Targeting galactose metabolic reprogramming overcomes immunotherapy resistance in KRAS-mutant lung adenocarcinoma: Integrative multi-omics and machine learning approaches

Highlights

• •Galactose-metabolic (GM) reprogramming is the top dysregulated pathway in KRAS-mutant LUAD and correlates with “cold” immune microenvironment.
• •Single-cell and bulk multi-omics reveal two GM-based subtypes: high-GM/C1 (poor survival, KEAP1 co-mutations) vs low-GM/C2 (high TMB, robust CD8⁺ T-cell infiltration).
• •101-algorithm machine-learning screen establishes a 37-gene GM signature whose risk score outperforms clinicopathologic factors for predicting overall survival.
• •KDM5A is identified as the key GM-hub gene driving ICB resistance; its inhibitor CPI-455 synergizes with anti-PD-1 to restore CD8⁺ T-cell influx and suppress tumor growth in vivo.
• •Targeting GM or its epigenetic regulator KDM5A offers a readily translatable strategy to overcome primary immunotherapy resistance in KRAS-driven lung adenocarcinoma.

Abstract

Background

KRAS-mutant lung adenocarcinoma (LUAD) is associated with aggressive phenotypes and therapy resistance, which highlights an urgent need to identify reliable biomarkers and therapeutic targets. This study aims to investigate the role of galactose metabolism (GM) reprogramming in shaping the tumor microenvironment (TME) and driving immunotherapy resistance in KRAS-mutant LUAD.

Methods

Bulk-tissue transcriptomics data were collected from public cohorts such as Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA). Single-cell RNA-sequencing (scRNA-seq) data, mutational profiles, and advanced bioinformatic approaches were integrated to analyze the correlation between mutation status and metabolic heterogeneity in LUAD. Analytical methods included dimensionality reduction, differential expression, intercellular communication, trajectory inference, gene co-expression network analysis, molecular subtyping, and machine learning-based prognostic modeling. Animal experiment and flow cytometry were performed to validate KDM5A inhibitor (CPI-455) can sensitize KRAS-mutant LUAD cells to immunotherapy.

Results

Single-cell analysis revealed reduced immune infiltration and upregulated GM in KRAS-mutant epithelial cells. Two GM-based molecular subtypes with distinct mutational profiles and immune patterns were identified. A prognostic model integrating GM-related genes demonstrated superior predictive performance. KDM5A was identified as a key predictor of ICB resistance. Targeting KDM5A can enhance the sensitivity to immunotherapy in KRAS-mutant LUAD.

Conclusion

Our findings highlight the therapeutic potential of targeting galactose metabolism-related hub genes to sensitize KRAS-mutant LUAD to immunotherapy.

Graphical abstract

Introduction

Lung adenocarcinoma (LUAD) that harbors KRAS mutation represents a significant subset of non-small cell lung cancer (NSCLC), which is always characterized by aggressive behavior and poor clinical outcomes [1,2]. Despite advances in immunotherapy, a substantial proportion of KRAS-mutant LUAD patients exhibit innate or acquired resistance to immune checkpoint blockade (ICB), and this dilemma limits therapeutic efficacy of ICB therapy [3]. Compounding this challenge is the absence of directly druggable targets for KRAS mutations, leaving clinicians with limited options and underscoring an urgent need to identify robust biomarkers and novel therapeutic targets for this aggressive subtype [4].

In this context, tumor metabolic reprogramming has emerged as a hallmark of cancer, profoundly influencing the tumor microenvironment (TME) and anti-tumor immunity [5,6]. In LUAD, altered metabolic pathways not only support rapid proliferation and survival of cancer cells but also contribute to an immunosuppressive milieu, thereby facilitating immune evasion and resistance to therapy [[7], [8], [9]]. Among various metabolic pathways, carbohydrate metabolism—including glycolysis, galactose metabolism, and other monosaccharide pathways—has recently attracted attention for its potential role in cancer progression and therapy response [8,[10], [11], [12]]. However, the specific implications of galactose metabolism in KRAS-mutant LUAD and immune evasion remain largely unexplored.

To address this critical challenge, our study employed a multi-omics integrative approach that combined single-cell RNA-sequencing (scRNA-seq), bulk transcriptomics, genetic mutational profiles, and advanced computational methods. We performed various bioinformatic analyses to systematically dissect the cellular heterogeneity, metabolic alterations, and intercellular communication within the TME of KRAS-mutant LUAD. Specifically, we focused on elucidating the role of galactose metabolism (GM) reprogramming as a key mediator of tumor progression and immune resistance, especially in KRAS-mutant LUAD.

Our comprehensive analysis reveals distinct metabolic and immune cellular landscapes between KRAS-mutant and wild-type tumors, identifies a GM-related gene signature with prognostic significance, and establishes a reliable predictive model for risk assessment using machine learning approaches. Furthermore, in the GM-related gene signature, we identified KDM5A as a critical regulator of immunotherapy response for KRAS-mutant LUAD. Animal treatment experiments demonstrated that inhibition of KDM5A synergizes with anti-PD-1 therapy could enhance T cell infiltration and suppress tumor growth in vivo.

Our study not only provides novel insights into the metabolic basis of immune evasion in KRAS-mutant LUAD but also proposes a promising biomarker and a therapeutic strategy targeting galactose metabolism-related hub gene to overcome resistance to immunotherapy.

Materials and methods

Data collection and processing

The bulk-tissue transcriptomic profiles of LUADs were obtained from two public databases: Gene Expression Omnibus (GEO, accession No: GSE72094)[13] and The Cancer Genome Atlas (TCGA). Somatic mutation data for TCGA-LUAD were retrieved from the GDC Data Portal (https://portal.gdc.cancer.gov/). To facilitate standardized comparison of RNA-seq data, fragments per kilobase of exon per million mapped reads (FPKM) values were converted to transcripts per million (TPM) values. The “maftools” R package was employed to analyze the mutation data [14], while the “Sigminer” package was used to extract mutational signatures from whole exome sequencing (WES) data [15].

Single-cell RNA-sequencing analysis

A single-cell RNA-sequencing (scRNA-seq) dataset (GSE171145) [16], comprising LUAD samples with mutational annotations, was utilized for single-cell transcriptome analysis using the R package “Seurat.” Our study included one KRAS-mutant sample (GBG) and three wild-type samples (CYD, TGS, XMS), resulting in a total of 10,410 eligible cells for further analysis. Principal component analysis (PCA) was conducted on the scaled data of the top 2000 highly variable genes (HVGs) to reduce dimensionality. The top 20 principal components were extracted for UMAP construction, and differentially expressed genes (DEGs) were identified using the “FindAllMarkers” function with the Wilcox rank-sum test. Intercellular communication networks were characterized across different cell types using the “CellChat” package [17]. Additionally, single-cell trajectory analysis was performed using the Monocle3 algorithm with default parameters. Cells were organized into a developmental trajectory to simulate their evolutionary paths. The detailed processing procedures of scRNA-seq analysis for GSE171145 can be found in a previous study [18]. Batch effects were effectively mitigated by Harmony, enabling robust integration of multi-sample scRNA-seq data.

To investigate the distinct transcriptional profiles and co-expression patterns of malignant epithelial cells with high and low galactose metabolism (high-GM and low-GM) in both KRAS-mutant and WT conditions, we employed the high-dimensional Weighted Gene Co-expression Network Analysis (hdWGCNA) approach [19], and different gene modules were identified and represented coordinated expression patterns correlated with different characteristics (WT.lowGM, WT.highGM, Mut.lowGM, Mut.highGM) were identified and further investigated.

Estimation of immune infiltration and evaluation of immune features

The Cibersort algorithm was employed to quantify the relative proportions of 22 tumor-infiltrating immune cell types using transcriptome profiling data [20]. Aneuploidy scores, fraction_altered values, homologous recombination deficiency (HRD) scores, interferon-γ (IFN-γ) response levels, non-silent mutation counts, and single-nucleotide variant (SNV) neoantigen burdens were derived from a comprehensive pan-cancer immune landscape study [21]. The TIDE algorithm (Tumor Immune Dysfunction and Exclusion, link as: tide.dfci.harvard.edu) was utilized to predict the potential response to immunotherapy among patients with KRAS-mutant lung adenocarcinoma, and these patients were categorized into non-responders and responders according to their TIDE scores [22].

Selection of optimal models from multiple machine learning approaches

A comprehensive framework involving 101 distinct machine learning combinations [23] was employed to analyze galactose metabolism-related candidate genes, with the goal of constructing prognostic models for patients with KRAS-mutant lung adenocarcinoma (LUAD). The algorithms encompassed CoxBoost, Gradient Boosting Machine (GBM), Lasso regression, partial least squares regression combined with Cox regression (plsRcox), Ridge regression, Random Survival Forest (RSF), Elastic Net (with alpha values ranging from 0.1 to 0.9), Stepwise Cox regression (including forward, backward, and bidirectional selection methods), and Super Principal Component Analysis (SuperPC). To ensure robustness and identify the most effective model, a 10-fold cross-validation strategy was utilized. The model achieving the highest concordance index (C-index) for predicting overall survival was selected. Furthermore, the risk scores generated by the machine learning model were compared with traditional clinicopathological features to evaluate their prognostic capacity.

A neural network model was constructed using the “neuralnet” R package with GM-related candidate genes to predict the immunotherapy response for KRAS-mutant LUAD patients. A confusion matrix was generated to evaluate the model’s performance, providing insights into the accuracy, precision, recall, and F1 score of the predictions. In addition, the random forest algorithm was applied to assess the importance of each candidate gene in predicting immunotherapy response referring to the metric of “mean decrease accuracy”.

Molecular docking analysis

Structure of active compounds were downloaded from PubChem database, and imported into ChemBio3D 14.0 software to adjust the spatial conformation of active ingredients, calculate the optimization of energy. The three-dimensional crystal structure of the target protein was downloaded from the Uniprot. The water molecule and organic matter in the target protein were removed by pymol, and then the target protein was imported into AutoDockTools1.5.6 for hydrogenation, charge distribution, and atomic type addition. AutoDock Vina was used for molecular docking, and the docking results were plotted with Pymol 2.6.1. In this study, docking scores ≤ −7.0 kcal/mol were considered indicative of strong binding affinity, consistent with commonly applied thresholds in structure-based drug design and previously reported AutoDock Vina scoring benchmarks [24].

Animal experiment

In the animal study, 1 × 10⁶ Lewis lung carcinoma (LLC) cells were subcutaneously injected into the right flank of six-week-old C57BL/6 mice to establish a subcutaneous tumor model. Once the tumor volume reached approximately 80–100 mm³, mice bearing tumors were randomly assigned into three groups (n = 5 per group). LLC cells are derived from a spontaneous lung tumor in C57BL/6 mice and carry an endogenous Kras^G12D mutation, making them a relevant model for studying immunotherapy resistance in KRAS-mutant LUAD. The treatment regimens included administration of PBS, a mouse anti-PD-1 monoclonal antibody (mAb) (BioXcell, BE0146, 100 µg per mouse), or a combination of the KDM5A inhibitor CPI-455 (MCE, HY-100421, 50 mg/kg) and the anti-PD-1 mAb. All injections were performed intraperitoneally using a 27-gauge needle, with a total volume of 200 µL per mouse. All treatments were administered via intraperitoneal injection, with each group of mice receiving injections every other day. Body weight and tumor dimensions were measured every three days. Tumor volumes were calculated using the formula: 0.5 × (length × width²). The animal experimental protocols were conducted in accordance with guidelines and approved by the Institutional Animal Care and Use Committee of Nanjing Medical University.

Flow cytometry analysis

C57BL/6 mice were sacrificed to harvest fresh tumor specimens, which were enzymatically digested for 2 hours and then filtered through 70-µm strainers to generate single-cell suspensions. After washing, the cells were resuspended in 40% Percoll and centrifuged on a 70% Percoll gradient to isolate lymphocytes. The isolated cells were subsequently stimulated for 4 hours with a leukocyte activation cocktail (BD Pharmingen) in preparation for intracellular cytokine staining. Anti-Cd45, anti-Cd3e, and anti-Cd8a monoclonal antibodies were obtained from BD Pharmingen. Cell analysis was performed using a BD FACS Canto II flow cytometer, and data were analyzed with FlowJo software. More detailed descriptions of these procedures can refer to our previous works [18].

Additional statistical and bioinformatics analyses

The non-negative matrix factorization (NMF) clustering was performed to identify molecular subtypes within KRAS-mutant LUAD based on galactose metabolism-related genes. The analysis was conducted using the NMF R package on the expression matrix of the galactose metabolism-related genes across 139 KRAS-mutant samples from the TCGA-LUAD cohort. The factorization rank (k) was set from 2 to 5, and the optimal number of clusters (k=2) was determined by evaluating the cophenetic correlation coefficient, dispersion, and residual sum of squares. Samples were subsequently categorized into distinct subtypes (C1 and C2) based on the final consensus membership. The “limma” package was used to identify differentially expressed genes (DEGs) between different groups with microarray data. A single-sample gene set enrichment analysis (ssGSEA) algorithm was applied to calculate the enrichment scores of different metabolic hallmarks for each sample or cell based on the scRNA-seq transcriptome data and corresponding gene signatures using the R package “gsva”. The concordance index (C-index) or receiver operating characteristic (ROC) analysis was used to measure the discriminative capacity. Additionally, a scoring nomogram that integrates risk score and clinicopathological variables was generated to predict overall survival of KRAS-mutant LUAD patients. Calibration analysis and decision curve analysis (DCA) were performed to evaluate the predictive performance and clinical utility of the nomogram. Student’s t-test was used to analyze the differences between two groups that were normally distributed. Two-sided p value less than 0.05 was considered statistically significant. All analyses were performed in the GraphPad Prism 9 and R 4.2.1 software.

Results

Single-cell RNA-seq analysis reveals cellular and metabolic heterogeneity in KRAS-mutant LUAD

We obtained the single-cell RNA-sequencing data of LUAD samples from the GSE171145 dataset, which includes one KRAS-mutant (GBG) and three wild-type (CYD, TGS, XMS) samples. A total of 10,410 high-quality cells were preserved, and we identified 12 distinct cell populations: Epithelial, Fibroblast, CD4_T, CD8_T, Endothelial, Epithelial, Myeloid, Mast, NK, Plasma, Treg, and Tprolif cells (Fig. 1A). The nebula plot further illustrates the density distribution of these 12 cell types (Fig. 1B). The marker genes associated with these populations were visualized in a bubble plot (Fig. 1C), illustrating the differential expression patterns that characterize each cell type. The UMAP plots for each sample (Fig. 1D), accompanied by pie charts, depicted the cellular composition and distribution across the samples. The Sankey diagram (Fig. 1E) further detailed the cellular distribution among the samples, highlighting the heterogeneity in cellular composition. A notable finding was the reduced immune cell infiltration in KRAS-mutant samples compared to wild-type samples, as shown in the stacked bar chart (Fig. 1F). This suggests a potential immune evasion mechanism in KRAS-mutant tumors. The volcano plot revealed significant differences in 114 metabolic pathways between mutant and wild-type samples, with galactose metabolism being the most prominently altered one (Fig. 1G). This metabolic shift was visually confirmed in the UMAP plots (Fig. 1H), which displayed differential GM scores in malignant epithelial cells between KRAS-mutant and wild-type groups. The pseudotime analysis using Monocle3 (Fig. 1I) and the distribution of cells along this trajectory (Fig. 1J) provided insights into the dynamic cellular processes. The GM scores in malignant epithelial cells (Fig. 1K) further underscored the metabolic reprogramming in KRAS-mutant cells. Finally, the violin plot (Fig. 1L) demonstrated a significant increase in GM scores in KRAS-mutant epithelial cells, indicating metabolic adaptation in these tumors. These results collectively highlight the complex cellular and metabolic landscape in KRAS-mutant LUAD. The observed metabolic reprogramming, particularly in galactose metabolism, suggests a role in KRAS-mutant LUAD progression and may serve as a promising biomarker and target.

Fig. 1.

Single-cell RNA-seq analysis reveals cellular and metabolic heterogeneity in KRAS-mutant LUAD. (A) UMAP visualization of 10,410 cells from LUAD patients, and 12 distinct cell types were identified. (B) The nebula plot further illustrates the density distribution of these 12 cell types. (C) Bubble plot shows the average expression levels of marker genes for each cell type. (D) UMAP plots for individual samples (CYD, TGS, XMS, and KRAS-mutant GBG), with pie charts indicating cellular composition. (E) Sankey diagram depicts cellular distribution across samples. (F) Stacked bar chart compares immune cell infiltration between KRAS-mutant (Mut) and wild-type (WT) samples. (G) Volcano plot shows significant differences in 114 metabolic pathways, and identified galactose metabolism (GM) as the most altered one. (H) UMAP plots shows differential GM scores in malignant epithelial cells between KRAS-mutant and WT groups. (I) Pseudotime analysis using Monocle3, and (J) distribution of cells along this trajectory. (K) GM scores in malignant epithelial cells, and (L) violin plot demonstrates increased GM scores in KRAS-mutant epithelial cells. ***p < 0.001.

Intercellular communication patterns of high-GM epithelial cells and other cell populations within KRAS-mutant LUAD microenvironment

Utilizing the CellChat R package, we conducted a comprehensive analysis of cell-to-cell communication in KRAS-mutant and wild-type samples. The comparative analysis of pathway strength revealed distinct signaling profiles between the two groups (Fig. 2A). Notably, the SPP1 signaling pathway was specifically enhanced in WT samples, whereas pathways such as GRN and EPHA were uniquely present in KRAS-mut LUADs. To further elucidate these differences, epithelial cells were stratified into high-GM and low-GM groups. The subsequent bubble heatmaps illustrated the intensity of these signaling pathways, with pathways in Mut samples showing increased activity when high-GM epithelial cells served as the source (Fig. 2B). Conversely, pathways in WT samples were more active when low-GM epithelial cells were the source (Fig. 2C). The interaction network among all cell subtypes including both low-GM and high-GM epithelial cells was depicted in Fig. 2D, highlighting the complex interplay between different cell types. Specifically, in KRAS-mut samples, the IFN-II signaling pathway was found to be significantly elevated in low-GM cells (Fig. 2E), suggesting a potential role in the mutated cellular context. In contrast, the GDF signaling pathway was specifically enhanced in low-GM cells of WT samples (Fig. 2F), indicating a distinct regulatory mechanism in the non-mutated state. These analyses revealed that certain signals were enhanced in KRAS-mut samples, each exhibiting unique signature profiles. These findings underscore the differential regulation of cell-to-cell communication pathways in KRAS-mut and WT LUADs, providing insights into the molecular mechanisms underlying these distinct cellular contexts.

Fig. 2.

Intercellular communication patterns of high-GM epithelial cells and other cell populations within KRAS-mutant LUAD microenvironment. (A) Pathway strength comparison between KRAS-mutant (mut) and wild-type (WT) samples. (B & C) Bubble heatmaps illustrate the intensity of signaling pathways in high-GM and low-GM epithelial cells for KRAS-mut and WT samples, respectively. (D) Interaction network among all cell subtypes, including both low-GM and high-GM epithelial cells. (E & F) IFN-II and GDF signaling pathways were specifically elevated in low-GM cells of KRAS-mut and WT samples, respectively.

Gene module analysis of KRAS-mutant high-GM LUAD epithelial cells

We extracted all malignant epithelial cells and performed gene module analysis on their transcriptomes using the hdWGCNA package. The optimal soft power threshold was determined to be 4 (Fig. 3A), which facilitated the construction of a dendrogram with 7,242 genes assigned to 14 distinct modules when β was set to 4 (Fig. 3B). The distribution of gene numbers across these modules revealed that module 3 contained the highest number of genes (Fig. 3C). Correlation analysis between the four features (WT.lowGM, WT.highGM, Mut.lowGM, Mut.highGM) and the 14 gene modules (NEWs) indicated that modules NEW2, 3, 4, 6, 7, 8, 9, 10, 11, and 14 were positively correlated with the Mut.highGM group (Fig. 3D). The scoring of genes within these modules across all epithelial cells was visualized in Fig. 3E. Using the “FindAllMarkers” function, we identified differentially expressed genes among the four groups, with a volcano plot illustrating their distribution (Fig. 3F). The threshold for selection was set at a q-value < 0.05 and a log2 fold change > 0.5, and a total of 453 upregulated genes in the Mut.highGM group were identified. An intersection analysis with the Mut.highGM-related gene modules identified 364 genes that were most associated with KRAS-Mut high-GM (Fig. 3G). Gene Ontology (GO) enrichment analysis of these 364 genes revealed significant enrichment in metabolic pathways, with the “generation of precursor metabolites and energy” process as the most prominent one (Fig. 3H). These findings suggest that the identified gene modules and their associated pathways may play crucial roles in the biology of KRAS-mutant high-GM malignant epithelial cells, potentially offering novel targets for therapeutic intervention.

Fig. 3.

Gene module analysis of KRAS-mutant high-GM LUAD epithelial cells. (A) Determination of the optimal soft power threshold for hdWGCNA analysis. (B) WGCNA dendrogram with 7,242 genes assigned to 14 distinct modules. (C) Distribution of gene numbers across the 14 modules. (D) Correlation analysis between features (WT.lowGM, WT.highGM, Mut.lowGM, Mut.highGM) and the 14 gene modules. (E) Gene scoring across all epithelial cells for modules positively correlated with Mut.highGM. (F) Volcano plot identifies differentially expressed genes among the four groups. (G) Intersection analysis identifies 364 genes most associated with KRAS-Mut high-GM. (H) GO enrichment analysis of these 364 genes.

Subtyping and characterization of KRAS-mutant LUAD based on galactose metabolism-related gene signature

From the TCGA-LUAD cohort, we extracted 139 KRAS-mutant lung adenocarcinoma samples and performed subtyping using Non-negative Matrix Factorization (NMF) based on the expression matrix of 394 genes. Bootstrap resampling with 100 iterations was performed to assess the stability of NMF clustering. The cophenetic correlation coefficients for different k values were calculated respectively, and k=2 yields the most stable clustering, and the entire KRAS-mutant cohort was divided into two distinct subtypes (C1 and C2) (Fig. 4A). Notably, subtype C1 exhibited a significantly higher Galactose Metabolism Score (Fig. 4B, p < 0.001), which was associated with a higher overall mortality rate of 45.7% (Fig. 4C). In contrast, subtype C2 showed a significantly elevated Tumor Mutation Burden (TMB) compared to C1 (Fig. 4D, p < 0.001). The most frequently mutated genes in each subtype were visualized using oncoplots (Fig. 4, Fig. 4), with KEAP1 being the most significantly mutated gene in C1 (Fig. 4G). Co-mutation analysis revealed that KEAP1 co-mutated with KRAS significantly more often in C1, while TP53 co-mutated with KRAS was enriched in C2. The mutually exclusive and co-occurring patterns were generated using the “somaticInteractions” function for each subtype (Fig. 4H). We observed that more co-occurrence mutation pairs in C2. The mutational profiles of KRAS-mutant LUAD were further characterized (Fig. 4I), revealing a significant enrichment of Signature 3 (spontaneous deamination of 5-methylcytosine) in C1 (Fig. 4J). The “drugInteractions” function checks for drug–gene interactions and gene druggability information compiled from the Drug Gene Interaction database. We then applied this function to perform a druggable gene analysis specifically for the C1 subtype (Fig. 4K). We compared various tumor heterogeneity and immune features between C1 and C2, including Aneuploidy score, Fraction altered, HRD score, IFN-γ response, non-silent mutation rate, and SNV neoantigens (Fig. 4L). These features were significantly higher in C2, suggesting a more aggressive and immunogenic phenotype. The distribution of cellular subpopulations was analyzed using CIBERSORT (Fig. 4M), showing that C2 had a significantly higher abundance of total T cells, CD8⁺ T cells, and M1 macrophages (Fig. 4N, p < 0.001). These findings highlight the distinct molecular and immunological landscapes of C1 and C2 subtypes derived from KRAS-mutant LUADs.

Fig. 4.

Subtyping and characterization of KRAS-mutant LUAD based on the galactose metabolism-related gene signature. (A) Non-negative Matrix Factorization (NMF) clustering of 139 KRAS-mutant LUAD samples into two subtypes, C1 and C2. (B) Subtype C1 exhibits a significantly higher galactose metabolism score compared to subtype C2 (p < 0.001). (C) Subtype C1 is associated with a higher overall mortality rate (45.7%) than subtype C2. (D) Subtype C2 shows a significantly elevated tumor mutation burden (TMB) relative to subtype C1 (p < 0.001). (E) Oncoplot displays the most frequently mutated genes in subtype C1. (F) Oncoplot displays the most frequently mutated genes in subtype C2. (G) KEAP1 is the most significantly mutated gene in subtype C1 (p < 0.001). (H) Co-mutation analysis reveals that KEAP1 co-mutates with KRAS more frequently in C1, whereas TP53 co-mutates with KRAS more frequently in C2; moreover, more co-occurring mutation pairs are observed in C2. (I) Mutational profiles of KRAS-mutant LUAD are further characterized across both subtypes. (J) Signature 3 (spontaneous deamination of 5-methylcytosine) is significantly enriched in subtype C1. (K) Druggable gene analysis using the drugInteractions function identifies potential drug-gene interactions specifically for subtype C1. (L) Comparison of tumor heterogeneity and immune features (aneuploidy score, fraction altered, HRD score, IFN-γ response, non-silent mutation rate, and SNV neoantigens) shows that all these features are significantly higher in subtype C2. (M) CIBERSORT analysis reveals the distribution of immune cell subpopulations in subtypes C1 and C2. (N) Subtype C2 has a significantly higher abundance of total T cells, CD8⁺ T cells, and M1 macrophages compared to subtype C1 (p < 0.001). *p < 0.05, **p < 0.01, ***p < 0.001, # not significant.

Establishment of prognostic models based on galactose metabolism-related genes using machine learning approaches

We conducted a comprehensive analysis to identify a robust predictive model for overall survival in KRAS-mutant LUAD. Initially, univariate Cox regression analysis was performed on 139 KRAS-mut LUAD samples from the GSE72094 dataset, yielding 37 candidate genes associated with overall survival. Subsequently, we employed 101 machine learning models, including Lasso, RSF, CoxBoost, and GBM, to calculate the concordance index (C-index) for overall survival in both the training set (KRAS-mut LUADs in GSE72094) and the testing set (KRAS-mut LUADs in TCGA-LUAD). As shown in Fig. 5A, the combination of Lasso and RSF demonstrated the highest C-index in both datasets, indicating its superior predictive performance.

Fig. 5.

Establishment of prognostic models based on galactose metabolism-related genes using machine learning approaches. (A) Performance of 101 distinct machine learning algorithms in predicting overall survival. (B) Determination of the optimal cutoff value for risk scores in the training set. (C & D) Overall survival curves for the training and testing sets, respectively. (E) Time-dependent ROC analysis for the training and testing cohorts. (F & G) ROC curves for the training and testing sets at 1, 2, and 3 years. (H) Nomogram was constructed with clinicopathological features and risk scores for predicting overall survival probabilities. (I) Calibration analysis of the nomogram. (J) Decision curve analysis (DCA) shows that the nomogram outperforms all the other traditional clinicopathological parameters.

To further validate this model, we utilized the Survminer package to determine the optimal cutoff value for risk scores in the training set (Fig. 5B). The resulting overall survival curves (Fig. 5C) and the corresponding analysis for the testing set (Fig. 5, Fig. 5) confirmed that higher risk scores were associated with poorer prognosis in KRAS-mut patients. Time-dependent ROC analysis was applied to both training and testing cohorts, demonstrating the model’s strong predictive capacity (Fig. 5, Fig. 5). Additionally, we constructed a nomogram that integrated clinicopathological features and risk scores in the TCGA-LUAD cohort to accurately predict overall survival probabilities (Fig. 5H). Calibration analysis indicated excellent predictive performance of the nomogram (Fig. 5I). Decision curve analysis (Fig. 5J) further showed that our scoring nomogram outperformed traditional clinicopathological indicators in overall survival prediction.

Targeting KDM5A enhances the sensitivity to immunotherapy in KRAS-mutant LUAD

We utilized the TIDE algorithm to assess the response of KRAS-mutant samples from TCGA-LUAD to immune checkpoint blockade (ICB) therapy, categorizing them into non-responder (NR) and responder (R) groups. Subsequently, we employed the DESeq2 algorithm to perform differential expression analysis on 364 identified genes, yielding 13 candidate genes with p-values less than 0.05 (Fig. 6A). Among these, KDM5A was significantly upregulated in the NR group. In addition, we found that KDM5A was widely expressed in KRAS-mutant epithelial cells (Fig. 6B). We then constructed a neural network model using the “neuralnet” R package with these 13 candidate genes (Fig. 6C), which demonstrated high accuracy in predicting immune therapy responses (accuracy of 0.9424) as shown in Fig. 6D. Further analysis using the random forest algorithm revealed that KDM5A was the most important gene for predicting immune therapy response (Fig. 6E). According to the study of Li et al [25], we found that KDM5A deficiency enhances the efficacy of immunotherapy in KRAS-mutant LUAD cell model, ranking 10th out of 727 candidate genes (Fig. 6F). Furthermore, we performed molecular docking analysis and identified a molecule inhibitor: CPI-455, which exhibits a high binding affinity to KDM5A (a docking score of -9.2 kcal/mol indicates a strong and stable binding interaction, as scores below -7.0 kcal/mol are generally considered indicative of high-affinity binding) (Fig. 6G).

Fig. 6.

Targeting KDM5A enhances the sensitivity to immunotherapy in KRAS-mutant LUAD. (A) A volcano plot shows differentially expressed genes between responders and non-responders, and 13 candidate genes were identified with p-values < 0.05. (B) Expression of KDM5A in KRAS-mutant epithelial cells. (C) Neural network model using the "neuralnet" R package with 13 candidate genes, shows high accuracy in predicting immune therapy responses. (D) Confusion matrix evaluates the model’s performance with various metrics. (E) Random forest algorithm identifies KDM5A as the most important gene for predicting immunotherapy response in KRAS-mutant LUAD. (F) KDM5A deficiency enhances the efficacy of immunotherapy in the KRAS-mutant LUAD cell model from a previous study by Li et al. (G) Molecular docking analysis shows high binding affinity (-9.2 kcal/mol) between CPI-455 and KDM5A protein. (H-J) Changes in tumor volume over 4 weeks in the control, anti-PD-1, and combination-treated groups. (K) Changes of mean tumor volumes for each group, and the difference was evaluated using the two‑way repeated‑measures ANOVA with Tukey’s post‑hoc test (p < 0.001). (L) Tumor samples from each group. (M) Body weight of mice in the three groups. (N) Flow cytometry measures T cell and CD8⁺ T cell subpopulation in tumors. (O & P) Inhibition of KDM5A increases total T cells and CD8⁺ T cells in the combination-treated group. ***p < 0.001.

To evaluate the impact of targeting KDM5A on immunotherapy for KRAS-mutant LUAD in vivo, we employed the mouse LUAD cell line LLC to establish subcutaneous xenografts in C57BL/6 mice. The mice were subjected to treatment with PBS control, anti-PD-1 monoclonal antibody (mAb), and a combination of the KDM5A inhibitor (CPI-455) and anti-PD-1 treatment. Fig. 6H-J respectively illustrate the changes in tumor volume of each mouse in the control group, anti-PD-1 group, and the combination-treated group over 4 weeks, and the mean values ± SD of the measured tumor volumes for each group were shown in Fig. 6K, and the difference among different groups was evaluated using the two‑way repeated‑measures ANOVA with Tukey’s post‑hoc test (p < 0.001). Tumor samples were harvested for further analysis. The tumors in the combination-treated group were significantly smaller than the PBS-control group and anti-PD-1 alone groups (Fig. 6L). The body weight of each mouse in the three groups was measured and recorded as well (Fig. 6M). Flow cytometry analysis was performed to quantify the T cell subpopulations in the tumors from each group (Fig. 6N). We observed that inhibition of KDM5A could significantly increase total T cells and CD8⁺ T cell proportion in the tumors of the combination-treated group (Fig. 6, Fig. 6). These findings demonstrated that KDM5A serves as a biomarker for immunotherapy response in KRAS-mutant LUAD individuals, and targeting KDM5A can enhance the sensitivity to immunotherapy in KRAS-mutant LUADs.

Discussion

KRAS-mutant LUAD poses a significant therapeutic challenge due to its aggressive behavior, lack of directly targetable drivers, and frequent primary or acquired resistance to immunotherapy such as ICB [26]. Despite the efficacy of PD-1/PD-L1 blockade in a subset of NSCLC patients, KRAS-mutant tumors often exhibit an immunologically “cold” phenotype, characterized by poor T cell infiltration and diminished response rates [27]. Accumulating evidence suggests that metabolic reprogramming—a well-established hallmark of cancer—not only sustains tumor proliferation but also actively shapes an immunosuppressive TME [6,8]. Altered glucose, fatty acid, and amino acid metabolism have been implicated in immune evasion [28]. However, the role of galactose metabolism in KRAS-driven LUAD remains poorly understood.

In this study, we focused on galactose metabolism as a previously underexplored metabolic pathway in KRAS-mutant LUAD and demonstrated its critical role in shaping an immunosuppressive TME and modulating response to immunotherapy. We leveraged multi-omics data to reveal that GM is significantly upregulated in KRAS-mutant tumors and correlates with reduced immune cell infiltration—particularly cytotoxic CD8⁺ T cells. Through an integrative approach combining scRNA-seq, bulk transcriptomics, and bioinformatic analyses, we identified distinct GM-derived molecular subtypes with divergent clinical outcomes, constructed a robust prognostic model using machine learning, and validated KDM5A as both a predictive biomarker and a therapeutic target capable of sensitizing KRAS-mutant tumors to anti-PD-1 therapy. These findings position GM as a key metabolic regulator in LUAD immune evasion and provide a compelling rationale for targeting associated epigenetic mediators to overcome ICB resistance.

One of the central findings of our study is the pronounced metabolic heterogeneity between KRAS-mutant and wild-type LUAD, with galactose metabolism emerging as the most significantly altered pathway. Single-cell transcriptomics revealed that malignant epithelial cells in KRAS-mutant tumors exhibit elevated GM activity, which correlated with reduced immune infiltration—a hallmark of immune evasion. This aligns with growing evidence that metabolic reprogramming is not merely a bystander effect but an active driver of immunosuppression. Tumors often exploit metabolic pathways to create a nutrient-depleted, acidic microenvironment that impairs T cell function and promotes regulatory immune cell populations. Our observation that high-GM epithelial cells in KRAS-mutant samples drive unique intercellular communication networks—particularly through pathways such as GRN and EPHA—further supports the notion that metabolic activity directly modulates cellular crosstalk within the TME.

The application of NMF clustering based on GM-related genes allowed us to stratify KRAS-mutant LUAD into two subtypes with distinct clinical and molecular features. The C1 subtype, characterized by high GM activity, was associated with higher mortality and prevalence of KEAP1 mutations—a gene known to cooperate with KRAS in promoting aggressive tumor behavior and therapy resistance. In contrast, the C2 subtype displayed higher tumor mutation burden, immune infiltration, and better overall survival, suggesting that GM may serve as a metabolic biomarker for immune-excluded phenotypes. These findings are particularly relevant in the context of immunotherapy, where patient stratification based on metabolic features could improve response prediction and therapeutic planning. Co-occurring mutations such as KEAP1 and STK11 are known to modulate immune landscapes and ICB response in KRAS-mutant LUAD. Our finding that KEAP1 mutations are enriched in the high-GM C1 subtype suggests a potential synergistic role in driving immune exclusion and resistance to immunotherapy.

Leveraging machine learning, we developed a robust prognostic model integrating GM-related genes, which outperformed traditional clinicopathological variables in predicting overall survival. The model’s generalizability across independent cohorts underscores the reliability of GM-related signatures as prognostic tools. Moreover, our neural network and random forest analyses identified KDM5A as a top predictor of immunotherapy resistance. KDM5A, a histone demethylase involved in epigenetic regulation, has been implicated in immune evasion and therapy resistance across multiple cancer types. In ovarian cancer, KDM5A suppresses antigen presentation via downregulating MHC class I genes [29]. In breast cancer, it promotes endocrine therapy resistance by erasing H3K4me3 at estrogen-responsive genes [30]. In hepatocellular carcinoma, KDM5A overexpression correlates with poor prognosis and immune exclusion [31]. These findings contextualize our discovery that KDM5A is a critical regulator of immunotherapy response in KRAS-mutant LUAD. Its upregulation in non-responders to anti-PD-1 therapy suggests a role in establishing an immune-suppressive niche, possibly through altering chromatin accessibility of immune-related genes. KDM5A is a druggable epigenetic target, with small-molecule inhibitors such as CPI-455, KDM5-C70, and YUKA1 showing preclinical efficacy. Its translational value lies in overcoming immunotherapy resistance in ‘cold’ tumors like KRAS-mutant LUAD. However, challenges remain: (1) suboptimal selectivity over other KDM5 family members; (2) potential on-target toxicities in normal hematopoietic stem cells; (3) need for predictive biomarkers. To date, no KDM5A inhibitor has entered clinical trials, but several are in preclinical development.

The translational significance of our work is strongly supported by in vivo experiments showing that pharmacological inhibition of KDM5A with CPI-455 synergizes with anti-PD-1 therapy to enhance T cell infiltration and suppress tumor growth. This combination strategy effectively reversed the immune-excluded phenotype characteristic of KRAS-mutant LUAD, providing a compelling rationale for targeting epigenetic-metabolic crosstalk in immunotherapy-resistant tumors. The high binding affinity between CPI-455 and KDM5A, as confirmed by molecular docking, further supports the feasibility of this therapeutic approach.

Several limitations should be acknowledged. First, our scRNA-seq analysis included only one KRAS-mutant LUAD sample (Patient number: GBG) and three wild-type samples. While the findings were robust in this dataset, the small number of mutant samples limits the generalizability of the cellular heterogeneity observed. Larger-scale single-cell studies with multiple KRAS-mutant donors are warranted to confirm the metabolic and immune cell atlas described here. Second, the precise mechanism by which GM influences immune cell function and how KDM5A modulates this process remains to be fully elucidated. Functional studies using genetic knockdown or overexpression models will help dissect the causal relationships between GM, KDM5A, and immune evasion. Third, although our neural network model achieved high predictive accuracy, the risk of overfitting cannot be entirely excluded given the modest sample size of KRAS-mutant LUAD patients. Future prospective studies with larger independent cohorts are needed to validate the generalizability of this model. Moreover, the training (GSE72094, microarray) and testing (TCGA-LUAD, RNA-seq) cohorts were derived from different platforms, and although we converted FPKM to TPM, no additional cross-platform batch correction (e.g., ComBat) was applied; this may introduce batch effects that could affect the model's performance, and should be addressed in future multi-cohort validations. Finally, the LLC syngeneic model, while widely used, carries a Kras^G12D mutation, which is less prevalent than KRAS^G12C in human LUAD. Additionally, subcutaneous tumor models do not fully recapitulate the native lung microenvironment. Future studies should employ orthotopic models and patient-derived xenografts (PDXs) with different KRAS alleles to validate our findings.

In conclusion, our study identified galactose metabolism as a key metabolic regulator in KRAS-mutant LUAD that drives immune suppression and resistance to immunotherapy. We provide a machine learning-based prognostic framework and nomogram for overall survival prediction, and identify KDM5A as both a predictive biomarker and a therapeutic target for immunotherapy for KRAS-mutant LUAD. The combination of KDM5A inhibition with anti-PD-1 therapy represents a promising strategy to reprogram the immunosuppressive TME and improve outcomes in this challenging subset of lung cancer. These findings highlight the potential of targeting metabolism-related hub genes to overcome resistance to immunotherapy in KRAS-driven LUADs.

Ethics statement

The animal experimental protocols were confirmed and approved by the Institutional Animal Care and Use Committee of Nanjing Medical University (IACUC of NMU).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Disclosure

All the authors have read and approved the manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

CRediT authorship contribution statement

Long Qian: Methodology, Conceptualization. Juan Zhao: Writing – review & editing, Supervision, Investigation. Mengtao Fan: Writing – review & editing, Validation, Software, Investigation. Chuanxian Hu: Writing – review & editing, Validation, Investigation. Zhuqing Ji: Methodology, Conceptualization.

Declaration of competing interest

The authors declare no conflicts of interest.