NRF2 activation promotes aggressive lung cancer and associates with poor clinical outcomes

Keap1 inactivation synergizes with Stk11 loss to promote early and diffuse tumor formation, impaired lung function and mortality in Kras-driven lung cancer mouse models in vivo

To characterize the impact of combined loss of Keap1 and Stk11 in the context of a Kras^G12D activating mutation in vivo, we crossed Keap1^fl/fl mice to Kras^LSL-G12D mice and Stk11^fl/fl mice followed by intra-tracheal administration of adenovirus encoding CMV-Cre-recombinase to delete exons 2–3 of Keap1 and exons 3–6 of Stk11 in the mouse lungs (13–16). The following genotypes were generated: (a) Kras^LSL-G12D, (b) Keap1^fl/flKras^LSL-G12D, (c) Stk11^fl/flKras^LSL-G12D, (d) Keap1^fl/flStk11^fl/flKras^LSL-G12D. Real-time PCR (Q-PCR) confirmed reduced expression of Keap1 and Stk11 in lungs with the respective genotypes (Fig. S1A). The four described genotypes of our novel GEMMs exhibited striking differences in overall survival (OS), most strikingly a reduced median OS to only 59 days in mice with combined loss of Keap1 and Stk11 that became progressively moribund with severe respiratory distress (Fig. 1A).

Transcriptional characterization of the murine lungs further confirmed robust induction of Nrf2 upon Keap1 deletion. While Stk11 loss impacted the expression of 19 genes consistently in the absence or presence of concurrent Keap1 loss, loss of Keap1 resulted in 304 up- and 147 down-regulated genes, irrespectively of Stk11 status (>2 fold expression change, adjusted p <0.05; Fig. 1B, Table S1). Keap1^fl/flStk11^fl/flKras^LSL lungs were uniquely characterized by 59 higher and 169 lower expressed genes compared to all other genotypes, implying synergistic transcriptional implications from co-occurring Stk11 and Keap1 loss. Gene set enrichment analysis (GSEA) comparing Keap1 deficient (Keap1^fl/fl) to wild-type (Keap1^+/+) tumor bearing murine lung samples showed induction of glutathione, xenobiotics, pentose phosphate pathway and arachidonic acid metabolism genes in Keap1^fl/fl-murine lung samples (Fig. S1B, Table S1).

To evaluate tumor burden and identify a cause for the animals’ morbidity and short lifespan, we evaluated lung density, function and histology. Direct measurement of tumor burden by micro-computed tomography (CT) was hindered by the diffuse and widely dispersed growth of neoplastic lesions in Keap1^fl/flStk11^fl/flKras^LSL-G12D animals; thus, we assessed lung density as a surrogate measurement of transformed lung tissue across genotypes. The Keap1^fl/flStk11^fl/flKras^LSL-G12D genotype displayed high lung density as early as 4 weeks post-viral infection (pvi) and maintained it at 6 and 8–9 weeks when they succumbed. By comparison, Stk11^fl/flKras^LSL-G12D lungs showed a lower lung density at 4 weeks, progressively increasing at 6 and 8–9 weeks while Keap1^fl/flKras^LSL-G12D lungs reached high density at 8–9 weeks. Kras^LSL-G12D mice were analyzed only at 8–9 weeks pvi at which timepoint they displayed the lowest lung density across all genotypes, consistent with their more indolent disease course (Fig. 1C–D, Fig. S1C).

By 6–8 weeks pvi, Keap1^fl/flStk11^fl/flKras^LSL-G12D mice were already demonstrating severe respiratory distress. Based on three key functional pulmonary readouts: (i) carbon monoxide diffusion (DF_CO); (ii) functional residual capacity (FRC); and (iii) baseline respiratory resistance (Rrs), the Keap1^fl/flStk11^fl/flKras^LSL-G12D mice exhibited maximal lung dysfunction compared to the other genotypes (Fig. 1E–F). These data, combined with the severe respiratory distress and labored breathing in the triple mutant mice, indicated that lung failure, associated with inspiratory muscle fatigue, contributed to the reduced survival of this group.

Histological evaluation of mouse lungs at 7 weeks pvi identified significant differences in both the number and distribution of tumor nodules in Stk11^fl/flKras^LSL-G12D lungs with versus without Keap1 loss. In Keap1^fl/flStk11^fl/flKras^LSL-G12D animals, tumors affected all lung lobes and occurred throughout the lung parenchyma from the airways to the pleura (Fig. 1G, Fig. S1D). These lungs contained dozens of distinct foci ranging from atypical adenomatous hyperplasia (AAH) to poorly differentiated adenocarcinomas containing dysplastic cells with abnormal mitotic figures. In contrast, lobes of Stk11^fl/flKras^LSL-G12D animals contained large, but fewer tumors that appeared to be radiating from the airways into the adjacent parenchyma. Furthermore, such tumors were often surrounded by pools of mucin admixed with necrotic debris. Overall, these data point towards a tumor-promoting effect of Keap1 loss in the context of Stk11 loss and Kras^G12D activation, facilitating earlier onset and multi-focal tumor initiation compared to Stk11 loss alone.

Keap1^fl/flStk11^fl/flKras^LSL-G12D mice display tumor growth within airways, severe airway obstruction, and mucinous differentiation linked to Stk11 loss

To gain a better understanding of the impact of Keap1 and Stk11 loss on lung cancer histology, we performed a more detailed evaluation of lungs at 6–12 weeks pvi. At this early timepoint, one- to two-third of Stk11^fl/flKras^LSL and Keap1^fl/flStk11^fl/flKras^LSL-G12D mutant lungs bore adenomas and adenocarcinomas. Furthermore, for both genotypes lungs developed AAH and both mucinous as well as non-mucinous bronchiolar epithelial hyperplasia (BEH). Kras^LSL-G12D and Keap1^fl/flKras^LSL-G12D lungs developed carcinomas only at later timepoints, 12–24 weeks pvi (Fig. 2A, Fig. S1E, Table 1). Unique among triple mutant lungs was the presence of intra-bronchiolar papillary carcinomas, sometimes completely obstructing the airways (Fig. 2A, green crosses). The combination of these features, absent from the other three genotypes, likely explains the respiratory distress and early-onset mortality in triple mutant mice.

Interestingly, all adenomas and adenocarcinomas in Stk11^fl/flKras^LSL-G12D mice showed evidence of mucinous differentiation; this phenomenon was only apparent in 43% of carcinomas in Keap1^fl/flStk11^fl/flKras^LSL-G12D lungs (Fig. 2A, Fig. S1F, Table 1). Staining for mucins using Alcian Blue/PAS at 6–7 weeks pvi confirmed mucin production in Stk11^fl/flKras^LSL-G12D and Keap1^fl/flStk11^fl/flKras^LSL-G12D lungs; in contrast, Kras^LSL-G12D and Keap1^fl/flKras^LSL-G12D lungs displayed only non-mucinous AAH (Fig. 2B). The association of a mucinous differentiation phenotype with Stk11 loss was further supported by a 6–10-fold increase in the mRNA expression of the canonical mucinous marker genes Muc5b and Agr2 in Stk11^fl/fl versus Stk11 wild-type lungs (p = 0.0002 and 0.002, respectively; Fig. 2C, Table S1). Consistent with the mouse data, a significantly higher expression of the mucinous markers MUC5AC, MUC5B and AGR2 was present in STK11-mutant human tumors from TCGA LUAD (7), independent of KEAP1/NRF2 status (Fig. 2D). Mucinous differentiation is often associated with deleterious mutations and/or decreased expression of the NKX2 homeobox 1 (NKX2–1 or TTF1) protein, a lineage-specific transcription factor that negatively regulates mucinous marker expression (17). Yet, our mouse lung adenocarcinomas stained positive for TTF1/NKX2–1, including the areas of mucinous differentiation and independently of genotype (Fig. S1E–F). By contrast, there was a modest decrease in Nkx2–1 expression in Keap1^fl/flStk11^fl/flKras^LSL-G12D lungs at the mRNA level (1.5-fold, p = 0.018), but we found no evidence of a downregulation in Stk11^fl/flKras^LSL-G12D lungs (Fig. 2C, Table S1). Consistent with the mouse phenotypes, STK11 status had no impact on NKX2–1 levels in the human LUAD TCGA dataset, whereas NKX2–1 mRNA levels were ~2-fold lower among KEAP1/NRF2-mutant tumors (Fig. 2E). An additional characteristic of the Stk11^fl/flKras^LSL-G12D genotype was the presence of squamous carcinomas in 40% of lungs at 12–19 weeks psi, consistent with previous reports (18). Finally, quantification of Ki67-positive nuclei across tumor histotypes revealed an overall higher proliferation rate among Keap1^fl/flStk11^fl/flKras^LSL-G12D compared to Stk11^fl/flKras^LSL-G12D tumors of any histology (Fig. S1G). Differences by genotype per histology did not reach significance due to the small sample size.

In summary, our GEMMs and TCGA analyses provide additional insights into the impact of Keap1 versus Stk11 loss on LUAD differentiation and progression, revealing phenotypes specific to Stk11 loss (namely a high frequency of mucinous adenocarcinomas arising independently of changes in Nkx2–1 mRNA levels) and those exacerbated by combined Keap1 loss (early onset tumorigenesis, high proliferation rates, presence of intra-bronchiolar carcinomas and association with moderately lower Nkx2–1 mRNA expression). The cross-species concordance suggests that our mouse LUAD model faithfully recapitulates human LUAD phenotypes.

Keap1 loss counteracts the increase in redox stress induced upon Stk11 inactivation in a Kras-driven lung cancer mouse model in vivo

We have previously shown that STK11 loss limits the flexibility of tumor cells to overcome bioenergetic and redox stress, due to defective AMPK activation and decreased NADPH levels (19). Hence, we evaluated the impact of Keap1 loss on lung redox homeostasis in the context of combined Stk11 loss in vivo, making use of the described GEMMs. Using a flow cytometry-based measurement of intracellular ROS (CM-H₂ DCFDA), we observed a doubling of epithelial cells high in ROS levels (defined as EpCAM⁺CM-H₂ DCFDA^high) in Stk11^fl/flKras^LSL-G12D lungs, relatively to Kras^LSL-G12D lungs, that was dramatically decreased upon concurrent Keap1 loss (Fig. 2F). Assessing lipid peroxidation, a process linked to free oxygen radicals modifying unsaturated fatty acids, revealed a similar pattern: two lipid peroxidation byproducts, thiobarbituric acid-reactive substances (TBARS) and 4-hydroxynonenal (4-HNE), displayed highest levels in Stk11^fl/flKras^LSL-G12D mouse lungs, whereas concurrent loss of Keap1 was able to completely reverse the Stk11-null phenotype (Fig. 2G–H). Keap1 loss also significantly increased glutathione (GSH) levels, irrespectively of Stk11 expression (Fig. 2I). Collectively, these data support the model that Keap1 loss with ensuing constitutive Nrf2 transcriptional activation - as evidenced through the upregulation of several Nrf2 target genes (Fig. 1B, Fig. S1B) - results in a robust ROS detoxification capacity, neutralizing the increased redox stress induced upon Stk11 loss in vivo.

We confirmed the cooperativity between NRF2 activation and STK11 loss in maintaining redox balance in human LUAD HCC515 isogenic cell lines with either inducible knockdown of KEAP1 and/or re-introduction of STK11 (Fig. S2A–D). Under suspension growth, STK11-null HCC515 cells showed decreased viability compared to STK11-reconstituted cells and this phenotype was rescued by KEAP1 knockdown. Accordingly, KEAP1 knockdown reversed the dramatic increase in ROS levels observed in STK11-null cells under suspension conditions (Fig. S2E–F). In the case of KEAP1 knockdown this phenotype was associated with increased total and de novo (M+5, shaded bar) synthesis of the anti-oxidant tripeptide GSH (Fig. S2G). These experiments support the model that KEAP1 loss, leading to NRF2 activation, can compensate for the loss of STK11 to maintain low ROS levels via increased total GSH synthesis and enable survival under redox stress conditions.

A transcriptional signature of NRF2 activation is prognostic in primary human lung adenocarcinoma and correlates with low STK11 mRNA expression

We set out to assess the prevalence of functional NRF2 activation, beyond merely genomic alteration, and its clinical impact across human NSCLC. We and others previously released KEAP1/NRF2 signatures (12,20–23). All but one of those signatures were curated from literature or originated from in vitro models, with signatures containing genes that associated with KEAP1 genomic alterations or correlated with known NRF2 targets. Only one KEAP1-mutant-specific signature was previously derived from TCGA lung adenocarcinoma tumors; yet, its whole-transcriptome nature limits clinical practicality (11). Herein, we employed a different approach aimed at identifying a core set of KEAP1/NRF2 associated genes from human tumors that are independent from histology, hence considering both adenocarcinoma and squamous cell carcinoma (7,24), since NRF2 activation is also prevalent in LUSC (35%) (24); further we aimed at identifying genes that were independent from recurrent oncogenic events, namely alterations in KRAS and STK11 in LUAD and TP53 in LUSC (Fig. 3A, Fig. S3A). As described in Supplementary Materials & Methods, we devised a 96-gene signature of KEAP1/NRF2 pathway activation (Fig. 3B, Fig. S3B–D, Table S2). This signature captures metabolic processes such as glutathione, xenobiotics as well as arachidonic acid metabolism (Fig. 3C, Table S2). A subset of these 96 genes was consistently upregulated in both clinical and pre-clinical NRF2-active tumor samples (Fig. 3D–E, Table S3), and likely represents a set of universal NRF2 targets. Other signature genes were only upregulated in clinical samples, suggesting they may either represent context dependent NRF2 targets in primary human lung cancer and/or features of the tumor microenvironment associated with NRF2 activation, both important to capture in human tumor samples.

By calculating a signature score based on the 96 signature genes and defining histology-specific score thresholds above which a tumor is more likely to have aberrant KEAP1/NRF2 pathway activity, we found that the majority of patients with documented genomic KEAP1/NRF2 alterations [63% of LUAD and 89% of LUSC] possessed a high signature score (Fig. S3E–F). Reversely, 7% of wild-type LUAD and 24% of wild-type LUSC also had a high signature score (Fig. S3E–F), indicating that a transcriptional signature of NRF2 activation can capture patients displaying functional NRF2 pathway activation in the absence of a documented pathway mutation. Furthermore, our signature was associated more strongly with poor OS in LUAD (multivariate HR = 1.82, p = 0.0011; adjusting for KRAS and STK11 status) (Fig. 3F) than genomic KEAP1/NRF2 alterations (HR = 1.31, p = 0.14) and remained a poor prognostic factor, even when accounting for genomic KEAP1/NRF2 alteration status (HR = 2.07, p = 0.0035). While the signature added independent prognostic information beyond genomic KEAP1/NRF2 alterations (Likelihood Ratio (LR) test, p = 0.0043), the reverse was not true (LR test, p = 0.45). These characteristics were generally true for the various KEAP1/NRF2 signatures, as indicated by a comparative analysis in Fig. S3G–H, Table S4 and S9, and Supplementary Materials & Methods. Finally, in support of a synergistic relationship between concurrent KEAP1/STK11 loss, a high NRF2 activation signature was associated with low STK11 mRNA expression in primary human LUAD, even among tumors with no documented STK11 genomic alteration (multivariate model, signature p = 6e-11, STK11 mutation status p = 0.8, Fig. 3G). Overall, this analysis suggests that a transcriptional signature of NRF2 activation has the potential to be a superior clinical diagnostic tool to KEAP1/NRF2 mutation status in NSCLC.

NRF2 activation and STK11 mutations associate with shorter survival in docetaxel- and atezolizumab-treated advanced non-squamous lung cancer

To further dissect the impact of NRF2 activation on tumor immune contexture and clinical outcomes, we made use of a clinical lung cancer dataset of second-line (2L+) metastatic or locally advanced NSCLC patient participants in OAK (NCT02008227), a phase III trial comparing standard of care treatment (docetaxel) to the immune checkpoint inhibitor atezolizumab (anti-PD-L1) (25). We first confirmed (i) that the biomarker evaluable populations (BEP) were representative of the intent-to-treat populations (Table S5) and (ii) the robust performance of the NRF2 activation signature in 2L+ non-squamous tumors, including a strong enrichment among KEAP1/NRF2-mutant tumors (Fig. 4A, Fig. S4A–B) and a negative association with OS (HR = 1.79, p = 7e-5; Fig. S4C). Importantly, our NRF2 signature captured significant prognostic signal beyond mere genomic KEAP1/NRF2 alterations (LR test p = 7.8e-4), yet, the reverse was not true (LR test p = 0.77; Table S6). We also confirmed that STK11-null tumors were more likely to be NRF2 signature-high (41% of signature-high vs. 13% of signature-low; Fisher’s exact test p = 1.8e-7; Fig. 4A; Table S7).

Among patients with KRAS-altered tumors, we found that either STK11 loss and/or NRF2 activation significantly reduced overall survival (Fig. 4B), consistent with our in vivo mouse model (Fig. 1A). Yet, a similar negative impact of these two events on overall survival in KRAS-wildtype tumors indicated independence from KRAS status (Fig. 4C); hence, we further assessed the impact of STK11 loss and NRF2 activation in the combined sample set, i.e. both KRAS-altered and -wildtype tumors.

We found both the NRF2 activation signature and STK11 mutation status to be prognostic in 2L+ LUAD, irrespectively of treatment arm (NRF2 signature: atezolizumab HR = 1.72, Wald test p = 0.0078; docetaxel HR = 1.98, p = 0.002; STK11 mutations: atezolizumab HR = 1.54, p = 0.051; docetaxel HR = 2.28, p = 7.3e-5; Fig. 4D–E). In addition, the NRF2 signature and STK11 mutation status each contributed independently to poor prognosis, across both treatment arms (Fig. S4C–D, see Table S4 for statistics). In contrast, KRAS mutation status was not prognostic in either treatment arm (p = 0.76; Fig. 4F, Fig. S4E, Table S6). Instead, while the BEP as a whole trended towards longer survival on atezolizumab compared to docetaxel, KRAS-altered tumors were the ones for which this benefit reached statistical significance (HR = 0.53 [CI = 0.35; 0.81], Fig. 4F–G, Table S6). Similar associations with outcome were observed for alternative end points (progression free survival and overall response, Table S6, Fig. S4F–L). Further, observed associations with OS were independent from previously described molecular correlates of response to atezolizumab in NSCLC, PD-L1 protein expression on either immune cells (IC) or tumor cells (TC) as measured by immunohistochemistry (IHC) and tumor mutation burden (26–28) (Supplementary Materials & Methods).

In OAK, patients with high PD-L1 TC expression derived the greatest benefit from atezolizumab, although OS improvement versus docetaxel was noted for the complete range of positive PD-L1 expression (25). Hence, we evaluated PD-L1 protein expression in the context of alterations in KRAS, STK11 and the NRF2 activation signature (Fig. 4H–I). KRAS alterations were associated with increased PD-L1 TC staining (LR test p = 2.2e-6, Fig. 4H), which may partially explain the improved survival of patients with KRAS-altered tumors when treated with atezolizumab, relative to docetaxel (Fig. 4F–G). In contrast, STK11 mutations were negatively associated with PD-L1 TC expression, however only in the context of a co-existing KRAS alteration: while 3% of STK11-mutant compared to 51% of STK11-wildtype tumors stained positive for PD-L1 on TCs among KRAS-altered tumors (Fisher’s exact test p = 7.5e-5), there was no significant difference in TC PD-L1 expression by STK11 status among KRAS-wildtype tumors (16% vs. 27%, Fisher’s exact test p = 0.73; Fig. 4H). In the case of PD-L1 IC staining, thought to correlate with IFNγ–induced adaptive immune evasion, there was also a negative association with STK11 mutations and this was observed irrespectively of KRAS status (LR test p = 3.8e-4, Fig. 4I). The NRF2 activation signature was also negatively associated with PD-L1 TC staining, however this was likely driven by the enrichment for STK11-mutant tumors among NRF2 signature-high patients. Of note, despite their negative association with PD-L1 expression and negative prognostic outlook, both STK11-mutant and NRF2 signature-high tumors derived benefit from atezolizumab to a similar extent as the overall evaluable patient population, irrespectively of KRAS status (Fig. 4G).

OAK being an immunotherapy trial, we assessed whether NRF2 activation as well as KRAS and STK11 alterations impacted the tumors’ immune contexture. To do so, we calculated scores for established immune cell signatures (Table S8). Expression of a signature estimating infiltration by T effector cells trended towards statistical significance (F-test p = 0.05) for being lowest in the STK11-mutant /NRF2 signature-high subgroup (Fig. S4M). The NK cell gene signature, however, showed significantly higher expression among STK11-mutant/NRF2 signature-low tumors (Tukey’s HSD test p = 0.035) (Fig. S4N). Expression of a dendritic cell (DC) signature was significantly reduced in tumors with high NRF2 signature score, without (p = 0.030) and with concurrent STK11 mutations (p < 0.001) (Fig. S4O). On the other hand, a myeloid gene expression signature was more highly expressed in STK11-mutant only tumors (p = 0.003) (Fig. S4P). Consistent with our mouse models, we observed a positive association of the mucin gene expression signature with STK11 loss (p <<< 0.001) (Fig. S4Q), and a negative association of NKX2–1 expression with NRF2 activation in the absence of STK11 loss (p = 0.001) (Fig. S4R).

To summarize, in 2L+ LUAD, we find that high expression of the NRF2 activation signature and STK11 inactivation are two independent negative prognostic factors, associated with inferior outcomes to both chemotherapy and immunotherapy regimens. In contrast, KRAS alterations are associated with a better response to atezolizumab versus docetaxel treatment in this patient population. NRF2 signature score, STK11 inactivation and KRAS alterations have different effects on tumor phenotype and immune contexture, manifested through their distinct impact on PD-L1 immune checkpoint expression on tumor and immune cells, and their association with immune cell gene expression signatures.

Advanced squamous lung cancer patients without NRF2 activation experience longer overall survival when receiving atezolizumab

Finally, we expanded our analyses to OAK participants with squamous tumors (2L+ LUSC), testing the association of our NRF2 activation signature with overall survival, PD-L1 protein expression and impact on tumor immune contexture (Fig. S5A–E, G–I). Strikingly, the NRF2 signature was negatively associated with OS only for patients who received atezolizumab (HR = 1.768 [CI = 1.094; 2.856]), not for those on docetaxel (HR = 1.068 [CI = 0.696; 1.639], Fig. S5C–D). While these data suggest that a low NRF2 signature score may be predictive of benefit from atezolizumab therapy in squamous lung (Fig. S5E), a formal test for the interaction between the NRF2 signature and treatment arm in affecting survival did not reach significance (LR test p = 0.111).

Intrigued by this finding, we validated this association in another independent clinical data set, IMpower131 (NCT02367794) (Fig. 5A, Fig. S5F). IMpower131 is a phase III trial in first line (1L) stage IV squamous NSCLC, comparing the efficacy of a combination of atezolizumab plus chemotherapy (ACP or ACNP) versus chemotherapy alone (CNP; Carboplatin + Nab-Paclitaxel) (29). While 54% of 2L+ patients were NRF2 signature-high (OAK, Fig. S5A–B), only 32% of tumors exhibited NRF2 activation in 1L (IMpower131, Fig. 5A). We confirmed in this 1L LUSC population that NRF2 signature-low patients derived unique survival benefit only when treated with atezolizumab-containing treatment regimens (ACP vs. CNP: HR = 0.754 [CI = 0.595; 0.955]; ACNP vs. CNP: HR = 0.724 [CI = 0.574; 0.913]; Fig. 5B). Still, a formal test for interaction between the NRF2 signature and treatment arm in affecting survival did not reach significance (LR test p = 0.60, comparing CNP to ACNP), possibly due to the large percentage (>50%) of patients in the control arm who received immunotherapy follow-on therapies (29).

All tested immune signatures were significantly more lowly expressed in NRF2 activated, 1L squamous tumors (Fig. 5C). The same associations were apparent in 2L+ LUSC, with the myeloid gene signature reaching significance (adjusted Wilcoxon test p = 0.011) (Fig. S5G). Interestingly, we did not observe an association of the NRF2 activation signature with PD-L1 TC or IC staining in LUSC in either the IMpower131 (Fig. 5D–E) or OAK (Fig. S5H–I) data sets.

We conclude that NRF2 activation impacts the tumor immune contexture in both 1L and 2L+ LUSC. Further, patients with low expression of our NRF2 activation signature show superior survival when treated with atezolizumab, while patients with a NRF2-activated tumor have poor prognosis and remain a high unmet clinical need population.