Lung cancer is the leading cause of cancer-related deaths worldwide. Lung adenocarcinoma (LUAD) accounts for approximately 50% of all cases [1]. Epidermal growth factor receptor (EGFR)-mutant non-small cell lung cancer (NSCLC) patients respond significantly to EGFR-tyrosine kinase inhibitors (TKIs). This illustrates a major success for precision medicine [2]. Despite initial efficacy, acquired resistance to EGFR-TKIs ultimately drives disease progression, even with sequential first- to third-generation regimens [3]. Well-established resistance mechanisms to all three generations of EGFR-TKIs include secondary EGFR mutations, cellular-mesenchymal epithelial transition factor (c-MET) amplification, and lineage transformation [4]. In contrast, the mechanisms underlying clinical drug tolerance remain elusive. Accumulating evidence supports the superior efficacy of combination targeted therapies for achieving more durable clinical responses [5]. Therefore, deciphering the molecular mechanisms of acquired resistance and identifying predictive biomarkers are imperative to optimizing therapeutic strategies and improving survival outcomes in NSCLC.

Cellular redox homeostasis is maintained by regulating the production of reactive oxygen species (ROS) and the activity of antioxidant defense systems [6]. Disrupted redox homeostasis is a cancer hallmark, regulated by conserved redox-related genes (RRGs) [7, 8]. Unlike normal cells, cancers exhibit a dysregulated redox state with concurrent ROS elevation and enhanced antioxidant defense, with a consequence of metabolic reprogramming, genomic instability, mitochondrial dysfunction, and tumor microenvironment (TME) adaptations [9, 10]. ROS play dual and context-dependent roles in carcinogenesis. In NSCLC, this redox interplay influences progression, drug resistance, and apoptosis. Here, differentially expressed RRGs are key regulators, particularly of resistance mechanisms, making them promising therapeutic targets to overcome NSCLC resistance, metastasis, and recurrence [11, 12, 13]. However, the prognostic role and functional mechanisms of RRGs in NSCLC lack clarity.

Single-cell RNA sequencing (scRNA-seq) deciphers tissue heterogeneity by categorizing cells, advancing precision medicine [14]. Beyond defining cell types, it precisely defines gene function within heterogeneous populations to reveal cancer-promoting pathways [15]. Leveraging these advantages, studies increasingly identify novel biomarkers by integrating scRNA-seq with bulk transcriptome data [16, 17]. Given the complex landscape of therapy resistance, we combined these approaches to develop a prognostic signature for improving NSCLC outcome prediction and guiding therapy.

Using single-cell resolution mapping, we identified differentially expressed redox-related genes (DE-RRGs) in malignant cell clusters of osimertinib-resistant LUAD. Integration with bulk transcriptomics established a six-gene prognostic signature (AGER, CYP2J2, FMO2, HSPA1B, SOD3, VASN). Immune profiling and molecular docking revealed the signature’s therapeutic potential, which was confirmed by the upregulated DE-RRG expression we observed in resistant cells. Crucially, SOD3 knockdown restored osimertinib sensitivity in vitro, and lower SOD3 levels were clinically associated with sensitivity to third-generation EGFR-TKIs. Our work defines clinically actionable biomarkers and reveals novel therapeutic vulnerabilities in treatment-resistant LUAD.

Osimertinib, a third-generation EGFR-TKI, serves as the first-line treatment for EGFR-mutant LUAD. However, acquired resistance remains the primary cause of treatment failure [23, 24]. Early detection and effective counterstrategies are crucial to improve patient outcomes. Thus, identifying novel biomarkers that predict prognosis and therapeutic benefits in osimertinib-treated LUAD patients is critically needed.

Oxidative stress, driven by ROS generation, promotes cancer progression and treatment resistance by disrupting genomic stability and altering signaling pathways [25, 26]. Moderate ROS levels can activate key oncogenic pathways, such as nuclear factor kappa-B (NF-$\kappa$B), phosphatidylinositol 3 kinase (PI3K)/AKT, and p53, while high ROS levels sensitize tumor cells to oxidative death. Many conventional and targeted therapies exert their effects by elevating intracellular ROS. However, cancer cells counteract this through robust antioxidant systems like nuclear factor erythroid-derived 2-like 2/kelch-like ECH-associated protein 1 (NRF2/KEAP1) and B cell lymphoma 2 (BCL2), which enhance survival and confer treatment resistance [27, 28]. Targeting oxidative stress has thus become a major therapeutic goal. Notably, redox dysregulation persisted even in models with canonical resistance mechanisms like EGFR-T790M, suggesting that RRG-mediated resistance may complement known pathways [29]. This persistence creates synergistic targeting opportunities, particularly through combining osimertinib with RRG inhibitors alongside MET or EGFR inhibitors, irrespective of the primary resistance mechanism. Tumor heterogeneity further underscores the need to map resistance mechanisms comprehensively. Single-cell RNA sequencing provides a powerful tool to dissect LUAD heterogeneity [30], uncover how redox balance shapes EGFR-TKIs resistance, and identify novel biomarkers.

We analyzed scRNA-seq data from matched OS- and OR-PDX models. The OR group exhibited a higher abundance of CAFs, corroborating prior reports [30]. Cell communication analysis revealed a remodeled interaction network in resistant tumors, with strengthened contacts among Tregs, epithelial cells, CAFs, and tumor cells, notably via the peptidylprolyl isomerase A-basigin (PPIA-BSG) axis, which Liu et al. [31] implicated in lung cancer metastasis. CAFs displayed especially enhanced crosstalk with multiple cell types in OR samples. Key hyperactivated ligand-receptor pairs included midkine-syndecan 1 (MDK-SDC1), known to induce immunosuppressive macrophages [32, 33], and WNT7B-(FZD2 + LRP5), which promotes tumor progression [34].

We also identified 10 DE-RRGs and assigned RRG activity scores to individual cells. High-scoring cells localized predominantly to CAF clusters, consistent with established roles of CAFs in driving osimertinib resistance through secreted factors like IL-6, periostin, hepatocyte growth factor (HGF), and insulin-like growth factors (IGFs) [35, 36, 37, 38]. Given substantial CAF heterogeneity—including inflammatory CAFs (iCAFs) that recruit immune cells via Eotaxin-1 (CCL11)/Chemokine (C-X-C Motif) Ligand 12 (CXCL12) and matrix-producing CAFs (mCAFs) that overexpress ECM-remodeling genes (like matrix metalloproteinase 11 (MMP11) and collagen type I alpha 1 (COL1A1)) [39]—we classified CAFs into six distinct subtypes. Pseudotime analysis indicated that panCAFs arise early and may differentiate into other CAF subtypes, paralleling findings in cholangiocarcinoma [40]. During this process, ABCA1, FMO2, and KLF2 were downregulated, while CYP2J2, SOD3, and VASN were upregulated, suggesting these RRGs help regulate CAF differentiation. We propose that these DE-RRGs operate during early CAF differentiation to enhance intercellular communication, thereby fostering metastasis, resistance, and disease progression. CAFs, as central components of the mesenchymal tumor microenvironment, may thus supply biomarkers, such as specific RRGs or their products, for predicting osimertinib resistance. Further studies should clarify the functional mechanisms of specific CAF subpopulations and RRGs in acquired resistance.

Using parallel scRNA-seq and bulk transcriptome data, we constructed a robust prognostic model based on six DE-RRGs. This model effectively stratified LUAD patients into distinct survival groups. Differential pathway activation revealed potential mechanisms underlying these survival disparities. High-risk scores associated with an immunosuppressive, tumor-progressive microenvironment, whereas low-risk scores indicated a more tumor-suppressive milieu. Incorporating TMB further improved prognostic accuracy. Mutation profiles also differed significantly between risk groups, suggesting these patterns could help guide personalized therapeutic strategies for LUAD patients.

Our findings reveal that these genes above are significantly up-regulated to varying degrees in LUAD-OR cell lines, indicating possible role in acquired EGFR-TKIs resistance. AGER binds toll-like receptor 4 (TLR4) and mediates AGE-RAGE signaling, contributing to diabetic angiopathy and thymic hyperplasia. Previous studies indicate that AGER overexpression suppresses proliferation, invasion, and migration in H1299 cells, suggesting a protective role in LUAD with significant prognostic value [41]. However, we found AGER positively correlates with M2 macrophage infiltration (p 0.001), a key driver of immunosuppressive TME. High AGER expression in LUAD-OR cells further suggests it might promote osimertinib resistance through cell crosstalk. CYP2J2 could epoxidize arachidonic acid into oncogenic epoxyeicosatrienoic acids (EETs). Kojima et al. [42] demonstrated that several TKIs suppress NSCLC growth and proliferation by inhibiting CYP2J2 activity. Consequently, the observed upregulation of CYP2J2 in OR cells may represent a compensatory mechanism driving therapeutic resistance. FMO2, a flavin-containing monooxygenase (FMO), acts as both a tumor suppressor that reduces paclitaxel resistance in LUAD [43] and a tumor promoter that facilitates macrophage infiltration in ovarian cancer [44]. FMO2 significantly correlated with M2 macrophage infiltration in our study, suggesting it might promotes osimertinib resistance in LUAD via immunosuppression. This reveals a paradoxical pro-tumor role contrasting its known protective functions. HSPA1B encodes HSP70 chaperones that maintain proteostasis through protein folding. These proteins promote tumor progression and poor outcomes in cancers including lung cancer. Notably, HSPA1B activation degrades EGFR, potentially delaying TKIs resistance [45, 46]. Paradoxically, LUAD-OR cells showed elevated HSPA1B mRNA versus parental lines, indicating HSPA1B may drive osimertinib resistance through EGFR-independent pathways like PI3K-mediated inflammatory upregulation. VASN, primarily expressed in vascular smooth muscle cells, critically regulates neointimal formation post-vascular injury. Its soluble form promotes tumorigenesis by inhibiting transforming growth factor-beta (TGF-$\beta$) signaling, modulating epithelial-mesenchymal transition, and enhancing angiogenesis. Functioning as an oncogene, VASN drives cancer cell proliferation and invasion in multiple malignancies including colorectal cancer and LUAD [47, 48]. We observed VASN overexpression in OR versus parental cells, indicating its potential significance role in EGFR-TKI resistance. Unlike other superoxide dismutase (SOD) isoforms, SOD3 localizes primarily to vascular smooth muscle cells and fibroblasts, with peak expression in lungs and blood vessels [49]. Studies have shown that demonstrated that elevated SOD3 levels in cancer cells confer resistance to diverse therapeutic strategies, such as chemotherapy, thereby promoting disease progression and correlating with poorer prognosis in NSCLC patients [49, 50]. Consistently, we noted increased SOD3 in OR cell models and NSCLC patients with third-generation EGFR-TKI resistance. Meng et al. [51] showed ROS accumulation reverses osimertinib resistance in NSCLC by inducing apoptosis. Our findings confirm this mechanism that SOD3 suppression overcomes resistance through ROS accumulation, consistent with its ROS-scavenging function. Understanding the functions of these prognostic genes and the factors that contribute to their negative prognostic significance will be instrumental in formulating combination therapies designed to overcome osimertinib resistance in NSCLC patients.

Drug repurposing represents an innovative therapeutic strategy enabled by deepening mechanistic insights into disease pathogenesis. Exemplified by compounds like EGCG, originally anti-inflammatory now investigated in oncology, this paradigm informed our targeted screening of prognostic genes to overcome osimertinib resistance and improve LUAD outcomes. Rotenone binds AGER to downregulate its expression while inhibiting mitochondrial complex I, activating NADPH-Oxidase 2 (NOX2)-mediated ROS generation that induces apoptosis in lung cancer cells [52]. This oxidative phosphorylation (OXPHOS) disruption may counter osimertinib adaptive resistance in NSCLC [53], though mechanisms require further elucidation. MRK003, a potent $\gamma$-secretase inhibitor, significantly reduced CYP2J2 expression while inhibiting Notch signaling through suppression of Notch intracellular domain generation. This Notch pathway blockade compromised tumor stemness by impairing self-renewal capacity, chemoresistance, invasive/migratory potential, and in vivo tumorigenicity [54]. Consequently, Notch inhibition via MRK003 represents a promising therapeutic strategy to improve survival in LUAD patients [55]. Lutein, a prominent xanthophyll carotenoid found in pigmented fruits, such as kale and spinach [56], demonstrated high-affinity docking with FMO2. Epidemiologically, lutein-rich diets correlate with reduced breast and gastric cancer incidence [57, 58]. Mechanistically, lutein suppresses NSCLC tumors by inducing DNA damage and activating the ATR/Chk1/p53 pathway [59]. MTX demonstrated strong binding affinity among compounds screened for HSPA1B mRNA suppression. This dihydrofolate reductase inhibitor depletes tetrahydrofolate pools essential for DNA synthesis and cellular replication [60]. Clinically, concomitant intrathecal MTX and oral afatinib (second-generation TKI) induced more than 10-month clinical remission in an osimertinib-resistant NSCLC patient [61]. Sunitinib demonstrated high-affinity binding to VASN in docking studies, effectively downregulating its expression. As a multi-targeted TKI inhibiting VEGFR and PDGFR, sunitinib exerts broad cytotoxic and anti-angiogenic effects across malignancies through receptor-specific mechanisms. Critically, it exhibits significant synergism with docetaxel against EGFR TKI-resistant NSCLC variants like T790M mutation [62], though further evidence is required to validate its osimertinib-enhancing potential. Our study demonstrated that the natural compound EGCG has a strong affinity for binding to SOD3. As the main biologically active component of green tea, EGCG is part of the polyphenol group and is well-known for its cancer-fighting properties, anti-inflammatory effects, and antioxidant capabilities. EGCG influences glucose metabolism and suppresses the AMPK/AKT/MAPK signaling pathway, which contributes to decreased resistance to osimertinib in NSCLC [63, 64]. However, these compounds might exist significant toxicities, such as rotenone causes mitochondrial dysfunction, sunitinib induces cardiovascular side effects, MRK003 promotes tumor neoangiogenesis, and MTX impairs hepatic/renal function. These may severely impair patient quality of life. Therefore, preclinical assessment remains essential to prevent treatment failure or adverse effects during repurposing [52, 54, 61, 62]. Although mechanisms underlying small-molecule enhancement of osimertinib efficacy remain unclear, acquired resistance to EGFR-TKIs persists as a major clinical challenge in EGFR-mutant NSCLC. Our current docking results provide a theoretical basis for the prioritization of compounds. To address this, we propose combining osimertinib with targeted small-molecule inhibitors as promising therapeutics for resistant NSCLC.

We integrated single-cell RNA sequencing to identify osimertinib resistance mechanisms at a subpopulation level, overcoming the resolution limitations of bulk-tissue analysis. To our knowledge, this study establishes the first gene signature specifically predictive of osimertinib resistance in NSCLC, distinct from previous cuproptosis-related prognostic models [65]. Furthermore, we provide novel functional insights into the therapeutic potential of SOD3—a key signature gene—through combined molecular docking and experimental validation, moving beyond purely predictive correlations. However, several limitations should be noted. First, although we show that CAF–tumor cell interactions promote osimertinib resistance, CAF heterogeneity necessitates further subtype-specific investigation using spatial transcriptomics at tumor–CAF niches and co-culture systems with purified CAF subsets. Second, while transcriptomic and functional data suggest metabolic adaptation contributes to resistance, direct validation through metabolic flux assays or lipidomics is needed to quantify changes in ATP generation, glycolysis, and lipid metabolism. Current single-cell datasets remain constrained by limited sample sizes and follow-up data due to high costs. Integrating bulk and single-cell RNA-seq data with advanced computational algorithms will help overcome these constraints. Although RRGs inhibition restored osimertinib sensitivity in vitro, future studies should include in vivo validation using shRNAs and osimertinib co-treatment in PDX models. Finally, although we identified RRGs mediating resistance, common genomic alterations in LUAD like TP53 mutations may influence redox signaling. Future work should stratify patients by mutation status to refine prognostic biomarkers, prospectively validate the signature’s clinical utility and investigate underlying molecular mechanisms.

Integrating single-cell and bulk transcriptome analyses, we identified key DE-RRGs as prognostic biomarkers and potential therapeutic targets in LUAD. Our results establish a critical link between redox homeostasis, osimertinib-resistant cell population, and patient survival outcomes. These findings provide novel mechanistic and therapeutic insights for treating lung adenocarcinoma.

LUAD, lung adenocarcinoma; TME, tumor microenvironment; OR, osimertinib resistance; RRGs, redox-related genes; CAFs, Cancer-associated fibroblasts; NSCLC, non-small cell lung cancer; EGFR, epidermal growth factor receptor; TKIs, tyrosine kinase inhibitors; ROS, reactive oxygen species; scRNA-seq, single-cell RNA sequencing; DE-RRGs, differentially expressed redox-related genes; PDX, patient-derived xenograft; TCGA, cancer genome atlas; GEO, gene expression omnibus; t-SNE, t-distributed stochastic neighbor embedding; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; GO, gene ontology; KEGG, kyoto encyclopedia of Genes and Genomes; ssGSEA, single-sample gene set enrichment analysis; TMB, tumor mutational burden; CNVs, copy number variations; IHC, immunohistochemistry; DEGs, differentially expressed genes; OS, osimertinib sensitivity; MTX, methotrexate; EGCG, epigallocatechin gallate; TLR4, toll-like receptor 4; AGER, advanced glycosylation end product-specific receptor‌; CYP2J2, cytochrome P450 family 2 subfamily J polypeptide 2; HSPA1B, heat shock 70 kDa protein 1B; FMO2, flavin-containing monooxygenase 2; VASN, vasorin; SOD3, superoxide dismutase 3; KLF2, kruppel like factor 2‌; ABCA1, ATP-binding cassette transporter A1; GSTA4, Glutathione S Transferase A4; AGE-RAGE, advanced glycation end products-receptor for advanced glycation end products; mesCAFs, mesothelial CAF; UMAP, Uniform Manifold Approximation and Projection; NF-$\kappa$B, nuclear factor kappa-B; PI3K, phosphatidylinositol 3 kinase; NRF2/KEAP1, nuclear factor erythroid-derived 2-like 2/Kelch-like ECH-associated protein 1; BCL2, B cell lymphoma 2; AKT, protein kinase B; FAK, focal adhesion kinase; PPIA-BSG, peptidylprolyl isomerase A-basigin; MDK-SDC1, midkine-syndecan 1; HGF, hepatocyte growth factor; IGFs, insulin-like growth factors; CCL11, Eotaxin-1; CXCL12, Chemokine (C-X-C Motif) Ligand 12; MMP11, matrix metalloproteinase 11; COL1A1, collagen type I alpha 1.

The data generated or analysed to support the findings during this study are available from the TCGA (https://portal.gdc.cancer.gov/) and GEO (https://www.ncbi.nlm.nih.gov/geo/) open database. The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

FQ designed the study. CF and DL made contribution to the data analyses and drew up the manuscript. CL and RQ finished statistical analysis for the data. GZ, FY and CS conducted molecular experiments. JD and TL processed figures and tables. FQ undertook support for the project. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

The information utilized in this research was sourced from publicly available databases. The samples were derived from the pathological specimen bank of the First Affiliated Hospital of Nanchang University [Ethics Approval Number: (2024) CDYFYYLK (10-005)] and abided by the rules of the Declaration of Helsinki. All patients or their guardians signed informed consent forms.

This study was funded by the National Natural Science Foundation of China (No. 82460598), Surface project of the Natural Science Foundation of Jiangxi Province (No. 20212ACB206029), Technology Supporting Program of Jiangxi Province (No. 2021-NCZDSY-009), Major science and technology research projects of Nanchang (No. 2022-KJZC-028).

The authors declare no conflict of interest.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/FBL40454.