Keap1 mutation renders lung adenocarcinomas dependent on Slc33a1

Abstract

Approximately 20-30% of human lung adenocarcinomas (LUAD) harbor loss-of-function (LOF) mutations in Kelch-like ECH Associated-Protein 1 (KEAP1), which lead to hyperactivation of the nuclear factor, erythroid 2-like 2 (NRF2) antioxidant pathway and correlate with poor prognosis^1-3. We previously showed that Keap1 mutation accelerates KRAS-driven LUAD and produces a marked dependency on glutaminolysis⁴. To extend the investigation of genetic dependencies in the context of Keap1 mutation, we performed a druggable genome CRISPR-Cas9 screen in Keap1-mutant cells. This analysis uncovered a profound Keap1 mutant-specific dependency on solute carrier family 33 member 1 (Slc33a1), an endomembrane-associated protein with roles in autophagy regulation⁵, as well as a series of functionally-related genes implicated in the unfolded protein response. Targeted genetic and biochemical experiments using mouse and human Keap1-mutant tumor lines, as well as preclinical genetically-engineered mouse models (GEMMs) of LUAD, validate Slc33a1 as a robust Keap1-mutant-specific dependency. Furthermore, unbiased genome-wide CRISPR screening identified additional genes related to Slc33a1 dependency. Overall, our study provides a strong rationale for stratification of patients harboring KEAP1-mutant or NRF2-hyperactivated tumors as likely responders to targeted SLC33A1 inhibition and underscores the value of integrating functional genetic approaches with GEMMs to identify and validate genotype-specific therapeutic targets.

Cancer genome sequencing studies have identified Kelch-like ECH-Associated Protein 1 (KEAP1), a negative regulator of the antioxidant transcription factor nuclear factor, erythroid 2-like 2 (NFE2L2; hereafter NRF2), to be mutated in approximately 20-30% of all lung cancers, with loss-of-function (LOF) mutations spanning the full length of the KEAP1 protein^3,6-8. NRF2 regulates a network of genes that together coordinate the cellular response to mitigate oxidative damage via detoxification of reactive oxygen species (ROS) and xenobiotic insults, as well as the regulation of glutathione (GSH) synthesis and recycling^2,9,10. The high frequency of mutations in the KEAP1/NRF2 pathway suggests a critical role for the oxidative stress pathway in lung tumorigenesis, and presents an opportunity to identify novel therapeutic strategies for this genetic subtype of LUAD.

Genetic inactivation of Nrf2 in Kras-mutant genetically engineered mouse models (GEMMs) of pancreatic cancer and LUAD leads to decreased tumor initiation and tumor proliferation, suggesting that a subset of RAS-driven tumors are dependent on this transcription factor^11-13. We recently showed that Keap1 inactivation cooperates with Kras mutation in driving lung adenocarcinoma (LUAD) initiation and progression, and that these Keap1-mutant tumors are sensitive to glutaminase inhibitors⁴. Tumors with KEAP1/NRF2 pathway mutations have also been shown to depend on non-essential amino acids, as well as exhibiting altered metabolism resulting from increased flux of the pentose phosphate pathway, as well as usage of glutamine, serine, and cysteine derived metabolites^14-18. Importantly, KEAP1 mutations are strongly correlated with decreased lung cancer mortality and shown to promote lung cancer metastasis in GEMMs of Keap1-mutant LUAD^19,20.

Given the large proportion of cancers with KEAP1 pathway mutations and the inherent difficulty in therapeutically targeting LOF mutations, we investigated whether Keap1-mutant tumors (or otherwise NRF2-hyperactivated tumors) harbored additional genetic dependencies that could be therapeutically targeted. Towards this goal, we performed a CRISPR-Cas9-based screen of all druggable genes in Keap1-mutant isogenic LUAD cell lines. Cross validation with publicly available CRISPR screens and our druggable genome screen uncovered several novel dependencies, including a marked reliance on the endoplasmic reticulum (ER) resident gene, Slc33a1. Using multiple validation approaches, including in vitro and in vivo approaches, as well as preclinical GEMMs of Kras-driven LUAD, we confirmed Slc33a1 as a Keap1-mutant-specific dependency. Finally, we demonstrate the power of genome-wide modifier genetic screens by identifying suppressors of Slc33a1 dependency and highlight an underappreciated relationship between ER redox biology and KEAP1 mutation. Altogether, these data implicate a unique vulnerability that could be therapeutically exploited in KEAP1/NRF2-mutant cancers with limited clinically approved targeted therapies.

Results:

Druggable genome CRISPR screen identifies Keap1-mutant dependencies

To uncover novel dependencies in a high-throughput manner, we performed pooled CRISPR-Cas9 screens in Keap1-mutant cell lines using a custom druggable-genome library (DGL) targeting 4,915 genes with high-confidence known or predicted druggability (Fig. 1a,b, Supplementary Fig. 1a,b, and Supplementary Table 1; see Methods)²¹. This library was used to transduce two isogenic pairs of Cas9-expressing murine LUAD cell lines with Kras^G12D/+; p53^−/− (KP) and Kras^G12D/+; p53^−/−; Keap1^−/− (KPK) genotypes⁴ (Fig. 1a,b). Library representation was determined by deep sequencing of sgRNAs amplified from genomic DNA harvested early (t₀) and 8 population doublings post transduction (t₈; Fig.1b). To quantify the impact of gene targeting, we calculated gene scores (GS) as the median log₂ fold-change in sgRNA abundance between t₈ and t₀. As expected, we observed significant genotype-independent depletion of sgRNAs targeting known essential genes, and enrichment of those targeting tumor suppressor genes at t₈ (Supplementary Fig. 1c and Supplementary Table 2).

Figure 1: A genetic screen for Keap1-deficient genetic vulnerabilities yields Slc33a1 as a potent dependency.

(a) A schematic for the generation of two isogenic cell line pairs. Plucked tumors from two individual KP mice were used to generate parental KP tumor-derived cell lines. Each parental line was transfected in vitro with plasmids containing sgKeap1, Cas9 and GFP. Transfected lines were single cell sorted and validated for deficiency (KPK) or maintenance of two wild-type Keap1 alleles (KP).

(b) CRISPR-Cas9 based screening strategy designed to identify Keap1-deficient genetic vulnerabilities. Two isogenic KP and KPK cell line pairs were infected with a Druggable Genome Library (DGL). Genomic DNA was collected at initial time points (t₀) and after 8 population doublings (t₈).

(c) Top common differentially depleted genes between each isogenic pair. Each dot represents the log₂ fold change of each sgRNA (sgRNA score) targeting the indicating gene (4 sgRNAs per gene). Bar represents median gene score across all sgRNAs targeting the indicated gene per genotype.

(d) Gene set enrichment analysis of the rank ordered differential gene scores across the entire DGL of isogenic pair 1.

(e) Fluorescence competition assays of wild-type versus sgCtrl or sgSlc33a1 cells in KP and KPK isogenic cell lines. Plot displays day 2 normalized %GFP⁺ (pUSPmNG) cells marking sgCtrl or sgSlc33a1 edited cells (mean ± SD; n = 3; n denotes technical replicate infections). Statistics derived from the comparison of the average KP- or KPK-sgCtrl values versus the average KP- or KPK-sgSlc33a1 values across all Slc33a1 targeting sgRNAs.

(f) Colony formation of KP and KPK isogenic cells targeted with sgCtrl or sgSlc33a1 (pUSEPR).

(g) Overexpression of an sgRNA mutant-Slc33a1 cDNA (*) or control cDNA (mScarlet) in both KPK cell lines. Plot displays %GFP⁺ (pUSPmNG) cells 10 days post infection relative to day 2 (mean ± SD; n = 3; n denotes technical replicate infections).

(h) Fluorescence competition assays of KP, KP-Nrf2 KO (KPN), KPK, KPK-Nrf2 KO (KPKN) isogenic pairs. Plot displays %RFP⁺ (pUSEPR) cells marking sgCtrl or sgSlc33a1 infected cells 6 days post infection relative to day 2 (mean ± SD; n = 3; n denotes technical replicate infections).

(i) Waterfall plot of the rank ordered Pearson correlation coefficient between NRF2 CERES scores versus all genes screened from the DepMap database version 19Q2.

(j) Fluorescence competition assay on a panel of KEAP1-mutant human LUAD cell lines. Plot displays day 2 normalized %RFP⁺ (pUSEPR) cells marking sgCTRL, sgSLC33A1, or sgRPA3 infected cells (mean ± SD; n = 3; n denotes technical replicate infections). Statistics derived from the comparison of the sgCTRL values versus the average sgSLC33A1 values across all SLC33A1 targeting sgRNAs.

(k) Differential CERES scores between core NRF2 core gene set high-ranking cell lines (n = 42) versus low-ranking cell lines (n = 32) from the DepMap database 19Q2. Horizontal dotted line represents p-value significance cut-off of p<0.05. Each dot represents the differential CERES score per gene. Blue dots represent genes that pass all set threshold values. Statistical analyses were performed using two-way ANOVA in (e) and (j) with Tukey’s post-hoc multiple comparisons test. Student’s two-tailed t-test was performed for (g), (h) and (k).

To identify Keap1-mutant-specific vulnerabilities, we calculated differential GS by subtracting KP GS from KPK GS. Using thresholds defined by the distribution of control sgRNA “pseudogene” scores, 53 genes reached significance (FDR = 0.05) for differential depletion in isogenic pair 1 (KP1 and KPK1), and 25 in pair 2 (KP2 and KPK2; see Methods; Supplementary Fig. 1d,e). Of these genes, Slc33a1, Arf4, Spcs2, Abcb7, Aifm1, and Rad51 reached significance in both isogenic pairs (Fig. 1c, Supplementary Fig. 1d,e, and Supplementary Table 2). Intriguingly, SLC33A1, ARF4, and SPCS2 are implicated in protein transport and the endoplasmic reticulum (ER)^22,23, while ER-stress has been shown to modulate levels of RAD51 and sensitivity to DNA damage^24,25. Reinforcing the importance of these pathways, gene set enrichment analysis (GSEA) of the ranked list of differential GS identified significant enrichment of the ER stress and unfolded protein response (UPR) gene set (FDR = 0.0073; Fig. 1d and Supplementary Table 3).

Reliance on Slc33a1 in Keap1-mutant cells is NRF2-dependent

To validate the most differentially depleted genes, Slc33a1 and Arf4, as Keap1 mutant-specific genetic vulnerabilities, we examined the top two scoring sgRNAs targeting these genes in fluorescence competition and colony formation assays (Supplementary Fig. 2a,b). We observed clone-specific depletion of sgArf4-targeted KPK1 but not sgArf4-targeted KPK2 cells, and thus deprioritized Arf4 for further evaluation (Supplementary Fig. 2c). On the other hand, sgSlc33a1-targeted KPK1 and KPK2 cells exhibited a robust decrease in growth; while modest growth defects were observed in sgSlc33a1-targeted KP cells, the magnitude and kinetics of the effect was significantly less pronounced than in KPK cells (Fig. 1e,f and Supplementary Fig. 2d). We observed similar decreases in colony formation using inducible shRNAs or constitutive CRISPR-interference (CRISPRi; Supplementary Fig. 2e-k). Transduction with an sgRNA-resistant cDNA of Slc33a1 fully rescued gene expression as well as the growth defect in KPK cells targeted with Slc33a1 sgRNAs, demonstrating on-target dependency (Fig. 1g and Supplementary Fig. 3a). Collectively, these data validate the CRISPR screening results and further support Slc33a1 as a potent target for inhibiting the proliferation of Keap1-mutant cancer cells.

To examine whether the Slc33a1 dependency in Keap1-mutant cells is explained by increased NRF2 activity versus other Keap1 clients^8,26, we targeted Nrf2 in KP and KPK cell lines and examined the effect of targeting Slc33a1. Inactivation of Nrf2 substantially rescued the growth defect caused by Slc33a1 mutation in KPK cells (Fig. 1h and Supplementary Fig. 3b-f). These data indicate that NRF2-hyperactive cells are dependent on SLC33A1, and that decreased fitness following Slc33a1 deletion is a consequence of the cellular programs controlled by the NRF2 transcription factor.

KEAP1-mutant human cancer cells are dependent on the ER-resident genes SLC33A1, SUCO, and TAPT1

To assess the effects of SLC33A1 inactivation in human cell line studies, we analyzed whole-genome CRISPR-Cas9 screening results from the publicly available DepMap dataset encompassing 563 human cancer cell lines spanning 29 cell lineages²⁷ (see Methods). We first confirmed that SLC33A1, on average, does not behave as a core-essential gene across the large majority of human cell lines in the DepMap (Supplementary Fig. 4a). However, average SLC33A1 CERES normalized GS (hereafter CERES scores) were ~3.5-fold lower in KEAP1/NRF2-mutant cancer cell lines (n = 63) compared to wild-type cancer cell lines (n = 500; Supplementary Fig. 4b and Supplementary Table 4).

To extend the analysis beyond the genes targeted by the DGL used in our studies, we explored previously published Cancer Coessentiality Networks, a meta-analysis of depletion screens across different cell types²⁸. We identified NRF2, SLC33A1, SUCO, TAPT1, DNAJB11, and ADPGK in a single co-essential gene cluster using the DepMap dataset as input (Supplementary Fig. 4c). Importantly, analogous to the top hits identified in the DGL, this co-essential cluster is composed entirely of known endomembrane-associated genes. Similarly, given that KEAP1 mutation results in increased NRF2 pathway activity², as well as data showing NRF2 dependency in KEAP1-mutant human cell lines^29,30, we calculated the Pearson correlation coefficient between NRF2 CERES scores in the DepMap compared to each other gene’s CERES scores as an alternative approach to identify genes that correlate most with NRF2 dependence on a relative basis. Using this approach, we identified SLC33A1 scores as having the second highest correlation to NRF2 scores (r=0.361), followed by MAFK (r = 0.287), a known NRF2 transcriptional co-activator³¹. Moreover, the ER-resident proteins SUCO and TAPT1, previously established as the mammalian orthologs of the SLP1:EMP65 yeast complex³², were also closely correlated with NRF2 dependency in this analysis (Fig. 1i, Supplementary Fig. 4a,b, and Supplementary Table 4; r = 0.283 and r = 0.258, respectively). Importantly, SLC33A1 dependency weakly correlated with activating transcription factor 4 (ATF4) dependency (r = −0.001), a downstream target of the UPR response known to exhibit cross-talk with the NRF2 pathway¹⁵ (Fig. 1i, Supplementary Fig. 4d,e, and Supplementary Table 4).

To further explore these associations, we first probed the effects of SLC33A1 mutation in a panel of human KEAP1-mutant cells by fluorescence competition assays (Fig. 1j). These data confirmed that SLC33A1 represents a KEAP1-mutant vulnerability across genetically defined human cancer cell lines. Next, to address whether other dependencies identified in the human NRF2-SLC33A1 co-essentiality network also validated, we targeted Suco or Tapt1 in the murine KP and KPK isogenic cell line pairs. Supporting the bioinformatic predictions, Suco and Tapt1 inactivation led to decreased fitness only in KPK cells (Supplementary Fig. 4f,g). These data support an increased dependence on endomembrane-associated protein function in the context of NRF2-hyperactivated cancer cells, corroborating the significant depletion of UPR target genes in the DGL by GSEA in KPK cell lines compared to KP cell lines (Fig. 1c,d and Supplementary Fig. 4c).

High NRF2-transcriptional signatures correlate with increased SLC33A1 dependency

Given that tumor cells may upregulate the NRF2 response in the absence of KEAP1/NRF2 mutation or NRF2 amplification, we questioned whether the degree of NRF2 transcriptional activity could predict SLC33A1 dependency. To that end, we used a previously established NRF2 core target gene set (hereafter NRF2 core) or KEAP1-mutant transcriptional signature as indicators of NRF2 activation state⁴. These signatures were used to stratify cell lines screened in the DepMap into high-scoring or low-scoring cohorts (NRF2 core high-scoring n = 42; low-scoring n = 32; KEAP1-mutant signature high-scoring n = 43; low-scoring n = 21; Supplementary Table 5; see Methods). By comparing the CERES scores between these high- and low-scoring groups, we identified 123 (NRF2 core, p<0.05) or 34 (KEAP1-mutant signature, p<0.05) statistically significant depleted genes in the high-scoring group, of which NRF2, SLC33A1, SUCO, and TAPT1, fell within the top 15 genes regardless of the signature used (Fig. 1k, Supplementary Fig. 5a, and Supplementary Table 5). Interestingly, although we could not validate a genotype-specific dependency for ARF4 in murine KPK cell lines, we observed significant depletion of ARF4 in the high-scoring groups of either signature. Moreover, despite the enrichment of NRF2- or KEAP1-mutant cell lines in either the NRF2 core or KEAP1-mutant high-scoring groups (n = 27, p = 2.47e-18 or n = 18, p = 5.95e-8, respectively; hypergeometric test), a substantial fraction did not harbor either NRF2 or KEAP1 mutations (n = 15, 35% and n = 25, 58%, respectively), demonstrating the importance of functionally classifying tumors by NRF2-transcriptional activity rather than KEAP1 or NRF2 mutational status. We observed a step-wise decrease in GS for SLC33A1, SUCO, and TAPT1, but not RPA3 or OR12D2 when grouping cell lines by KEAP1/NRF2-wild-type status (n = 424), mutant status (n = 58), or NRF2 core high-ranking cell lines that are genetically KEAP1/NRF2-wild-type (n = 31). However, we did not observe significant differences in CERES scores between wild-type and NRF2 core low-ranking cell lines that are genetically KEAP1/NRF2-wild-type (n = 45), as expected. Importantly. KEAP1/NRF2 mutation displayed the largest fold change compared to wild-type cells (fold change SLC33A1 = 2.49, SUCO = 1.41, TAPT1 = 1.59), followed by NRF2 core high-ranking cells (fold change SLC33A1 = 2.41, SUCO = 1.5, TAPT1 = 1.4), further demonstrating that dependencies towards SLC33A1, SUCO, and TAPT1 can be stratified by the magnitude of NRF2 transcriptional output independently of KEAP1/NRF2 mutation status (Supplementary Fig. 5b,c).

Loss of Slc33a1 results in induction of an unfolded protein response

SLC33A1 is an ER-localized transmembrane protein with reported acetyl-Coenzyme A (CoA) transport activity⁵ (Fig. 2a). We observed that targeting of Slc33a1 in KPK cells led to a striking increase in intracellular blebbing, consistent with previous reports of Slc33a1 loss^5,33 (Fig. 2b). Given the reported ER localization of SLC33A1 and its significant KPK-specific depletion in our screens, we performed whole-transcriptome analysis (RNA-seq) on isogenic pair #1 following sgCtrl (non-essential gene, Olfr102) or sgSlc33a1 targeting to define a transcriptional signature of Slc33a1 deficiency. We found that differentially-expressed genes resulting from Slc33a1 loss were enriched in the UPR pathway, further supporting a role for Slc33a1 in maintaining ER-homeostasis (NES=2.48; FDR=0; Fig. 2c and Supplementary Table 6).

Figure 2: Loss of Slc33a1 promotes the activation of an unfolded protein response.

(a) Schematic of the reported function of SLC33A1.

(b) Confocal images of cytoplasmic-tagRFP (pUSEPR; see methods) labeled isogenic pair #1 infected with sgCtrl or sgSlc33a1. Larger picture denotes increase magnification of KPK1-sgSlc33a1 cells. Far right: Phase contrast microscopy of KPK-Slc33a1 cells.

(c) GSEA enrichment plot of the Hallmark UPR signature enriched in sgSlc33a1 targeted cells compared to sgCtrl targeted cells.

(d) Western blot analysis of Cas9 expressing KP and KPK isogenic pairs targeted with sgCtrl or sgSlc33a1 for autophagy and UPR protein markers.

(e) Western blot analysis of dCas9-KRAB expressing KP and KPK isogenic pairs targeted with sgCtrl or sgSlc33a1 for autophagy and UPR protein markers.

(f) qPCR validation of UPR and NRF2 target genes in Cas9 expressing isogenic pair #1 targeted with sgCtrl or sgSlc33a1. (mean ± SD; n = 4; n denotes technical replicates). Statistical analyses were performed using two-way ANOVA in (f) with Tukey’s post-hoc test.

Using previously defined, conserved UPR transcriptional signatures³⁴, we observed significant positive enrichment of gene sets associated with ER-chaperones (NES=2.13; FDR=0), aminoacyl tRNA synthetase (NES=1.91; FDR=0), ER/Golgi trafficking (NES=1.68; FDR=0.0146), and endoplasmic reticulum-associated protein degradation (ERAD; NES=1.61; FDR=0.025) in sgSlc33a1-targeted cells (Supplementary Fig. 6a-c and Supplementary Table 7). Slc33a1 targeting results in increased p62 and LC3-II expression, PERK mobility shift, and an increased ratio of phospho-to-total eIF2A upon Slc33a1-targeting in KPK cells, with no change in GCN2 levels, all consistent with induction of autophagy and a UPR, respectively (Fig. 2e,f). Moreover, increased expression of UPR target genes was evident in KP- and KPK-sgSlc33a1 cells by both protein (BiP/HSPA5, ATF4; Fig. 2e,f) and targeted gene expression analysis (BiP/Hspa5, Chop/Ddit3, Atf4, and Xbp1s; Fig. 2f). Notably, the level of induction of Atf4, BiP and Chop/Ddit3 was greater in KPK-sgSlc33a1 cells compared to KP-sgSlc33a1 cells, suggesting that Keap1 loss results in enhanced induction of a UPR program in the face of this form of ER-stress (Fig. 2e-g).

Slc33a1 dependency is rescued by pharmacological inhibition of GSH synthesis

Our previous studies showed that KPK cells potently upregulate the NRF2 pathway, resulting in increased glutathione (GSH) synthesis⁴. This is likely due to the upregulation of NRF2-dependent enzymes directly involved in GSH synthesis (glutamate cysteine ligase catalytic subunit; Gclc)³⁵ or uptake of L-Cystine that is necessary for GSH synthesis (system xCT encoded by Slc7a11 and Slc3a2; Fig. 3a). Previous studies in yeast have demonstrated that enhanced extracellular GSH uptake results in growth inhibition and cell death proportional to the amount of GSH added to the media³⁶. We hypothesized that increased NRF2-dependent GSH synthesis in combination with genetic perturbation of endomembrane-associated genes might induce an UPR, which may explain the increased dependency of Keap1-mutant cells on Slc33a1, Suco, and Tapt1. To test this hypothesis, we probed the GSH synthesis and recycling pathway at multiple steps using both genetic and small molecule approaches. Reduction of GSH synthesis via inhibition of GCLC (L-buthionine sulfoximine; BSO)³⁷ reduced total GSH within 48 hours of treatment (Supplementary Fig. 7a). Remarkably, treatment with BSO completely rescued the fitness defect of sgSlc33a1-, sgTapt1-, and sgSuco-targeted KPK cells (Fig 3b, Supplementary Fig. 7b). Similarly, BSO treatment rescued the effects of shRNA-mediated Slc33a1 knockdown (Supplementary Fig. 7c) and blunted the enrichment of sgGclc cells over time, further supporting its on-target activity in rescuing the viability defects caused by Slc33a1 loss (Fig. 3c). Importantly, BSO treatment also rescued the fitness defects observed in a panel of SLC33A1-deficient KEAP1-mutant human LUAD cell lines (Fig. 3d). To determine whether additional NRF2-dependent metabolic changes could also modulate Slc33a1 dependence, we treated cells with the xCT inhibitor Erastin³⁸, N-acetyl cysteine (NAC), or dimethyl 2-oxoglutarate (DMG; cell permeable alpha-ketoglutarate). Similar to BSO, Erastin fully rescued the fitness defect in sgSlc33a1-targeted cells (Fig. 3e and Supplementary Fig. 7b) and, importantly, neither agent was able to rescue sgRpa3-targeted cells (Supplementary Fig. 4d). NAC and DMG were unable to rescue the fitness of Slc33a1-deficient KPK cells (Fig. 3e). These results suggest that Slc33a1 dependency is not due to increased ROS, decreased cysteine availability, or defects in central carbon metabolism¹⁴, respectively. However, these results do not rule out the possibility that organelle specific changes occur post Slc33a1 loss. Perturbation of the GSH:GSSG ratio to favor GSSG accumulation via menadione-induced NADPH oxidation³⁹ fails to rescue Slc33a1-deficient KPK cells (Supplementary Fig. 7e). Moreover, treatment with the thioredoxin reductase-1 (Txnrd1) inhibitor, Auranofin⁴⁰, failed to rescue the effects of Slc33a1 loss, suggesting that total pools or de novo synthesis of GSH, but not thioredoxin, mediate Slc33a1 dependence (Fig. 3e). Collectively, these data demonstrate that inhibition of GSH synthesis is sufficient for rescuing the effects of Slc33a1 loss.

Figure 3: Loss of Slc33a1 is rescued by inhibiting glutathione synthesis.

(a) Schematic of the biosynthetic pathways contributing to GSH synthesis. Enzymes marked in blue circles are direct NRF2 target genes. Small molecule inhibitors are denoted by red font.

(b) Fluorescence competition assay in KP and KPK isogenic cell lines infected with pUSEPR lentiviruses expressing sgCtrl, sgSlc33a1, sgSuco, or sgTapt1. All cell and sgRNA combinations are treated with vehicle or 50 uM BSO. Relative %RFP is normalized to %RFP at day 2 post infection. Statistics are derived as the mean ± SD; n = 3; n denotes technical replicate infections.

(c) Fluorescence competition assay of RFP⁺-Slc33a1-deficient cells (pUSEPR) targeted with sgCtrl or sgGclc containing pUSPmNG lentiviral vectors. Double positive %RFP:GFP values normalized to day 2 %RFP:GFP values. Cells were treated with vehicle or 50 uM BSO (mean ± SD; n = 3; n denotes technical replicate infections).

(d) Heat map of fluorescence competition assays in KEAP1-mutant human LUAD cell lines (pUSEPR) treated with vehicle or 50 uM BSO. Data normalization as in (b). Each block is derived as the mean ± SD; n = 3; n denotes technical replicate infections.

(e) Day 2 normalized %RFP (pUSEPR) marked cells in a fluorescence competition assay targeted with sgCtrl or sgSlc33a1. Cells were treated with vehicle, 50 uM BSO, 500 nM Erastin, 500 nM Auronafin, 1 mM NAC, or 2.5 mM DMG (mean ± SD; n = 3; n denotes technical replicate infections and treatment).

(f) qPCR validation of UPR and NRF2 target genes in isogenic pair #1 targeted with sgCtrl or sgSlc33a1. (mean ± SD; n = 4; n denotes technical replicates). Cells were treated with vehicle or 50 uM BSO for 48 hours post infection. Statistical analyses were performed using student’s two-tailed t-test in (b) and (c) and two-way ANOVA in (e) and (f) with Tukey’s multiple comparisons test. All error bars denote standard deviation.

To test whether induction of the UPR is prevented by GSH depletion, we treated KP- or KPK-sgSlc33a1 and -sgCtrl cells with BSO and quantified relative expression of a panel of UPR target genes. Consistent with recent observations, short-term treatment with BSO had no impact on UPR gene induction in sgCtrl cells⁴¹ (Fig. 3f). However, we observed increases in NRF2 pathway targets (Gclc, Nqo1) upon BSO treatment in KP and KPK cells, consistent with decreases in total GSH pools, which in turn lead to an oxidized environment that activates NRF2 signaling (Fig. 3f). Notably, BSO treatment rescued UPR gene induction in sgSlc33a1 cells, suggesting that high GSH synthesis is a critical mediator in UPR induction following Slc33a1 loss (Fig. 3f). To determine whether Keap1 loss renders cells more sensitive to ER stress and UPR induction, we treated KP and KPK cells with tunicamycin and thapsigargin, commonly used small-molecule inducers of the UPR. On average, KPK cells displayed slightly increased sensitivity to treatment (Supplementary Fig. 8a-c). KP1 cells displayed similar sensitivity to tunicamycin compared to their isogenic KPK1 counterpart. We speculate that this is due to the increase in overall NRF2 pathway activation, as we previously detected increased GCLC protein in this line compared to other KP cell lines⁴. Importantly, co-treatment with BSO led to significantly increased thapsigargin resistance as measured by the area under the curve (AUC), with a larger magnitude of rescue in KPK cells (Supplementary Fig. 8d-f). Together, these data suggest that increased GSH synthesis promoted by Keap1 deficiency leads to heightened sensitivity to genetic perturbations of ER-localized proteins (e.g. loss of Slc33a1, Suco, or Tapt1) or chemical perturbations (e.g. thapsigargin treatment).

Loss of Slc33a1 is associated with widespread metabolic changes

Slc33a1 has been reported to directly transport acetyl-CoA into the ER lumen⁴². As such, mutation of Slc33a1 could disrupt cellular metabolic processes that depend on acetyl-CoA transport. To characterize differences in polar metabolite levels and usage following Slc33a1 loss, we performed liquid chromatography-mass spectrometry (LC-MS) in KP- and KPK-sgCtrl or -sgSlc33a1 cells maintained in BSO for 48 hours or 48 hours post BSO withdrawal (a time point at which there is no significant cell death; Supplementary Table 8). The steady state levels of 45/167 detected metabolites were significantly different across all samples (Supplementary Table 8). Unsupervised hierarchical clustering revealed three major clusters, with all technical replicates of vehicle-treated KPK-sgSlc33a1 cells clustering separately from the rest (Fig. 4a). Consistent with earlier reports, we detected reduced TCA cycle intermediates in KPK-sgCtrl compared to KP-sgCtrl cells, with increased intracellular GSH, as well as concomitant decreases in intracellular α-ketoglutarate^4,14 (Supplementary Fig. 9a,b). Notably, TCA-derived metabolites were significantly increased in KPK cells targeted with sgSlc33a1, consistent with Slc33a1 loss either boosting TCA cycle utilization or causing increases in TCA cycle components via autophagy-mediated breakdown (Supplementary Fig. 9a). Interestingly, in addition to increased total levels of GSH (Fig. 4b), we observed increased oxidized GSH (GSSG) levels, resulting in decreased GSH:GSSG ratios in KPK-sgSlc33a1 cells compared to KPK-sgCtrl cells (Supplementary Fig. 9b).

Figure 4: Slc33a1 loss results in metabolic rewiring.

(a) Unsupervised clustering of significantly differentially detectable polar metabolites. Each row represents a different metabolite. Each column represents a different sample. Each cell line condition was completed in triplicate. Data is normalized by cell counts for each cell line.

(b) GSH concentrations of KP and KPK cells 72 hours post transduction with the above labeled sgRNAs.

(c) Acetyl-CoA and CoA relative abundances from data represented in (a). Cells were treated with or without 50 uM BSO for 48 hours. Each cell line condition was completed in triplicate. Data is normalized by cell counts for each cell line. Statistical analyses were performed using Mann-Whitney U-test in (b) and (c).

To assess whether the observed changes in intracellular metabolites were a result of changes in influx or secretion, we used gas-chromatography-mass spectrometry (GC-MS) to assess changes in metabolite consumption or secretion in tissue culture media. GC-MS-based detection of metabolites confirmed an increase in glutamine and cystine consumption and an increase in glutamate secretion in KPK-sgCtrl versus KP-sgCtrl cells, likely associated with increased demand for GSH synthesis resulting from Keap1 deficiency (Supplementary Fig. 9c and Supplementary Table 9). Of note, the increase in glutamine and cystine consumption, as well as glutamate secretion, was further increased upon Slc33a1 loss (Supplementary Fig. 9c).

Given that Slc33a1 is reported to be involved in cytoplasmic-to-ER-acetyl CoA transport⁵ and that steady state acetyl-CoA and CoA pools were lower in KPK-sgCtrl cells relative to KP-sgCtrl cells (Fig. 4c and Supplementary Fig. 9a), it is possible that KPK cells were more sensitive to Slc33a1 loss due to limitations in the transport of acetyl-CoA (and other undefined metabolites) into the ER. Moreover, we detected further decreases of intracellular CoA and transsulfuration intermediates in KPK-sgSlc33a1 cells relative to KPK-sgCtrl cells (Supplementary Table 8). A potential explanation for this observation is that GSH or GSH synthesis may directly impact the availability of intracellular acetyl-CoA. Supporting this, BSO-treated KPK-sgSlc33a1 cells clustered most closely with KP-sgCtrl and KPK-sgCtrl cells, and displayed an increased relative abundance of acetyl-CoA and CoA (Fig. 4c and Supplementary Fig. 9d,e). Altogether, these data demonstrate that Slc33a1 loss results in dramatic changes in the metabolic profiles of murine Kras-mutant LUAD cell lines, particularly in combination with Keap1 loss. However, whether acetyl-CoA is the relevant metabolite that explains Slc33a1 dependency remains unknown.

Transplant models and GEMMs of Keap1-mutant LUAD are dependent on Slc33a1

To address the impact of Slc33a1 loss in vivo, we first transplanted KP1 or KPK1 cell lines expressing sgCtrl or sgSlc33a1 into the flanks of immunocompromised mice and measured longitudinal growth, final subcutaneous tumor mass, and in vivo competitive fitness and confirmed a dependency for Slc33a1 in transplanted KPK tumors (Fig. 5a-c). Moreover, inducible shRNA-mediated knockdown (Supplementary Fig. 2e-h) of Slc33a1 post tumor initiation led to a significant decrease in tumor volume in KPK but not KP transplants compared to shRenilla controls (p = 0.0006; Fig. 5d,e). Additionally, sgSLC33A1-targeted A549 xenografts (KEAP1-mutant) grew more slowly (A549 p<0.001) and resulted in smaller tumors (A549: fold change = 2.711, p = 0.019) compared to sgCTRL targeted tumors (Fig. 5f,g). Importantly, we do not observe decreases in tumor size in KEAP1-wild-type sgSLC33A1-targeted H2009 or H441 xenografts (Fig. 5f,g). These data confirmed a targetable dependency of Keap1-mutant cells on Slc33a1 and prompted us to perform further validation in a genetically engineered mouse model (GEMM) of LUAD⁴³ that better recapitulates the differential nutrient availability between tumors and surrounding normal tissues, as well as the native lung microenvironment^44,45. We utilized the Kras^LSL-G12D/+ (LSL=loxP-STOP-loxP); p53^fl/fl GEMM of LUAD (hereafter KP)^46,47, in which tumors are induced upon intratracheal instillation of lentivirus expressing Cre recombinase^43,48. KP mice were crossed to germline conditional alleles of Keap1 (ref. 38) and Cas9 (R26^LSL-Cas9; ref. 39) for NRF2 pathway activation and somatic editing of genes within initiated lung tumors, respectively, generating Kras^LSL-G12D/+; p53^fl/fl; Keap1^fl/fl; R26^LSL-Cas9/+ mice (hereafter KPC or KPKC). To assess the effects of Slc33a1 targeting in Keap1-wild-type and Keap1-null tumors, we induced tumors in KPC or KPKC mice with pUSEC lentivirus⁴⁹ expressing Cre recombinase and sgRNAs targeting Slc33a1. To boost the efficiency of bi-allelic loss of function events⁵⁰, we generated pUSEC lentiviruses expressing two potent sgRNAs against Slc33a1 or Olfr102, hereafter referred to as sgSlc33a1 or sgCtrl, respectively (Fig. 6a and Supplementary Fig. 10a).

Figure 5: Slc33a1 loss validates as a Keap1-mutant-specific vulnerability in transplant models.

a) Final subcutaneous tumor masses obtained from the transplantation of indicated cell lines into immunocompromised mice.

b) Subcutaneous tumor volumes of KP and KPK cells transduced with the indicated sgRNAs injected into immunocompromised mice (n = 6 per group). KP sgCtrl vs. KP sgSlc33a1, p = 0.0074. KPK sgCtrl vs. KPK sgSlc33a1, p = 0.0003.

c)in vivo fluorescence based competition assay of GFP-labeled KPK cells injected into immunocompromised mice. KPK cells were transduced with pUSEPR constructs containing the indicated sgRNAs. Transduced cells representation was normalized to RFP⁺ cells at d2 (pre-implantation).

(d) Subcutaneous tumor volumes of KP or KPK cells transduced with doxycycline-inducible shRNAs targeting Renilla or Slc33a1 (LVt-TSTOP; see methods). Arrow indicates timepoint at which mice began doxycycline treatment. Error bars denote S.E.M. KPK-shRenilla versus KPK-shSlc33a1 groups p = 0.0006. KP-shRenilla versus KP-shSlc33a1 groups p = 0.8467).

(e) Tumor volumes related to (d) at the indicated time points. Day 10: KP, p = 0.9387, KPK, p = 0.0088. Day 12: KP, p = 0.5126, KPK, p < 0.0001.

(f) Xenografts of the KEAP1-mutant A549 or KEAP1-wild-type H2009 and H441 human lung cell lines infected with either sgCTRL or sgSLC33A1 lentiviral vectors (pUSEPR). Error bars denote S.E.M. A549 (n = 10 per group): p<0.001. H2009 (n = 9 sgCTRL; n = 10 sgSLC33A1): p = 0.292, H441 (n = 10 sgCTRL; n = 10 sgSLC33A1): p = 0.842.

(g) Final subcutaneous tumor masses obtained from xenografts of indicated cell lines into immunocompromised mice. A549 (n = 10 sgCTRL; n = 10 sgSLC33A1): p<0.0190. H2009 (n = 9 sgCTRL; n = 7 sgSLC33A1): p<0.8805. H441 (n = 10 sgCTRL; n = 10

sgSLC33A1): p<0.3061. Statistics obtained by: Student’s two-tailed t-test in (a), (c), and (g). Two-way ANOVA with Sidak’s multiple comparisons testing in (b), and (f). Two-way ANOVA with Dunnet’s multiple comparisons testing in (d). One-way ANOVA with Dunnet’s multiple comparisons testing in (e).

Figure 6: Keap1-deficient tumors harbor an increased dependency for Slc33a1 in an autochthonous model of murine lung adenocarcinoma.

(a) Schematic representation of Kras^LSL-G12D/+; p53^fl/fl; Rosa26^LSL-Cas9 (KPC) or Kras^LSL-G12D/+; p53^fl/fl; Keap1^fl/fl; Rosa26^LSL-Cas9 (KPKC) mice intratracheally infected with pUSEC lentiviruses containing dual sgRNAs targeting Slc33a1 or Olfr102 (sgCtrl).

(b) Distribution of histological tumor grades (G1-G4) in KPC or KPKC mice 20 weeks after infection with pUSEC lentiviruses expressing control (sgCtrl, KPC, n = 10 mice; KPKC, n = 20 mice) or sgSlc33a1 (KPC, n = 12 mice; KPKC, n = 22 mice). Error bars denote S.E.M. (KP-sgCtrl versus KPK-sgCtrl p = 0.049; KPK-sgCtrl versus KPK-sgSlc33a1 G2 p = 8.2e-9; KPK-sgCtrl versus KPK-sgSlc33a1 G3 p = 2.81e-6; KPK-sgCtrl versus KPK-sgSlc33a1 G4 p = 0.025).

(c) Combined quantification of tumor burden (total tumor area/total lung area) in KPC or KPKC mice after infection with pUSEC lentiviruses. Mouse numbers equivalent to (b). (KP-sgCtrl versus KP-sgCtrl p = 0.08; KP-sgCtrl versus KPK-sgCtrl p = 0.029; KPK-sgCtrl versus KPK-sgSlc33a1 p = 5.9e-7).

(d) Combined quantification of average tumor number in KPC or KPKC mice after infection with pUSEC lentiviruses.

(e) Quantification of phosphorylated histone H3 (pHH3)-positive nuclei per squared millimeter of tumor for assessment of the mitotic index of tumor cells from lung tumors in KPC or KPKC mice at 20 weeks after infection with the indicated pUSEC lentivirus.

(f) Fraction of mutant and wild-type reads within individually plucked tumors from 20-week pUSEC-sgSlc33a1 lentiviral infected KPC (mean of n = 15 tumors from 7 mice) and KPKC mice (mean of n = 14 tumors from 7 mice).

(g) Mean number of frameshift (FS) reads from USEC-sgSlc33a1 lentiviral infected KPC (mean of n = 15 tumors from 7 mice) and KPKC mice (mean of n = 14 tumors from 7 mice) * P = 0.0102.

(h) Representative H&E and IHC staining of serial sections from lung tumors of KPC or KPKC mice 20 weeks after infection with pUSEC-sgCtrl (top) or pUSEC-sgSlc33a1 (bottom). First panels (KPC mice), overall lung tumor burden (scale bar = 2mm). Second panels, higher magnification H&E staining of representative tumors (scale bar = 100 um; insets depict higher-magnification images, scale bar = 50 um). Third panels, representative IHC staining for NRF2 target gene, NQO1. This is repeated for KPKC mice. Statistical analyses were performed using Mann-Whitney U-test with Holm’s multiple comparisons correction used in (b-e). Student’s two-tailed t-test was performed for (g).

Histological analysis revealed an increase in high grade adenocarcinomas, tumor burden, tumor number, and proportion of proliferating cells in KPKC-sgCtrl mice compared to KPC-sgCtrl mice (p<0.05; Fig. 6b-e), consistent with previous studies⁴. We did not detect significant differences in grade, tumor burden, or tumor number in KPC-sgSlc33a1 compared to KPC-sgCtrl mice (Fig. 6b-d). However, there was a significant reduction in high-grade adenocarcinomas (G2 fold change = 6.6, p = 8.18e-9; G3 fold change = 4.7, p = 2.81e-6; G4 fold change = 23, p = 0.025, tumor burden (fold change = 4.36, p = 5.95e-7), and number in KPKC-sgSlc33a1 versus -sgCtrl mice (Fig. 5b-d). We were unable to detect significant differences in the mitotic index of KPKC-sgSlc33a1 tumors compared to KPKC-sgCtrl tumors, as assessed by phosphorylated histone H3 immunohistochemistry (IHC; Fig. 6e and Supplementary Fig. 10b). These results suggest that the reduced tumor burden of KPKC-sgSlc33a1 mice is likely due to decreased tumor initiation and/or a heightened barrier for transition to higher grade LUAD (Fig. 6e). Importantly, early lesions may have been enriched for incomplete Keap1 recombination (resulting in decreased NRF2 activation) or incomplete deletion of Slc33a1.

Supporting the latter possibility, deep sequencing of the Slc33a1 locus in micro-dissected tumors from sgSlc33a1-targeted mice revealed a significantly greater ratio of wild-type sequence reads and non-frameshift mutations (which may preserve gene function) in KPKC tumors compared to KPC tumors, indicating strong selection against combined deletion of Keap1 and Slc33a1 (Fig. 6f,g and Supplementary Fig. 10c-e). Furthermore, while Keap1-recombined tumors displayed enhanced expression of the NRF2 target gene, NQO1 (refs. 4, 42; Fig. 6h), we observed a nonsignificant trend towards decreased NQO1 positive tumors in KPKC-sgSlc33a1 animals (Supplementary Fig. 10f). These results suggest a potential mechanism of tumor escape via failure to completely recombine Keap1, or alternative mechanisms to dampen the NRF2 program.

Collectively, these data validate Slc33a1 as a potential target in Keap1-mutant LUAD in vivo. Furthermore, Keap1-wild-type tumors that have upregulated the NRF2 pathway via other genetic or epigenetic mechanisms may also show sensitivity to Slc33a1 loss or inhibition.

Whole genome CRISPR screen identifies suppressors of Slc33a1 dependency

To gain further insights into the underlying biology of Slc33a1, Suco, and Tapt1 loss in a Keap1-mutant context, as well as identify patient genotypes that may not respond to SLC33A1 inhibition, we performed an unbiased whole genome CRISPR screen using a previously established library⁵¹ (hereafter Brie; Fig. 7a). To that end, we leveraged BSO-mediated rescue to isolate KPK-Slc33a1, KPK-Suco, and KPK-Tapt1 knockout clones (Fig. 7a and Supplementary Fig. 11a,b). Four KPK-Slc33a1 single cell clones were pooled together for subsequent screening. We transduced the Brie library into KPK-Slc33a1 cells and separated into BSO (3 replicates) or vehicle (6 replicates) treated conditions. Library representation was determined by deep sequencing of sgRNAs amplified from BSO or vehicle treated cells at early (t₀) and 8 population doublings post transduction (t₈; Fig.7b, Supplementary Fig. 11c and Supplementary Table 10). We assessed median GS within the vehicle treated groups (t₈-t₀) and identified Nrf2, Gclc, and Gclm as the top enriched genes, thus validating our previous data, now in a fully unbiased manner (Fig. 7c). Given that Slc33a1 loss results in a potent growth-disadvantage in KPK cells, we hypothesized that any gene at t₈ with a positive sgRNA count across all replicate infections, and that was represented by ≥ 2 individual sgRNAs, would represent true biological hits capable of rescuing KPK-Slc33a1 deficiency. Using these thresholds (see methods), we identified a number of additional enriched genes, including ER-resident oxidoreductase 1A (Ero1l), protein disulfide isomerase 1a (Pdia/P4hb), and the NRF2 target genes Osgin1, Amotl2, and Pgd⁵² (Fig. 7c). GSs of the top enriched genes were decreased in BSO-treated samples (Supplementary Fig. 12a,b).

Figure 7: Genome-wide CRISPR screen identifies suppressors of Slc33a1 deficiency.

(a) Schematic of the screen conditions. KPK-Slc33a1 knockout cells were generated and infected with the Brie library to identify hits that would rescue Slc33a1 dependency.

(b) Schematic representing screen conditions. KPK-Slc33a1 knockout cells were infected with Brie library in the presence of BSO and subsequently placed in BSO or vehicle treated conditions for 8 cumulative population doublings.

(c) Ranked median gene scores of genes in the Brie library that passed all thresholds (see methods). Dotted lines represent the spread of the control sgRNAs and mark the 25^th and 75^th gene score percentile boundaries.

(d) Waterfall plot of the normalized median sgRNA counts across all technical replicates at t₈. The median is plotted and the error bars denote the range of the data.

(e) Above: denotes the schematic for the validation pipeline. Below: Relative cell numbers post infection with pUSEPR vectors expressing the indicated sgRNA. Cell morphology related to Supplementary Fig. 13a.

To validate the top enriched candidate genes and to further investigate underlying biology that might link the observed dependency associated with targeting Slc33a1, Suco, and Tapt1, we infected KPK-Slc33a1, KPK-Suco, and KPK-Tapt1 knockout cells with lentiviruses encoding two independent top scoring sgRNAs from this screen, and assessed growth and phenotypic consequences post-BSO withdrawal. Remarkably, loss of the majority of the candidate genes rescued the growth of Slc33a1 cells, but not Suco or Tapt1 cells, as measured by cell numbers and colony formation assay (Fig. 7e and Supplementary Fig. 12c). Importantly, although Nf2, Cul3, Cul5, Rnf7, Amotl2, and Eif2ak3 (Perk) sgRNAs rescued the growth of KPK-Slc33a1 cells, Ero1l, P4hb, Osgin1, Serp1, and Yipf5 sgRNAs rescued both the growth and blebbing phenotype associated with Slc33a1 loss (Supplementary Fig. 13a). Moreover, Ero1l, P4hb, Serp1, and Yipf5 are all known ER-resident genes, supporting the original observation that KPK cells have a dysfunctional ER. Additionally, loss of Osgin1 rescued the growth and phenotype of Slc33a1-deficient cells, suggesting that other genes downstream of NRF2 might contribute to Slc33a1 dependency (Fig. 7e, Supplementary Fig. 12c and Supplementary Fig. 13a). Interestingly, Suco and Tapt1 knockout cells were only rescued by sgRNAs targeting Nf2, Cul5, Rnf7, Osgin1, and Pgd, suggesting that Suco and Tapt1 dependency is biologically distinct from Slc33a1 dependency, but overlap in redox biology via rescue through GCLC inhibition (Fig. 3b, Fig. 7e and Supplementary Fig. 12c). In summary, these data suggest that Slc33a1 loss results in a dysfunctional ER state that significantly affects the growth of Keap1-mutant cells. ERO1L is positively regulated by GSH to generate oxidation events used by P4HB for proper protein-folding^53,54. Therefore, loss of either gene is predicted to result in decreased ER-oxidation or decreased GSSG production and raises the possibility that SLC33A1 may be involved in regulating ER redox potentials.

Discussion:

We have demonstrated that Keap1 mutation results in an increased NRF2-dependent vulnerability to Slc33a1, Suco, and Tapt1 loss. Although several studies have identified a number of putative KEAP1/NRF2-mutant-specific synthetic lethal interactions that could be exploited for therapeutic purposes^{4,14,16,55,56}, all studies so far have been limited in scale (e.g. focused on interrogating canonical NRF2 target genes as candidate synthetic lethal factors) and do not rule out the possibility that more comprehensive and systematic queries using increasingly complex functional genomic approaches will uncover novel factors that could be targeted for therapeutic benefit. Importantly, although we focused this study on Slc33a1, we hypothesize that Keap1-mutant cells are universally sensitive to dysfunction of an ER-stress pathway of which Slc33a1 loss imposes the largest selective disadvantage. In line with this observation, we demonstrate increased dependency for the endomembrane associated genes SUCO and TAPT1³² in KEAP1-mutant human cancer cell lines observed in the DepMap, which we validated in our murine Keap1-mutant cell lines. While SLC33A1 mutations are associated with developmental defects and childhood mortality, suggestive of a critical role in development⁴², pooled genetic screens in embryonic and cancer cell lines suggest that SLC33A1 is not a pan-essential gene^57,58. Consistent with a non-essential role, we demonstrate that the Slc33a1 dependence of Keap1-mutant cancer cells can be rescued by loss of a number of genes with roles in the integrated stress response network. Specifically, loss of the ER-resident genes Erol1, P4hb, Serp1, and Yipf5 or the NRF2-target genes Gclc, Gclm, and Osgin1 rescue both the growth disadvantage and cellular morphology associated with Slc33a1 loss in NRF2-hyperactivated cells. Although the direct link between GSH and SLC33A1, SUCO, and TAPT1 remains unclear, we hypothesize that the ER of Keap1-deficient cells have been shifted into a dysfunctional state that likely increases proteotoxic stress due to challenges in protein folding that can be rescued by decreasing ER-specific oxidation or ER-specific GSSG production. Consistent with this idea, we observe increased intracellular GSSG upon Slc33a1 loss in Keap1-mutant cells. However, ER-specific accumulation of GSSG upon SLC33A1 inhibition remains to be experimentally validated, and it is possible that other mechanisms underlie the dependence of KEAP1-deficient cells on SLC33A1 and other ER-resident proteins.

Our results are particularly interesting in light of a recent meta-analysis in human cancer cell lines identifying a presumed connection between SLC33A1 dependency, the NRF2 pathway, and high GSH³⁰. Here, we demonstrate that Slc33a1 dependency is rescued via inhibition of GSH synthesis and also behaves as a potent Keap1-mutant specific genetic dependency in a GEMM of Kras-mutant LUAD, validating the cell-autonomous requirement for Slc33a1 function in Keap1-mutant LUAD in vivo. Overall, our study provides a strong rationale for substratification of patients harboring KEAP1-mutant or NRF2-hyperactivated tumors as likely responders to targeted SLC33A1 inhibition and underscore the value of integrating functional genetic approaches with GEMMs to identify and validate genotype-specific therapeutic targets.

Materials and Methods:

Mice

All animal studies described in this study were approved by the MIT Institutional Animal Care and Use Committee. Kras^LSL-G12D and Trp53^fl/fl mice have already been described^46,47. Keap1^fl/fl mice were generated and shared by S. Biswal⁵⁹. Rosa26-LSL-Cas9 mice were obtained from Jackson labs⁶⁰. For all animal studies, >3 animals were used for each experimental cohort per specified genotype. All mice were maintained on a mixed C57BL/6:SV129 genetic background. Total burden, and grading analysis were conducted on >3 mice per genotype. Animals lacking detectable tumors by histopathology were excluded from the analysis to ensure all animals were properly infected with pUSEC lentiviruses. Animals with the appropriate genotypes between the ages of 6-10 weeks were randomly selected to begin tumor initiation studies with pUSEC-sgOlfr102 (sgCtrl) or pUSEC-sgSlc33a1. Mice were infected intratracheally with 20,000 TU’s of lentiviruses as described⁴³. Total lung area occupied by tumor was measured on hematoxylin and eosin (H&E) stained slides using NIS-elements software. Histological quantification of mouse lung tumor burden by grade was performed by an automated deep neural network (unpublished) developed by Aiforia Technologies in collaboration with the Jacks lab, and in consultation with veterinarian pathologist Dr. Roderick Bronson. The algorithm performed consistently and with high correlation with human graders across multiple validation datasets independent of the training dataset. For murine cell sgRNA transplant experiments, a total of 1E6 pUSEPR-infected and puromycin selected cells were injected into the flanks of immunocompromised mice in a total volume of 100 uL of 50:50 PBS:Matrigel mix. For murine cell doxycycline inducible shRNA transplant experiments, a total of 500E5 LVt-TSTOP infected and puromycin selected cells were injected into the flanks of immunocompromised mice in a total volume of 100 uL PBS. Mice were placed on doxycycline chow when tumors reached 10-25mm³. For xenograft transplant experiments (A549, H2009, H441), a total of 1E6 pUSEPR-infected and puromycin selected cells were injected into the flanks of immunocompromised mice in a total volume of 100 uL of 50:50 PBS:Matrigel mix. Subcutaneous tumor volumes were calculated according to the following formula: mm³ = (a²*b)*(π/6), where a is the smaller dimension and b is the larger dimension. 8-week female J/nu, athymic nude mice (Jackson 007850) were used for all transplant experiments.

Cell culture

HEK293T cells were maintained in DMEM (Corning #10-013-CV) supplemented with 10% fetal bovine serum, 2 mM L-glutamine (Gibco #25030) and 50 ug/mL gentamycin (Gibco #15710). Parental cell lines from KP mice⁶¹ were previously established and described. Human cell lines were acquired from ATCC. All lines tested negative for mycoplasma and routinely tested every 3-4 months. Murine KP derivative cells were maintained in DMEM or RPMI supplemented with 10% Fetal Bovine Serum and 50 ug/mL gentamicin. Murine cell lines transduced with Puro^R-containing constructs were kept under 6ug/mL Puromycin selection. Cell lines transduced with lenti-Cas9-Blast (Addgene #52962) remained under Blasticidin selection (10ug/mL). Cells were treated with inhibitors: L-Buthionine-sulfoximine (BSO, Sigma Aldrich), Auranofin (AUR, TOCRIS bioscience), Erastin (ERA, Sigma Aldrich), N-acetyl-L-cysteine (NAC, Sigma Aldrich), and 2mM dimethyl-2-oxoglutarate (DMG, Sigma Aldrich) at concentrations denoted in figure legends. Viability in the presence of all compounds was assessed by Cell Titer Glo (Promega #G7570). Plates were quantified using a Tecan infinite M200 Pro plate reader or a SpectraMax M5 Microplate reader (Molecular Devices). For clonogenic and low-density assays, cells were stained with Crystal Violet solution (25% Methanol 75% H₂O). Crystal violet stains were analyzed by ImageJ software.

Design and cloning of the druggable genome sgRNA library

As previously described¹¹, the 4,915 genes included in the sgRNA library were selected from the Drug-Gene Interaction Database^{62, 63}. Human gene names were first updated using the HUGO Gene Nomenclature Committee (HGNC) multi-symbol checker tool, and then converted to the corresponding mouse orthologs using the HGNC Comparison of Orthology Predictions tool (accessed 7/21/2016)^64,65. sgRNA sequences (four per gene, 19,660 total) were obtained from the Brie whole-genome sgRNA library designed by the Broad Institute sgRNA Designer tool⁵¹. 500 non-targeting control sgRNA sequences from the Asiago whole-genome sgRNA library⁵¹ were included, bringing the total library size to 20,160 sgRNAs. The sgRNA library was cloned into lentiGuide-Puro (Addgene #52963; lentiGuide-Puro was a gift from Feng Zhang), using a previously established protocol⁶⁶. In brief, 5’ and 3’ flanking adapter sequences, corresponding to the U6 promoter and tracrRNA sequences, respectively, were appended to the sgRNA sequences. The oligo library was synthesized by Twist Bioscience, PCR amplified with Oligo-Fwd and Oligo-knockout-Rev primers using PfuUltra II Fusion HS DNA Polymerase (Agilent Technologies #600670), purified using the QIAquick PCR Purification Kit (QIAGEN #28104), then assembled into the BsmBI-digested lentiGuide-Puro vector using the Gibson Assembly Master Mix (New England Biolabs #E2611S). A total of 800 ng of assembled plasmid DNA was electroporated into Endura electrocompetent cells (Lucigen #60242), the plated onto LB-ampicillin plates overnight at 37C. The total number of colonies was quantified to ensure a representation of >100X (>2 million colonies for a 20,000 sgRNA library). Plasmids were isolated and purified using the QIAGEN Plasmid Plus Maxi Kit (QIAGEN #12963). To determine sgRNA distribution in the cloned library, the sgRNA target region was amplified with NGS-Lib-Fwd (primers 1-10) and NGS Lib-KO-Rev (primer 1) primers, size selected in a 2% agarose gel, purified using the QIAquick Gel Extraction Kit (QIAGEN #28704), then submitted for sequencing on an Illumina HiSeq 2500 system (100-nt single end reads) at the Whitehead Institute Genome Technology Core.

Infection of cells with the druggable genome library

All transductions of KP and KPK lines were performed by spin-infection. Cells were plated at a concentration of 2 million cells/well in 12-well plates, together with the appropriate amount of viral supernatant in a total volume of 2 mL per well, then centrifuged at 1,500 RPM for 2 hours at room temperature. Cells were then incubated 37C overnight.

To ensure that most cells harbor single integration events, a multiplicity of infection (MOI) of 0.3 was used for transduction of cell lines with the sgRNA library⁶⁶. To determine the viral titer for each cell line, all cells were transduced with decreasing volumes of viral supernatant, selected with 6 ug/mL puromycin (Gibco #A11138) for 24 hours, then quantified by live cell numbers and normalized to a control, unselected sample.

For the actual screen, a minimum of 20 million cells was maintained at each step to ensure a coverage of >1000 cells per sgRNA. For each cell line, 200 million cells were transduced with the sgRNA library in 12-well plates at an MOI of 0.3, then pooled into 15 cm dishes the next day. 24 hours post re-plating, 100 million cells were snap frozen in liquid nitrogen, then stored at −80C (for the preselection t₀ population), while 100 million cells were selected with 6 ug/mL puromycin for 24 hours. Subsequently, cells were passaged every two days, and at least 20 million cells were plated after each passage. After 8 population doublings (t₀), 20 million cells were snap-frozen in liquid nitrogen, then stored at −80C.

Infection of cells with the Brie library

Brie library was obtained from Addgene (cat. #73633). KPK-Slc33a1 knockout clones were pooled (5 western validated clones) and plated at a concentration of 2 million cells/well in 12-well plates, together with the appropriate amount of viral supernatant in a total volume of 2 mL per well, then centrifuged at 1,500 RPM for 2 hours at room temperature. Cells were then incubated 37C overnight. To ensure that most cells harbor single integration events, a multiplicity of infection (MOI) of 0.3 was used for transduction of cell lines with the Brie library⁶⁶. A minimum of 5 million cells was maintained at each step to ensure a coverage of >50 cells per sgRNA (78,637 sgRNAs x 50X coverage = ~3.9E6 cells). Cells were transduced in technical triplicates in 50 uM BSO with the Brie library in 12-well plates at an MOI of 0.3, then pooled into 15 cm dishes containing 50 uM BSO the next day. 48 hours post infection, cells were selected in 6 ug/mL Puromycin for 48 hours. Following selection, 5E6 cells per triplicate infection were harvested for t₀ sgRNA representation. 5E6 cells were then plated into vehicle (water; 6 replicates) or 50 uM BSO (3 replicates) for the duration of the experiment. Subsequently, cells were passaged every two days or when confluency was reached, and at least 5 million cells were plated after each passage to maintain 50X library coverage. After 8 population doublings (t₈), 10 million cells were snap-frozen in liquid nitrogen, then stored at −80C.

Genomic DNA isolation

Cell pellets were resuspended in 6 mL NK cell lysis buffer (50 mM Tris HCl, 50 mM EDTA, 1% SDS, pH 8) per 30-50 million cells. 30 uL of 20 mg/mL proteinase K (QIAGEN #19131) was added to the sample, which was then rotated overnight at 55C. The next day, 30 uL of 10 mg/mL RNase A (QIAGEN #19101) was added to the sample, which was then inverted 25 times and incubated for 30 minutes at 37C. Samples were cooled on ice, then 2 mL of pre-chilled 7.5 M ammonium acetate was added. Samples were vortexed at high speed for 20 seconds, then centrifuged at 4,000 x g for 10 minutes. The supernatant was decanted into a new tube, 6 mL of 100% isopropanol was added to the cleared supernatant. The sample was inverted 50 times and centrifuged at 4000 x g for 10 minutes. The supernatant was discarded, leaving genome DNA as a small white pellet. 6 mL of 70% ethanol was added to wash the pellet, then the tube was inverted 10 times and centrifuged at 4000 x g for 1 minute. The supernatant was discarded, then the DNA pellet was re-dissolved in 500 uL of TE Buffer (Sigma #T9285) and incubated overnight at 55C.

Screen deconvolution and analysis (druggable genome library)

Integrated sgRNAs were amplified by PCR to attach sequencing adapters and barcodes, as described⁵¹. In brief, a mix of 1 P5_XPR/LKO1 forward primers and 15 P7-index reverse primers (see supplemental sequences) was used to amplify sgRNAs from each genome DNA sample, using TaKaRa Ex Taq DNA polymerase (RR001C) with 25 cycles at 62C annealing temperatures. Assuming that each cell contains approximately 6.6 pg of genome DNA⁶⁷, a total of 160 ug of genome DNA (corresponding to >20 million cells) was used as a template for t₀ samples, while 720 ug of genomic DNA (>100 million cells) was used for t₀ samples. Amplified PCR products of around 350 bp in length were size-selected on a 2% agarose gel, then purified using the QIAquick Gel Extraction Kit (QIAGEN #28704). All samples were submitted for sequencing on an Illumina NextSeq system (75-nt single end reads) at the MIT BioMicro Center.

Sequencing results were analyzed using R (Version 3.1.1, www.r-project.org). Flanking adapter sequences were stripped to retrieve sgRNA target sequences from FASTQ file. Target sequences were collapsed by identity and quantified in terms of exact matches to library sequences. Sequencing read counts per sample were normalized to the total number of reads to yield counts per million sequences, followed by a log₂ transformation⁵¹.

The log₂ fold change (L2FC) score for each sgRNA (sgRNA score) was calculated as the difference between t₈ and t₀ log₂ transformed counts. Gene scores (GS) were then calculated as the median of the four sgRNA scores associated with each gene⁶⁸. Differential GS for each isogenic pair were calculated by subtracting KP line GS from the KPK line GS. To determine thresholds for significant differentially depleted genes in the KPK versus KP cell line pairs, the 500 control sgRNAs were randomly binned into groups of 4, median L2FC of t₈ versus t₀ sgRNA read counts calculated, and differential GS calculated for these “pseudogenes”. This process was repeated 10,000 times to define a confident range of possible pseudogene differential GS, and thresholds set at the lower 5% of scores (false-discovery rate = 0.05. −0.64 in pair #1, and −0.63 in pair #2). To filter out genes that substantially deplete in KP lines, only genes with scores above the median GS of all essential genes (defined as the union of two previously described essential gene sets^{27,69, 70}, Supplemental Table 1) in each KP line (−1.77 in KP1, −2.25 in KP2) were included. To ensure that identified genes are significantly depleted in KPK lines (rather than enriched in KP lines but neutral in KPK lines), the population of pseudogene scores determined previously were used to set thresholds for significant depletion in the KPK lines (FDR = 0.05. −0.43 in KPK1, −0.34 in KPK2). Genes with scores above these thresholds in the KPK lines were excluded. Finally, we excluded essential genes from the analysis^{27,69, 70}.

Screen deconvolution and analysis (Brie library)

gDNA extraction and library PCR amplification was performed as above. A total of 1 ug of genome DNA was used as a template for t₀ and t₈ samples. All samples were submitted for sequencing on an Illumina NextSeq system (75-nt single end reads) with 50% Phi-X DNA spike in at the MIT BioMicro Center. Sequencing results were analyzed using R (Version 3.1.1, www.r-project.org). Flanking adapter sequences were stripped to retrieve sgRNA target sequences from FASTQ file. Target sequences were collapsed by identity and quantified in terms of exact matches to library sequences. Sequencing read counts per sample were normalized to the total number of reads to yield counts per million sequences, followed by a log₂ transformation⁵¹. The log₂ fold change (L2FC) for each sgRNA (sgRNA score) was calculated as the difference between t₈ and t₀ log₂ transformed counts for either vehicle treated or BSO treated samples. Gene scores (GS) were then calculated as the median of the four sgRNA scores associated with each gene across all replicates⁶⁸. Differential GS were calculated by subtracting median BSO treated GS from the median vehicle treated GS. Thresholds set for screen analysis: For t₈ analysis, all sgRNAs with a normalized sgRNA count of 0 at t₀ across 2 or more replicates were removed from the analysis. Genes for which >2 gRNAs met these criteria across all replicates were used for further analysis. For GS analysis: To determine thresholds for significantly enriched genes in the t₈ samples, we isolated all control sgRNAs and binned them into “pseudogenes” as described above. With these control pseudogene scores, we defined the upper and lower boundaries as the 25th-percentile (GS = −3.445) and 75th-percentile (GS = −4.311). Genes falling above the 25^th-percentile threshold were deemed significantly enriched in the dataset compared to the spread of the pseudogene scores. Finally, we excluded essential genes from the analysis^{27,69, 70}

In vitro competition assays

For in vitro competition assays, lentiviral vectors expressing a target sgRNA together with mNeonGreen (pUSPmNG) or tagRFP-P2A-Puro (pUSEPR) was generated by Gibson assembly as described previously⁷¹. The U6-filler sequence was amplified by PCR using lentiGuide-puro as a template, the PGK promoter sequence was obtained from the GMAP collection⁷¹, and the mNeonGreen sequence was obtained as a gBlock gene fragment (Integrated DNA Technologies). sgRNAs targeting Slc33a1, Tapt1, Suco, Rpa3, and Olfr102 were cloned into the BsmBI-digested pUSPmNG or pUSEPR vectors, using a previously described protocol⁷². Dual sgRNA sequences were ordered as a gBlock gene fragment (Integrated DNA Technologies) contained an sgRNA (#1) H1 sgRNA (#2) sequence flanked by BsmBI-digest sites. Dual sgRNA sequences were BsmBI-digested and cloned into BsmBI-digested pUSEPR vectors. Cas9-expressing cells were infected with pUSPmNG or pUSEPR vectors expressing each sgRNA at an infection efficiency of ~50%. The percentage of mNeonGreen/tagRFP-positive cells was quantified over time using a Guava easyCyte HT flow cytometer (Millipore). For each sgRNA, the percentage of mNeonGreen/tagRFP-positive cells was normalized to respective values at 2 days post infection (defined as day 2).

Colony formation assays Brie library

Cells were infected at a high MOI with pUSEPR-sgRNA constructs by spin-infection in a 12 well plate as described above in the screen format. Cells were selected and maintained in BSO as described above in the Brie library infection. After infection, cells were withdrawn from BSO and plated in 6 well plates at a density 50E5 cells. At 7 days post seed, cells were passaged into new 6 well plates at a density of 50E5 cells. At 14 days, pictures were taken using a Nikon Eclipse Ti2 microscope at 20X magnification. At day 15, plates were processed for crystal violet stain.

RNA knockdown studies

Lentiviral vectors encoding TRE3GS-mScarlet-miR30a-hPGK-TetOne-P2A-Puro (LVt-TSTOP) were generated by Gibson assembly as shown previously⁷¹. Briefly, TRE3GS promoter and TetOne (rtTA) sequences were PCR amplified from Clontech plasmid pLVX-TetOne inducible expression systems (Cat # 631844). mScarlet sequences containing miR30a cloning site in the mScarlet 3’-UTR were obtained as a gBlock gene fragment (Integrated DNA Technologies). PGK promoter sequence was obtained from the GMAP collection⁷¹. P2A-Puro sequence was PCR amplified from lentiCRISPR v2 (Addgene #52961). 97-mer sequences targeting Renilla and Slc33a1 were obtained from http://www.splashrna.mskcc.org and cloned as previously described⁷³ (see supplemental sequences). Cells were infected at an M.O.I. ~0.5 and subsequently selected in puromycin (6ug/mL) for 72 hours. Cells were induced with doxycycline (Sigma #D9891; 10 ng/mL) for 7 days and continuously passaged in the presence or absence of doxycycline or BSO (50 uM). CRISPR-interference (CRISPRi) constructs were obtained Addgene (pLV hU6-sgRNA hUBC-dCas9-KRAB-T2A-Puro; #71236). Promoter targeting sgRNA’s for Slc33a1 or Olfr102 were obtained from previously published whole genome CRISPRi datasets.⁷⁴ For CRISPRi experiments, a dual sgRNA setup was used to obtain maximal promoter silencing. sgRNA binding sites were selected to be >50 bp apart in genomic space.

Immunoblotting

Cells were lysed in 250 μL ice-cold RIPA buffer (Pierce, #89900) supplemented with 1× Complete Mini inhibitor mixture (Roche, #11836153001) and mixed on a rotator at 4°C for 30 minutes. The protein concentration of the cell lysates was quantified using the Bio-Rad DC Protein Assay (Catalog #500-0114). 30–50 μg of total protein was separated on 4–12% Bis-Tris gradient gels (Bio-Rad) by SDS-PAGE and then transferred to nitrocellulose membranes. The following antibodies were used for immunoblotting: anti-GAPDH (Santa Cruz, sc-25778, 1:500), anti-HSP90 (BD, #610418, 1:10,000), anti-NRF2 (Santa-Cruz, sc-722, 1:200), anti-KEAP1 (CST, #8047, 1:1000), anti-GCLC (Santa Cruz, sc-22755, 1:200). anti-p62 (CST #23214, 1:1000), anti-LC3 (CST #12741, 1:1000), anti-BECLIN (CST #3495, 1:1000), anti-pS6K T421/424 (CST #9204, 1:1000), anti-BiP (CST 3177, 1:1000), anti-FLAG (CST #14793, 1:1000), anti-HH3 (CST #4499, 1:1000), anti-GCN2 (CST #65981S, 1:1000), anti-PERK (CST #5683S, 1:1000), anti-ATF4 (CST #11815S, 1:1000), anti-eIF2A (CST #5324S, 1:2000), anti-phospho-eIF2A S51 (CST #3398S, 1:1000), anti-SLC33A1 (Sigma HPA042430, 1:1000).

Immunohistochemistry and immunofluorescence

Mice were euthanized by carbon dioxide asphyxiation. Lungs were perfused through the trachea with PBS and then 4% paraformaldehyde (PFA), fixed overnight, transferred to 70% ethanol and subsequently embedded in paraffin. Sections were cut at a thickness of four micrometers and stained with H&E for pathological examination. Chromogenic immunohistochemistry (IHC) was performed on a Ventana Medical Systems Discovery XT instrument with online deparaffinization using Ventana’s reagents and detection kits and antigen retrieved in Ventana Cell Conditioner 1 or 2. The following antibodies were used for IHC: anti-phospho-Histone H3 (pHH3) (Ser10; Cell Signaling, 9701, 1:200), anti-NQO1 (Sigma Aldrich, HPA007308, 1:200). Horseradish peroxidase (HRP) detection was used for NQO1, and pHH3. NQO1 and pHH3 was antigen retrieved in Ventana Cell Conditioner 1 (Tris-Borate-EDTA). Pictures were obtained using a digital whole slide scanner Leica SCN400F and Slidepath software version 4.0.8.

Lentiviral production

Lentiviruses were produced by co-transfection of 293T cells with lentiviral backbone constructs and packaging vectors (psPAX2 and pMD2.G; Addgene #12260 and #12259) using TransIT-LT1 (Mirus Bio #MR 2306). Supernatant was collected 48- and 72-hours post-transfection, concentrated by ultracentrifugation at 25,000 RPM for 90 minutes and resuspended in an appropriate volume of OptiMEM (Gibco #31985-062).

Lentiviral vectors and sgRNA cloning

pUSEC lentiviral vector and cloning strategy was previously described⁴⁹. For in vitro CRISPR experiments, pUSEPR or pUSPmNG lentiviral vectors were used. For sgRNA cloning, all vectors were digested with BsmBI and ligated with BsmBI-compatible annealed oligos for sgRNAs (Supplementary Sequences). For dual sgRNA vectors, sgRNA dual sequences containing the H1 promoter were ordered as a gBlock gene fragment (Integrated DNA Technologies), BsmBI digested, and cloned as previously described⁷⁵.

DepMap analysis

All data analyzed were accessed from the Broad Institute (www.depmap.org) whole genome CRISPR screens (DepMap 19Q2 Public). Gene co-essentiality datasets were obtained from ref. 27. Ranking DepMap cell-lines by their transcriptional correlation with the NRF2 core target gene set (and the KEAP1-mutant signature) was performed using ssGSEA⁷⁶. Separately, in order to search for targets that correlate well with NRF2 dependence, we calculated the Pearson Correlation coefficient between select genes and every other gene. The genetic background of DepMap cell lines was assessed from mutational data made available by the CCLE. Results of these analyses were used to compare CERES scores between KEAP1/NRF2-mutant lines versus wild-type lines, as well as across the ranked list of cell lines based on their transcriptional correlation with the NRF2 core gene set.

Transcriptome analysis

RNA was collected from cells as before⁴ with RNeasy plus mini kit (Qiagen). For Real Time qPCR analysis, complementary DNA (cDNA) was synthesized from RNA with the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems #4368814). Genes Gclc, Hmox1, Nqo1, Atf4, Hspa5, Xpb1s, Ddit3 were analyzed by quantitative reverse transcription polymerase chain reaction on LightCycler 480 II (Roche). RT qPCR primers (Supplementary Sequences).

Extracellular Flux Measurements

Extracellular flux measurements were calculated by extracting fresh and spent medium supernatant from tracing experiments after 24 hours of growth. Cells were assumed to grow exponentially over the culture period. Metabolites were measured using YSI biochemistry analyzer (Yellow Springs Instruments, Yellow Springs, OH).

Liquid chromatography-mass spectrometry (LC/MS) analysis

Cells were seeded in the morning in DMEM supplemented with 10% dialyzed FBS at 1 mL/well containing 50 uM BSO in 6-well plates (100,000 cells/well). The following day, 2 mL of fresh DMEM was added to each well with or without 50 uM BSO. 48 hours after media replenishment, cell numbers were counted for another set of BSO-treated and untreated samples (3 wells per condition, each well counted twice). Samples in each well were washed twice with 4 mL of ice-cold Blood Bank Saline (Azer Scientific, #ES1244G), then extracted on ice using ice-cold methanol:water (80:20) containing norvaline as an internal standard. Extracts were collected, transferred to 1.5 mL tubes, vortexed at maximum speed for 15 min at 4°C, then centrifuged at 13,000 RPM for 10 min at 4°C. The supernatant was collected, transferred to a new 1.5 mL tube, then evaporated to dryness under nitrogen. Samples were analyzed by LC/MS (Whitehead Metabolite Profiling Core Facility, Cambridge, MA) as described previously⁷⁷.

Glutathione measurements

Reduced (GSH) and oxidized glutathione (GSSG) was measured with a GSH/GSSG-Glo Assay kit (V6611, Promega) for the indicated amount of time.

Statistics

For statistical analyses, we used GraphPad Prism software v.6.03, variance was similar between the groups that were compared: P-values were determined by two-tailed Student’s t-test for all measurements comparing untreated to treated samples of single time points. Mann Whitney U-test with Holm’s multiple comparisons testing was used for tumor burden and grade quantifications. One-way analysis of variance (ANOVA) with Dunnet’s post hoc test were used for comparisons between multiple groups; for analysis between groups over multiple time measurements (growth curves) two-way ANOVA was used with the appropriate multiple comparisons test listed in the figure legends. Figure legends denominate statistical analysis used. All error bars denote standard deviation, unless otherwise noted in the figure legend. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Bioinformatic analysis of CRISPR-targeted loci

For PCR amplicons (sequenced at the MGH sequencing facility), 150-300 bp paired-end reads were used in downstream analyses. The reference sequence of the target locus was supplemented with 10bp genomic flanks and was indexed using an enhanced suffix array⁷⁸. Read ends were anchored in the reference sequence using 10bp terminal segments for a suffix array index lookup to search for exact matches. A sliding window of unit step size and a maximal soft-clip limit of 10bp was used to search for possible anchors at either end of each read. For each read, optimal Smith-Waterman dynamic programming alignment⁷⁹ was performed between the reduced state space of the read sequence and the corresponding reference sequence spanning the maximally distanced anchor locations. Scoring parameters were selected to allow for sensitive detection of short and long insertions and deletions while allowing for up to four mismatches and the highest scoring alignment was selected. Read pairs with both reads aligned in the proper orientation were processed to summarize the number of wild-type reads and the location and size of each insertion and deletion event. Overlapping reads within pairs were both required to support the event if they overlapped across the event location. Additionally, mutation events and wild-type reads were summarized within the extents of the sgRNA sequence and PAM site by considering read alignments that had a minimum of 20bp overlap with this region. Mutation calls were translated to genomic coordinates and subsequently annotated using Annovar⁸⁰. The alignment and post-processing code was implemented in C++ along with library functions from SeqAn⁸¹ and SSW and utility functions in Perl and R (www.R-project.org). Mutation calls were subjected to manual review using the Integrated Genomics Viewer⁸² (IGV).

Gene expression signature analyses

Illumina HiSeq 2000 50-nt single-ended reads were mapped to the UCSC mm9 mouse genome build (http://genome.ucsc.edu/) using RSEM⁸³ Raw estimated expression counts were upper-quartile normalized to a count of 1000⁸⁴. Keap1-mutant (n = 2), WT (n = 2), and WT plus sgSlc33a1 transduced (n = 2) samples were jointly analyzed to derive a murine signature of Slc33a1-mutant gene expression changes. Given the complexity of the database in terms of a mixture of genotypes and treatment, a high-resolution signature discovery approach (Independent Component Analysis) was employed to characterize global gene expression profiles, as described previously^{4,61, 85}. This unsupervised blind source separation technique was used on this discrete count-based expression dataset to elucidate statistically independent and biologically relevant signatures. ICA is a signal processing and multivariate data analysis technique in the category of unsupervised matrix factorization methods. Conceptually, ICA decomposes the overall expression dataset into independent signals (gene expression patterns) that represent distinct signatures. High-ranking positively and negatively correlated genes in each signature represent gene sets that drive the corresponding expression pattern (in either direction). Each signature is thus two-sided, allowing for identification of upregulated and down-regulated genes for each signature within each sample. Formally, utilizing input data consisting of a genes-samples matrix, ICA uses higher order moments to characterize the dataset as a linear combination of statistically independent latent variables. These latent variables represent independent components based on maximizing non-gaussianity, and can be interpreted as independent source signals that have been mixed together to form the dataset under consideration. Each component includes a weight assignment to each gene that quantifies its contribution to that component. Additionally, ICA derives a mixing matrix that describes the contribution of each sample towards the signal embodied in each component. This mixing matrix can be used to select signatures among components with distinct gene expression profiles across the set of samples. The R implementation of the core JADE algorithm⁸⁶ (Joint Approximate Diagonalization of Eigenmatrices) was used along with custom R utilities. Statistical significance of biologically relevant signatures was assessed using the Mann-Whitney-Wilcoxon test (alpha = 0.05). A murine Slc33a1-mutant signature was derived from this analysis, identifying genes with a differential expression pattern between wild-type and all other samples. All RNA-seq analyses were conducted in the R Statistical Programming language (http://www.r-project.org/). Gene set enrichment analysis (GSEA) was carried out using the pre-ranked mode with default settings^87,88. This Slc33a1 knockout signature was compared to publicly available datasets³⁴.