Metabolic pathway signatures define prognostic subtypes of lung adenocarcinoma Authors: Doris Kafita, Zitha Mukwamba, Vanity Mapulanga, Adon Chawe, Musalula Sinkala Published: 2026 ABSTRACT AbstractMetabolic reprogramming is a core hallmark of cancer, yet how it contributes to the clinical heterogeneity of lung adenocarcinoma (LUAD) remains poorly understood. Here, by applying unsupervised clustering to the metabolic transcriptomes of 510 LUAD tumours, we identify two robust and prognostically significant subtypes. We show that these subtypes have divergent clinical outcomes, with subtype-1 patients exhibiting significantly longer overall and disease-specific survival compared to the more aggressive subtype-2. This clinical distinction is underpinned by fundamentally different metabolic architectures: the aggressive subtype-2 is characterized by the upregulation of central carbon metabolism, including the citric acid cycle and respiratory electron transport, whereas the less aggressive subtype-1 shows a distinct enrichment for choline catabolism. Consistent with this, Cox regression analysis reveals that high expression of the TCA cycle enzyme OGDH is a top predictor of increased disease risk, while the glycolytic enzyme PGK1 is associated with a decreased risk. Our findings establish a direct link between transcriptional metabolic states and patient survival in LUAD, defining a framework for prognostic stratification and identifying subtype-specific metabolic vulnerabilities for therapeutic targeting.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s12672-026-05480-5.Keywords: Lung adenocarcinoma, Metabolic reprogramming, Bioinformatics, Machine learning, Transcriptome, Tumour subtyping, Prognosis, Survival analysis, Citric acid cycle, Choline metabolism Introduction Lung adenocarcinoma (LUAD), the most prevalent subtype of non-small cell lung cancer, represents a major clinical challenge due to its profound molecular and pathological heterogeneity [1, 2]. This heterogeneity drives variable responses to therapy and contributes to a bleak prognosis, with five-year survival rates for advanced stages often falling below 20% [3]. A deeper understanding of the biological drivers underpinning this clinical variability is therefore critical for developing more effective, stratified therapeutic strategies. Among the core biological capabilities acquired by cancer cells, the reprogramming of cellular metabolism has emerged as a fundamental hallmark [4]. Tumour cells dynamically rewire their metabolic pathways to meet the heightened bioenergetic and biosynthetic demands of continuous proliferation and survival in challenging microenvironments [5–8]. In lung cancer, this metabolic dysregulation is not merely a consequence of oncogenic transformation but is an active contributor to disease progression and therapeutic resistance [9, 10]. This positions metabolic pathways as a rich source of potential biomarkers for patient stratification and as targets for novel therapeutic interventions [7, 9, 11]. Moreover, recent studies have highlighted the intimate link between metabolic reprogramming and the tumour immune microenvironment in LUAD, with evidence that metabolic subtypes exhibit distinct immune contextures and differential responses to immunotherapy [12–14] . Given the central role of metabolism in cancer, several studies have leveraged transcriptomic data to explore the metabolic heterogeneity of LUAD [15–17]. For instance [15], identified four metabolic subtypes of LUAD based on glycolysis and glutaminolysis gene expression patterns, each exhibiting distinct immune microenvironment features and clinical outcomes. Wu et al. [16] characterised metabolic phenotypes in LUAD and demonstrated their association with the tumour immune microenvironment, including differential immune cell infiltration and PD-L1 expression levels. Similarly, Huang et al. [18] developed a metabolism-related gene classification that stratified LUAD patients into prognostically distinct groups with different pathway activation patterns. While these efforts have provided valuable insights, a consensus classification that robustly links metabolic subtypes to distinct clinical outcomes has yet to be established, likely due to variations in cohort selection and analytical methodologies. A large-scale, systematic analysis is therefore needed to delineate the metabolic landscape of LUAD and its prognostic implications. Results Unsupervised clustering of metabolic gene expression reveals two distinct LUAD subtypes To investigate the metabolic heterogeneity in LUAD, we analysed the transcriptomes of 510 tumours from The Cancer Genome Atlas (TCGA). Our analysis focused on the expression profiles of a curated set of 2,325 genes representing the major human metabolic pathways catalogued in the Reactome database. After applying a multi-step feature selection procedure to reduce the initial 2,325 metabolic genes to approximately 500 of the most informative (see Methods), we performed unsupervised k-means clustering on this filtered gene set and identified two robust and distinct molecular subtypes, hereafter designated subtype-1 (n = 356) and subtype-2 (n = 154) (Fig. 1). The PCA visualisation in Fig. 1 was generated using the expression of all 2,325 metabolic genes to provide a comprehensive view of the metabolic landscape, while the cluster assignments shown were derived from k-means clustering on the filtered ~ 500-gene set. Transcriptomic analysis reveals two distinct metabolic subtypes of lung adenocarcinoma. A scatter plot of the first two principal components (PC1 and PC2) based on the mRNA expression of all 2,325 metabolic genes. Cluster assignments (colours) were derived from k-means clustering on the filtered ~ 500-gene subset (see Methods). The analysis separates 510 LUAD tumours into two distinct clusters: Subtype-1 (n = 356, orange) and Subtype-2 (n = 154, blue). The percentage of variance explained by each principal component is shown in parentheses on the axes We performed multiple complementary analyses to validate the robustness of this two-cluster solution. Silhouette analysis across k = 2, 3, and 4 demonstrated that k = 2 yielded the highest mean silhouette score (0.393), substantially exceeding k = 3 (0.217) and k = 4 (0.208), indicating that the two-cluster partition produces the most internally cohesive and well-separated groups. A per-sample silhouette plot for k = 2 confirmed that most samples in both clusters had positive silhouette coefficients, with very few misassigned samples. Principal component analysis revealed that PC1 alone explains 80.7% of the variance in metabolic gene expression, and kernel density estimation along PC1 showed a clear bimodal distribution with two distinct peaks corresponding to each subtype, confirming that the subtypes represent genuinely separable populations rather than an arbitrary bisection of a continuous distribution. A Hartigan’s dip test on PC1 projections rejected the null hypothesis of unimodality. Bootstrap cluster stability analysis (100 resampled datasets) demonstrated 97.2% mean agreement with the original cluster assignments, indicating high reproducibility of the subtype designations. Finally, the gap statistic confirmed k = 2 as optimal (gap ≈ 1.416) (Supplementary Figures S1–S5). Metabolic subtypes are associated with distinct clinical and pathological features To determine if the metabolic subtypes corresponded to different clinical phenotypes, we compared their associated pathological and outcomes data (Fig. 2). This analysis revealed a consistent pattern of increased disease aggressiveness in subtype-2. Patients with subtype-2 tumours had a significantly higher proportion of deceased individuals (44.2%) compared to subtype-1 (32.9%; p = 0.023) (Fig. 2c). This poorer outcome was associated with more advanced disease, as subtype-2 tumours were enriched for higher pathological stages (Fig. 2b; p < 0.001) and were more likely to be present after the first course of treatment (35.2% vs. 22.2% in subtype-1; p = 0.016) (Fig. 2a). Although not reaching statistical significance, a trend towards higher rates of disease recurrence or progression was also observed in subtype-2 (Fig. 2d; p = 0.35). The patient demographics were broadly similar across subtypes, with most individuals in both groups being White (Subtype-1: 267 White, 33 Black or African American, 7 Asian; Subtype-2: 117 White, 19 Black or African American, 1 Asian). Race was not specified for 49 and 16 patients in subtype-1 and subtype-2, respectively. Taken together, these clinical characteristics establish subtype-2 as a more aggressive form of LUAD associated with more advanced disease and poorer patient outcomes. Metabolic subtypes are associated with distinct clinical phenotypes. Bar plots and pie charts comparing the clinical features of Subtype-1 and Subtype-2. a Proportion of patients with or without tumours after the first course of treatment. b Distribution of AJCC pathological tumour stages. c Percentage of patients classified as living or deceased at the time of data collection. d Proportion of patients who were disease-free versus those who had recurred or progressed after initial treatment Subtype-1 exhibits a significant survival advantage To assess the clinical relevance of the metabolic subtypes, we performed Kaplan-Meier survival analysis [19]. This revealed a striking difference in patient outcomes between the two groups (Fig. 3). Patients with subtype-1 tumours demonstrated a significant survival advantage, with a median overall survival (OS) of 52.6 months compared to just 37.7 months for patients with subtype-2 tumours (log-rank test, p = 0.007) (Fig. 3a). A similar and even more pronounced advantage was observed for disease-specific survival (DSS), with a median of 88.1 months for subtype-1 versus 49.8 months for subtype-2 (p = 0.004) (Fig. 3b). In contrast, no significant differences were observed for disease-free survival (DFS; p = 0.48) or progression-free survival (PFS; p = 0.28) (Fig. 3c, d). These results firmly establish that our metabolic classification stratifies LUAD patients into distinct prognostic groups, with subtype-1 representing a less aggressive disease course. Subtype-1 is associated with a significant survival advantage. Kaplan-Meier survival curves comparing the two metabolic subtypes. Patients in Subtype-1 show significantly better. a Overall Survival (OS; p = 0.007) and b Disease-Specific Survival (DSS; p = 0.004) compared to patients in Subtype-2. No significant differences were observed for c Disease-Free Survival (DFS; p = 0.477) or d Progression-Free Survival (PFS; p = 0.275). P-values were calculated using the log-rank test Individual metabolic genes are predictive of patient survival To pinpoint specific metabolic drivers of the observed survival differences, we performed a Cox proportional hazards regression analysis on a per-gene basis. Following Benjamini–Hochberg false discovery rate (FDR) correction of all 187 univariate Cox p-values, 6 genes remained significant at FDR < 0.05, while an additional 44 genes were nominally significant (p < 0.05). This identified 50 metabolic genes whose expression was associated with overall survival (p < 0.05; Supplementary File 1). Of the six FDR-significant genes, three were associated with an increased risk of death (GDE1, RPS25, OGDH; positive beta coefficient), while three were associated with a protective effect (PGK1, DCN, LUM; negative beta coefficient). An additional 44 genes were nominally significant (p < 0.05) and are reported as exploratory findings. Notably, high expression of OGDH, a key enzyme in the TCA cycle, was the most significant predictor of poor survival (p = 0.0006, FDR = 0.048), followed by GDE1 and LPGAT1. Conversely, high expression of the glycolytic enzyme PGK1 was most significantly associated with improved survival (p = 0.00015, FDR = 0.041), along with DCN and RPL6. The list of significant genes also included other key enzymes of glycolysis and the TCA cycle, such as CS, HK2, ENO1, PFKL, and ALDOA, further underscoring the central role of these pathways in shaping LUAD prognosis. To assess whether the metabolic subtype classification serves as an independent prognostic factor beyond established clinical variables, we performed multivariable Cox proportional hazards regression with subtype, pathological stage, age at diagnosis, and sex as covariates. For overall survival (N = 483), Subtype-2 was independently associated with worse prognosis (HR = 1.39, 95% CI: 1.01–1.90, p = 0.041), while pathological stage was also a significant predictor (HR = 1.64, p < 0.001). Age showed a trend toward significance (p = 0.095), whereas sex was not a significant covariate (p = 0.99). For disease-specific survival (N = 482), Subtype-2 demonstrated an even stronger independent prognostic effect (HR = 1.57, 95% CI: 1.05–2.35, p = 0.029). Forest plots depicting these multivariable Cox regression results are presented in Fig. 4b and c. Cox regression analysis of metabolic genes and subtype prognostic value. a Univariate Cox volcano plot for 187 metabolic genes. FDR-significant genes (FDR < 0.05) are labelled: risk genes (red) and protective genes (blue). b, c Multivariable Cox forest plots for overall survival (b) and disease-specific survival (c), adjusting for subtype, stage, age, and sex. Subtype-2 is independently associated with worse OS (HR = 1.39, p = 0.041) and DSS (HR = 1.57, p = 0.029) Subtypes exhibit distinct gene expression signatures To understand the molecular basis for the different clinical outcomes, we compared the transcriptomic profiles of the two metabolic subtypes. Differential expression analysis revealed that 129 genes were significantly upregulated in the less aggressive subtype-1, while 149 genes were upregulated in the more aggressive subtype-2 (see Fig. 5 and Supplementary File 2). Among the genes upregulated in subtype-2, 18 were directly involved in metabolism. The most significantly upregulated metabolic genes included ALB (albumin; fold-change [FC] = 10.6), FABP7 (Fatty Acid Binding Protein 7; FC = 23.9), and G6PC (Glucose-6-Phosphatase Catalytic Subunit; FC = 6.8). While typically a marker of liver function, ectopic ALB expression has been documented in non-hepatic tumours and albumin processing within tumours contributes nutrients to the tumour microenvironment [20, 21]. FABP7, which facilitates fatty-acid handling, promotes proliferation and metastasis in non-small cell lung cancer [22]. Similarly, aberrant G6PC expression can rewire glucose metabolism, supporting metabolic plasticity and tumour growth [23, 24]. Conversely, the top upregulated genes in the less aggressive subtype-1 included KRT20 (Keratin 20; FC = 9.1) and REG4 (Regenerating Family Member 4; FC = 24.0). KRT20 is a well-established marker of gastrointestinal differentiation, and its expression in lung adenocarcinoma is enriched in mucinous histological patterns that show distinct behaviour and variable prognosis across cohorts [25–27]. REG4 is involved in epithelial regeneration and has been identified as a biomarker in lung adenocarcinoma, with context-dependent links to prognosis across cancers [28, 29] (Fig. 5). Transcriptomic signatures distinguish the two metabolic subtypes. Volcano plot showing the results of the differential expression analysis between Subtype-1 and Subtype-2. The analysis identified 129 genes significantly upregulated in Subtype-1 (red) and 144 genes significantly upregulated in Subtype-2 (blue), based on log2 fold change and false discovery rate (FDR) To further contextualise the subtype-specific signatures, we compared the expression profile of each subtype against normal lung tissue samples. Both subtypes displayed extensive transcriptomic dysregulation compared to healthy tissue. Subtype-1 exhibited 2,397 differentially expressed genes (DEGs), comprising 1,151 upregulated and 1,246 downregulated genes (Fig. 6a). Subtype-2 showed a similar number of DEGs (2,649), with 1,120 genes upregulated and 1,529 downregulated (Fig. 6b). A substantial overlap was observed between the two subtypes, with 895 DEGs being common to both, while 256 and 225 DEGs were uniquely dysregulated in subtype-1 and subtype-2, respectively (Fig. 6c). When examining the downregulated genes specifically, we found 1,223 genes were commonly suppressed in both subtypes, while 751 and 382 genes were uniquely downregulated in subtype-1 and subtype-2, respectively (Fig. 6d). Transcriptomic comparison of LUAD subtypes with healthy lung tissue. a, b Volcano plots illustrating the landscape of differentially expressed genes in Subtype-1 and Subtype-2 relative to normal tissue. Red and blue points indicate significantly upregulated and downregulated genes, respectively. c Venn diagram showing 895 commonly dysregulated genes, with 256 and 225 genes uniquely altered in Subtype-1 and Subtype-2. d Overlap analysis of downregulated genes, highlighting 1223 commonly suppressed genes Aggressive subtype-2 is defined by upregulation of central carbon metabolism To elucidate the functional metabolic programmes defining the two subtypes, we performed Gene Set Enrichment Analysis (GSEA) [30] using the Reactome pathway [31] database. This analysis revealed a striking metabolic dichotomy between the subtypes. The aggressive subtype-2 was significantly enriched for pathways central to energy production and biosynthesis. Notably, the citric acid (TCA) cycle and respiratory electron transport—including ATP synthesis by chemiosmotic coupling—were among the most significantly upregulated pathways (Supplementary File 3). This signature of heightened mitochondrial activity was complemented by an enrichment for glucose and glycogen metabolism, as well as pathways for the biosynthesis of nucleotides and amino acids. In stark contrast, the less aggressive subtype-1 was distinguished by a significant enrichment in choline catabolism (Supplementary File 3). Collectively, these results paint a clear picture: the poor prognosis of subtype-2 is driven by a metabolic phenotype geared towards rapid energy production and the synthesis of biomass, a classic hallmark of aggressive cancers, while subtype-1 relies on a distinct metabolic programme. Subtypes are governed by distinct signalling pathway activities To explore the broader signalling networks associated with each metabolic state, we performed GSEA using the KEGG pathway [32] database. This revealed that the subtypes are governed by fundamentally different cellular programmes (Supplementary File 4). The less aggressive subtype-1 was significantly enriched for pathways related to core cellular machinery and maintenance, including DNA replication, cell cycle, nucleotide excision repair, the proteasome, and the spliceosome. This suggests a phenotype characterised by controlled proliferation and robust cellular housekeeping. In contrast, the aggressive subtype-2 was defined by pathways reflecting a state of high metabolic stress and demand. It showed significant enrichment for oxidative phosphorylation and ribosome biogenesis, consistent with its high energy and protein synthesis requirements. Intriguingly, subtype-2 was also enriched for pathways related to neurodegenerative diseases such as Alzheimer’s, Parkinson’s, and Huntington’s disease. These pathways are mechanistically linked to cellular stress responses, protein misfolding, and mitochondrial dysfunction, suggesting that the aggressive metabolic phenotype of subtype-2 tumours may render them vulnerable to these forms of cellular damage. Distinct genomic alterations underlie the metabolic subtypes To identify potential genomic drivers of the metabolic phenotypes, we analysed the landscape of somatic mutations and copy number alterations. The overall mutational profile was consistent with previous LUAD cohorts, with TP53, KRAS, KEAP1, EGFR, and STK11 being among the most frequently altered genes (Fig. 7). A comparative bar chart of mutation frequencies across subtypes is provided in Supplementary Figure S6. However, a comparative analysis revealed a starkly different distribution of key driver mutations between the subtypes (Fig. 8). The aggressive subtype-2 was significantly enriched for alterations in the tumour suppressor TP53 (67.3% in subtype-2 vs. 45.7% in subtype-1; p < 0.001). This subtype also showed a higher frequency of alterations in KEAP1 (26.8% vs. 16.0%; p = 0.005), a critical regulator of the cellular antioxidant response. Conversely, the less aggressive subtype-1 was significantly enriched for mutations in the oncogene KRAS (36.3% in subtype-1 vs. 26.1% in subtype-2; p = 0.026). These findings suggest that the divergent metabolic and clinical phenotypes are underpinned by distinct genomic events, with the aggressive subtype-2 being driven primarily by the loss of key tumour suppressors, while the less aggressive subtype-1 is more frequently associated with oncogenic KRAS activation. Genomic alteration landscape of the LUAD cohort. The oncoprint illustrates the distribution of somatic mutations and copy number alterations across the cohort for the most frequently altered genes. Each column represents a patient, and each row a gene. Clinical annotations for subtype, age, and sex are shown at the top. The bar plot on the right indicates the alteration frequency for each gene Metabolic subtypes are defined by distinct driver mutations. A scatter plot comparing the percentage of samples with genomic alterations for each gene in Subtype-1 (x-axis) versus Subtype-2 (y-axis). The size of each bubble is proportional to the statistical significance (Fisher’s exact test), and the colour indicates the log ratio of the frequencies. Key enriched genes, such as TP53 and KEAP1 in Subtype-2 and KRAS in Subtype-1, are labelled Discussion In this study, we leveraged an integrative analysis of transcriptomic, genomic, and clinical data to deconstruct the metabolic heterogeneity of lung adenocarcinoma. Our unsupervised analysis of metabolic gene expression successfully stratified LUAD tumours into two robust subtypes with starkly different clinical trajectories. We demonstrate that the less aggressive subtype-1 is associated with significantly better overall and disease-specific survival, establishing a powerful, metabolism-centric framework for prognosticating LUAD. Our subsequent analyses reveal that these subtypes represent distinct biological entities, defined by coherent differences across their transcriptomic, metabolic, and genomic landscapes, consistent with prior systems-level work using multi-omics integration and network analysis by us and others [5, 6, 11, 33]. The poor prognosis of subtype-2 is underpinned by a classic oncogenic phenotype of heightened energy metabolism and biomass production. Gene-set enrichment pointed to upregulation of central carbon metabolism, including the tricarboxylic acid (TCA) cycle and oxidative phosphorylation (OXPHOS), a state frequently linked to aggressive disease biology and worse outcomes in LUAD and other cancers [34, 35]. In vivo metabolic tracing further supports active mitochondrial metabolism in human NSCLC tumours, underscoring clinical relevance for OXPHOS-dominant programmes [36]. This metabolic state is likely shaped by the specific genomic alterations enriched in this subtype. We found that subtype-2 tumours have a higher frequency of TP53 mutations; wild-type p53 is a master metabolic checkpoint that restrains anabolic flux and coordinates redox control, so its disruption removes key brakes on biosynthesis and survival pathways [11, 37, 38]. Furthermore, enrichment for KEAP1 alterations—a central sensor of oxidative stress whose loss constitutively activates NRF2—suggests adaptation to the high-ROS environment of intensive mitochondrial respiration and a shift toward NRF2-driven metabolic and cytoprotective programmes [39–41]. In contrast, the less aggressive subtype-1 is characterised by a distinct metabolic programme and genomic landscape. This subtype was uniquely enriched for choline/phospholipid metabolism, implicating membrane biogenesis and lipid handling rather than a purely glucose-centric phenotype; choline pathway activity is a recognised axis of cancer metabolism in lung and other tumours, though with context-dependent prognostic associations [42, 43]. Genomically, subtype-1 was defined by a higher frequency of oncogenic KRAS in the relative absence of cooperating suppressor loss. Oncogenic RAS can impose intrinsic stresses and senescence programmes that constrain progression when canonical failsafe pathways remain intact, offering a plausible basis for the more indolent course we observe in this setting [44, 45]. Together, these observations support a model in which clinical behaviour in LUAD emerges from the interplay between specific driver genotypes and the metabolic states they engender, aligning with our earlier cross-cancer analyses linking altered metabolic gene programmes to clinical endpoints [6, 11]. Our findings have direct clinical implications. First, metabolic subtyping provides prognostic information that complements traditional staging and histopathology. Numerous LUAD signatures derived from mitochondrial/oxidative, lipid and broader metabolism-related genes demonstrate independent prognostic utility across TCGA and external cohorts, supporting the translational promise of metabolism-aware risk models [46, 47]. Second, the reliance of aggressive subtype-2 on OXPHOS nominates vulnerabilities to inhibitors of the mitochondrial electron transport chain; complex I inhibitors such as IACS-010759 exemplify this strategy, with strong preclinical activity and emerging (if challenging) clinical experience [48, 49]. Third, lipid/choline-biased programmes in subtype-1 suggest orthogonal opportunities in phospholipid metabolism and membrane biogenesis pathways [42]. More broadly, our systems perspective dovetails with our prior work on integrating multi-omic and network features for stratification and response prediction, including linking receptor/transcript profiles to therapeutic sensitivity—an avenue that could be extended to metabolism-targeted agents in LUAD [33, 34, 50]. Our findings align with emerging evidence linking metabolic reprogramming to the tumour immune microenvironment and therapeutic responses in LUAD. Zhang et al. [12] recently constructed a metabolic reprogramming signature that predicted immunotherapy response in LUAD, reinforcing the clinical utility of metabolism-based classifiers for guiding treatment decisions. Xia et al. [14] identified disulfidptosis-related molecular subtypes in LUAD with distinct tumour microenvironment compositions, further underscoring the interplay between metabolic states and immune contexture that we observe between our two subtypes. Additionally, Gong et al. [13] employed machine learning to define EGFR-related immune subtypes in LUAD, demonstrating that integrative molecular classification can refine treatment stratification beyond single-marker approaches. Collectively, these studies and our own results support the view that metabolic subtyping captures biologically meaningful heterogeneity with direct therapeutic relevance. This study has limitations. Our analysis uses bulk tumour transcriptomics, which may obscure metabolic heterogeneity within tumours and non-malignant compartments. Single-cell and in vivo tracer studies show pronounced intra-tumour metabolic diversity and microenvironmental influences that bulk data cannot resolve [36, 37]. Importantly, the transcriptomic signatures reported here remain correlative and require functional validation through metabolomic profiling, metabolic flux analysis, and experimental perturbation in genotype-matched models to confirm that transcriptional patterns reflect bona fide metabolic activity and therapeutic vulnerabilities [36]. Future work applying single-cell multi-omics and metabolic flux imaging will be essential to delineate cell-type-specific states and interactions. Additionally, as our inferences are correlative, functional perturbation studies—particularly in genotype-matched models (e.g., KEAP1/NRF2-high; TP53-deficient)—are needed to validate causal links between genomic alterations and metabolic phenotypes [39, 40]. Finally, while OXPHOS-targeting is mechanistically compelling, clinical translation will require biomarkers of dependency and tolerability-optimised combinations [48, 49]. In conclusion, we delineate two fundamentally different metabolic subtypes of LUAD with distinct genomic drivers and clinical outcomes. This provides a new lens on LUAD heterogeneity and a clear rationale for developing metabolism-based biomarkers and subtype-specific therapeutic strategies. Our results build on and extend systems-level insights from prior multi-omics studies—including our own—supporting the integration of metabolic state, genotype and network context to improve risk stratification and guide treatment in lung adenocarcinoma. Methods Data curation and pre-processing Lung adenocarcinoma (LUAD) patient data were obtained from The Cancer Genome Atlas (TCGA) via the NCI Genomic Data Commons (GDC) Data Portal (https://portal.gdc.cancer.gov/). The primary references for the GDC pipelines and TCGA Pan-Cancer resources are Zhang et al. and Weinstein et al. [51, 52]. The clinical cohort comprised 584 TCGA LUAD patients. Available data types included treatment outcomes, whole-exome sequencing (WES) mutations for 365 samples, RNA-seq for 517, copy-number alterations (CNA) for 516, and RPPA proteomics for 460 samples; due to assay failures and specimen quality, not all data types were available for all patients. TCGA RPPA generation and processing and the TCPA resource are described by Akbani et al. and Li et al. [53, 54]. All data are publicly accessible from the GDC repository. For comparative analyses with normal tissue, RNA-seq from GTEx v8 was used [55]. When we required uniformly processed counts across TCGA and GTEx, we used recount2 (Rail-RNA–based harmonisation) [56, 57]. Where noted, we additionally referenced large-scale reprocessed RNA-seq compendia from ARCHS4 [58]. Metabolic pathways and genes were defined from Reactome (v68.25) [31]. We focused on the top-tier “Metabolism” term (R-HSA-1430728), which hierarchically covers 16 curated first-tier metabolic pathways encompassing 2325 genes. KEGG pathway collections were used for complementary enrichment analyses [32]. Unless stated otherwise, expression values were log2-transformed normalized counts. Specifically, RNA-Seq v2 RSEM-normalised counts were downloaded from cBioPortal and log2-transformed for clustering, PCA, GSEA, and Cox regression analyses. For differential expression between the two tumour subtypes, we used these log2-transformed values with fold-change and FDR-based thresholds.For integrative analysis downstream, we retained the 510 LUAD tumours with complete data for the relevant analyses. Machine learning identifies two metabolic LUAD subtypes We first restricted the expression matrix to the 2325 Reactome metabolism genes [31]. To identify robust metabolic subtypes, we first implemented a feature selection workflow in MATLAB R2022b (The MathWorks Inc.) to reduce noise and retain the most informative genes. This multi-step process involved sequentially removing genes with low variance (bottom 10th percentile), those with an absolute log2 fold-change of less than 1, and those with low entropy (bottom 80th percentile). This filtering cascade yielded the approximately 500 most variable metabolic genes, which were used as input for k-means clustering. For PCA visualisation (Fig. 1), the expression of all 2,325 metabolic genes was used to provide a comprehensive representation of the metabolic landscape, with cluster assignments from the filtered gene set overlaid. To determine the optimal number of clusters in an unbiased manner, we employed the evalclusters function to systematically evaluate several internal validation indices, including the Silhouette, Calinski–Harabasz, Davies–Bouldin, and GAP statistics. This evaluation was performed across three distinct clustering algorithms: k-means, Gaussian mixture models, and hierarchical linkage [59, 60]. This evaluation was performed across three distinct clustering algorithms: k-means, Gaussian mixture models, and hierarchical linkage. A consensus of these criteria robustly indicated that the optimal number of clusters was two, with the k-means algorithm providing the solution with the highest silhouette score. To further validate the two-cluster solution, we computed per-sample silhouette score distributions for k = 2, 3, and 4, performed kernel density estimation along PC1 to assess bimodality, conducted Hartigan’s dip test for unimodality, performed a bootstrap cluster stability analysis (100 resamples), and evaluated the gap statistic across k = 2–6. All metrics consistently supported k = 2 as the optimal and stable partitioning of the data (Supplementary Figures S1–S5). We then performed k-means clustering on the filtered expression data using the squared Euclidean distance metric and 100 replicates to ensure a stable solution [61]. The resulting cluster assignment was used for all downstream analyses. Survival and Cox regression analysis Subtype differences in survival (OS, DSS, PFS, DFS) were assessed using Kaplan–Meier estimators with log-rank tests [62]. For gene-level associations, we fit univariate Cox proportional hazards models for each metabolic gene [63]. Reported p-values reflect Wald tests unless specified. Multiple-testing correction was applied using the Benjamini–Hochberg procedure; genes with FDR < 0.05 were considered statistically significant, while those with uncorrected p < 0.05 only are reported as exploratory. Multivariable Cox models included metabolic subtype, AJCC pathological stage (ordinal I–IV), age, and sex as covariates. Differential expression and pathway enrichment For the tumour–versus–normal comparison, we used harmonised RNA-Seq counts from the recount2 resource (which uniformly processes TCGA and GTEx data via Rail-RNA) and applied the DESeq2 negative binomial framework. We required a minimal expression threshold (≥ 1 FPKM in at least one group) and we conducted differential gene expression analysis using the negative binomial test [64]. Multiple testing was controlled using false-discovery-rate procedures (Benjamini–Hochberg family); for a PubMed-indexed reference describing FDR concepts and BH control see Benjamini & Yekutieli and related tutorials [65]. Unless stated otherwise, significance was set at FDR < 0.05 and |log2FC| > 1. Gene Set Enrichment Analysis (GSEA) identified pathways enriched between subtypes [30], using curated Reactome and KEGG collections [32]. For pathway-level visualization, we mapped z-scored expression onto PathVisio 3 templates (Low: z < − 0.5; No change: − 0.5 ≤ z ≤ 0.5; High: z > 0.5) [66]. Genomic alteration analysis Significantly mutated genes were called with MutSigCV, which accounts for context-specific background mutation rates [67]. Significant focal copy-number alterations were identified using GISTIC 2.0 [68]. Alteration frequencies were computed per subtype, and differences were evaluated using Fisher’s exact test [69]. For pathway-centric visualization of subtype-specific genomic changes, we used PathwayMapper to overlay mutation/CNA data onto signalling pathways [70]. Statistical analyses Unless stated otherwise, all analyses were performed in MATLAB R2022b. Two-sided p-values are reported throughout. Associations between categorical variables were tested with Fisher’s exact test. For continuous variables, we used independent-samples Student’s t-test (two groups) or one-way ANOVA (≥ 3 groups) when distributional assumptions were approximately met; otherwise we used the Wilcoxon rank-sum test (two groups) or the Kruskal–Wallis test (≥ 3 groups). For single comparisons, significance was set at p < 0.05. For families of comparisons, p-values were adjusted using the Benjamini–Hochberg false-discovery-rate procedure, and significance was assessed at FDR < 0.05. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Author contributions M.S., V.M, Z.M., and D.K. conceptualised the study, designed the methodology, performed the formal analysis, curated the data, created the visualisations, and wrote the original manuscript draft. V.M. and A.C. methodology, and formal analysis. M.S. conceptualised, and visualisation, and supervised the study. All authors reviewed and edited the manuscript and approved the final version. Funding This research was conducted without financial support from any external funding agency in the public, commercial, or non-profit sectors. Data availability The data that underlie the results reported in this study are available from the following repositories: The Cancer Genome Atlas (https://portal.gdc.cancer.gov/repository), Reactome pathways (https://reactome.org/PathwayBrowser/#/R-HSA-1430728), GTEx portal (https://gtexportal.org/home/), and ARCHS4 (https://amp.pharm.mssm.edu/archs4/). These data sources were accessed under appropriate licensing and ethical requirements. Therefore, access to these repositories may require registration and approval from the respective repositories. Declarations Ethics approval and consent to paricipate To ensure ethical compliance, the protocol of this study was approved by the University of Zambia Biomedical Research Committee. The study involved the analysis of publicly available datasets collected by the TCGA, Reactome Pathways, ARCHS4, and other databases from participants who provided informed consent. All methods adhered to the relevant policies, regulations, and guidelines provided by the TCGA and other databases for analysing their datasets and reporting the findings. Consent for publication Not applicable. Competing interests Musalula Sinkala declares they are an Editorial Board Member of Discover Oncology and confirms that they were not involved in the handling or decision-making of their own submission. The rest of the authors declare that they have no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Data Availability Statement The data that underlie the results reported in this study are available from the following repositories: The Cancer Genome Atlas (https://portal.gdc.cancer.gov/repository), Reactome pathways (https://reactome.org/PathwayBrowser/#/R-HSA-1430728), GTEx portal (https://gtexportal.org/home/), and ARCHS4 (https://amp.pharm.mssm.edu/archs4/). These data sources were accessed under appropriate licensing and ethical requirements. Therefore, access to these repositories may require registration and approval from the respective repositories.