KEAP1/NFE2L2 Mutations Predict Lung Cancer Radiation Resistance That Can Be Targeted by Glutaminase Inhibition

More than 40% of patients with lung cancer are diagnosed with localized disease and are potential candidates for curative treatment (1). For stage III non–small cell lung cancer (NSCLC), concurrent chemoradiotherapy (CRT) with conventionally fractionated radiotherapy (RT) is a mainstay of treatment. Although distant recurrence is the most common pattern of failure, ∼30% of patients experience local recurrence (LR), which is associated with worse overall survival (OS; refs. 2, 3). For patients with stage I–II NSCLC who are not candidates for surgery, stereotactic ablative radiotherapy (SABR) is the preferred treatment and is associated with ∼90% local control (4, 5). For NSCLC treated by CRT or SABR, tumor burden is the only variable that has been repeatedly associated with risk of LR (6–10).

Although somatic mutations play a critical role in personalizing systemic therapy in advanced NSCLC, this is not the case for patients undergoing RT. Recently, mutations in several genes have been suggested to be associated with clinical LR, but these findings remain preliminary and not validated (11–13). Mutations leading to activation of the KEAP1–NFE2L2 pathway are particularly promising candidates for contributing to clinical radioresistance because NFE2L2 is a transcription factor that drives expression of free radical defense genes which could interfere with radiation-induced DNA damage. KEAP1 is an adaptor protein that targets NFE2L2 for ubiquitination and proteasomal destruction under normal homeostasis (14). Mutations in KEAP1 or NFE2L2 occur in approximately 20% of NSCLC (15, 16) and lead to constitutive activation of the pathway. Previous studies have demonstrated that activation of the pathway in vitro (14, 17–19) or in vivo promotes radioresistance (17). Furthermore, our group previously reported an association of KEAP1/NFE2L2 mutations with high risk of LR in a heterogeneous pilot cohort of patients with localized NSCLC treated with RT (17).

In the current study, we sought to identify recurrent mutations in localized NSCLC that are associated with LR after RT. We found that KEAP1/NFE2L2 mutations strongly increase the risk of LR after RT but not surgery. Functional classification of mutations using isogenic cell line systems to distinguish pathogenic and passenger mutations further strengthened the association of mutations with LR. Finally, we demonstrate that glutaminase inhibition is a potential approach for personalized radiosensitization of KEAP1/NFE2L2-mutant tumors.

We identified 232 consecutive patients who were treated at Stanford University with curative intent with RT or surgery and who underwent tumor genotyping using a clinical hybrid capture–based sequencing assay covering 130 to 198 genes (20). Our study included three cohorts: (i) 47 patients with locally advanced NSCLC treated with conventionally fractionated RT (CRT cohort), (ii) 50 patients with early-stage NSCLC treated with high-dose SABR (SABR cohort), and (iii) 135 patients with early-stage NSCLC treated with surgical resection (surgery cohort; Fig. 1A). None of these were included in our previously published cohort of RT-treated patients with KEAP1/NFE2L2 tumor genotyping (17). Differences in baseline patient characteristics were reflective of the types of patients routinely selected for each type of treatment (Supplementary Table S1).

In order to identify tumor mutations associated with clinical radioresistance, we determined the association between recurrent mutations and LR in the combined cohort of patients treated with CRT or SABR. LR, defined as tumor regrowth within the radiation field, was chosen as the endpoint of interest because it is most likely to reflect responses of tumor cells exposed to RT. Distant recurrence outside of the prior radiation field could be due to micrometastases that were present at the time of treatment and that were therefore not exposed to RT. Ten mutations met our predetermined recurrence frequency threshold of >5% (Fig. 1B; see Methods for power calculation). Mutations in KEAP1 and NFE2L2 were considered as one group because they result in the same biochemical phenotype (i.e., NFE2L2 overexpression) and display mutual exclusivity (14, 17). Strikingly, only KEAP1/NFE2L2 mutations were significantly associated with LR (Fig. 1C; adjusted P = 0.005), and these mutations were present in nearly half of tumors that had LR (Fig. 1D). Mutations in KEAP1 were distributed throughout the protein, whereas NFE2L2 mutations were located in the DLG and ETGE hotspots (Fig. 1E). Furthermore, in exploratory analyses, we found that KEAP1 mutations were significantly associated with LR on their own (P = 0.002) and that neither mutations leading to ERK activation (EGFR/KRAS/BRAF; P = 0.49) nor E2F activation (CDKN2A/RB1; P = 0.15) were associated with LR when considered together. The frequency of comutations in driver genes was similar in KEAP1/NFE2L2-mutant tumors with and without LR (5 of 7 with LR vs. 4 of 10 without LR; P = 0.33). We also did not observe a significantly different frequency of co-occurring TP53 mutations in KEAP1/NFE2L2-mutant tumors with (3 of 7) versus without LR (8 of 10; P = 0.16; Fig. 1B). Thus, among recurrent mutations in NSCLC, KEAP1/NFE2L2 mutations appear to be the dominant cause of clinical radioresistance.

Next, we compared LR based on KEAP1/NFE2L2 mutation status in patients with locally advanced NSCLC treated with conventionally fractionated CRT. For the entire CRT cohort, 2-year OS was 65.6% [95% confidence interval (CI), 46.9%–79.1%] and 2-year LR was 22.8% (95% CI, 20.8%–24.8%; Supplementary Fig. S1A and S1B), similar to the rates reported for the standard RT dose arm in RTOG 0617 (21). Patients with KEAP1/NFE2L2-mutant tumors had a significantly increased incidence of LR compared with those without, with 2-year LR of 50.0% (95% CI, 36.3%–63.7%) versus 16.9% (95% CI, 14.6%–19.2%), respectively (P = 0.01; Fig. 1F). There were no significant differences in clinical characteristics between KEAP1/NFE2L2^MUT and wild-type (WT) cases (Supplementary Table S2). Metabolic tumor volume (MTV) measured using FDG PET/CT scans and stage did not differ significantly between patients with and without KEAP1/NFE2L2 mutations (P = 0.45 and P = 0.67, respectively). Furthermore, KEAP1/NFE2L2 mutation status was the only predictor of LR in both univariable analysis (UVA; HR = 4.74; 95% CI, 1.30–17.20; P = 0.02) and multivariable analysis (MVA; HR = 5.17; 95% CI, 1.30–20.58; P = 0.02; Supplementary Table S3). We did not observe a significant association between KEAP1/NFE2L2 mutations and distant recurrence as first recurrence after CRT (P = 0.69; Supplementary Table S4 and Supplementary Fig. S1C), suggesting that patients with locally advanced NSCLC and KEAP1/NFE2L2 mutations are specifically at high risk of LR.

Patients with stage I–II NSCLC treated with hypofractionated SABR receive significantly higher radiation doses than patients treated with conventionally fractionated CRT. We therefore assessed whether the “ablative” doses delivered during SABR might overcome KEAP1/NFE2L2 mutation–associated radioresistance. Patients in the SABR cohort had 2-year OS of 82.8% (95% CI, 68.5%–91.0%) and 2-year LR of 12.4% (95% CI, 10.9%–13.9%; Supplementary Fig. S2A and S2B). Surprisingly, as in the CRT cohort, we observed a significantly higher incidence of LR after high-dose SABR for patients with KEAP1/NFE2L2 mutations versus those without, with 2-year LR of 35.2% (95% CI, 23.5%–46.9%) versus 7.0% (95% CI, 5.6%–8.4%), respectively (Fig. 1G; P = 0.009). There were no significant differences in clinical characteristics between KEAP1/NFE2L2^MUT and wild-type cases (Supplementary Table S2). MTV and stage were not significantly different between patients with and without KEAP1/NFE2L2 mutations (P = 0.88 and P = 0.71, respectively). On both UVA and MVA, KEAP1/NFE2L2 mutations (HR = 8.50; 95% CI, 1.56–46.30; P = 0.01 and HR = 17.92; 95% CI, 2.05–156.67;P = 0.009, respectively) and MTV (HR = 1.42 per 10 cc; 95% CI, 1.25–1.62; P = 1.2e−7 and HR = 1.68; 95% CI, 1.27–2.22; P = 0.0003, respectively) were significantly associated with LR (Supplementary Table S3). We did not observe a significant association between KEAP1/NFE2L2 mutations and distant recurrence as first recurrence after SABR (P = 0.17; Supplementary Table S5 and Supplementary Fig. S2C). Thus, KEAP1/NFE2L2 mutations are predictive of LR after SABR.

Lastly, we hypothesized that KEAP1/NFE2L2 mutations are not associated with LR after surgery, because the expression of free radical defense genes induced by these mutations would not be expected to affect tumor resectability. In our 135-patient surgery cohort, LR at the staple line or bronchial stump was a rare occurrence (1- and 2-year rates of 0.9 and 2.3%, respectively) and, consistent with our hypothesis, did not differ based on KEAP1/NFE2L2 mutation status (Fig. 1H). There were no significant differences in patient age and tumor size between KEAP1/NFE2L2^MUT and wild-type cases, but mutant cases were more likely to have non-adenocarcinoma histology and a more significant smoking history (Supplementary Table S2). Our findings are consistent with prior reports showing that LR is not a primary pattern of failure after surgery (22). Additionally, we validated there is no difference in prognosis by mutation status within patients with stage I–II lung adenocarcinoma and squamous cell carcinoma treated with primary surgery in The Cancer Genome Atlas (TCGA; P = 0.71; Supplementary Fig. S3). Taken together, our results indicate that KEAP1/NFE2L2 mutations are predictive biomarkers for LR after RT but not surgery.

Although our clinical results indicate that KEAP1/NFE2L2 mutations are strongly associated with LR after RT, 58.8% (10 of 17) of patients with these mutations did not develop tumor regrowth within the radiation field. We therefore hypothesized that a subset of the mutations were passengers that did not affect protein function. To test this hypothesis, we developed isogenic KEAP1 wild-type and knockout cell lines in which we could express open reading frame (ORF) constructs containing the mutations we observed in the two RT cohorts and functionally evaluated these for KEAP1/NFE2L2 pathway activity (Fig. 2A). Specifically, we targeted KEAP1 in H1299 human NSCLC cells using CRISPR/Cas9-mediated gene editing by directly introducing the single guide RNA (sgRNA)–Cas9 complex (Supplementary Fig. S4A and S4B; ref. 23). KEAP1^NULL H1299 cells displayed significant overexpression of NFE2L2 and its targets, including NQO1 and SQSTM1 (also known as p62), compared with parental KEAP1^WT H1299 cells (Fig. 2B).Additionally, these cells displayed significant resistance to ionizing radiation (Fig. 2C).

To explore the mechanism of radiation resistance induced by KEAP1 loss, we next examined levels of DNA damage induced by ionizing radiation in the isogenic lines. Although baseline levels of DNA double-strand breaks as measured by γH2AX foci were similar, there were significantly fewer foci in KEAP1^NULL compared with KEAP1^WT H1299 cells following 2 Gy of ionizing radiation (Fig. 2D). These findings suggest that KEAP1^NULL cells are radioresistant due to elevated levels of free radical scavengers and decreased DNA damage production by the indirect effect of ionizing radiation. To test this hypothesis, we examined levels of glutathione (GSH) and found that KEAP1^NULL H1299 cells had significantly higher levels of reduced GSH than KEAP1^WT cells (Supplementary Fig. S4C). Next, we examined the effects of the free radical scavenger N-acetylcysteine (NAC) and the GSH biosynthesis inhibitor buthionine sulfoximine (BSO) on radiation sensitivity. Although treatment with NAC had no effect in unirradiated cells, it significantly enhanced survival of KEAP1^WT but not KEAP1^NULL H1299 cells after 4 Gy, suggesting that KEAP1^NULL H1299 cells are already maximally protected by endogenous free radical scavengers (Fig. 2E). Conversely, BSO treatment had no effect on radiosensitivity of KEAP1^WT cells but radiosensitized KEAP1^NULL cells to levels similar to those of KEAP1^WT cells. These data suggest that elevated levels of GSH are a major driver of the radioresistant phenotype of KEAP1^NULL cells.

In order to evaluate the functional effects of the KEAP1 mutations we observed in the RT patients, we individually expressed ORFs carrying the 15 KEAP1 mutations in KEAP1^NULL H1299 cells and tested multiple aspects of KEAP1/NFE2L2 pathway activity. We reasoned that pathogenic (i.e., inactivating) mutations should maintain KEAP1/NFE2L2 pathway activity, whereas passenger mutations should act like wild-type KEAP1 and decrease it. We first confirmed that expression levels of both mutant and wild-type KEAP1 constructs were similar (Supplementary Fig. S5A). Next, we analyzed expression of the NFE2L2 target gene HMOX1 by qRT-PCR and observed that all 6 constructs containing KEAP1 mutations from LR cases did not alter expression, consistent with a loss-of-function phenotype (Fig. 2F). Conversely, 6 of 9 (67%) constructs containing mutations from non-LR cases decreased expression similarly to wild-type KEAP1, suggesting that these mutations do not alter KEAP1 function. Similar results were observed when examining expression of HMOX1 and SQSTM1 at the protein level (Supplementary Fig. S5B–S5F). Furthermore, we performed in vitro clonogenic survival assays using 5 Gy of ionizing radiation and observed that all 6 constructs containing KEAP1 mutations from LR cases did not resensitize KEAP1^NULL H1299 cells, whereas 6 of 9 (67%) constructs containing mutations from non-LR cases did (Fig. 2G). Finally, we observed the same pattern when examining resistance to hydrogen peroxide (Supplementary Fig. S5G).

In an analogous fashion, we developed isogenic NFE2L2 wild-type and knockout cell lines. NFE2L2^NULL H1299 cells displayed loss of NFE2L2 protein, reduced target gene expression, and greater sensitivity to ionizing radiation compared with parental NFE2L2^WT H1299 cells (Supplementary Fig. S6A and S6B). We expressed ORFs carrying each of the three NFE2L2 mutations from our RT cohorts in NFE2L2^NULL H1299 cells (Supplementary Fig. S6C and S6D). To remove any native NFE2L2, we generated a KEAP1 overexpression cell model using the NFE2L2^NULL H1299 cells transfected with multiple copies of wild-type KEAP1 (Supplementary Fig. S6D). Although wild-type NFE2L2 did not significantly increase NFE2L2 target gene expression, each of the three mutant constructs did, suggesting KEAP1-mediated degradation of wild-type but not mutant NFE2L2 (Fig. 2G). Analysis of protein expression of NFE2L2 and targets HMOX1 and GCLM confirmed stabilization of the mutant but not wild-type NFE2L2 constructs (Supplementary Fig. S6D). Finally, all 3 mutant but not wild-type NFE2L2 constructs induced increased protein expression of HMOX1 (Fig. 2H) and resistance to ionizing radiation and hydrogen peroxide in NFE2L2^NULL H1299 cells (Fig. 2I; Supplementary Fig. S6E). Figure 2J provides a summary of functional assay results and mutation classification for each KEAP1 and NFE2L2 mutation found in the CRT and SABR cohorts.

Based on these molecular analyses, we classified the KEAP1 and NFE2L2 mutations observed in our RT cohorts as either pathogenic (i.e., loss-of-function for KEAP1 or gain-of-function for NFE2L2) or passenger (i.e., neutral). For KEAP1, 9 of 15 (60%) mutations were pathogenic, whereas 6 (40%) were passengers, and for NFE2L2, all 3 mutations were pathogenic (Fig. 2I). Notably, we observed LR only in patients with pathogenic KEAP1 or NFE2L2 mutations and in none of the patients with passenger mutations (Fig. 3A). Additionally, we found that patients with pathogenic mutations who did not develop LR had significantly smaller tumors (median 2.7 cc) than those who did develop LR (median 64.1 cc; P = 0.03; Fig. 3B). Among patients with pathogenic mutations, 28.6% (2 of 7) with tumors <20.4 cc experienced an LR, whereas 100% (5 of 5) with tumors ≥20.4 cc experienced an LR. These findings suggest that doses of RT delivered during CRT or SABR can control some tumors carrying pathogenic mutations if tumors are small and fewer clonogens are present.

Reanalyzing LR rates in CRT- and SABR-treated patients based on the presence of pathogenic mutations increased separation between mutant and wild-type patients (Supplementary Fig. S7A and S7B). Strikingly, rates of LR did not differ significantly between patients treated with CRT or SABR when mutation functional classification was considered (Fig. 3C; P = 0.54). Patients with pathogenic KEAP1/NFE2L2 mutations who were treated with SABR had a significantly worse OS than nonmutant patients (Fig. 3D; P = 0.002). Similarly, the presence of KEAP1/NFE2L2 mutations was associated with inferior survival in patients with stage I–IIA NSCLC from TCGA who were treated with radiation and not surgery (Fig. 3E; P = 0.01; ref. 24). Thus, functional classification of KEAP1/NFE2L2 mutations improves their association with LR, and the lack of local control associated with pathogenic mutations appears to lead to higher rates of death.

Because prior studies have suggested an association of the expression of NFE2L2 target genes and resistance to radiation (25), we next explored whether NFE2L2 target gene expression is also associated with LR after RT. To do so we performed RNA sequencing (RNA-seq) on formalin-fixed, paraffin-embedded biopsy samples from 41 (42.3%) patients in the SABR and CRT cohorts who had sufficient tissue remaining after DNA sequencing. In order to focus on gene expression of tumor cells, we used CIBERSORTx digital cytometry to infer tumor cell–specific gene-expression profiles and scored samples for the presence of a previously defined NFE2L2 target gene signature (Fig. 3F; refs. 26, 27). We observed significantly higher expression of NFE2L2 target genes in tumors harboring pathogenic KEAP1/NFE2L2 versus wild-type/passenger mutations (P = 0.01; Fig. 3G). Expression of NFE2L2 targets was not significantly associated with LR by Cox regression (HR = 2.33; 95% CI, 0.06–84.30; P = 0.64) and did not stratify risk of LR in competing risk analysis (P = 0.93; Fig. 3H). Similarly, there was no difference in OS in the TCGA radiation cohort (P = 0.48; Fig. 3I). Taken together, these results suggest that although KEAP1/NFE2L2 mutation status is predictive of clinical response to radiation, expression of NFE2L2 target genes is not.

Finally, we wished to identify a potential approach for radiosensitizing KEAP1/NFE2L2-mutant tumors. Recent work has demonstrated that KEAP1 loss promotes dependence on glutamine metabolism and sensitivity to glutaminase inhibition (28, 29). Because glutamine metabolism includes production of GSH, a critical antioxidant that has been linked to resistance to ionizing radiation, we hypothesized that glutaminase inhibition can preferentially radiosensitize KEAP1/NFE2L2-mutant cells (Fig. 4A). To test this hypothesis, we first examined expression of the glutamine transporter alanine–serine–cysteine transporter 2 (ASCT2) in our isogenic cell lines and observed increased expression in KEAP1^NULL and decreased expression in NFE2L2^NULL H1299 cells (Fig. 4B and Supplementary Fig. S8A). Similarly, siRNA-mediated knockdown of NFE2L2 in H1975 NSCLC cells (KEAP1/NFE2L2 wild-type) resulted in decreased ASCT2 expression. Furthermore, RNA-seq analysis of both KEAP1^NULL compared with KEAP1^WT H1299 cells and tumor biopsy specimens from patients with pathogenic KEAP1/NFE2L2 mutations (n = 5) compared with patients without (n = 36) revealed significant overexpression of genes involved in glutamine metabolism (Fig. 4C and D). Thus, both KEAP1-mutant human tumors and our KEAP1^NULL cell lines display upregulation of genes involved in glutamine metabolism.

To explore if targeting of glutamine metabolism can preferentially radiosensitize KEAP1/NFE2L2-mutant cells, we tested the combination of RT and the small-molecule glutaminase inhibitor CB-839, which is currently in phase I/II clinical trials, in NSCLC cell lines of differing genetic backgrounds (Fig. 4E). Although KEAP1^NULL and KEAP1^WT H1299, H1437, and A549 cells displayed minimal sensitivity to CB-839 doses in the range of 10 to 100 nmol/L, the combination of ionizing radiation and CB-839 preferentially killed KEAP1^NULL cells (Fig. 4F–I). The sensitization effect was substantial, with a dose-modifying factor for CB-839 in H1299 KEAP1^NULL cells of 3.4 (at 63% survival; Fig. 4F). Importantly, we observed this effect starting with parental cell lines that were either KEAP1 wild-type (H1299 and H1437) or mutant (A549). For the latter, treating A549 cells with 100 nmol/L CB-839 did not have a dramatic effect on cell survival but sensitized cells to ionizing radiation (Fig. 4I and Supplementary Fig. S8B and S8C). However, expression of wild-type KEAP1 in A549 cells increased radiation sensitivity and abolished the sensitizing effect of CB-839. Thus, CB-839 preferentially radiosensitizes KEAP1^NULL cells.

Lastly, we investigated the mechanism by which CB-839 preferentially radiosensitizes KEAP1^NULL cells and hypothesized that glutaminase inhibition counteracts the enhanced free radical defenses caused by loss of KEAP1. We first examined the effects of CB-839 and ionizing radiation on intracellular reactive oxygen species (ROS) levels. KEAP1^NULL H1299 cells displayed significantly lower baseline levels of ROS than their wild-type counterparts. However, unlike in KEAP1^WT H1299 cells, the combination of CB-839 and radiation significantly increased ROS levels in KEAP1^NULL H1299 cells compared with radiation alone (Fig. 4J). Additionally, CB-839 treatment decreased GSH to glutathione disulfide (GSSG) ratios significantly more in KEAP1^NULL than in KEAP1^WT H1299 cells (Fig. 4K). This suggests that CB-839 leads to decreased free radical scavenging capacity in KEAP1^NULL cells. We next tested if exogenous addition of a free radical scavenger can rescue KEAP1^NULL cells from CB-839–mediated radiosensitization. Treatment of H1299 KEAP1^NULL cells with the ROS scavenger NAC did not significantly affect cell survival by itself or when combined with either 2 Gy or CB-839 alone. However, NAC significantly rescued the enhanced cell killing caused by the combination of 2 Gy and CB-839 (Fig. 4L), suggesting that CB-839 radiosensitization is mediated by depletion of free radical scavengers. Finally, we reasoned that if this mechanism is correct, CB-839 should increase the amount of DNA damage induced by ionizing radiation. To test this, we measured DNA double-strand breaks via γH2AX before and after ionizing radiation in the presence or absence of CB-839. Irradiation with 2 Gy induced approximately 2-fold higher levels of γH2AX in KEAP1^WT compared with KEAP1^NULL cells (Fig. 4M). Strikingly, 24-hour pretreatment with CB-839 resulted in elevation of γH2AX in KEAP1^NULL but not KEAP1^WT cells. These data indicate that CB-839 preferentially radiosensitizes KEAP1^NULL cells by decreasing free radical scavenging capacity and therefore increasing the amount of DNA damage caused by the indirect effect of ionizing radiation.

In summary, in this study we demonstrate that KEAP1/NFE2L2 mutations are a common cause of LR in patients with localized NSCLC treated with RT but not surgery. These mutations are therefore clinically relevant predictive biomarkers of radioresistance. Using functional analyses including radiation clonogenic assays, we found that 60% of KEAP1 mutations and all NFE2L2 mutations in our cohort were pathogenic, and that pathogenic mutations were associated with a ∼60% rate of LR after RT but not surgery. Additionally, we show that the glutaminase inhibitor CB-839 preferentially sensitizes KEAP1-mutant NSCLC cells to RT.

Of the recurrent mutations found in our cohort, only KEAP1/NFE2L2 mutations were statistically significantly associated with LR. This finding both validates and extends our prior analysis of a smaller, independent cohort in which we examined KEAP1/NFE2L2 mutations and found that these were associated with LR after RT (17). Overall, nearly half of all LRs occurred in tumors with these mutations, suggesting that KEAP1/NFE2L2 mutations are a dominant biological driver of clinical radioresistance in localized NSCLC. Comutations in other frequently mutated genes such as lung cancer drivers were not associated with LR. Additionally, unlike other prior studies, we did not observe an association of KRAS and/or TP53 mutations with LR (12, 13). One potential explanation for this discrepancy is that KEAP1 mutations often co-occur with both KRAS and TP53 mutations, and KEAP1/NFE2L2 genotyping was not performed in the prior studies. In an exploratory analysis, we also examined if STK11 mutations were associated with LR after RT because these have been implicated in treatment resistance to other therapies (28, 30, 31). However, only one of five STK11-mutant tumors in our RT cohorts developed LR, and this was not statistically significantly different from wild-type tumors (Supplementary Fig. S9; P = ns). Of note, this tumor also carried a pathogenic KEAP1 mutation, suggesting that it was the likely cause of LR. However, we cannot rule out that there may be additional mutations beyond KEAP1/NFE2L2 that are associated with radioresistance and that our study was underpowered to detect. That said, our findings suggest that discovering such mutations will require very large cohorts and that other mutations are unlikely to match KEAP1/NFE2L2 mutations in recurrence frequency and/or effect size.

We were surprised to find high rates of LR in patients with KEAP1/NFE2L2 mutations who were treated with SABR, which delivers significantly higher doses of ionizing radiation to tumors than conventionally fractionated RT. We found no statistically significant difference in the 2-year LR rate after CRT (67%) and SABR (56%), suggesting dose escalation alone is not sufficient to significantly improve local control in these patients. However, it is possible that our study was underpowered to detect a modest benefit of dose escalation. Indeed, the fact that smaller tumors with KEAP1/NFE2L2 mutations had better local control suggests dose escalation may be useful in some settings. Notably, our results and prior studies have shown that increasing rates of LR in patients with stage I–II NSCLC are associated with worse OS, highlighting the need to achieve better local tumor control in this setting (2, 32).

By expressing mutant alleles in knockout cell lines, we found that ∼40% of KEAP1 mutations in our RT cohorts have neutral effects on KEAP1 function and are therefore likely passenger mutations. This is somewhat higher than in a prior study in which 4/18 (22%) KEAP1 mutations found in a lung squamous cell carcinoma cohort were determined to have neutral effects on KEAP1 activity (33). We compared four functional assays for KEAP1/NFE2L2 pathway activity, including NFE2L2 target mRNA expression, NFE2L2 target protein expression, resistance to hydrogen peroxide, and resistance to ionizing radiation. Variants demonstrated similar patterns in all four assays and behaved either as loss of function or neutral alleles. Notably, two of the mutations (R204P and R320Q) found in our SABR cohort had also been included in prior studies and were classified as hypomorphic variants (33, 34). One of these (R204P) behaved like a loss-of-function allele in our assays and was found in a patient who developed LR, while the other (R320Q) behaved like a passenger mutation and was found in a patient who did not have LR. Thus, functional classification of KEAP1 variants may be context dependent, and the knockout cell line system we established faithfully recapitulates radiation resistance phenotypes observed clinically.

Because pathogenic KEAP1/NFE2L2 mutations result in overexpression of NFE2L2 target genes, we tested whether analysis of gene expression could also identify tumors at highest risk of LR. Interestingly, we observed that whereas tumors with KEAP1/NFE2L2 mutations did tend to overexpress NFE2L2 target genes, a significant subset of wild-type tumors also displayed overexpression. Unlike genotyping for KEAP1/NFE2L2 mutations, high expression of NFE2L2 target genes did not stratify LR or OS. One potential explanation for this observation is that stress induced by ischemia and/or tissue handling might affect expression of NFE2L2 target genes, as has been demonstrated for NQO1 (35). Separately, NFE2L1 (also called NRF1) shares many downstream target genes with NFE2L2, and it is possible that KEAP1/NFE2L2 wild-type tumors with high expression of NFE2L2 target genes may have enhanced NFE2L1 activity (36). Although more work is needed in this area, our results suggest that tumor genotyping for KEAP1/NFE2L2 mutations is the most reliable method for identifying radioresistant NSCLCs.

A key implication of our findings is the need to develop strategies for overcoming KEAP1/NFE2L2-mediated radioresistance. Our analysis of DNA double-strand breaks after ionizing radiation suggests that KEAP1 loss leads to radioprotection by decreasing DNA damage via enhanced free radical scavenging. Consistent with this hypothesis, KEAP1^NULL cells display lower ROS levels and higher levels of GSH, a key intracellular antioxidant that has previously been implicated in contributing to radioresistance (29, 37). Furthermore, depletion of GSH via BSO radiosensitized KEAP1^NULL cells to levels similar to those in KEAP1^WT cells, suggesting that the mechanism of radioresistance is largely mediated by increased GSH production. We therefore tested the hypothesis that glutaminase inhibition, which decreases GSH production, preferentially radiosensitizes KEAP1-mutant cells. Our interest in glutaminase was in part based on prior observations that KEAP1-mutant NSCLC is dependent on glutamine metabolism (28, 29) and that inhibition of glutamine metabolism may be of therapeutic benefit in NSCLC (38–42). Strikingly, we found that low doses of the glutaminase inhibitor CB-839 (0.1–500 nmol/L), which is currently in clinical trials in other contexts, can preferentially radiosensitize KEAP1-mutant but not isogenic wild-type cells in three different NSCLC cell line models. CB-839 sensitized KEAP1^NULL cells to similar levels as KEAP1^WT cells, suggesting that this approach could improve local control of KEAP1-mutant tumors to similar levels as that of wild-type tumors. Additionally, we expect that the ∼2-fold increase in cell killing we observed in KEAP1^NULL cells treated with CB-839 after a single dose of RT would be compounded over a multifraction course, leading to potentially dramatic increases in total cell killing.

Mechanistic analyses demonstrated that CB-839 pretreatment leads to depletion of free radical scavengers and increased DNA damage via the indirect effect of radiation. Although our analyses were focused on in vitro experiments, a recent study by another group demonstrated that CB-839 can radiosensitize KEAP1-mutant cells in vivo. Specifically, Boysen and colleagues treated H460 NSCLC xenografts grown in nude mice with CB-839, 12 to 18 Gy of radiation, or both (43). The combination of CB-839 plus radiation significantly delayed tumor growth compared with radiation alone. Although not mentioned by Boysen and colleagues, H460 cells carry the pathogenic KEAP1 D236H mutation (44). Thus, CB-839 radiosensitization appears to be a promising approach for overcoming KEAP1/NFE2L2 mutation-mediated radioresistance.

We have now validated the association of KEAP1/NFE2L2 mutations with inferior outcomes after RT in four independent cohorts of patients with NSCLC (one from our prior report and three in the current report; ref. 17). Our findings therefore have potential implications for the clinical management of patients with KEAP1/NFE2L2-mutant NSCLC. For medically operable patients with node-negative disease who are candidates for surgery or SABR, surgery may be preferable because it does not appear to be associated with increased rates of LR. For medically operable patients with lymph node–positive NSCLC, the presence of KEAP1/NFE2L2 mutations is more complicated because these patients usually receive RT and/or surgery plus chemotherapy, and these mutations have also been linked to chemoresistance (45, 46). Such patients may therefore be ideal candidates for trimodality therapy consisting of surgery, RT, and chemotherapy, because surgical debulking should increase the chance that RT and chemotherapy can eliminate any clonogens remaining after surgery. Our finding that local control of KEAP1/NFE2L2-mutant tumors appears to be better in smaller tumors with fewer clonogens supports this idea. For medically inoperable early-stage patients, a radiosensitization strategy such as treatment with CB-839 might be beneficial. Lastly, based on several randomized phase II trials demonstrating survival benefits, use of SABR in patients with NSCLC with oligometastatic disease appears to be a promising approach for improving survival (47–49). Our findings suggest that patients with oligometastatic NSCLC with KEAP1/NFE2L2-mutant tumors will likely have inferior local control if treated with SABR alone and therefore might benefit from radiosensitization with CB-839. Because RT is frequently delivered with systemic therapies, future studies will also need to ensure that CB-839 does not increase toxicity in combination therapy settings.

Limitations of our study include its retrospective nature and that all patients were treated at a single institution. Thus, although we have validated the association of KEAP1/NFE2L2 mutations with LR after RT in multiple cohorts, it would be useful to repeat these analyses in cohorts from other institutions in the future. Additionally, LR was relatively uncommon and thus forced us to limit our analyses to mutations with sufficient recurrence frequency. Furthermore, because our cohort was genotyped using a targeted assay, it is possible that mutations in genomic regions that were not covered might also be associated with LR. Lastly, there were significant differences in patient characteristics between the SABR and surgery cohorts that reflect the types of patients who receive each therapy and that could have confounded the comparison of the two groups. However, the fact that tumor size/stage, which to our knowledge is the only consistently reported parameter associated with LR after either treatment, did not differ between the two cohorts decreases this risk (9, 50).

In conclusion, in a large cohort of patients with localized NSCLC treated with definitive radiation or surgery, we have demonstrated that KEAP1/NFE2L2 mutations are strongly predictive of LR after RT but not surgery. Furthermore, functional classification of KEAP1/NFE2L2 mutations revealed that only pathogenic mutations that lead to radioresistance in vitro were associated with LR clinically. Finally, glutaminase inhibition may offer a strategy for a precision radiation oncology approach for radiosensitizing KEAP1/NFE2L2-mutant tumors.

Using prospective registries of patients with NSCLC treated with RT (2009–2018) or surgery (2015–2018) at Stanford University School of Medicine, we performed a retrospective study with institutional review board approval of all consecutive patients with American Joint Committee on Cancer (AJCC) 8th edition stage IA1–IIIC NSCLC who had clinical next-generation sequencing (NGS) performed on tumor tissue with baseline characteristics shown in Supplementary Table S1. We excluded patients who were included in our previously published cohort of RT-treated patients with KEAP1/NFE2L2 tumor genotyping in order to ensure analysis of a completely independent cohort (17).

Our primary goal was to evaluate the association between somatic tumor mutations and LR in patients with localized NSCLC receiving CRT or SABR. Given that these cohorts totaled 97 patients, we calculated that in order to achieve 80% power to detect an absolute difference in LR rate of ≥15% at 2 years (alpha = 0.05), a gene would need to be mutated in ≥6 patients. We therefore only considered genes above this recurrence threshold.

The CRT cohort (Supplementary Table S4) consisted of patients with stage IIB–IIIC NSCLC, and the SABR cohort (Supplementary Table S5) consisted of patients with stage IA1–IIB NSCLC. We have previously reported our RT treatment protocols (10, 51). Follow-up post-RT consisted of CT and/or PET-CT imaging at 3-month intervals during the first 2 years, 6-month intervals during the next 2 years, and yearly thereafter.

Patients in the surgery cohort had stage IA1–IIB NSCLC and underwent surgery without neoadjuvant or adjuvant chemotherapy. Patients underwent wedge resection, segmentectomy, or lobectomy with or without selective ipsilateral hilar and mediastinal nodal dissections. Follow-up visits occurred at 6-month intervals, with CT performed at each visit.

LR was defined as tumor regrowth within the RT planning target volume (PTV) for the RT cohorts and recurrence at the staple line or bronchial stump for the surgery cohort. LR was scored using pathologic confirmation or by radiologic changes consistent with tumor regrowth. Radiologic criteria for LR were: (i) interval increase in size of a mass-like lesion on CT and/or (ii) interval increase in FDG uptake in a focal pattern on PET. Whenever possible, we confirmed radiologic findings on serial imaging in the context of overall disease progression. Radiologic scoring of LR was done by a single investigator (M.S. Binkley) blinded to the tumor-genotyping results.

We used the “PET Edge” tool in MIMvista software, a gradient-based algorithm, to delineate the MTV of individually distinct lesions on attenuation-corrected PET scans as previously described (10, 52). For lesion-specific MTV analyses in CRT patients, the MTV of the lesion that recurred was used in patients with LR and the MTV of the largest targeted lesion was used in patients who did not develop LR.

Tumor genotyping was performed in the Stanford Molecular Pathology CLIA laboratory on formalin-fixed, paraffin-embedded tumor tissue via a laboratory-developed test NGS assay called the “STanford Actionable Mutation Panel” (STAMP), as previously described (20). Two versions of the assay were used during the years of the study covering 130 or 198 genes (both including all exons of KEAP1, NFE2L2, TP53, and KRAS). Mutation calls were extracted from clinical genotyping reports. Lung cancer driver genes were defined as EGFR, KRAS, NRAS, PIK3CA, HER2, BRAF, MET, ALK, ROS1, RET, and NTRK1/2/3. Copy-number alteration analysis was performed as previously described considering both on-target and off-target reads (53). Activating mutations in KRAS (codons 12 and 13) and EGFR (exons 18–21; ref. 54) were considered pathogenic. All other mutations with Combined Annotation Dependent Deletion (CADD) PHRED scores ≥20 and with a population frequency of <1% in the genome aggregation database (Broad Institute, v3; ref. 55) were considered pathogenic. Mutations in NFE2L2 were considered pathogenic if they were located in the Neh2 domain (amino acids 16–86) that interacts with KEAP1 and had CADD PHRED scores ≥20. Final tumor genotyping used for analysis is summarized in Supplementary Table S6.

Isolation of RNA from formalin-fixed, paraffin-embedded tumor tissue was performed for cases with sufficient tissue remaining using the RNAstorm kits (CELLDATA) with DNAse digestion (Qiagen) and subsequent purification with RNA Clean and Concentrator kits (Zymo Research). Sequencing libraries were prepared using the SMARTer Stranded total RNA-seq Kit v2 (Takara Bio USA, Inc.) and sequenced on the Illumina HiSeq4000 platform with paired-end 150-bp reads and 14 to 16 samples per lane. Gene-expression levels were quantified using Salmon (v0.8.2) under quasi-mapping mode to Gencode version 27 (56). Transcript per million (TPM; Supplementary Table S7) values were used as input for CIBERSORTx digital cytometry, allowing for inference of tumor cell–specific gene-expression profiles using a reference transcriptome of flow-sorted malignant, endothelial, immune, and fibroblast cell expression profiles (26, 57). Differential gene-expression analysis and normalization were conducted with the package “DESeq2.” Single and bulk sample gene set enrichment and gene set variation analyses were performed with GSVA.

H1299 and A549 cells were purchased from the ATCC. The H1299 cell line was cultured in RPMI-1640 supplemented with 10% fetal bovine serum (FBS), and A549 cells were maintained in DMEM with 10% FBS. All cell lines were tested for Mycoplasma and were negative using PlasmoTest (Invivogen). KEAP1 or NFE2L2 plasmids were purchased from Addgene (#87545 for KEAP1 and #21555 for NFE2L2), and the KEAP1 or NFE2L2 cDNA was subcloned into the pcDNA4-V5-His vector (Invitrogene, #V861-20). Site-directed mutagenesis was performed to generate KEAP1 and NFE2L2 alleles bearing mutations identified in patients from the RT cohorts using QuickChange II site-Directed Mutagenesis Kit (Agilent, #200524) and confirmed with Sanger sequencing. Transient expression of these constructs as well as siRNAs was performed using Lipofectamine 3000 as per the manufacturer's protocol (Invitrogen). Cell viability was measured using the CellTiter-Glo Luminescent Cell Viability Assay kit (Promega, #G7572) or clonogenic assays. For the latter, cells were cultured for 1 to 2 weeks, and colonies greater than 50 densely packed cells were counted using crystal violet staining. Quantitation was performed using ImageJ software as previously described (58). The survival fraction was calculated based on the plating efficiency, number of colonies, and number of seeded cells [SF = colonies counted/(cells seeded × plating efficiency)]. For experiments involving mutant alleles, cells were transfected with plasmids ≥72 hours prior to treatment with ionizing radiation and/or CB-839.

Anti-ASCT2 (#8057), anti-GAPDH (#5174), anti-HMOX1 (#5061), anti-KEAP1 (#8047, used to detect endogenous KEAP1 protein), anti-NQO1 (#3187), anti-P62 (#88588), anti–phospho-Histone H2AX (#9718), anti-ubiquitin (#3936), and anti-Vinculin (#13901) antibodies were purchased from Cell Signaling Technology and used at 1:1,000 dilution for Western blot analysis and 1:100 dilution for immunoprecipitation and immunofluorescence analysis. Polyclonal anti-KEAP1 antibody (#10503), used to detect KEAP1-mutant proteins because some lacked binding site for the monoclonal anti-KEAP1 antibody, was purchased from Protein-Tech and used at 1:1,000 dilution for Western blot analysis. Anti-V5 antibody (#MA5-15253) was purchased from Invitrogen and used at 1:5,000 dilution for Western blot analysis. Anti-flag M2 antibody (F3165) was purchased from Sigma-Aldrich and used at 1:100 dilution for immunoprecipitation analysis. All antibodies were monoclonal except anti-HMOX1 and anti-KEAP1 (#10503).

The sgRNAs for KEAP1 or NFE2L2 gene were designed using the Genetic Perturbation Portal (Broad Institute). To generate knockout cell lines, sgRNAs and Cas9 protein (Integrated DNA Technologies, Inc., # 1081058) complexes were nucleofected into H1299 cell. After 48 hours, single cells were sorted into 96-well plates. After 2 to 3 weeks, colonies were harvested and KEAP1 or NFE2L2 knockout was confirmed by qRT-PCR, Western blot, and gDNA sequencing analyses. In order to prevent exogenous overexpression of wild-type NFE2L2 from overcoming the capacity of endogenous KEAP1 to degrade it, we cotransfected a wild-type KEAP1–containing plasmid with plasmids containing NFE2L2 constructs into NFE2L2^NULL H1299 cells (Fig. 2I and J; Supplementary Fig. S6C–S6E). To generate A549-KEAP1 cells, KEAP1 cDNA was subcloned into the pCDH lentiviral vector (System Bioscience). A549 cells were then transduced with the virus, followed by selection using puromycin (1 μg/mL) for 7 days. Stable cell line generation was confirmed by qRT-PCR and Western blot analyses. The sequences of all sgRNAs and siRNAs are listed in Supplementary Fig. S4.

KEAP1 mutations were scored as being neutral (i.e., passengers) in the various assays if they displayed: (i) decreased HMOX1 mRNA expression compared with KEAP1^NULL cells transfected with empty vector (P < 0.05), (ii) increased H₂O₂ sensitivity in CellTiter-Glo assays compared with KEAP1^NULL H1299 cells transfected with empty vector (P < 0.05), (iii) decreased NFE2L2 target gene protein expression compared with KEAP1^WT H1299 cells by visual scoring, and (iv) increased radiation sensitivity in clonogenic assays compared with KEAP1^NULL H1299 cells transfected with empty vector (P < 0.05). KEAP1 mutations not meeting these criteria were scored as loss-of-function (i.e., pathogenic). NFE2L2 mutations were scored as being gain-of-function (i.e., pathogenic) in the various assays if they displayed: (i) increased HMOX1 mRNA expression compared with NFE2L2^NULL cells transfected with NFE2L2^WT vector (P < 0.05), (ii) increased NFE2L2 target gene protein expression compared with NFE2L2^NULL cells transfected with NFE2L2^WT vector by visual scoring, and (iii) increased radiation resistance in clonogenic assays compared with NFE2L2^NULL cells transfected with NFE2L2^WT vector (P < 0.05).

A549, H1299, and their isogenic cell lines (1.0 × 10³ cells) were subcultured into 96-well plates. After 24 hours, the cells were treated with CB-839 for 72 hours or hydrogen peroxide for 24 hours as indicated concentration in the figure legend; the cells were subsequently subject to CellTiter-Glo Luminescent Cell Viability Assay (Promega, #G7572) to determine cell viability as per the manufacturer's protocol.

Cells were preincubated with vehicle, CB-839, or N-acetylcystein (Abcam, #ab143032) for 24 hours and subsequently exposed to ionizing radiation. After an additional 24 hours, cells were stained with 25 μmol/L DCFDA for 45 minutes (Abcam) and evaluated via FACS analysis. The GSH:GSSG ratio was measured using the GSH/GSSG Ratio Detection Assay Kit (Abcam, # ab138881) according to the manufacturer's protocol and quantified using a microplate reader (E_x/E_m = 490/520 nm).

Follow-up was measured using the reverse Kaplan–Meier method from completion of treatment until last thoracic imaging. OS was estimated using the Kaplan–Meier method from the time of completion of RT. All other statistical analyses were conducted with adjustment for the competing risk of death. The cumulative incidence of LR and out-of-field recurrence (recurrence outside of the RT PTV) were measured using the R package “cmprsk.” Univariable and multivariable competing risk regressions were conducted with the R package “crrSC.” Variables with P < 0.1 on UVA were included in MVA. Adjusted Benjamini and Hochberg P values were calculated with the R package “stats.” Sample size calculation was conducted with the R package “samplesize.” For comparison of experimental data, the Mann–Whitney U test was used to calculate P values unless otherwise specified. All statistical analyses were conducted with R version 3.6 and PRISM version 6. Unless otherwise specified, all error bars represent standard deviation. All P values were two-sided and considered significant at P < 0.05.

E.J. Moding reports personal fees from DeciBio (for consultation on the use of biomarkers in cancer) outside the submitted work. J.J. Chabon reports ownership interest in Foresight Diagnostics, paid consultancy from Factorial Diagnostics, and patent filings related to cancer biomarkers. D.M. Kurtz reports personal fees from Roche Molecular Diagnostics (consultancy) and other from Foresight Diagnostics (ownership interest) outside the submitted work, as well as a patent for Methods and Systems for Assessment and Treatment of Cancer pending, a patent for Methods of Treatments Based Upon Updated Probabilities of Clinical Outcome pending, and a patent for Methods of Analyzing Cell Free Nucleic Acids and Applications Thereof pending. N.S. Lui reports grants from Intuitive Surgical Foundation (research grant) outside the submitted work. L.M. Backhus reports personal fees from Johnson and Johnson (speaker) and Guidpoint Consulting (expert consultant) and grants from the Department of Veterans Affairs (HSR&D Grant), and NIH (co-principal investigator on RO1) outside the submitted work. K.J. Ramchandran reports personal fees from Dhristi Inc (advisor), GTX (advisory board meeting), and Varian (advisor) outside the submitted work. S.K. Padda reports grants from Epicentrx (research funding), Bayer (research funding), and Boehringer Ingelheim (research funding) and personal fees from Blueprint (advisory board), G1 Therapeutics (advisory board), CME Solutions (honoraria), Pfizer (advisory board), AstraZeneca (advisory board), PER (honoraria), and 47 Inc. (research funding) outside the submitted work. M. Das reports personal fees from Jazz Pharmaceuticals (consultant), Bristol Myers Squibb (consultant), and AstraZeneca (consultant) and grants from Varian (research funding), AbbVie (research funding), United Therapeutics (research funding), and Novartis (research funding) outside the submitted work. J.W. Neal reports personal fees from Amgen (advisory board), AstraZeneca (advisory board), Calithera Biosciences (advisory board), Eli Lilly and Company (advisory board), Iovance Biotherapeutics (advisory board), Jounce Therapeutics (advisory board), Biomedical Learning Institute (honoraria), CME Matters (honoraria), Medscape (honoraria), MJH CME (honoraria), MLI Peerview (honoraria), Prime Oncology (honoraria), Research to Practice (honoraria), and Rockpointe (honoraria); grants from Adaptimmune (research funding), Boehringer Ingelheim (research funding), GlaxoSmithKline(research funding), Merck (research funding), Nektar Therapeutics (research funding), and Novartis (research funding); and grants and personal fees from Exelixis (advisory board, research funding), Genentech/Roche (advisory board, research funding), and Takeda Pharmaceuticals (advisory board, research funding) outside the submitted work. H.A. Wakelee reports grants from NIH and Ludwig institute for Cancer Research during the conduct of the study, as well as grants and personal fees from Novartis (advisory board, clinical trial support to institution), AstraZeneca (advisory board, clinical trial support to institution), and Xcovery (advisory board, clinical trial support to institution), personal fees from Janssen (advisory board), Mirati (advisory board), Daiichi Sankyo (advisory board), Helsinn (advisory board), and Blueprint (advisory board), and grants from ACEA Biosciences (clinical trial support to institution), Arrys Therapeutics (clinical trial support to institution), Bristol Myers Squibb (clinical trial support to institution), Celgene (clinical trial support to institution), Clovis (clinical trial support to institution), Genentech/Roche (clinical trial support to institution), Gilead (clinical trial support to institution), Merck (clinical trial support to institution), and Pfizer (clinical trial support to institution) outside the submitted work. A.A. Alizadeh reports grants and personal fees from Celgene; personal fees from Karyopharm, Roche, and Phamacyclics; personal fees and other from Gilead (stock); other from Forty Seven (stock), CiberMed (founder equity), and Foresight Diagnostics (founder equity) outside the submitted work. B.W. Loo reports grants and nonfinancial support from Varian Medical Systems (research funding and technical support) outside the submitted work. M. Diehn reports grants from NIH and Ludwig Institute for Cancer Research during the conduct of the study, as well as personal fees from Roche, AstraZeneca, RefleXion, Novartis, Gritstone, and BioNTech, personal fees and nonfinancial support from Illumina, and grants from Varian Medical Systems outside the submitted work. No disclosures were reported by the other authors.

M.S. Binkley: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. Y.-J. Jeon: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. M. Nesselbush: Data curation, formal analysis, visualization, writing–original draft, writing–review and editing. E.J. Moding: Data curation, investigation, visualization, writing–original draft, writing–review and editing. B.Y. Nabet: Data curation, investigation, writing–original draft, writing–review and editing. D. Almanza: Data curation, writing–original draft, writing–review and editing. C. Kunder: Data curation, investigation, writing–original draft, writing–review and editing. H. Stehr: Data curation, investigation, writing–original draft, writing–review and editing. C.H. Yoo: Data curation, writing–original draft, writing–review and editing. S. Rhee: Data curation, investigation, writing–original draft, writing–review and editing. M. Xiang: Data curation, writing–original draft, writing–review and editing. J.J. Chabon: Data curation, writing–original draft, writing–review and editing. E. Hamilton: Data curation, writing–original draft, writing–review and editing. D.M. Kurtz: Data curation, methodology, writing–original draft, writing–review and editing. L. Gojenola: Resources, data curation, writing–original draft, writing–review and editing. S.G. Owen: Resources, data curation, writing–original draft, writing–review and editing. R.B. Ko: Data curation, writing–original draft, writing–review and editing.J.H. Shin: Data curation, investigation, writing–original draft, writing–review and editing. P.G. Maxim: Writing–original draft, writing–review and editing. N.S. Lui: Data curation, writing–original draft, writing–review and editing. L.M. Backhus: Data curation, writing–original draft, writing–review and editing. M.F. Berry: Data curation, writing–original draft, writing–review and editing. J.B. Shrager: Data curation, writing–original draft, writing–review and editing. K.J. Ramchandran: Data curation, writing–original draft, writing–review and editing. S.K. Padda: Data curation, writing–original draft, writing–review and editing. M. Das: Data curation, writing–original draft, writing–review and editing. J.W. Neal: Data curation, writing–original draft, writing–review and editing. H.A. Wakelee: Data curation, writing–original draft, writing–review and editing. A.A. Alizadeh: Data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. B.W. Loo Jr: Resources, data curation, validation, investigation, writing–original draft, writing–review and editing. M. Diehn: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.

This work was supported by grants from the NCI (M. Diehn and A.A. Alizadeh, R01CA188298), the U.S. NIH Director's New Innovator Award Program (M. Diehn, 1-DP2-CA186569), the Ludwig Institute for Cancer Research (M. Diehn and A.A. Alizadeh), the CRK Faculty Scholar Fund (M. Diehn), and in part by PA-14-015, grant number T32 CA 121940, awarded by the Ruth L. Kirschstein National Research Service Award (NRSA), and the National Research Foundation of Korea (NRF) grant funded by the Korean government (Y.-J. Jeon, NRF-2020R1F1A1071579). The schematics for Fig. 1A, Fig. 2A, and Fig. 4A were produced using Servier Medical Art (https://smart.servier.com). Servier Medical Art by Servier is licensed under a Creative Commons Attribution 3.0 Unported License (https://creativecommons.org/licenses/by/3.0/).

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