Lewis lung carcinoma (LLC) cells (gift from C. Li, Chinese University of Hong Kong) were confirmed mycoplasma-negative (Beyotime, China) and cultured in Dulbecco’s Modified Eagle Medium (Gibco, USA) supplemented with 10% fetal bovine serum (Gibco, Brazil) and 1% penicillin/streptomycin (Fdbio science, China) at 37°C with 5% CO2.

Female C57BL/6N mice (5–6 weeks old, Zhejiang Weitong Lihua) were housed in a specific pathogen-free setting (23 ± 1°C, 55% humidity, 12-hour light/dark cycle) with ad libitum access to irradiated chow and autoclaved water. All procedures complied with ARRIVE guidelines and were approved by the Animal Ethics Committee of Shenzhen Third People’s Hospital (Protocol #2024-001).

Radiation delivery was performed using an X-RAD SmART Irradiator (Precision X-ray, North-Branford, CT, USA), equipped with cone-beam computed tomography (CBCT) capabilities, delivering 225-kV photon beams at a dose rate of 4.62 Gy/min. Mice were randomly assigned to receive either sham irradiation (0 Gy) or 24 Gy in three daily fractions (8 Gy/fraction) to the left lung.

Six days post-radiation, 3.75×10⁴ LLC cells in 25 μL PBS/Matrigel (ABW, Shanghai, China; 1:1 v/v) were stereotactically injected into the left lung parenchyma. Tumor growth was confirmed by computed tomography (CT) scan at day 6.

Anti-mouse PD-1 (Bio X Cell, West Lebanon, NH, USA) was administered intraperitoneally (200μg/mouse, i.p., twice weekly) starting at day 7 post-implantation. Tumor volume (V) was calculated using the formula V = π/6 × L × W × H from CT scan at day 14 and 21, based on the tumor’s maximum length (L), width (W), and height (H) (16).

RP was assessed at 7 days post-tumor implantation (pre-immunotherapy). Left lungs were fixed in 10% neutral-buffered formalin (Solarbio, China) for 24 hours, paraffin-embedded, sectioned at 4μm thickness, and stained with hematoxylin & eosin (H&E). RP severity was graded by two board-certified pathologists (blinded to treatment groups) based on the percentage of lung parenchyma affected: 1 score, less than 25%; 2 score, 25-50%; 3 score, more than 50% (8).

Neutrophil quantification was performed post-treatment completion through standardized Ly6G immunohistochemistry. Tissue sections underwent EDTA-based antigen retrieval (pH 9.0, 98°C), sequential blocking (3% H₂O₂-methanol/5% BSA), overnight anti-Ly6G antibody incubation (Servicebio GB11229, 1:400, 4°C), and PV-9000/DAB detection (Zhongshan Goldenbridge, China) with hematoxylin counterstain. Five random fields of view were photographed at 40x, and quantitatively analyzed for Ly6G-positive area percentage using ImageJ 1.54m software.

Arginase 1 (Arg1) is a well-established marker of the pro-tumor phenotype of neutrophils (17). To quantify Arg1⁺ Ly6G⁺ cells in tumor tissue following immunotherapy, we performed tyramide signal amplification (TSA) immunofluorescence staining. Briefly, after antigen retrieval and blocking, tissue sections were incubated overnight with the anti-Arg1 antibody (Servicebio GB115724, 1:2500, 4°C). The following day, secondary antibody (Servicebio GB23303, 1:500) and TSA reagent (Servicebio G1223, 1:500) were sequentially added. Subsequently, after repeat antigen retrieval and blocking, the sections were incubated overnight with the anti-Ly6G antibody (Servicebio GB11229, 1:500, 4°C). On the third day, secondary antibody (Servicebio GB22403, 1:200) and DAPI (Servicebio G1012) were sequentially applied. The number of Arg1⁺Ly6G⁺cells was then quantified in three fields under 20X magnification using an LSM900 confocal laser scanning microscope (ZEISS, Germany).

In addition, liver samples were prepared for paraffin embedding. Serial sections were examined, and the section exhibiting the maximal cross-sectional area of the liver was selected for H&E staining. Liver metastases in this section were then quantified.

Tumors (n = 6–7/group) were snap-frozen in liquid nitrogen, homogenized in 80% ice-cold methanol (−20°C) and centrifuged (14,000 rpm, 15 min, 4°C). Lyophilized supernatants were reconstituted in 50% ice-cold methanol and analyzed using an ACQUITY UPLC System I Class (Waters, USA) coupled to a Q-Exactive Plus high-resolution mass spectrometer (Thermo Fisher Scientific, Germany). Chromatographic separation was performed on a T3 column (100 × 2.1 mm, 1.8 µm) at 40°C with a mobile phase of (A) 5 mM ammonium acetate/5 mM acetic acid in water and (B) acetonitrile, delivered at 0.3 mL/min. The gradient program was as follows: 5% B (0-0.3 min), 5-70% B (0.3–2 min), 70-99% B (2-6.2 min), 99% B (6.2-7.5 min), 99-5% B (7.5-8.0 min), 5% B (8.0-10.0 min).

Mass spectrometry data were recorded in both positive (+4000 V) and negative (-2800 V) ion modes, converted to mzXML format and analyzed for peak extraction, retention time adjustment, metabolite annotation, and quantification. Gene Set Enrichment Analysis (GSEA) was performed using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database to assess intergroup metabolic pathway differences. To identify significant metabolic alterations, we employed a combination of univariate and multivariate statistical analyses. Initially, t-tests were performed to evaluate the statistical significance of intergroup differences in metabolite abundance. Subsequently, partial least squares discriminant analysis (PLS-DA) was conducted, and variable importance in projection (VIP) scores were extracted for each metabolite. Significantly altered metabolites were defined as those with fold change (FC) > 1.20 or < 0.83, VIP > 1, and P-value < 0.05. Significantly altered metabolites were then used for the KEGG enrichment analysis.

LLC cells (1 × 10⁵/well) were treated for 24 hours with linoleic acid (low-dose: 50 μM; high-dose: 100 μM; Sigma-Aldrich, Germany). Intracellular ROS was visualized using a ROS Assay Kit (Beyotime, China) and imaged via confocal laser scanning microscopes (Carl Zeiss, Germany).

Public datasets included two scRNA-seq datasets related to immunotherapy in LUAD: GSE207422 (Gene Expression Omnibus [GEO], https://www.ncbi.nlm.nih.gov/geo) and HRA002509 (Zenodo, https://zenodo.org/). Bulk RNA-seq datasets comprised TCGA-LUAD (The Cancer Genome Atlas [TCGA], https://portal.gdc.cancer.gov), GSE30219, GSE31210, GSE72094, GSE126044, GSE207422, OAK, and POPLAR, with the latter four being immunotherapy cohorts. OAK and POPLAR were obtained from previous study (18). By searching the Genecards database with the term “oxidative stress”, we identified 1663 OSRGs with a relevance score ≥ 3. Detailed gene information is available in Supplementary Table 1 .

Two scRNA-seq datasets (GSE207422 and HRA002509) were integrated using Seurat v4. Low-quality cells (<200 genes/cell) and genes expressed in <3 cells were filtered. Batch effects were mitigated during integration using the SCTransform method (https://github.com/ChristophH/sctransform) with default parameters. Following integration, principal component analysis was performed on the scaled and centered data, and the top 30 principal components were used for dimensionality reduction via Uniform Manifold Approximation and Projection (UMAP). Cells were annotated based on the expression of known marker genes from the original publications.

Based on pathologic response to immunotherapy, patients were classified as responders (major pathologic response, MPR) or non-responders (non-MPR, NMPR). We identified differentially expressed genes (DEGs) significantly upregulated (logFC > 0.25, adjusted P value < 0.05) in each cell subset of non-responders relative to responders. By intersecting these DEGs (GSE207422 and HRA002509) and the OSRGs gene set, we identified OSRGs potentially linked to immunotherapy response. The optimal modeling algorithm was explored and validated across six bulk RNA-seq cohorts (TCGA-LUAD, GSE30219, GSE31210, GSE72094, OAK, POPLAR) using 118 machine learning algorithms. Random survival forest (RSF) analysis of TCGA-LUAD data (R software, randomForestSRC package) was then used to assess the relationship between the selected genes and prognosis.

The AUCell algorithm quantified OSRGs activity in cell subpopulations and tumor tissue spatial transcriptomic sections. Correlations between OSRGs score and immunotherapy response were validated in four bulk RNA-seq cohorts (GSE126044, GSE207422, OAK, POPLAR), and neutrophil infiltration in the TME was assessed via CIBERSORT in TCGA-LUAD cohort. LUAD patients were stratified into high/low OSRGs score groups (median cutoff) across four cohorts (TCGA, GSE31210, OAK, and POPLAK). Kaplan-Meier analysis evaluated survival differences. Immune cell fractions (22 subsets) were deconvoluted using CIBERSORT in TCGA-LUAD databases, and pathway activity was profiled via the R package ‘GSVA’ (v 1.30.0).

Statistical analyses used Mann-Whitney U tests for group comparisons (P < 0.05), with data presented as mean ± SEM (standard error of the mean). Analyses were performed in GraphPad Prism 8 (San Diego, California, USA) and R v4.3.1.

The animal experimental workflow was illustrated in Figure 1A . Mice receiving fractionated left lung radiation exhibited marked histopathological features of RP at 7 days post-tumor implantation, including significant diffuse congestion, alveolar wall thickening, and inflammatory cell infiltration ( Figure 1B ). Tissue injury scores were significantly elevated in irradiated lungs compared to unirradiated lungs ( Figure 1C ). These findings established an experimental basis for our subsequent investigation of the impact of RP on lung cancer immunotherapy.

Longitudinal CT imaging revealed no significant difference in primary tumor volume between the two groups at baseline or week 1 post-treatment ( Figures 1D-F ). By week 2, mice in the RP + anti-PD-1 group showed suppressed primary tumor growth (P < 0.05, Figure 1G ), and endpoint tumor-bearing lung weight was reduced in the RP + anti-PD-1 group ( Figure 1H ). However, mice in the RP + anti-PD-1 group developed significantly higher incidence (70.00% vs. 36.36%, P > 0.05) and burden (1.09 ± 1.76 vs. 9.90 ± 11.52, P < 0.05) of liver metastases ( Figures 1I, J ). Consequently, survival outcomes were comparable between the anti-PD-1 group and the RP + anti-PD-1 group (median survival: 34 vs. 35 days, P > 0.05, Figure 1K ).

PLS-DA revealed distinct separation between two groups before and after immunotherapy ( Figure 2A ). Untargeted metabolomics identified 1943 metabolites across tumor tissue, predominantly lipids and lipid-like molecules, organic acids and derivatives, and organoheterocyclic compounds. The partial metabolic profiles of tumor tissues were shown in Figure 2B . Immunotherapy induced significant alterations in the metabolic landscape of tumor tissues independent of RP.

Immunotherapy alone upregulated 128 metabolites (e.g., sorbitol, ergothioneine, lysophosphatidylcholine [LPC P-18:0], uridine5-diphosphate, and sclareol) and downregulated 206 metabolites (e.g., naringenin, guanosine-5’-monophosphate, lysophosphatidylglycerol [LPG 18:0, LPG 18:1], and L-Tryptophan) ( Figure 2C ). Top 5 enrichment pathways of differential metabolites included glycerophospholipid metabolism, nucleotide metabolism, pyrimidine metabolism, alpha-linolenic acid metabolism, and linoleic acid metabolism ( Figure 2D ). GSEA further revealed activation of ascorbate/aldarate metabolism and unsaturated fatty acids biosynthesis in tumor tissues after immunotherapy ( Figure 2E ), alongside suppression of alpha-linolenic acid metabolism, aldosterone synthesis and secretion, arachidonic acid metabolism, and phenylalanine, tyrosine and tryptophan biosynthesis ( Figure 2F ).

In RP mice ( Figure 2G ), metabolic changes induced by immunotherapy were more abundant, with 239 upregulated metabolites (e.g., L-kynurenine, prostaglandin E2, histamine, lysophosphatidic acid [LPA O-18:0], cis-8,11,14,17-Eicosatetraenoic acid), and 200 downregulated metabolites (e.g., guanosine-5’-monophosphate, glutamine, 7-Oxocholesterol, LPG 16:2, LPG 17:0). Pathway analysis revealed unique enrichment in glycerophospholipid metabolism, arachidonic acid metabolism, inflammatory mediator regulation of TRP channels, choline metabolism in cancer, and linoleic acid metabolism ( Figure 2H ). GSEA demonstrated activation of biosynthesis of unsaturated fatty acids, histidine metabolism, arginine biosynthesis, arginine and proline metabolism, and alanine, aspartate and glutamate metabolism in tumor tissues after immunotherapy ( Figure 2I ), concurrent with suppression of alpha-linolenic acid metabolism, aldosterone synthesis and secretion, nucleotide metabolism, vascular smooth muscle contraction, and phenylalanine, tyrosine and tryptophan biosynthesis ( Figure 2J ). These results indicated that immunotherapy may trigger more pronounced and distinct metabolic reprogramming in the RP microenvironment.

Metabolomic profiling of tumors at 2 weeks post-immunotherapy revealed profound metabolic rewiring in RP + anti-PD-1 mice compared to anti-PD-1 group ( Figure 3A ). A total of 120 metabolites were significantly elevated, while 59 metabolites were decrease. KEGG enrichment analysis identified lysosome, retrograde endocannabinoid signaling, neuroactive ligand-receptor interaction, purine metabolism, and the FoxO signaling pathway as top altered pathways ( Figure 3B ).

Among the above 120 upregulated metabolites, 57 were also significantly elevated compared with the RP + anti-PD-1 group pre-immunotherapy, with seven metabolites—phosphatidylcholine (PC(18:1/14:0)), PC(18:0/20:5), PC(P-18:1/20:4-OH), PC(20:3/20:3), PC(18:3/18:2), PC(20:3/16:0), and cis-8,11,14-Eicosatrienoic acid—enriched in linoleic acid metabolism ( Figure 3C ). Furthermore, oxidation stress-related lipids, including hydroxy fatty acids, LPCs, branched fatty acid esters of hydroxy fatty acid (FAHFA), and epoxydocosapentaenoic acid were predominant ( Figure 3D ). This evidence indicates that the synergistic effect of immunotherapy and RP lead to severe dysregulation of lipid metabolism and pronounced oxidative stress response in the tumor tissue.

Given the enrichment of linoleic acid metabolism in RP + anti-PD-1 group post-immunotherapy, we interrogated its functional role in driving oxidative stress. As expected, LLC cells treated with high-dose linoleic acid had significantly higher intracellular ROS levels compared to both control and low-dose linoleic acid ( Figure 3E ), supporting linoleic acid metabolism as a latent driver of redox imbalance in lung cancer cells.

We integrated two scRNA-seq cohorts of immunotherapy-treated LUAD patients (GSE207422, n=15; HRA002509, n=19), resolving eight major cell types—epithelial cells, endothelial cells, fibroblasts, T cells, B cells, macrophages, mast cells, and neutrophils ( Figures 4A, B ). Comparative analysis of NMPR versus MPR identified 944 (GSE207422, Supplementary Table 2 ) and 1493 (HRA002509, Supplementary Table 3 ) significantly regulated DEGs, respectively ( Figures 4C, D ). Intersecting these DEGs with 1663 OSRGs yielded 99 candidate genes potentially implicated in immunotherapy-driven redox regulation ( Figure 4E ). Machine learning optimization across 118 algorithms identified CoxBoost + RSF as the optimal model (mean C-index = 0.678) ( Figure 4F ; Supplementary Figure 1 , Supplementary Table 4 ). Then final OSRGs score was derived from 14 genes: OSRGs score = 0.182 × LDHA + 0.129 × ITGB1 + 0.473 × PPIA + 0.034 × KRT8 + 0.098 × DDIT4 + 0.171 × FURIN + 0.102 × ADM + 0.188 × TIMP1 + 0.090 × PABPC1 + 0.084 × CFL1 - 0.026 × GAPDH - 0.141 × GPX3 - 0.026 × KRT18 - 0.185 × CXCR4 ( Figure 4G ; Supplementary Figure 2 ).

In scRNA-seq cohorts, the highest OSRG scores were observed in macrophages and fibroblasts ( Supplementary Figure 3 ). OSRGs scores of tumor cells were significantly elevated in non-responders compared to responders (P < 0.0001, Figures 5A, B ). Furthermore, this redox imbalance extended to tumor-infiltrating neutrophils, T cells, and macrophages (P < 0.0001, Figures 5A, B ), suggesting broad impairment of anti-tumor immunity. Simultaneously, analysis of spatial/bulk RNA-seq among immunotherapy cohorts also revealed a relevance between high OSRGs score in tumor tissue and adverse immunotherapy response ( Figures 5C, D ). Correlation analysis revealed a positive association (R = 0.38, P < 0.0001) between OSRGs score and neutrophil infiltration in the TME ( Figure 5E ). Consistent with this, our in vivo experiments detected a marked increase in tumor-infiltrating neutrophils and the pro-tumor phenotype in the RP + anti-PD-1 group, alongside unbalanced oxidative stress ( Figures 5F-I ). These findings suggest that the interplay between oxidative stress and neutrophils may contribute to immunotherapy resistance.

Stratification by median OSRGs score revealed divergent TME composition between high- and low-score LUAD patients. As depicted in Figure 6A , high-OSRGs tumor exhibited elevated infiltration of neutrophils, activated mast cells, activated dendritic cells, activated memory CD4+ T cells, and resting NK cells, alongside reduced activated NK cells, follicular helper T cells, naive B cells, memory B cells, monocytes, and resting dendritic cells.

Functional enrichment analysis linked high-OSRGs scores to activation of response to oxidative stress, glycolysis, and EMT ( Figures 6B-D ; Supplementary Table 5 ). Oxidative stress may reshape TME into a hostile yet permissive niche that paradoxically accelerates clonal selection of evolved cancer cells. Critically, LUAD patients with high-OSRGs scores had significantly worse prognoses in the TCGA, GSE31210, OAK, and POPLAR datasets ( Figures 6E-H ).