSG mice for expansion in vivo (Fig. 3A). Successfully transduced cells from returning tumors were enriched by fluorescence-activated cell sorting (FACS) based on ZsGreen expression and were immediately reinjected into NSG mice (Fig. 3A). With two rounds of propagation in vivo and sorting, we developed transduced models (>90% ZsGreen positive; see Fig. 3A) and used Western blotting to successfully confirm KEAP1 protein knockout upon expression of the KEAP1-targeting sgRNA (sgKEAP1) (Fig. 3B). We also used the Inference of CRISPR Edits (ICE) tool (Synthego) to characterize KEAP1 alterations in the model (fig. S2A). This analysis revealed a polyclonal population of indels. The top two indels were out of frame deletions, while an in-frame indel also present removes five amino acids from KELCH domain, including the Y334 residue reported to be important for binding to NRF2 (32). We also confirmed expected increased protein levels of both NRF2 and a canonical target, NQO1, in sgKEAP1 tumors [Fig. 3B and see fig. S2 (B and C)]. To test the impact of KEAP1 deletion on response to chemotherapy, we set up parallel cohorts of NSG mice with subcutaneous right flank injections of sgCtrl or sgKEAP1 FHSC14 tumors. Once tumors reached ~150 mm³, we treated mice with three weekly cycles of saline or cis-eto and measured tumor volume over 21 days. While sgCtrl tumors slightly regressed in response to cis-eto, sgKEAP1 tumors grew at comparable rates to the saline-treated group, exhibiting almost complete resistance to chemotherapy (Fig. 3C). To assess the impact of KEAP1 deletion on the short-term response to cis-eto, in a parallel experiment, we collected tumors from each isogenic model for molecular analyses 48 hours after the start of treatment. We performed immunohistochemistry on fixed tumor samples from this experiment. sgCtrl tumors exhibited decreased mitotic cells [as assessed by Ser¹⁰ phospho-histone H3 (pH3)] and increased apoptosis [as assessed by cleaved caspase-3 (CC3)] in response to cis-eto (interaction P < 0.0001 for both comparisons; Fig. 3D). However, no changes in pH3 or CC3 positivity were observed with 48-hour cis-eto treatment in the sgKEAP1 PDX model. We noted high expression of NQO1 in the vast majority of the sgKEAP1 transduced cells (fig. S2B) consistent with about two-thirds of detected KEAP1 indels being out of frame and one-third deleting five amino acids including Y334 (fig. S2A), important for KEAP1-NRF2 interaction. Western blot analysis also revealed a KEAP1 band in the sgKEAP1 model (fig. S2C), likely reflecting a combination of lack of complete purity of the model (Fig. 3A), the presence of KEAP1 wild-type stromal cells and the presence of in-frame deleted KEAP1 as detected by Synthego ICE analyses (fig. S2A); we also observed a slight decrease in the level of KEAP1 protein with chemotherapy treatment (fig. S2C), suggesting selection for fully deleted KEAP1 with chemotherapy.

Fig. 3. Isogenic PDX model reveals KEAP1 loss as a driver of chemoresistance in SCLC.

(A) Schematic of isogenic PDX model generation and flow cytometry with lentiviral sgRNA expression showing purity of the resulting models. PE, phycoerythrin area; FITC, fluorescein isothiocyanate area. (B) Western blot confirming KEAP1 deletion with sgKEAP1 expression as well as corresponding NRF2 and NQO1 increase. (C) Tumor volume growth curves of nontargeting sgRNA versus sgKEAP1-expressing FHSC14 PDX models over a 21-day treatment period with saline or three weekly cycles of cis-eto. Data are means ± SEM; n = 7 animals total per treatment group. (D) Immunohistochemistry analyses and representative images for pH3 and CC3 to assess proliferation and cell death, respectively, in sgCtrl versus sgKEAP1 FHSC14 PDX models at a 48-hour time point after treatment with either saline or cis-eto. Scale bars, 10 μm. Statistical analyses in (C) and (D) were performed by two-way analysis of variance (ANOVA) with Tukey’s multiple comparisons test. *P < 0.05, **P < 0.01, and ***P < 0.001. GAPDH, glyceraldehyde phosphate dehydrogenase; n.s., not significant.

Glutamine metabolism represents a targetable vulnerability for KEAP1-deleted SCLC

To analyze changes in gene expression with KEAP1 deletion, we extracted RNA from snap-frozen tumor tissue at 48 hours following treatment and performed RNA sequencing (RNA-seq). Querying the molecular signatures database (MSigDB) using gene set enrichment analysis (GSEA) (33–35), we identified the top six Hallmark gene pathways positively enriched with KEAP1 deletion (Fig. 4A). The reactive oxygen species pathway was the most enriched, consistent with the KEAP1-NRF2 pathway’s role in coordinating the antioxidant response (29–31) with consistent up-regulation of key genes NQO1, TXNRD1, and GCLC (Fig. 4B). Genes in several additional metabolic pathways, including hypoxia and glycolysis, were also enriched. For the pathways shown, including reactive oxygen species, we observed a general increase in expression for both saline- and cis-eto–treated sgKEAP1 tumors relative to the sgCtrl tumors, with fewer differences between the two treatment conditions within the same genotype. To further explore the gene expression changes with KEAP1 deletion, we applied log fold change and FDR cutoffs to the edgeR differential expression analysis and ran Enrichr (36, 37) on the resulting gene sets. Several of the top positively enriched Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways with KEAP1 deletion are related to metabolism, including glutathione (GSH) metabolism (Fig. 4C). Given the enrichment of metabolic pathways with KEAP1 deletion, we performed metabolomic analysis using liquid chromatography mass spectrometry analyses on untreated tumors which revealed an increase in reduced GSH in sgKEAP1 relative to sgCtrl tumors, as expected (Fig. 4D). Glutamate levels relative to the control were lower in sgKEAP1 consistent with prior reports of increased glutamate usage to support cystine uptake through the cystine/glutamate antiporter cystine/glutamate transporter (xCT) and for increased synthesis of GSH resulting in a higher dependence on glutamine catabolism for KEAP1-mutant tumor models (27, 38–40). Given this reported dependency, we tested whether sgKEAP1 SCLC is sensitive to inhibition of glutaminase, the enzyme that hydrolyzes glutamine to glutamate using the glutaminase inhibitor telaglenastat also known as CB-839 (27, 38). We propagated parallel cohorts of sgKEAP1 and sgCtrl tumors in NSG mice and treated with vehicle or CB-839 twice daily via oral gavage once tumors reached 200 to 250 mm³ (Fig. 4E). Tumors were measured daily for 1 week before collection. While FHSC14 sgCtrl treated with CB-839 increased in tumor volume similar to the sgCtrl and sgKEAP1 vehicle–treated tumors, sgKEAP1 tumors treated with CB-839 exhibited partial tumor regression. Therefore, glutamine metabolism represents a targetable vulnerability for KEAP1-deleted SCLC.

Fig. 4. Glutamine metabolism represents a targetable vulnerability for KEAP1-deleted SCLC.

(A) Top six MSigDB Hallmark pathways that are up-regulated in sgKEAP1 FHSC14 tumors relative to sgCtrl tumors, based on GSEA of RNA-seq data after treatment for 48 hours with saline or cis-eto. Point size is scaled to FDR, and point color is scaled to normalized enrichment score (NES). (B) Heatmap showing select gene enrichment for the indicated MSigDB pathways in sgKEAP1 versus sgCtrl FHSC14 tumors treated for 48 hours with saline or cis-eto. (C) Top six KEGG pathways positively enriched in sgKEAP1 relative to sgCtrl tumors, based on Enrichr analysis of RNA-seq data after treatment for 48 hours with saline or cis-eto. The input list of genes was defined on the basis of applying a cutoff of log fold change of >2 and an FDR of <0.05 to the edgeR data. (D) Bar charts showing relative ion counts (normalized to sgCtrl) for reduced GSH and glutamate for FHSC14 sgCtrl and sgKEAP1. (E) Tumor volume growth curves of nontargeting sgRNA versus sgKEAP1-expressing FHSC14 PDX models over a 1-week treatment period with vehicle or twice daily with CB-839 (200 mg/kg). Data are means ± SEM; n = 5 animals total per treatment group. Statistical analysis in (E) was performed by two-way ANOVA with Tukey’s multiple comparisons test. Statistical analysis in (D) was performed by unpaired t test. *P < 0.05, **P < 0.01.

KEAP1/NRF2 pathway signal is associated with patient outcomes

With validation that deletion of KEAP1 promotes resistance to SCLC chemotherapy, we sought to determine clinical relevance. The IMpower133 trial, with placebo + chemotherapy and chemotherapy plus Programmed Death Ligand 1 (PD-L1) inhibitor atezolizumab arms, included both RNA-seq and exome analyses of pretreatment samples for most patients. Whole-exome sequencing (WES) analysis on 191 patients revealed expected high frequency of mutations in RB1 and TP53 and demonstrated KEAP1 deep deletions and predicted pathogenic mutations in 6% of samples (Fig. 5A). We also observed a case of NFE2L2 high-level amplification. We queried the IMpower133 RNA-seq data for activation of an 11 gene KEAP1/NRF2 pathway signature (see Materials and Methods). Transcriptome data from the IMpower133 trial data depict a subset of patients with robust KEAP1/NRF2 signature score (Fig. 5B). NRF2 signature activation was more evident in ASCL1-high samples, while we did not observe clear activation of this signature in the Neurogenic differentiation factor 1 (NEUROD1)–high samples, saw a modest enrichment in non-neuroendocrine samples, and no strong association with MYC family members. Overlaying DNA alteration status and KEAP1/NRF2 expression signature revealed KEAP1/NFE2L2 genetic alterations in many of the samples that exhibited strong transcriptional activation of KEAP1/NRF2 target genes (Fig. 5B). Last, we asked whether the KEAP1/NRF2 signature correlated with patient survival in the chemotherapy arm. Stratifying patients by lower (<median) versus higher (>median) KEAP1/NRF2 pathway signature revealed shorter progression-free survival (PFS) with higher KEAP1/NRF2 signature expression [hazard ratio (HR) = 1.41, 95% confidence interval (CI) [1.00 to 1.99], P = 0.049] and similarly shorter overall survival (OS) (HR = 1.44, 95% CI [0.99 to 2.09], P = 0.056) (Fig. 5, C and D). While we focused on a median KEAP1 signature score cutoff, the trend toward worse survival in the chemotherapy arm holds across most cutoffs set higher than the median, with a limitation that sample sizes passing the highest thresholds for signature score were small (Fig. 5E). We also examined the impact of high KEAP1 signature on survival in the atezolizumab (anti–PD-L1) arm of the trial, where we did not see a clear difference in survival (fig. S3A). In keeping with the overall benefit observed with atezolizumab + cisplatin/carboplatin-etoposide (CE) chemotherapy compared to placebo + CE overall in IMpower133, we see a similar trend toward atezolizumab benefit irrespective of high or low KEAP1 signature (fig. S3, B and C). One of the best characterized targets of NRF2 is NQO1, which was strongly activated in our KEAP1-deleted PDX model (Fig. 4B); we found that high expression of NQO1 also correlated strongly with worse patient survival in the chemotherapy arm (fig. S4). These data indicate that KEAP1 mutations are present in a subset of patients with SCLC, and increased activity of the KEAP1/NRF2 pathway associates with worse clinical outcomes for patients with SCLC treated with chemotherapy.

Fig. 5. KEAP1/NRF2 pathway is altered in a subset of patients with SCLC and associated with SCLC clinical response to chemotherapy treatment.

(A) WES alteration data for patient tumors at baseline in the IMpower-133 trial. Mutations for TP53, RB1, KEAP1, and NFE2L2 are shown. Each vertical bar corresponds to one patient tumor. (B) KEAP1 signature score and WES alteration data for patient tumors at baseline in the IMpower-133 trial organized by the transcription factor (TF) subset. Each vertical bar corresponds to one patient tumor. Gene expression data for transcription factor subsets and genes comprising the KEAP1/NRF2 signature and MYC family members shown. NE, neuroendocrine. (C) PFS curves stratified by the median KEAP1/NRF2 pathway signature score. Dashed curve indicates less than the median KEAP1/NRF2 pathway signature, and solid curve indicates greater than the median KEAP1/NRF2 pathway signature. (D) OS curves stratified by the KEAP1/NRF2 pathway signal. Dashed curve indicates less than the median KEAP1/NRF2 pathway signature, and solid curve indicates greater than the median KEAP1/NRF2 pathway signature. Vertical dashed lines indicate median OS in each group. (E) Forest plot showing HR of IMpower133 patients in the placebo + carboplatin + etoposide arm stratified by KEAP1/NRF2 signature low versus high status. The expression percentile cutoff threshold used to determine signature high versus signature low are shown in the population column, and the number (N) and percentage of patients belonging to the signature-high group with the given cutoff are shown. Higher OS or PFS HR indicate increased likelihood of PFS or OS event in the signature-high population. WT, wild type; Inf, infinity.

DISCUSSION

To elucidate the functional genomic landscape of chemotherapy response and resistance in SCLC, we developed a novel screening platform focused on PDX models. Validating our approach, cDNA overexpression screen revealed top hits known to confer chemoresistance in SCLC, such as MYCN, MYCL, and MYC (5, 10, 41). We focused our functional studies on KEAP1, revealing this gene as a novel driver of resistance to cis-eto chemotherapy in SCLC. Mining patient data from the IMpower133 trial (2) revealed a previously unappreciated role for KEAP1/NRF2 pathway transcriptional activation and genetic disruption in human SCLC, with relevance for clinical impact on chemotherapy response.

Our functional screens and subsequent hit validation focused on genetic perturbation of PDX models that recapitulate the chemosensitivity of patients with SCLC. Deletion of KEAP1 switched the FHSC14 SCLC PDX model to become chemoresistant. We observed activation of canonical NRF2 targets and transcriptional alteration of redox metabolism in the KEAP1-deleted model. In NSCLC, the KEAP1/NRF2 pathway is well appreciated as being functionally important, with inactivating mutations in KEAP1 and activating mutations in NFE2L2 being prevalent (42, 43). Also, KEAP1 loss has a known role in conferring chemoresistance in NSCLC (28). However, initial genomic analyses of SCLC, which focused on earlier-stage resected tumors, did not highlight roles for mutational alteration in KEAP1 (15, 16). The first hint of KEAP1/NRF2 pathway roles in SCLC came with recent observations that ~3% of a large, real-world cohort of 3600 SCLC samples subject to targeted sequencing [largely from blood circulating tumor DNA (ctDNA) analyses] harbored KEAP1 mutations, but the functional importance was unknown (44). By analyzing exome data from IMpower133 patients, we identified functional mutations/deep deletions in KEAP1 in 6% of the samples. We also identified a subset of SCLC with clear activation of a KEAP1/NRF2 gene expression signature. This subset appeared in ASCL1-positive SCLC but was not prevalent in patients with NEUROD1-expressing SCLC. Moreover, strong transcriptional activation of the pathway was often coupled with the presence of genetic alterations in KEAP1 supporting functional importance of the observed mutations. KEAP1/NRF2 alteration in SCLC may be underappreciated as many genomic analyses of SCLC focused on earlier-stage cases that underwent surgical resection (15, 16). Notably, the genomic analyses presented here were of IMpower133 patient samples collected before treatment, and, on the basis of our functional studies, we predict higher rates of KEAP1/NRF2 pathway mutation and/or transcriptional activation in SCLC tumors that have become chemoresistant. Critically, activation of a KEAP1/NRF2 gene signature was associated with shorter PFS and OS in the chemotherapy arm of IMpower133, highlighting the clinical relevance of our genetic observations. We further demonstrate that KEAP1 loss in a SCLC PDX model confers sensitivity to glutaminase inhibition with CB-839, with glutamine metabolism inhibition providing a potential therapeutic avenue for KEAP1-mutant SCLC. It will be important for future studies to explore combining inhibitors of glutamine metabolism with chemotherapy to determine whether this would lead to long-term, durable responses.

Our cDNA overexpression screen demonstrated that in vivo functional screens are possible; however, the library tested was not designed to target genes specifically relevant to SCLC biology. We noted that NFE2L2, encoding NRF2, and key target gene GCLC were included in the cDNA overexpression library but did not emerge as screen hits in the SC151 PDX model tested. Expression of a mutant stabilized version of NFE2L2 might better mimic the consequences of KEAP1 deletion, and it will be important for future studies to test the overexpression of NFE2L2 and of key targets across additional PDX models of different genetic backgrounds, including FHSC14, the model for which KEAP1 was found to be a CRISPR deletion screen hit. A limitation of our CRISPR screen results is that the focused library screens will miss genes that were not included among the 400 targeted in our library. Also, the presence of in-frame alterations will limit screen performance if an essential region of the protein is not targeted. While we prioritized genes for library inclusion that are mutated in human SCLC or present in regions of deletion, the vast majority of published genomic data on human SCLC are from untreated patients, which will not reveal genetic alterations that are acquired during the transition to chemoresistance. Furthermore, features of specific PDX models including genomic alterations and transcriptional heterogeneity are likely to influence the spectrum of gene alterations capable of driving resistance, and it is critical that a wide panel of heterogeneous PDX models be interrogated using these approaches.

Clinical outcomes for SCLC are dismal as chemoresistance is a persistent challenge. Our work identified new drivers of resistance to chemotherapy with previously unappreciated roles in SCLC, and we provide evidence of clinical relevance for impact of KEAP1 pathway alterations on response to chemotherapy in patients. Despite limited progress over recent decades in understanding factors that contribute to SCLC chemoresistance, high throughput study of chemosensitive PDX models now enables systematic interrogation of genes that modulate this process, providing critical information for future efforts to target relapsed SCLC in the clinic.

MATERIALS AND METHODS

Study design

This study was designed to identify genetic perturbations that confer chemoresistance in SCLC. We used lentiviral vectors to overexpress cDNAs to increase gene expression or CAS9 and sgRNAs to delete genes relevant to SCLC in PDX models and treated mice with chemotherapy to identify genetic alteractions that allowed for tumor cell growth under chemotherapy pressure. Focusing on one screen hit, KEAP1, we deleted this gene in a PDX model of SCLC and studied response of the model to chemotherapy and to inhibition of glutaminase. We analyzed genomic data from a clinical trial, IMpower133, to identify mutations/deletions in KEAP1 and to link expression of a KEAP1 gene signature to patient response to chemotherapy. Numbers of replicates and statistical analyses are reported in figure legends and below. No outliers were removed.

Mice

We obtained male or female immunodeficient NSG mice (JAX strain 005557) from Translational Research Model Services at Fred Hutch. Mice were housed in group settings of no more than five mice per cage in a temperature controlled, pathogen-free environment with a standard 12-hour light/dark cycle. All procedures performed in this study were approved by the Institutional Animal Care and Use Committee at Fred Hutch (protocol #50783).

PDX models

PDX models used in the study include FHSC14 (20) and SC151 (the Jackson Laboratory, model TM00269). Only adult mice ~12 weeks of age received xenografts. Mice were monitored weekly while awaiting tumor growth and multiple times per week when tumors were present.

Cell lines

293TN producer cells (LV900A-1, System Bioscience) were used for production of lentivirus in this study. We grew the cells in D10 medium: 500 ml of Dulbecco’s modified Eagle’s medium (DMEM), fetal bovine serum (FBS) at 10% (w/v), penicillin-streptomycin/l-glutamine mix at 1% (w/v), 100 mM sodium pyruvate at 1% (w/v), and 7.5% sodium bicarbonate at 1% (w/v). This medium was changed to virus production medium (VPM) at 16 hours posttransfection: 500 ml of DMEM, 100 mM sodium pyruvate at 1% (w/v), 7.5% sodium bicarbonate at 1% (w/v), 0.5 M sodium butyrate at 1% (w/v), FBS at 2% (w/v), and insulin human recombinant zinc (2 μg/ml). The cells were cultured in a humidified incubator at 37°C, 95% air, and 5% CO₂.

Focused library generation

The SCLC sgRNA library was designed to target 400 genes mutated/deleted in SCLC or implicated in chemotherapy response or found enriched in a pilot genome scale screen. For each gene, six sgRNA sequences were generated using GUIDES (Graphical User Interface for DNA Editing Screens) (45). One hundred twenty nontargeting sgRNAs were sourced as controls from the GeCKO v2.0 library (46) for a total of 2520 sgRNAs (table S2). The SCLC sgRNA library was synthesized (Twist Biosciences, San Francisco, CA) and cloned into lentiCRISPRv2 (Addgene #52961). Oligo pools were amplified using Phusion High-Fidelity DNA Polymerase [New England Biolabs (NEB)] combined with 1 ng of pooled oligo template, primers ArrayF and ArrayR (ArrayF primer: TAACTTGAAAGTATTTCGAT-TTCTTGGCTTTATATATCTTGTGGAAAGGACGAAACACCG and ArrayR primer: ACTTTTTCAAGTTGATAACGGACTAGCC-TTATTTTAACTTGCTATTTCT AGCTCTAAAAC), an annealing temperature of 59°C, an extension time of 20 s, and 25 cycles. Following polymerase chain reaction (PCR) amplification, a 140-bp amplicon was gel-purified and cloned into Bsm BI (NEB, R0580) digested lentiCRISPRv2 using Gibson Assembly (NEB, E2611S). Each Gibson reaction was carried out at 50°C for 60 min. Five microliters of the reaction was used to transform 25 μl of Endura electrocompetent cells (Lucigen, 60242-2) according to the manufacturer’s protocol using a Gene Pulser (Bio-Rad). To ensure adequate representation, sufficient parallel transformations were performed and plated onto carbenicillin containing LB agarose 245 mm–by–245 mm plates (Thermo Fisher Scientific) at 300 times the total number of oligos of each library pool. After overnight growth at 37°C, colonies were scraped off, pelleted, and used for plasmid DNA preps using the Endotoxin-Free Nucleobond Plasmid Midiprep kit (Takara Bio, 740422.10). About 400 clones were sequenced validated as an initial quality control measure.

Library propagation

We followed the protocol provided by the Zhang Lab for pooled library propagation for both our CRISPRko library and our cDNA library (46, 47). Briefly, each library was diluted to 50 ng/μl in nuclease-free water and electroporated into electrocompetent Lucigen Endura cells (VWR, 71003-038) according to the manufacturer’s guidelines. The transformations were then plated on 24.5-cm² bioassay plates with ampicillin. After 14 hours of incubation at 32°C, the amplified libraries were harvested from transformed bacteria using the PureLink HiPure Plasmid Maxiprep Kit (Invitrogen, K210007).

Lentiviral production and concentration

We produced concentrated lentivirus following a protocol modified from the Deisseroth Lab (48). 293TN producer cells were plated on poly-l-lysine–coated (MilliporeSigma, P4832-50ML) 500-cm² plates (Corning, 431111). Cells were cotransfected once approximately 65 to 75% confluent with the lentiviral vector of interest, psPAX2 (Addgene #12260), and pMD2.G (Addgene #12259) at a 2:2:1 ratio via calcium chloride. Media was changed to VPM (see recipe above under the “Cell lines” section) at 16 hours posttransfection. Viral supernatants were collected 64 hours posttransfection, filtered through 0.45-μm Millipore low-protein binding filter units (MilliporeSigma, S2HVU02RE), and then concentrated 70-fold in Millipore Centricon 70 Plus cartridges (Thermo Fisher Scientific, UFC710008) by centrifuging at 3300g for 45 min. The concentrated viral supernatant was recovered by inverting the upper filter cartridge from the Centricon, placing it on an included collection cup, and then centrifuging at 1000g for 2 min.

Chemotherapy treatments

For treatment experiments, tumors were measured daily or every other day. Saline treatments were performed via intraperitoneal injection once per week, while cis-eto treatments were administered intraperitoneally on weekly cycles as follows: cisplatin (5 mg/kg; MilliporeSigma, 479306) and etoposide (10 mg/kg; MilliporeSigma, E1383) day 0, followed by 2 more days of etoposide, and then 4 days off until the next cycle. CB-839 (MedChemExpress, HY-12248) was administered twice daily via oral gavage at a dose of 200 mg/kg. At the end of experiments, tumor tissue was collected for molecular and histopathological analyses.

CRISPR knockout and cDNA screening in vivo

Tumor processing and transduction with library

PDX tumors were dissociated into single-cell suspensions and transduced with concentrated lentivirus carrying the screening library of interest in the presence of polybrene (4 μg/ml) for no more than 18 hours ex vivo. A subset of the cells was left in culture for 72 hours to serve as an initial time point (P0), while the majority of the cells were spun down to remove virus and injected subcutaneously at ~1 (cDNA screen) or 4 (CRISPRko screen) million cells per flank in NSG mice. Once tumors reached a model-dependent threshold size, mice were treated with saline or 2 to 3 cycles of cis-eto to ablate tumors. Tumors treated with saline and tumors that returned following cis-eto treatment were collected and snap-frozen on dry ice. gDNA extraction was performed as previously described (49).

cDNA screen library preparation

For all PCR reactions, we used Phusion Flash High Fidelity Master Mix (Thermo Fisher Scientific, F548L). To prepare libraries from gDNA for sequencing, a two-step protocol was performed. The first PCR step (30 cycles) amplified the cDNA barcodes while preserving library complexity and added an additional barcode to identify individual samples. Following gel extraction of the 110-bp PCR product with a PureLink quick gel extraction kit (Invitrogen, K210012), individually barcoded samples were combined into pools such that no barcodes were repeated within a pool. The second PCR step followed the manufacturer’s instructions for the NEBNext Ultra II RNA Library Prep Kit (NEB, E7770L) end prep of cDNA library, with each pool of samples getting a secondary barcode. This second step also added Illumina sequencing adapters. Individual samples were identified by a unique combination of sample and pool barcode. Samples were pooled and submitted at 2 nM for sequencing.

CRISPRko screen library preparation

gDNA libraries were prepared for sequencing in a two-step protocol as described previously (47). For all PCR reactions, we used Phusion Flash High Fidelity Master Mix. For the first PCR step (16 cycles), a total of 8 μg of gDNA per sample (8 × 100 μl reactions with 1 μg of gDNA per reaction) was amplified and pooled to preserve library complexity. A second PCR step (16 cycles) was performed on 10 μl of the previous pooled PCR product to add Illumina sequencing adaptors and barcodes to identify individual samples. Following gel extraction of the 230-bp PCR product, individual samples were quantified, pooled, and submitted at 2 nM for sequencing.

Sequencing

The cDNA screen library consisted of single-end reads of 50 bp, and the CRISPRko screen library consisted of single-end reads of 57 bp. Libraries were loaded at 11.5 pM and sequenced on an Illumina HiSeq2500 using a high-output flow cell.

Generating single-gene perturbed PDX models

We collected PDX tumors from NSG mice and dissociated them into single-cell suspensions in collagenase (1 mg/ml; MilliporeSigma, C5138) before infecting the cells with concentrated lentivirus carrying a lentiCRISPRv2-ZsGreen vector with KEAP1-targeting or control nontargeting sgRNA. Transduced cells were collected after no more than 18 hours ex vivo, mixed with Matrigel at a 1:1 ratio, and injected subcutaneously into the flanks of NSG mice. Once tumors formed, we again dissociated them and used FACS to isolate ZsGreen⁺ (lentiCRISPRv2) cells successfully transduced with the vectors of interest. Typically, two to three rounds of sorting followed by reimplantation were needed to generate models that were nearly completely composed of transduced cells.

Western blot analysis

Protein extracts were prepared via homogenization of snap-frozen tumor tissue with the AgileGrinder Tissue Grinder (Thomas Scientific-ACTGene, ACT-AG3080) in ice-cold 1× radioimmunoprecipitation assay cell lysis buffer (Cell Signaling, 9806S) with protease and phosphatase inhibitor cocktail (Thermo Fisher Scientific, 78444). Homogenized tumor samples in lysis buffer were then rotated for 60 min at 4°C followed by centrifugation at max speed for 20 min at 4°C. We quantified protein extracts with the Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, 23227) and loaded ~10 μg of samples in 1× Laemmli sample buffer + β-mercaptomethanol on 4 to 20% Mini-PROTEAN TGX Precast Protein Gels (Bio-Rad, 4561096). We used 0.45-μm nitrocellulose membranes (GE Healthcare Life Sciences, 10600002) for the protein transfer and imaged membranes with a LI-COR Odyssey Fc. The following antibodies were used: anti-KEAP1 (Proteintech, 105032-AP), anti-NQO1 (Cell Signaling, 3187), anti-NRF2 (Cell Signaling, 12721), anti-pRb (Santa Cruz, 102), anti-ASCL1 (BD Biosciences, 556604), anti-NEUROD1 (Cell Signaling, 4373), anti-rabbit immunoglobulin G (IgG) horseradish peroxidase (HRP)–linked (Cell Signaling, 7074), and anti-mouse IgG HRP-linked (Cell Signaling, 7076). We used glyceraldehyde phosphate dehydrogenase (Santa Cruz, sc-32233) or α-tubulin (Cell Signaling, 3873) for protein loading normalization.

Immunohistochemistry

PDX tumor tissue fragments were fixed in 10% neutral-buffered formalin for 72 hours and transferred to 70% ethanol before processing to paraffin blocks sectioned at a thickness of 4 μm. Paraffin sections were dewaxed in xylene and rehydrated by passage through a series of ethanol to water. Antigen masking was performed in citrate buffer [10 mM citric acid and 0.05% Tween 20 (pH 6.0) in tris-buffered saline solution] preheated in a steamer for 20 min. Endogenous peroxidases were blocked with 3.5% H₂O₂. Tissue sections were blocked in 5% goat serum and then incubated overnight in a primary antibody within a humidified chamber at 4°C. Primary antibodies used were anti-phospho Ser¹⁰ histone H3 (Millipore, catalog no. 06-570), anti-CC3 (Cell Signaling, catalog no. 9661), and anti-NQO1 (Thermo Fisher Scientific, PA5-21290). Sections were then incubated in biotinylated goat anti-rabbit IgG secondary antibody (Vector Laboratories, BP-9100-50) for 1 hour. Biotin-peroxide complexes were formed using the VectaStain ABC kit (Vector Laboratories, PK-4000) according to the manufacturer’s protocol. Development was performed using the 3,3′-diaminobenzidine Substrate Kit Peroxidase (Vector Laboratories, SK-4100) according to the manufacturer’s protocol. Slides were imaged with a Nikon Ni-UI microscope and manually quantified at ×400 magnification with two to three fields per section.

RNA sequencing

RNA was isolated and purified from snap-frozen PDX tumor tissue by mechanical disruption in TRIzol reagent according to the manufacturer’s protocol. mRNA was polyadenylate [poly(A)] enriched using the NEBNext Poly(A) mRNA Magnetic Isolation Module (NEB, E7490) kit. Indexed libraries were generated using the NEBNext Ultra II RNA Library Prep Kit for Illumina (NEB, E7770) according to the manufacturer’s protocol. Single-end sequencing (50 bp) was performed on an Illumina NextSeq P3. STAR v2.7.7a (50) with two-pass mapping was used to align single-end reads to human genome assembly hg38 and GENCODE gene annotation v38 and mouse genome assembly mm10 and GENCODE gene annotation vM23. R package XenofilteR (51) was used to remove reads of mouse origin with parameters recommended by authors. FastQC 0.11.9, RNA-SeQC 2.3.4, and RSeQC 4.0.0 were used to check various quality control metrics including insert fragment size, read quality, read duplication rates, ribosomal RNA rates, gene body coverage, and read distribution in different genomic regions. FeatureCounts in Subread 2.0.0 was used to quantify gene-level expression by counting reads in an unstranded fashion. Bioconductor package edgeR 3.34.1 (13) was used to detect differential gene expression between sample groups. Genes with low expression were excluded using edgeR function filterByExpr with min.count = 10 and min.total.count = 15. The filtered expression matrix was normalized by the trimmed mean of M values (TMM) method (52) and subject to significance testing using the Generalized Linear Model Likelihood Ratio Test method and paired donor as blocking factor (if applicable). Genes were deemed differentially expressed if Benjamini-Hochberg (BH) adjusted P values were less than 0.05. STAR v2.7.1a and GENCODE gene annotation v31 were used. GSEA (version 4.2.3 software) was performed by querying the Hallmark gene set in the MSigDB (33). Commonly up-regulated and down-regulated genes in sgKEAP1 versus sgCtrl FHSC14 model were analyzed using the Enrichr (36, 37) website application querying the MSigDB Hallmark gene set (33).

Metabolomics

Lung tumors frozen in liquid N2 were pulverized to powder and extracted with 1000 μl of ice-cold 80% high-performance liquid chromatography–grade methanol (80:20 methanol:water) (Sigma-Aldrich, 646377 and 270733). Each tube was vortexed briefly and then kept at −20°C for 30 min. Samples were then vortexed at 4°C for 5 min and then centrifuged at 17,000g for 5 min at 4°C. The supernatant volumes equivalent to 8 mg of extracted tissue was transferred to a fresh microcentrifuge tube and dried on a Centrivap vacuum concentrator (Labonco, 10269602) at 4°C overnight. The next day, samples were resuspended in 40 μl of 80% methanol containing an isotopically labeled canonical amino acid mix standard (Cambridge Isotopes, MSK-CAA-1). Samples were then vortexed for 5 min at 4°C and centrifuged at 17,000g for 5 min at 4°C, and the top 20 μl of the supernatant was transferred to an liquid chromatography–mass spectrometry (LC-MS) vial (Thermo Fisher Scientific, 03-452-259) ran for LC-MS analysis. Metabolite quantitation of resolubilized tumor extracts was performed using a Q Exactive HF-X Hybrid Quadrupole-Orbitrap Mass Spectrometer equipped with an Ion Max API source and an H-ESI II probe, coupled to a Vanquish Flex Binary Ultra-High Performance Liquid Chromatography system (Thermo Fisher Scientific). Mass calibrations were completed at least every 5 days in positive and negative polarity modes using LTQ Velos ESI Calibration Solution (Pierce). Metabolites were chromatographically separated by injecting 1 μl into a SeQuant ZIC-pHILIC Polymeric column (2.1 mm by 150 mm, 5 mM; EMD Millipore). The flow rate was set to 150 ml/min, with autosampler temperature set to 10°C and column temperature set to 30°C. Mobile phase A consisted of 20 mM ammonium carbonate and 0.1% (v/v) ammonium hydroxide, and mobile phase B consisted of 100% acetonitrile. The sample was gradient-eluted (%B) from the column as follows: 0 to 20 min: linear gradient from 85 to 20% B; 20 to 24 min: hold at 20% B; 24 to 24.5 min: linear gradient from 20 to 85% B; 24.5 min to end: hold at 85% B until equilibrated with 10 column volumes. Mobile phase was directed into the ion source with the following parameters: sheath gas = 45, auxiliary gas = 15, sweep gas = 2, spray voltage = 2.9 kV in the negative mode or 3.5 kV in the positive mode, capillary temperature = 300°C, RF level = 40%, auxiliary gas heater temperature = 325°C. Mass detection was conducted with a 240,000 mass resolution in full scan mode, with an AGC target of 3,000,000 and a maximum injection time of 250 ms. Metabolites were detected over a mass range of 70 to 1050 mass/charge ratio. Quantitation of metabolites was performed using Tracefinder 4.1 (Thermo Fisher Scientific) referencing an in-house metabolite standards library using a mass error of ≤5 parts per million.

IMpower133 analyses

RNA-seq data and WES data from IMpower133 samples were processed as previously described (53). Briefly, RNA-seq reads were aligned to the human reference genome (National Center for Biotechnology Information Build 38) and adjusted for gene length using transcript per million normalization and subsequently log₂-transformed. Sample-level gene signature scores were calculated as the mean cohort-wide z score for all the genes in the signature. For Kaplan-Meier analyses, samples were dichotomized by the cohort-wide median using the KEAP1/NRF2 11 gene (G6PD, ABCC2, TXNRD1, GSR, PRDX1, TXN, HMOX1, ME1, GCLC, NQO1, and GCLM) signature score. This NRF2 target set was derived from the 14 gene GSEA set SINGH_NFE2L2_TARGETS, based on RNA interference studies in NSCLC cells (54) but modified to include the top 11 genes with consistent up-regulation upon KEAP1 deletion in our SCLC PDX model RNAseq analysis. KEAP1 and NFE2L2 mutations were called from whole-exome libraries from tumor DNA and matched germline DNA as previously described (53). Variants were called by three variant callers (Mutect2, LoFreq, and Strelka) (55–57), and nonsynonymous variants were only reported if identified by two of three variant callers. AlphaMissense annotation for hg38 genome was used to assess potential pathogenicity of the missense alterations (58).

Statistical analysis

cDNA screen analysis

Differential abundance analysis between the cis-eto, saline, and P0 groups was performed using edgeR version 3.36.0 based on the quasi-likelihood negative binomial regression pipeline (59). cDNA counts were normalized using the TMM method (52). Estimate of dispersion and regression parameters were weighted according to a robust regression algorithm to minimize the influence of potential outliers (60). Multiple testing was accounted for using the BH procedure to control the FDR (61).

CRISPRko screen analyses

Differential abundance analysis between the cis-eto, saline, and P0 groups in pilot and focused CRISPRko screens used MAGeCK version 0.5.9 (23). Samples were pooled to increase gRNA coverage, defined as the proportion of gRNAs present. In the SCLC focused library screens, individual samples were computationally pooled after read alignment. Samples were randomly assigned to pools, and the samples from each pool were treated as technical replicates and combined using the mageck count command. To help select the number of samples per pool, 10 random assignments were generated for each pool size, and the relationship between gRNA coverage and pool size was visualized using boxplots. Pool size was selected on the basis of the saline group where gRNA coverage was at least 90% across all cases and predominantly >99%. After pooling, gRNA counts were median-normalized by MAGeCK. The gRNA-level differences were summarized and ranked at the gene-level using robust rank aggregation score generated by the MAGeCK pipeline and the median log₂FC. Gene- and gRNA-level P values were adjusted for multiple testing using the BH FDR procedure.

Supplementary Materials

The PDF file includes:

Figs. S1 to S4

Legends for tables S1 to S5

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S5