Introduction

Lung cancer is the leading cause of cancer-related mortality on a global scale, accounting for an estimated 1.6 million lives lost annually. Among its heterogeneous subtypes, non-small cell lung cancer (NSCLC) represents a significant proportion, with diverse histological manifestations. Notably, NSCLC encompasses two predominant subtypes, lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), each presenting distinct molecular profiles and clinical behaviors1.

Over the years, considerable efforts have been devoted to deciphering the intricate genomic landscape of NSCLC, shedding light on the critical role of driver gene mutations in its pathogenesis and progression. Notable among these driver genes we find: the epidermal growth factor receptor (EGFR), the Kirsten rat sarcoma viral oncogene homolog (KRAS), the tumor protein p53 (TP53), and the Kelch-like ECH-associated protein 1 (KEAP1), along with its downstream effector, the nuclear factor erythroid 2-related factor 2 (NRF2). Mutations in these key genes orchestrate a complex cascade of events, influencing tumor evolution, therapeutic response, and patient prognosis. While EGFR mutations are prevalent in a subset of NSCLC cases and can typically be associated with a more favorable prognosis, also due to targeted therapies2, KRAS and TP53 mutations are associated with aggressive tumor behavior and resistance to most therapies, except for immunotherapy, where these mutations may actually indicate a better response to immune checkpoint blockade3. Furthermore, the dysregulation of the KEAP1-NRF2 pathway, primarily attributed to somatic mutations in KEAP1 or Nuclear factor erythroid-derived 2-like 2 (NFE2L2), has emerged as a significant determinant of disease progression and therapeutic resistance in NSCLC4. Recent evidence indicates that mutations in the KEAP1 gene, a key regulator of cellular response to oxidative stress, exert a detrimental impact on the prognosis of NSCLC patients5 and are correlated with resistance to immunotherapy in LUAD6 and to KRAS inhibitors7. The KEAP1 gene plays a critical role in maintaining cellular homeostasis. It oversees the cellular defense against oxidative damage and metabolic stress by modulating the activity of NRF2, a master transcription factor that controls the expression of antioxidant and detoxification enzymes. Mutations in KEAP1 prevent its binding to the NRF2 degron motifs, thereby impeding the targeting for proteasomal degradation. This results in NRF2 entering the nucleus, where it triggers the expression of target genes containing antioxidant response elements (ARE) in their promoters. Consequently, this facilitates metabolic reprogramming and detoxification8.

Similarly, mutations in the NFE2L2 gene occur in the KEAP1 binding sites, resulting in the constitutive activation of the NRF2 pathway. This activation contributes to disease progression, metastatic spread, and enhanced resilience against cytotoxic substances8.

The comprehensive understanding of the intricate genomic landscape of complex cancers such as NSCLC may greatly benefit from the effective integration of both genomic and transcriptomic data. This integrative approach may allow for a deeper examination of the interplay between somatic mutations and gene expression, thereby potentially offering invaluable insights into tumor biology and guiding therapeutic strategies. To address this critical need, we here introduce the Association of SomaTic mUtaTions to gene Expression profiles (ASTUTE) framework, designed to characterize genotype-phenotype associations in cancer. Through the integration of genomic and transcriptomic datasets, ASTUTE provides a sophisticated analytical tool to uncover novel molecular mechanisms driving cancer progression and treatment response, thus contributing to the advancement of precision oncology. Employing ASTUTE, we here systematically analyzed distinct NSCLC datasets, followed by an exploration across different cancer types harboring KEAP1 or NFE2L2 mutations. This comprehensive investigation allowed us to elucidate the intricate correlation between KEAP1/NFE2L2 mutations and the activation of the NRF2 pathway, shedding light on novel NRF2-related functionalities associated with specific genetic alterations. Additionally, we identified an expression signature associated with mutations in KEAP1 or NFE2L2 genes across different cancer types, showing a robust association with prognosis. Our discoveries offer significant insights into the underlying mechanisms of NSCLC pathogenesis and drug resistance. While several NRF2-related gene signatures have been previously reported, our approach differs in that it derives a mutation-driven expression signature through a rigorous statistical framework. By leveraging a robust approach to associate somatic mutations in KEAP1/NFE2L2 to gene expression, rather than expression-based clustering or pathway annotations, we aim to delineate a mechanistically grounded, reproducible transcriptional program associated with NRF2 pathway activation. These results present a promising avenue for the development of tailored therapeutic interventions and prognostic biomarkers. Furthermore, the relevance of this study becomes even more evident as specific inhibitors of the NRF2 axis become available in clinical practice.

Results

Genotype-phenotype mapping in cancer: the ASTUTE framework

Understanding how genetic mutations influence observable traits is crucial. ASTUTE is a novel computational framework capable of performing genotype-phenotype mapping between somatic mutations and expression data (Fig. 1).

Fig. 1: Overview of the ASTUTE framework.
figure 1

In A–C we illustrate how our framework can efficiently integrate mutations with gene expression data to perform the extraction of dysregulated genes associated with KEAP1 or NFE2L2 mutations in distinct NSCLC datasets. In D we highlight the consistent association of the identified genes with the NRF2 pathway. In E we show that ASTUTE can stratify patients based on the identified expression signatures, thus enhancing the prognostic insights returned by the approach. Finally, in F we showcase that ASTUTE could determine a set of genes consistently dysregulated across in NSCLC andother cancer types, emphasizing their role in prognosis at the pan-cancer level. ASTUTE’s multidimensional analysis enables a deeper understanding of genotype-phenotype associations in cancer.

Full size image

Leveraging regularized regression with the LASSO penalty, ASTUTE employs a sophisticated approach based on regularized linear regression and the bootstrap9 that incorporates a penalty term into the loss function, thus effectively mitigating overfitting and performing feature selection10. The LASSO penalty is determined by summing the absolute values of the model coefficients and then multiplying this sum by a regularization parameter, whose selection is optimized through cross-validation to determine the degree of penalty applied to the model. This process induces feature selection by encouraging coefficients associated with the less influential variables to shrink toward zero, resulting in a more interpretable model that emphasizes the most significant variables for predictive purposes. Consequently, the resulting models may include coefficients denoting features lacking a significant association with driver gene mutations, while other genes may exhibit significantly altered expression levels due to the presence of specific mutations. The application of the ASTUTE framework to distinct NSCLC datasets extracted a set of genes consistently upregulated or downregulated in association with mutations in either KEAP1 or NFE2L2 genes in lung cancer (Fig. 1A–C). Most of these genes are known to be associated with the NRF2 pathway (Fig. 1D).

Furthermore, ASTUTE offers the capability to estimate baseline gene expression levels, calculate fold changes with respect to the baseline, and compute p values using the bootstrapping technique to determine whether a fold change significantly indicates over- or under-expression. The selected features can be exploited for further analyses, such as Gene Set Enrichment Analysis (GSEA), to elucidate the biological implications of genetic somatic mutations (Fig. 1C).

Unlike multi-omics latent factor approaches, ASTUTE is specifically tailored to infer direct genotype-to-phenotype associations. While multi-omics factor analysis methods typically focus on identifying latent variables that explain joint variation across data types, ASTUTE is designed to capture direct, interpretable associations between somatic mutations and gene expression changes. This makes ASTUTE particularly well-suited for mechanistic studies and biomarker discovery, rather than exploratory dimensionality reduction. It employs LASSO regularization to extract sparse, interpretable gene sets whose expression changes are directly linked to specific somatic mutations. This design makes ASTUTE well-suited for biomarker discovery and mechanistic inference, rather than broad unsupervised factor analysis.

Moreover, ASTUTE enables patient stratification based on the identified expression signatures, enhancing prognostic insights (Fig. 1E). The signatures, associated with specific gene mutations, provide a phenotypic characterization of the related somatic mutations, which, in turn, can serve as clinical biomarkers. In the case of the identified NRF2 expression signatures, our approach discovered a set of genes consistently dysregulated in both NSCLC and other cancer types. These genes could effectively stratify patients based on prognosis in all the analyzed cancers, which presented frequent mutations in the KEAP1/NFE2L2 genes and were consistently overexpressed in cancers with worse prognoses (Fig. 1F). We describe these results in detail in the next sections.

Identification of a NRF2 expression signature associated with KEAP1 and NFE2L2 mutations in NSCLC

We applied the ASTUTE framework to analyze five distinct datasets providing both somatic mutations and gene expression data from LUAD and LUSC patients. Specifically, for LUAD, we considered data from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) consortium (110 samples)11, the Pan-Cancer Atlas (510 samples)12, and the study by Chen et al. (169 samples)13, for a total of 789 LUAD patients. For LUSC, data from the CPTAC consortium (108 samples)14 and the Pan-Cancer Atlas (481 samples)15 were included, comprising 589 LUSC patients. In total, our analysis encompassed 1378 NSCLC samples. ASTUTE was independently executed on each dataset, and the results were compared to identify consistent findings across all datasets (Supplementary Data 1 and 2).

ASTUTE revealed significant upregulation of a specific set of genes strongly associated with the NRF2 pathway activation in LUAD patients harboring KEAP1 mutations and in LUSC patients with either KEAP1 or NFE2L2 mutations. The low frequency of NFE2L2 mutations in LUAD patients precluded extensive exploration within this cohort.

These upregulated genes, previously identified as NRF2 targets, fall into three main functional classes crucial for cellular processes (Table 1). First, genes involved in glutathione synthesis, such as GCLM and GCLC, play significant roles in maintaining redox balance16,17. Second, a substantial group of genes implicated in cellular oxidative response, detoxification, or inhibition of ferroptosis (Table 1), underscores the critical role of NRF2 in protecting cells from oxidative stress and promoting survival under toxic conditions18. Third, genes associated with carbohydrate metabolism and NADPH generation (Table 1), highlight the NRF2 pathway’s involvement in sustaining metabolic flexibility and anabolic growth in cancer cells19. To experimentally validate these findings, the expression of 10 NRF2 targets was investigated by quantitative PCR in LUAD cells carrying comparable genetic makeup but different NFE2L2 mutational status. Specifically, H2228 cells harbor a gain-of-function mutation in NFE2L2 (G31A), while H3122 cells are wild-type for both KEAP1 and NFE2L2 (see Methods). All tested genes were strongly upregulated in KEAP1-mutated cells compared with KEAP-WT cells (Supplementary Fig. 1).

Table 1 LUAD/LUSC NRF2 expression signature. Genes identified by the ASTUTE framework in association with the presence of NFE2L2 or KEAP1 mutations in both LUAD and LUSC
Full size table

We further validated the identified signatures using proteomics data from CPTAC. Specifically, we compared protein levels of the genes listed in Table 1 between LUAD (Supplementary Fig. 2) and LUSC (Supplementary Fig. 3) patients, stratified by the presence or absence of KEAP1/NFE2L2 mutations. Patients with mutations consistently showed higher expression of the selected genes at the protein level as well.

In addition to these categories, we identified other NRF2 targets with roles extending beyond these classes (Table 1). TRIM16 stands out among these genes, which has been reported to associate with the p62-KEAP1-NRF2 complex, suggesting a potential positive feedback loop that may enhance NRF2 signaling activation, offering insights into the regulatory mechanisms driving this pathway20,21.

Association of other driver genes with the LUAD/LUSC NRF2 expression signature

Some genes listed in Table 1 also exhibited significant associations with other driver mutations. Notably, in LUAD, the genes AKR1C1, AKR1C2, AKR1C3, and ALDH3A1 were positively associated with STK11 mutations (0.2 < log2FC < 1), while GPX2 showed a similar association with SMARCA4 mutations (0.3 < log2FC < 0.5). In LUSC, the genes AKR1C1, AKR1C2, AKR1C3, and GPX2 were associated with KMT2D mutations (0.2 < log2FC < 0.7), and AKR1C1, AKR1C2, ALDH3A1, CYP4F3, GCLC, GCLM, and GPX2 were associated with TP53 mutations (0.2 < log2FC < 1.3).

These results underscore the varying degrees of association between different mutations and the expression of the NRF2 signature genes in both LUAD and LUSC. While KEAP1/NFE2L2 mutations are primary drivers of the NRF2 pathway activation, these findings suggest the existence of alternative mechanisms influencing NRF2 expression independently of these mutations.

Interestingly, ABCC222, SLC7A1123, AIFM224, and NEIL325 genes were identified as NRF2 target genes associated with KEAP1 mutations in LUAD, while in LUSC, these genes were linked with NFE2L2 mutations. This divergence might reflect distinct molecular pathways underlying NRF2 activation in these cancer types.

Moreover, our analysis revealed two genes involved in the NRF2 pathway selectively upregulated in LUAD but not LUSC in the presence of KEAP1 mutations: CPLX2, known for its role in regulating NRF2 expression in hepatocellular carcinoma (HCC)26 and recognized as a potential prognostic biomarker in lung cancer27, and KYNU, extensively linked to NRF2 pathway activation28,29. Additionally, we observed other genes associated with promoting lung cancer progression, which exhibited positive associations with KEAP1, SMARCA4, and STK11 mutations in LUAD but not LUSC: S100P, which encodes a calcium-binding protein implicated in KEAP1/NRF2 signaling that regulates the mobility of lung cancer cells30,31, and SERPINB5, acting as a prognostic biomarker and promoter of proliferation in LUAD32.

These findings not only deepen our understanding of NRF2 pathway regulation in lung cancer but also highlight potential biomarkers and therapeutic targets for personalized treatment strategies in different molecular subtypes of LUAD and LUSC.

Identification and characterization of a LUSC-specific NRF2 signature

Furthermore, our analysis has revealed the upregulation of 25 genes (see Table 2) specific for LUSC patients harboring either KEAP1 or NFE2L2 mutations. Notably, this upregulation has not been directly observed in association with KEAP1 mutations in LUAD patients. Particularly noteworthy among these genes are two classes involved in detoxification processes: (i) the Glutathione S-transferase Mu (GSTM) gene family (GSTM2, GSTM3, and GSTM4), critical in eliminating electrophilic compounds by conjugating with glutathione33,34, and (ii) the UDP-glucuronosyltransferases family, essential for drug clearance through the glucuronidation process35,36. We will refer to the genes listed in Table 1 as LUSC NRF2 expression signature, to indicate the NRF2 target genes whose expression was identified by ASTUTE to correlate with the presence of NFE2L2 and KEAP1 mutations exclusively in LUSC.

Table 2 LUSC NRF2 expression signature
Full size table

As for the NRF2 identified expression signature common to LUAD and LUSC, some of the genes associated with KEAP1/NFE2L2 mutations selectively in LUSC also exhibit significant associations with other driver mutations. For instance, UGT1A6, ADH7, and NDRG4 genes are positively associated with TP53 mutations (log2FC > 0.5), while also ADAM23, UGT1A1, and GSTM3 show a positive association but a lower strength (log2FC < 0.5). Additionally, ADH7 correlates strongly (log2FC = 0.75) with KMTD2 mutations, and UGT1A7 and NDRG4 are associated with CDKN2A mutations (0.2 < log2FC < 0.5). ABCA4 and NELL1 gene expression is linked with EGFR mutations (0.2 < log2FC < 0.7) specifically in LUSC. These findings suggest that while LUAD and LUSC patients with KEAP1 or NFE2L2 mutations share a common NRF2 pathway activation signature, LUSC also exhibits a distinct NRF2 signature. Notably, many of these genes show increased upregulation in LUSC patients with TP53 mutations, suggesting a potential synergistic effect between these mutations.

Interestingly, in addition to canonical regulators, we identified several non-canonical alterations—such as CUL3 mutations, AKT2 amplifications, and PTEN deletions—that are associated with upregulation of NRF2 target genes in specific tumor types. These findings suggest that the NRF2 pathway may also be modulated through broader signaling mechanisms, including ubiquitination, phosphorylation, and metabolic reprogramming. We explored the impact of CUL3 mutations, detected in approximately 3–4% of both LUSC and LUAD patients, on the NRF2 signature. CUL3, a member of the E3 ubiquitin ligase complex, plays a role in NRF2 degradation. Notably, we observed the presence of numerous genes reported in LUAD/LUSC NRF2 expression signature or LUSC NRF2 expression signature (Supplementary Data 1 and 2) in LUSC patients carrying CUL3 mutations, whereas this signature was not corroborated in LUAD.

Impact of the NRF2 signatures across cancer types

We applied ASTUTE to other cancer types to investigate whether the NRF2 gene expression signatures identified in NSCLC were also observed in association with NFE2L2 or KEAP1 mutations at the pan-cancer level. We considered cancers where either KEAP1 or NFE2L2 mutations were present in at least 5% of the patients for a total of 2258 distinct samples. In particular, we selected HCC (361 samples), head and neck squamous cell carcinoma (HNSCC, 507 samples), uterine corpus endometrial carcinoma (UCEC, 515 samples), cervical squamous cell carcinoma (CSCC, 288 samples), bladder urothelial carcinoma (BLCA, 406 samples), and esophageal adenocarcinoma (EAC, 181 samples) from the Pan-Cancer Atlas studies37.

In HCC, KEAP1 and NFE2L2 mutations occur respectively in about 5% and 3% of the patients. Of note, the activation of the NRF2 pathway has been reported to induce tumor cells to immune escape in HCC38. Additionally, NRF2 was identified as a prognostic factor associated with decreased survival in HCC patients39. ASTUTE confirmed the correlation between the upregulation of many members of the NRF2 signature and NFE2L2 mutations in HCC (Supplementary Data 3). In particular, the six genes most upregulated (log2FC > 1.9) were NRF2 targets: AKR1B15, NQO1, AKR1B10, CABYR, TRIM16L, and CPLX2. Similar results were observed for KEAP1 mutations, which positive correlated (log2FC > 1.9) with CPLX2, CABYR, TRIM16L, TRIM16, AKR1B15, and AKR1B10.

In head and neck squamous carcinoma, KEAP1 and NFE2L2 mutations are observed, respectively, in about 4% and 5% of patients. It was reported that the activation of NRF2 signaling promotes the acquisition of resistance to cisplatin and metastasis in HNSCC40. ASTUTE correlated both mutations with the expression of many genes of the LUAD/LUSC NRF2 expression signature (Supplementary Data 4). Interestingly, these genes appear to be among the most upregulated ones in the presence of these mutations (log2FC > 1) in this cancer type. We also identified a strong correlation between KEAP1 and NFE2L2 mutations and the expression of genes identified in the LUSC NRF2 expression signature.

In CSCC, 1% and 6% of patients harbor, respectively KEAP1 and NFE2L2 mutations. The expression of NRF2 was found to be higher in CSCC patients with lymph node metastasis, and in addition NRF2 pathway was positively associated with epithelial to mesenchymal transition41.

KEAP1 and NFE2L2 mutations were identified to be strongly associated with both the expression of genes within the LUAD/LUSC NRF2 expression signature and the ones related to LUSC NRF2 expression signature (Supplementary Data 5).

About 4% and 8% of the patients with UCEC harbor respectively KEAP1 or NFE2L2 mutations. NRF2 overexpression was found to be associated with endometrial neoplasms with serous differentiation42. In this tumor type, ASTUTE was able to correlate these mutations to the expression of ALDH3A1, AKR1C2, GPX2, AKR1C1, NQO1, TRIM16L, CYP4F3, JAKMIP3, AKR1B10, AKR1C3, and GCLC (log2FC > 0.5) (Supplementary Data 6). Additionally, other LUAD/LUSC NRF2 signature genes were identified to correlate with a major impact with NFE2L2 mutations in comparison to KEAP1 mutations. While among the genes identified the LUSC NRF2 signature, we found only that CES1, and UGT1A6 consistently correlated with KEAP1/NFE2L2 mutations (log2FC > 0.7).

In BLCA, KEAP1 mutations are present in 2% of patients, while NFE2L2 mutations occur in 6%. NRF2 expression was associated with cisplatin resistance in BLCA43. Among the most upregulated genes in the context of NFE2L2 mutations (log2FC > 0.8) in bladder cancer, ASTUTE identified 35 genes belonging to the LUAD/LUSC NRF2 expression signature, and 8 genes belonging to the LUSC NRF2 expression signature (Supplementary Data 7). While for the KEAP1 mutations 13 genes of the LUAD/LUSC NRF2 signature, and 6 genes specific for the LUSC NRF2 signature (log2FC > 0.8) were found.

Finally, 3% of patients with esophageal adenocarcinoma (EAC) harbor KEAP1 mutations, while 10% of patients harbor NFE2L2 mutations. In EAC, NRF2 expression promotes tumor cells' survival, and, in addition, it was demonstrated that NRF2 has a protective role against stress-triggered apoptosis and ferroptosis44. ASTUTE revealed the upregulation of 35 genes of the LUAD/LUSC NRF2 signature in association with NFE2L2 mutations (log2FC > 0.9), and 13 genes of LUSC NRF2 signature (log2FC > 1) in the EAC (Supplementary Data 8). We were not able to find similar results for KEAP1 mutations, in which only the upregulation of KYNU and B4GALNT1 was found.

Overall, ASTUTE’s results demonstrate high consistency across cancer types and reveal a widespread expression of the identified NRF2 signatures at the pan-cancer level.

NRF2 expression as a prognostic biomarker at the pan-cancer level

To assess the prognostic implications of the identified expression signatures, we considered the overall survival (OS) data provided by the Pan-Cancer Atlas studies37. We initially focused on genes associated with KEAP1 or NFE2L2 mutations in LUAD and LUSC, filtering for log2 fold change values greater than 1 and less than −1. We applied the same filtering criteria to the six other considered cancer types as described before, and we narrowed down our analysis to the genes identified by ASTUTE in LUAD or LUSC and in these cancer types. Subsequently, we conducted standard univariate Cox regression analysis across all eight cancers (LUAD, LUSC, and the six other cancer types), considering the selected genes. Only genes exhibiting consistent correlation with risk, as indicated by ASTUTE, were considered. For instance, if ASTUTE suggested that KEAP1/NFE2L2 mutations positively impacted gene expression, such genes were classified as oncogenes, and its higher expression had to positively correlate with worse prognosis. This initial screening yielded a list of genes strongly associated with both ASTUTE results and prognosis. We then proceeded to conduct multivariate regularized Cox regression analysis, identifying 14 genes whose expression significantly correlated with negative prognosis in at least one of the analyzed cancers (Table 3 and Supplementary Data 9). Several of the 14 genes comprising the prognostic NRF2 signature, such as SRXN1 and CABYR, are known canonical NRF2 targets, while TRIM16 has been reported to modulate NRF2 activity via the p62-KEAP1 complex. Notably, these genes were identified through our mutation-centric framework independently of prior pathway annotation, reinforcing the biological validity of the ASTUTE-derived signature and its ability to recover functionally relevant targets. Importantly, this signature was derived through a de novo mutation-centric approach and validated across multiple cancer types, offering both mechanistic insight and potential clinical applicability.

Table 3 NRF2 prognostic biomarkers
Full size table

Interestingly, several of the identified genes have been reported as NRF2 target genes in the literature and belong to the two signatures identified in NSCLC. Notable examples include CABYR, which is upregulated in HCC and suggested as a cancer-testis antigen in lung cancer45. GCLM has been implicated as a tumor promoter and immunological biomarker in bladder cancer46 and linked to cisplatin resistance in NSCLC47. ME1 is associated with poor prognosis in HCC48 and breast cancer49, while NQO1 is significantly associated with prognosis, immune infiltrates, and drug resistance across multiple cancer types50. SRXN1 is identified as an inducer of hepatocellular carcinogenesis and metastasis, correlating with poor prognosis in HCC patients51. TXNRD1 is an unfavorable prognostic biomarker in HCC52, breast cancer53, and NSCLC54. Additionally, SPP1, associated with NFE2L2 mutations in LUSC, is reported as a prognostic biomarker in urothelial bladder cancer55 and ovarian cancer56. These findings highlight the relevance of these genes in NRF2 pathway activation and their potential as significant biomarkers across various cancer types.

We then used the hazard ratios estimated by the regularized Cox multivariate regression for the identified 14 genes to compute a risk score for each patient of the 8 cancer subtypes (see “Methods”). Hierarchical clustering was performed considering the computed risk scores, resulting in the classification of patients into two risk groups within each cancer subtype. Further Kaplan-Meier analysis confirmed significant differences in prognosis for the risk groups in all cancers. In particular, in BLCA (p = 0.0024), CSCC (p = 0.004), EAC (p = 0.035), HNSCC (p = 0.0082), HCC (p = 0.0013), LUAD (p = 0.017), LUSC (p = 0.035), and UCEC (p = 0.0019). Moreover, all the 14 genes were overexpressed in the patients with worse prognosis in all cancer types (Fig. 2).

Fig. 2: Impact of the NRF2 expression signatures based on the identified 14 prognostic genes across eight cancer types.
figure 2

Bladder Cancer (BLCA), Cervical Squamous Cell Carcinoma (CSCC), Esophageal Adenocarcinoma (EAC), Head and Neck Squamous Cell Carcinoma (HNSCC), Hepatocellular Carcinoma (HCC), Lung Adenocarcinoma (LUAD), Lung Squamous Cell Carcinoma (LUSC), and Uterine Corpus Endometrial Carcinoma (UCEC). We show boxplots displaying the expression levels of the 14 prognostic genes in patients categorized into two risk groups based on hierarchical clustering of computed risk scores. The higher expression levels of these genes are associated with the worse prognosis group across all cancer types. Kaplan–Meier survival curves illustrate significant differences in overall survival (OS) between the two risk groups within each cancer type. The p-values for each cancer type are as follows: BLCA (p = 0.0024, low-risk group n = 246, high-risk group n = 107), CSCC (p = 0.004, low-risk group n = 173, high-risk group n = 89), EAC (p = 0.035, low-risk group n = 124, high-risk group n = 37), HNSCC (p = 0.0082, low-risk group n = 38, high-risk group n = 448), HCC (p = 0.0013, low-risk group n = 229, high-risk group n = 102), LUAD (p = 0.017, low-risk group n = 336, high-risk group n = 117), LUSC (p = 0.035, low-risk group n = 383, high-risk group n = 45), and UCEC (p = 0.0019, low-risk group n = 220, high-risk group n = 247).

Full size image

We finally validated our findings on external cohorts by replicating the analysis using the LUAD dataset from Chen et al13. (Supplementary Fig. 4), primarily comprising patients with EGFR mutations, and the one from Pleasance et al.57 (Supplementary Fig. 5), comprising metastatic lung cancers. Hierarchical clustering was performed based on the risk scores computed from the 14 identified prognostic genes. Kaplan–Meier analysis confirmed the presence of two distinct risk groups. Although the two curves did not reach statistical significance in the dataset from Chen et al. possibly due to the overrepresentation of patients carrying EGFR mutations, which correlated with a limited number of patients in the cluster exhibiting the NRF2 signature and the shortest OS (p value = 0.064), the trend was consistent. Instead, in the dataset from Pleasance et al., including metastatic cancers, the discovered NRF2 prognostic signature identified two clusters with clearly different prognosis (p = 0.026). In the cluster with the worse prognosis, the 14 genes comprising the NRF2 signature were significantly upregulated, and a significantly higher frequency of KEAP1 mutations was observed (0% vs 33%, p value adjusted for false discovery rate = 0.027), hence validating our results.

Mutational landscape associated with the NRF2 prognostic signature

To further elucidate the mutational landscape underlying the identified NRF2 prognostic signature, we examined the clusters with significantly different survival across the considered eight cancer types. Patients were stratified into two groups in each cancer type—those with better prognosis and those with worse prognosis—based on their pan-cancer NRF2 prognostic signature scores. Consistently, we observed a significantly higher expression of the pan-cancer NRF2 prognostic signature in the worse prognosis clusters across all cancer types. Subsequently, we conducted a differential mutational analysis between the two survival-based clusters within each cancer type. Using a z-score test for proportions, corrected for false discovery rate (p < 0.05), we identified specific genetic alterations that were significantly enriched in the worse prognosis clusters. These alterations include mutations in key driver genes and other genetic events potentially contributing to the aggressive phenotypes observed (Fig. 3). Detailed results of the differential mutational analysis are provided in Supplementary Data 10.

Fig. 3: Differential mutation frequencies associated with the NRF2 prognostic signature across eight cancer types.
figure 3

In A we show barplots representing the frequency of KEAP1 and NFE2L2 mutations in patients with worse prognosis clusters compared to those with better prognosis clusters across the eight studied cancer types: Lung Adenocarcinoma (LUAD), Lung Squamous Cell Carcinoma (LUSC), Bladder Cancer (BLCA), Cervical Squamous Cell Carcinoma (CSCC), Esophageal Adenocarcinoma (EAC), Head and Neck Squamous Cell Carcinoma (HNSCC), Hepatocellular Carcinoma (HCC), and Uterine Corpus Endometrial Carcinoma (UCEC). KEAP1 and NFE2L2 mutations were particularly enriched in LUAD, LUSC, and EAC, indicating a significant role in activating the NRF2 pathway and contributing to worse prognosis in these cancer types. In B we show the aggregated frequency of other significant genetic alterations in the worse prognosis clusters compared to better prognosis clusters in the same eight cancer types. These mutations include, but are not limited to, STK11, ROS1, IDH2, ARID5B, TAF1, ADAMTS9, AFF1, MST1R, ZNF217, CDKN2A, TP53, ARID2, AKT2, AXIN1, and TRIM31. These alterations might suggest alternative mechanisms of NRF2 activation and regulation, highlighting the complexity of NRF2 pathway dysregulation across different cancers and emphasizing the necessity for diverse therapeutic strategies.

Full size image

In LUAD, mutations in KEAP1 (5.8% vs 53.4%) and NFE2L2 (2.1% vs 6.9%) are enriched in the cluster with higher NRF2 prognostic signature, thus playing a central role in activating the NRF2 pathway, potentially disrupting its regulatory mechanisms. This disruption likely leads to enhanced antioxidant responses, contributing to drug resistance and metabolic reprogramming that supports tumor cell survival under oxidative stress. Additionally, enriched mutations in STK11 (10.1% vs 26.7%) and ROS1 (4.3% vs 12.1%) may further contribute to NRF2 pathway activation, although the precise mechanisms require further investigation. Furthermore, we found enrichment of copy number gains in IDH2 (9.8% vs 22.4%) in the LUAD patients with high NRF2 expression, a gene potentially impacting tumor metabolism and survival58,59. Similarly, in LUSC, we observed mutations in KEAP1 (9% vs 23.3%) and NFE2L2. This activation enhances antioxidant defenses, which could support tumor progression. Concurrent observed enriched mutations in genes like ARID5B (2.2% vs 11.6%) and TAF1 (4.6% vs 16.3%) might augment NRF2 activity, facilitating tumor growth and resistance to therapeutic interventions60,61.

Other cancer types, such as BLCA, CSCC, and EAC, also exhibit NRF2 activation influenced by specific mutational profiles. In BLCA, mutations in NFE2L2 (2.9% vs 15.1%) are enriched, suggesting a potential enhancement of antioxidant defenses. Additionally, enriched copy number loss in PPARG (9.9% vs 20.8%) may influence NRF2 activity indirectly through its regulatory networks, affecting metabolic and stress response pathways62,63. In CSCC, enriched mutations in ADAMTS9 (0.6% vs 6.2%) and AFF1 (0% vs 6.2%) might augment NRF2-mediated antioxidant responses, aiding in cell survival under oxidative stress conditions. The enriched demethylation of CST6 (9.9% vs 21%) could potentially affect NRF2 regulation, although its specific interaction requires further investigation. In EAC, NFE2L2 (3.2% vs 37.8%) mutations are also enriched. Additionally, enriched mutations in MST1R (0% vs 8.1%) and ZNF217 (0.8% vs 10.8%) might suggest potential pathways through which NRF2 activity could be modulated, influencing cellular proliferation and survival mechanisms64.

In HNSCC patients with high NRF2 signature show higher frequencies of point mutations in CDKN2A (5.7% vs 22.4%) and TP53 (34.3% vs 74.4%). While in HCC, the enriched mutations in KEAP1 (1.4% vs 12.5%) and NFE2L2 (1.4% vs 8.3%), along with copy number loss in ARID2 (6.9% vs 20.8%), suggest a significant association with NRF2 activation and its regulatory pathways65,66.

Finally, in UCEC, several key genetic alterations are enriched, underscoring the potential implications of NRF2 pathway activation in this cancer. Mutations in TP53 (24.2% vs 51.5%) are known to disrupt redox homeostasis, leading to increased oxidative stress within cells. This can activate NRF2 as a protective mechanism, enhancing antioxidant defenses and promoting cell survival. Concurrently, amplification of AKT2 (6.6% vs 21.5%), a regulator of growth factor signaling, may directly phosphorylate NRF2, stabilizing its protein levels and promoting transcription of antioxidant genes67. Additionally, copy number loss in AXIN1 (5.7% vs 21.9%), associated to the activation of the Wnt signaling pathways, might suggest a potential indirect modulation of NRF2 activity through crosstalk mechanisms68. Furthermore, methylation of TRIM31 (10% vs 20.7%), through its direct interaction with NRF2, requires further elucidation, likely influences gene expression profiles involved in oxidative stress responses69. Together, these genetic alterations in UCEC collectively highlight diverse pathways through which NRF2 overactivation may contribute to tumor progression, emphasizing the need for targeted therapeutic strategies aimed at disrupting NRF2-dependent oncogenic processes.

Overall, while NRF2 pathway activation appears to be a common feature across multiple cancer types, the specific genetic alterations influencing its activity can vary. These findings highlight the complexity of NRF2-mediated mechanisms in cancer and underscore the need for further research to elucidate these pathways fully, potentially informing targeted therapeutic strategies aimed at disrupting NRF2-dependent oncogenic processes.

Discussion

In this study, we employed the ASTUTE framework to conduct a comprehensive analysis of KEAP1 and NFE2L2 mutations in NSCLC, focusing on their influence on NRF2 expression and their potential as prognostic biomarkers. Our findings suggest a significant association between genetic alterations in the KEAP1/NRF2 pathway and NSCLC progression, offering valuable insights into the underlying mechanisms of pathogenesis and drug resistance.

Our analysis revealed a robust association between KEAP1/NFE2L2 mutations and the upregulation of NRF2 target genes critical in antioxidant response, detoxification, and metabolism. Notably, KEAP1 mutations were prevalent in both LUAD and LUSC subtypes, whereas NFE2L2 mutations were more frequent in LUSC. This highlights the necessity of subtype-specific considerations in understanding NRF2’s role in NSCLC and suggests implications for personalized therapeutic interventions.

Furthermore, we identified specific gene signatures associated with KEAP1/NFE2L2 mutations that could serve as prognostic biomarkers in NSCLC. For instance, upregulation of the GSTM family members and UDP-glucuronosyltransferases in LUSC patients harboring these mutations suggests enhanced detoxification capacity, potentially influencing treatment response and outcomes. Moreover, genes selectively upregulated in LUAD or LUSC due to KEAP1 mutations, such as CPLX2 and KYNU, underscore the subtype-specific activation of NRF2 and its implications for disease prognosis.

We also unveiled intriguing associations between NRF2 expression signatures and other key genetic alterations, notably TP53 mutations. Several genes upregulated in LUSC patients with KEAP1/NFE2L2 mutations showed positive associations with TP53 mutations as well, hinting at potential crosstalk between the NRF2 and TP53 pathways in driving tumorigenesis and drug resistance. Understanding these interactions could unveil novel therapeutic targets and combination strategies for NSCLC treatment.

Importantly, we conducted a comprehensive survival analysis revealing a set of 14 genes consistently associated with prognosis across different cancers, underscoring the NRF2 signature’s robustness as a prognostic biomarker at the pan-cancer level.

While our pan-cancer analysis demonstrates that the NRF2 expression signature identified by ASTUTE is consistently associated with prognosis across diverse tumor types, we acknowledge that NRF2 signaling might be context- and tissue-specific. Previous studies have shown that super-enhancer regulation might play a critical role in modulating NRF2 transcriptional activity, with implications for lineage-restricted gene expression programs. Additionally, hypoxia and other tumor microenvironmental factors are known to influence NRF2 activity and its downstream effects. Although our current analysis does not incorporate enhancer-state data or microenvironmental features such as oxygen tension, the reproducibility of our mutation-derived signature across squamous and adenocarcinoma subtypes suggests that it captures a broadly conserved core of NRF2 pathway activation. Future studies incorporating chromatin accessibility and single-cell resolution datasets could further refine the context-specific components of NRF2 signaling.

Several NRF2-related gene expression signatures have been proposed in previous studies, often derived via clustering or curated pathway annotations. In contrast, our work builds a prognostic signature through a mutation-centric approach that directly links KEAP1/NFE2L2 alterations to expression profiles, enabling mechanistic attribution. The identification of well-established NRF2 targets such as SRXN1 and CABYR, as well as NRF2 modulators like TRIM16, within our expression signature, derived solely from mutation-based associations, supports the capacity of ASTUTE to capture mechanistically grounded transcriptional programs without relying on prior pathway knowledge. This methodological distinction enhances the mechanistic interpretability and relevance of the signature for precision oncology.

The mutational landscape associated with the NRF2 prognostic signature revealed specific genetic alterations across various cancer types that significantly enrich NRF2 activity, promoting tumor progression, drug resistance, and metabolic reprogramming. These findings not only deepen our understanding of KEAP1 and NFE2L2 mutations in NSCLC but also hold promise for guiding therapeutic strategies and advancing precision oncology tailored to individual patient subtypes. Further validation in external patient cohorts is warranted to substantiate these observations and elucidate the molecular mechanisms driving NRF2 activation in NSCLC.

It is important to note that in certain tumor types, such as EAC, KEAP1 and NFE2L2 mutations are relatively infrequent. This low prevalence poses challenges in terms of statistical power and increases the risk of spurious associations. To address this, we applied bootstrapped regularized regression and false discovery rate (FDR) correction to reduce false positives. Moreover, gene-level associations identified in low-frequency mutation contexts were interpreted cautiously and primarily emphasized when supported by consistent trends across multiple datasets or cancer types. This limitation underscores the need for validation in larger or prospectively curated cohorts.

Another limitation of our study is its reliance on retrospective, publicly available datasets, which may introduce confounding variables such as heterogeneity in sequencing technologies, treatment regimens, and clinical annotations. These sources of variability may affect the generalizability of our findings and should be carefully considered when interpreting the results.

Moreover, while our 14-gene NRF2 prognostic signature demonstrates robust associations with patient outcomes across multiple cancer types, its clinical utility must ultimately be evaluated in the context of established biomarkers such as PD-L1 expression and tumor mutational burden. Unfortunately, the retrospective datasets used in this study lack standardized and complete annotations for these variables, limiting our ability to perform comparative multivariate Cox regression models. Future prospective studies with harmonized clinical, immunological, and molecular profiling will be essential to assess the incremental prognostic value and potential complementarity of our NRF2 signature relative to existing biomarkers. Such efforts will also help determine its cost-effectiveness and practical feasibility for integration into clinical decision-making.

In conclusion, our study demonstrates the clinical relevance of NRF2 activation pathways in NSCLC and highlights the potential for personalized treatment approaches based on subtype-specific molecular profiles. By elucidating the complex interplay between genetic alterations and the NRF2-driven pathways, our findings pave the way for innovative therapeutic avenues and enhance our broader understanding of cancer biology and precision medicine.

Methods

Input datasets

We considered genomic and transcriptomic data from five distinct datasets comprising NSCLC patients, specifically LUAD and LUSC. For LUAD, datasets included genomic profiles from the CPTAC (110 samples)11, the Pan-Cancer Atlas (510 samples)12, and Chen et al. (169 samples)13, totaling 789 patients. LUSC datasets encompassed samples from CPTAC (108 samples)14 and the Pan-Cancer Atlas (481 samples)14, comprising 589 patients. To minimize batch effects, each dataset was analyzed independently using ASTUTE to identify associations between somatic mutations and gene expression profiles, without direct data merging or cross-cohort normalization. This ensured that dataset-specific technical variability did not confound the genotype-to-expression associations.

Additionally, ASTUTE was also utilized to investigate the pan-cancer implications of NRF2 gene expression signatures in cancers where KEAP1 or NFE2L2 mutations were observed in at least ≥5% of the patients, including HCC, HNSCC, UCEC, CSCC, BLCA, and EAC from the Pan-Cancer Atlas studies37.

To validate our findings, we conducted a further analysis using the LUAD datasets from Chen et al.13 and from Pleasance et al.57. The data are publicly available and were downloaded from cBioPortal70.

The ASTUTE framework

ASTUTE is a computational framework designed to integrate somatic mutation data with gene expression profiles to elucidate their functional implications in cancer biology. It employs a regularized regression model using the LASSO penalty within a linear regression framework10. The LASSO penalty introduces a regularization parameter (lambda), which is optimized through k-fold cross-validation, typically with k = 10, to control model complexity, promote feature selection, and minimize prediction error. In particular, the regularization parameter used in the LASSO model by ASTUTE was selected via 10-fold cross-validation. All regularized regression procedures were implemented using the glmnet R package.

To enhance robustness, ASTUTE repeats this procedure multiple times (e.g., 100 iterations) using bootstrap resampling of the input data. The algorithm is reapplied in each iteration, and the final results are computed through bootstrap aggregation. This approach enables ASTUTE to identify gene expression features most significantly associated with somatic mutations in driver genes. Additionally, ASTUTE estimates baseline expression levels for each gene and calculates fold changes to quantify the impact of specific mutations. Finally, the fold changes computed at each bootstrap iteration are used to estimate whether they are significantly lower or greater than 1, enabling the calculation of p values that provide confidence estimates of the associated mutations either increasing or decreasing the expression of specific genes. In particular, for each gene, p values were derived from the distribution of fold changes obtained across bootstrap iterations. These p values were adjusted for multiple testing using the Benjamini-Hochberg FDR correction method. The R glmnet packages was used to implement the regularized regression procedures.

Survival analysis

Prior to survival modeling, we applied a preliminary log2 fold-change threshold (±1) as a conservative filter to retain genes with biologically meaningful expression changes associated with KEAP1/NFE2L2 mutations. This step reduced noise and focused the analysis on robust candidates. We then employed regularized Cox regression with LASSO penalty to investigate the prognostic impact of gene expression profiles on OS in various cancer types. We utilized the R glmnet package to implement regularized Cox regression. This technique facilitates variable selection by shrinking the coefficients of the less relevant predictors to zero, thereby highlighting the most significant associations between gene expression and survival outcomes. Cross-validation was employed to determine the optimal penalization parameter (lambda) that minimized prediction error and improved the robustness of our survival predictions. For each patient, we calculated risk scores based on the estimated hazard ratios of gene expression changes identified through LASSO regression. These risk scores were used to stratify patients into different risk groups, enabling the identification of high-risk subgroups with poorer survival outcomes. The clustering analysis was performed by hierarchical clustering using the dynamic tree cut approach71 implemented by the dynamictreecut R package to group patients based on their risk scores. Kaplan–Meier survival curves were generated to assess the differences in survival between risk groups, and statistical significance was determined using the log-rank test.

Mutations enrichment analysis

We conducted an enrichment analysis to assess differences in the proportions of mutation occurrences across predefined clusters. We consider the clusters obtained from the survival analysis and computed the proportion of mutations in the set of genes of interest. To determine statistical significance, we employed a z-test to compare these proportions, adjusted for multiple comparisons using false discovery rate correction. Genes were considered significantly enriched if they exhibited a q value < 0.05, indicating significance after controlling for false discovery. We further assessed the alterations that can directly influence gene expression, such as copy number gains and demethylations, which can result in increased gene expression and copy number losses and methylations, leading to decreased gene expression. To validate their impact on gene expression levels, we conducted t-tests (p < 0.05) to confirm the significant impact of the enriched alterations to expression. This approach allowed us to systematically evaluate and compare mutation patterns across distinct clusters associated to the NRF2 pathway expression.

RNA extraction, reverse transcription, and quantitative real-time PCR

H2228 and H3122 NSCLC cell lines were cultured in biological triplicates and seeded at a density of 3 × 106 cells per T75 flask. Cells were harvested at ~70% confluence. Total RNA was extracted using the RNeasy Mini Kit (QIAGEN, Germany) following the manufacturer’s instructions. Two micrograms of RNA were reverse-transcribed using the LunaScript RT SuperMix (Euroclone, Italy) in a final reaction volume of 40 µL. Quantitative PCR was performed in technical duplicates using 2 µL of cDNA and 2X Mastermix (GeneSpin, Italy) on a QuantStudio™ Real-Time PCR system (Life Technologies). Gene expression levels were normalized to GAPDH and calculated using the 2−ΔCt method. Target genes included CYP4F11, AKR1C1, AKR1C2, AKR1C3, AKR1B10, CYP4F3, GPX2, CABYR, JAKMIP3, and UCHL1, using commercially available TaqMan assays.

Software and statistical analysis

ASTUTE is available as an R package and can be installed from GitHub (https://github.com/ramazzottilab/ASTUTE). All computational analyses were conducted using R (version 4.4.1). Key R packages included glmnet for the regularized regression analyses, dynamicTreeCut for the clustering analysis, and the survival package for survival analysis. Statistical significance was determined at p < 0.05, unless otherwise specified.