Data for The Cancer Genome Atlas (TCGA)-LUAD cohort were obtained from the University of California Santa Cruz Xena platform (https://xenabrowser.net/datapages/) for clinical and survival data. Raw count data were downloaded from the GDC data portal (https://portal.gdc.cancer.gov/) using the STAR-Counts workflow. The dataset included gene expression profiles from 515 LUAD tissue samples and 59 adjacent normal tissue samples, together with survival, clinical and mutation data for the LUAD samples. Tumour mutational burden (TMB) data for LUAD samples were retrieved from TCGA database (https://tcga-data.nci.nih.gov/tcga/). In addition, the GSE31210 dataset was downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE31210), maintained by the National Center for Biotechnology Information (NCBI). This dataset contained gene expression profiles and survival information for 226 LUAD tissue samples and was generated using the GPL570 platform (Affymetrix Human Genome U133 Plus 2.0 Array). A total of 1,793 URGs were obtained from the iUUCD 2.0 database (integrated annotations for Ubiquitin and Ubiquitin-like Conjugation Database; http://iuucd.biocuckoo.org/) (Table SI). TRGs were subsequently identified through immune infiltration analysis and weighted gene co-expression network analysis (WGCNA), as described further below. All data were downloaded on 9th December 2024.

The DESeq2 package (v. 1.42.0; http://bioconductor.org/packages/DESeq2/) was used to identify differentially expressed genes (DEGs) between LUAD and normal tissues in the TCGA-LUAD cohort. Raw count data were used as input. DEGs were defined using the thresholds |log2 fold change (FC)| >0.5 and adjusted P<0.05. Volcano plots were generated using the ggplot2 package (v. 3.5.1; http://ggplot2.tidyverse.org) to visualise the DEG results, with the 10 most upregulated and 10 most downregulated genes labelled. The expression patterns of these 20 selected DEGs were subsequently visualised using the pheatmap package (v. 1.0.12; http://CRAN.R-project.org/package=pheatmap). To identify T-cell subsets, the CIBERSORT algorithm (v. 0.1.0; http://cibersort.stanford.edu/) was applied with 1,000 permutations and relative mode enabled to estimate the proportions of 22 immune cell types in TCGA-LUAD samples. Only samples with a CIBERSORT P<0.05 were included in subsequent immune infiltration analyses. The Wilcoxon rank-sum test, performed using R software (v. 4.3.1; R Foundation for Statistical Computing; http://www.r-project.org/), was then used to identify T-cell subsets that differed significantly between LUAD and normal samples (P<0.05).

WGCNA was performed using the TCGA-LUAD dataset to identify TRGs. The WGCNA package [v. 1.7.1; Comprehensive R Archive Network (CRAN), https://CRAN.R-project.org/package=WGCNA] was used, with the differentially infiltrated T-cell subsets defined as phenotypic traits. First, samples were hierarchically clustered to identify and exclude potential outliers. Network topology analysis was then performed using the pickSoftThreshold function to determine the optimal soft-thresholding power (β) and assess the scale-free topology fit index (R²). A gene co-expression network was subsequently constructed. Gene modules were identified using the dynamic tree-cutting algorithm, with a minimum module size of 100 genes. Modules with similar expression profiles were then merged based on the hierarchical clustering of the topological overlap matrix, using a MEDissThres value of 0.4. Correlations between the resulting modules and T-cell subsets were assessed using Pearson correlation analysis with the psych package (v. 2.4.3; Comprehensive R Archive Network (CRAN); http://CRAN.R-project.org/package=psych). The correlation coefficient was denoted as cor, and its absolute value (|cor|) was used to indicate the strength of the correlation. Associations with P<0.05 and |cor| >0.3 were considered significant. The modules showing the strongest positive and negative correlations with T-cell subsets were selected as key modules, and the genes within these modules were defined as TRGs.

The overlapping genes among the DEGs, TRGs and URGs were identified as candidate genes using the ggvenn package (v. 0.1.10; CRAN; http://CRAN.R-project.org/package=ggvenn). Gene Ontology (GO) and Kyoto Encyclopaedia of Genes and Genomes (KEGG) enrichment analyses were subsequently performed using the clusterProfiler package (v. 4.8.3; Bioconductor; http://bioconductor.org/packages/clusterProfiler/) to investigate the biological functions and pathways associated with these candidate genes in LUAD. P<0.05 was considered statistically significant for both analyses. The 10 pathways with the highest gene counts were selected for visualisation. In addition, a protein-protein interaction (PPI) network was constructed using the STRING database (https://string-db.org/) with an interaction score >0.4 and subsequently visualised using Cytoscape (v. 3.10.2; Cytoscape Consortium; http://cytoscape.org/).

Prognostic analysis of patients from the TCGA-LUAD dataset with available survival data was conducted using a systematic analytical workflow. First, univariate Cox regression analysis was performed for the candidate genes using the survival package (v. 3.7–0; CRAN; http://CRAN.R-project.org/package=survival). Genes meeting the criteria of P<0.05 and hazard ratio (HR)≠1 were retained for subsequent analyses. The results were visualised using a forest plot generated with the forestplot package (v. 3.1.3; CRAN; http://CRAN.R-project.org/package=forestplot). The proportional hazards (PH) assumption was then assessed for genes identified by univariate Cox regression using the cox.zph function in the survival package (v. 3.7–0; CRAN; http://CRAN.R-project.org/package=survival). Genes without significant violations of the PH assumption (P>0.05) were retained. Finally, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was performed using the glmnet package (v. 4.1–8; CRAN; http://CRAN.R-project.org/package=glmnet) to identify prognostic genes.

A patient-specific risk score was calculated using the linear combination formula ∑(β gene × expression gene), incorporating the LASSO-derived coefficients and expression levels of the prognostic genes. The prognostic model was developed using TCGA-LUAD samples with available survival data. Patients were stratified into a high-risk group (HRG) and low-risk group (LRG) according to the maximally selected rank statistic. Risk score distributions, survival status plots and prognostic gene expression heatmaps were generated to characterise heterogeneity within the cohort. The model was systematically validated in terms of survival discrimination and predictive accuracy. Differences in survival between the two risk groups were evaluated using Kaplan-Meier analysis with the log-rank test (P<0.05), implemented with the survminer package (v. 0.4.9; CRAN; http://CRAN.R-project.org/package=survminer). Time-dependent receiver operating characteristic (ROC) curves were generated using the survivalROC package (v. 1.0.3; CRAN; http://CRAN.R-project.org/package=survivalROC) to assess predictive performance for 1-, 3- and 5-year survival. Area under the curve (AUC) values >0.6 were considered indicative of satisfactory discriminatory performance. External validation using the GSE31210 cohort was performed to further assess the generalisability of the model. Decision curve analysis (DCA) was conducted using the rmda package (v. 1.9; CRAN; http://CRAN.R-project.org/package=rmda) across a range of threshold probabilities to evaluate the clinical utility of the nomogram. Its net benefit was compared with that of the risk score alone and with each individual clinical factor.

It was then assessed whether the risk score and clinicopathological characteristics, including age, sex, T category, N category, M category and overall stage, could serve as independent prognostic indicators. The risk signature was integrated with clinical information from patients from the TCGA-LUAD dataset with complete survival data. Univariate Cox regression analysis was initially performed using the survival package (v. 3.7–0) and included the risk score and six clinical variables: Age, sex, T stage (T1-T4), nodal status (N0/N1), metastatic status (M0/M1) and overall stage (stages I–IV). Variables meeting the criteria of P<0.05 and HR≠1 were retained for further analysis. The PH assumption was assessed using the cox.zph function in the survival package (v. 3.7–0), with P>0.05 indicating that the assumption was satisfied. Variables that met the PH assumption were subsequently included in a multivariate Cox regression analysis using the survival package (v. 3.7–0). Factors with P<0.05 in the multivariate analysis were considered independent prognostic factors. A nomogram incorporating these independent factors was then constructed using the rms package (v. 6.5.0; CRAN; http://CRAN.R-project.org/package=rms) to estimate 1-, 3- and 5-year mortality probabilities in patients with LUAD. Calibration curves at 1, 3 and 5 years were also generated using the rms package (v. 6.5.0) to assess agreement between predicted and observed outcomes.

The patients from the TCGA-LUAD dataset were stratified into an HRG and LRG according to the prognostic risk model. Differential expression analysis between the two groups was performed using DESeq2 (v. 1.42.0) with raw count data to obtain log2FC values. Genes were ranked in descending order according to their log2FC values, where positive and negative values indicate upregulation and downregulation, respectively. GSEA was subsequently performed using clusterProfiler (v. 4.8.3) with the c2.cp.kegg.v7.5.1.symbols gene set obtained from the Molecular Signatures Database (MSigDB; http://www.gsea-msigdb.org/gsea/msigdb/index.jsp). Pathways with |normalized enrichment score (NES)| >1, P<0.05 and false discovery rate (FDR) <0.25 were considered significantly enriched. NES is the enrichment score normalized for gene set size and correlation with the ranked gene list. A positive or negative NES indicates enrichment at the top or bottom of the ranked list, respectively, and |NES| >1 indicates significant enrichment regardless of direction. Functional associations involving the prognostic genes were further investigated using the GeneMANIA database (http://genemania.org), which was used to predict related genes, molecular interactions and functional associations. A co-expression network linking the prognostic genes with their functionally associated genes was subsequently constructed.

For TCGA-LUAD samples with available survival data, four tumour microenvironment (TME)-related metrics were calculated using the ESTIMATE algorithm (v. 1.0.13; MD Anderson Cancer Center; http://r-forge.r-project.org/projects/estimate/): StromalScore, ImmuneScore, ESTIMATEScore and tumour purity. Differences in these metrics between the HRG and LRG were evaluated using the Wilcoxon test, with P<0.05 considered statistically significant. Immune cell infiltration involving 22 immune cell subtypes was estimated using the CIBERSORT algorithm (v. 0.1.0). The Wilcoxon rank-sum test (Mann-Whitney U-test) was then used to identify immune cell populations that differed significantly between the two risk groups (P<0.05), which were defined as differential immune cells. The test was performed using R software (version 4.3.1; R Foundation for Statistical Computing; http://www.r-project.org/). Spearman correlation analyses were subsequently performed using the psych package (v. 2.4.3). Correlations among differential immune cells were considered significant when |cor| >0.3 and P<0.05.

The TMB was calculated for each TCGA-LUAD sample with available survival data using the maftools package (v. 2.18.0; http://bioconductor.org/packages/maftools/). Waterfall plots were generated to illustrate the mutation profiles of the 20 most frequently mutated genes separately in the HRG and LRG, as well as the mutation profiles of the prognostic genes across all samples. To evaluate potential differences in response to immunotherapy between the two risk groups, Tumour Immune Dysfunction and Exclusion (TIDE) scores were obtained from the TIDE database (http://tide.dfci.harvard.edu) for TCGA-LUAD samples with available survival information. Differences in TIDE scores between the HRG and LRG were assessed using the Wilcoxon rank-sum test (Mann-Whitney U-test), with P<0.05 considered statistically significant, and visualised using violin plots generated with the ggplot2 package (v. 3.5.1; http://ggplot2.tidyverse.org).

LUAD gene expression data and corresponding drug half-maximal inhibitory concentration (IC₅₀) values were initially obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) database (http://www.cancerrxgene.org/). The oncoPredict package (v. 1.2; Huang Laboratory, University of Minnesota; http://github.com/HuangLabUMN/oncoPredict) was subsequently used for data processing and drug sensitivity prediction. Ridge regression models were trained using the GDSC2 dataset, with drug-related gene expression data specified as the training expression dataset (training ExprData), IC₅₀ values as the training phenotype (training Ptype) and the prognostic gene expression matrix as the test expression dataset (test ExprData). Predicted IC₅₀ values were then calculated for each TCGA-LUAD tumour sample with available survival data. The Wilcoxon rank-sum test (Mann-Whitney U-test) (P<0.05), performed using R software (v. 4.3.1; R Foundation for Statistical Computing; http://www.r-project.org/), was used to compare the predicted IC₅₀ values of different antineoplastic agents between the HRG and LRG. The GDSC2 training dataset contained gene expression profiles and corresponding IC₅₀ values for >200 drugs across >800 cancer cell lines. Ridge regression models were constructed using the default parameters of the oncoPredict package.

The Wilcoxon rank-sum test (Mann-Whitney U-test) (P<0.05), performed using R software (v. 4.3.1; R Foundation for Statistical Computing; http://www.r-project.org/), was used to compare the expression levels of prognostic genes between LUAD and adjacent normal samples in the TCGA-LUAD cohort. For independent validation, 20 LUAD tumour tissues and 20 paired adjacent normal tissue samples were collected from The Second Hospital of Lanzhou University (Lanzhou, China). The tissue samples were collected between July 2025 and December 2025. The median age of the patients was 65 years (range, 54–72 years) and the patient population included 4 men and 16 women (Table SII). None of these samples overlapped with those included in the TCGA-LUAD dataset. All patients had pathologically confirmed LUAD and had not received any chemotherapy or radiotherapy before sample collection. Written informed consent was obtained from all participants before enrolment. The study protocol was approved by the Medical Ethics Committee of The Second Hospital of Lanzhou University (Lanzhou, China; approval no. 2025A-648), and all procedures were conducted in accordance with applicable guidelines and regulations. GraphPad Prism 10 (Dotmatics) was used for data visualisation and P-value calculation. Total RNA was extracted from the 40 tissue specimens (20 tumour and 20 paired adjacent normal tissues) using TRIzol reagent (Vazyme Biotech Co., Ltd.) according to the manufacturer's instructions. Complementary DNA (cDNA) was synthesised from total RNA using the Hifair^® III 1st Strand cDNA Synthesis SuperMix for qPCR kit (Shanghai Yeasen Biotechnology, Co., Ltd) according to the manufacturer's instructions. qPCR was subsequently performed using 2X Universal Blue SYBR Green qPCR Master Mix (Wuhan Servicebio Technology Co., Ltd.) according to the manufacturer's instructions. Each 10-µl reaction contained 5 µl of 2X Universal Blue SYBR Green qPCR Master Mix, 3 µl of diluted cDNA and 1 µl each of forward and reverse primers (10 µM). qPCR was performed on an Archimed R4-2 Real-Time PCR System (RocGene Technology Co., Ltd.) with the following thermocycling conditions: Initial denaturation at 95°C for 3 min, followed by 40 cycles of 95°C for 15 sec and 60°C for 30 sec. A melting curve analysis was performed after the amplification cycles to verify the specificity of the PCR products. GAPDH was used as the endogenous reference gene for quantification of prognostic gene expression. Primers were obtained from a commercial supplier and are listed in Table SIII. Relative gene expression differences between LUAD tumour and adjacent normal tissues were calculated using the 2^−∆∆Cq method (14).

All bioinformatics analyses were performed using R software (v. 4.3.1; R Foundation for Statistical Computing; http://www.r-project.org/). Wilcoxon rank-sum test (Mann-Whitney U-test) for independent samples, the Wilcoxon signed-rank test for paired samples or Student's t-test, as appropriate, with P<0.05 considered to indicate a statistically significant difference. To ensure reproducibility, a fixed random seed [set.seed(123456)] was applied to all stochastic procedures, including LASSO cross-validation and CIBERSORT permutations. The same seed was also used for fold assignment during LASSO regression. The principal R packages used in the present study included DESeq2 (v. 1.42.0), WGCNA (v. 1.7.1), glmnet (v. 4.1–8), survival (v. 3.7–0), survivalROC (v. 1.0.3) and rmda (v. 1.9).

Transcriptomic comparison between LUAD and adjacent normal tissues in the TCGA-LUAD cohort identified 9,580 DEGs, including 6,167 upregulated genes and 3,413 downregulated genes in tumour tissues (Fig. 1A and B; Table SIV). Immune infiltration analysis revealed substantial T-cell infiltration in both LUAD and normal samples, including CD8⁺ T cells and several other T-cell subsets. With the exception of CD8⁺ T cells, the infiltration levels of other T-cell subsets, including gamma delta T cells and naïve CD4⁺ T cells, differed significantly between the two groups (Fig. 1C and D). WGCNA confirmed that no outlier samples were present in the TCGA-LUAD cohort, and the optimal soft-thresholding power (β) was determined to be 5 (R²>0.9) (Fig. 1E and F, Table SV). A total of 12 co-expression modules were identified using a module-merging threshold of MEDissThres=0.4 (Fig. 1G and H). Differentially infiltrated T-cell subsets were used as phenotypic traits. For resting CD4⁺ memory T cells, the MEbrown module showed the strongest positive correlation (cor=0.47, P=5×10⁻³⁰), whereas the MEmagenta module showed the strongest negative correlation (cor=−0.47, P=2×10⁻²⁹) (Fig. 1I). These two modules were therefore considered to contain genes strongly associated with T-cell infiltration. The MEbrown module contained 6,130 genes, whereas the MEmagenta module contained 597 genes (Tables SVI and SVII). Accordingly, the genes in these two modules were defined as TRGs, yielding a total of 6,727 genes.

The intersection of the DEGs, TRGs and 1,793 URGs yielded 97 candidate genes (Fig. 2A and Table SVIII). GO enrichment analysis identified 171 biological process (BP), 20 cellular component (CC) and 42 molecular function (MF) terms (Fig. 2B and Table SIX, Table SX, Table SX, Table SXI). The candidate genes were significantly enriched in MF terms such as ‘ubiquitin-like protein transferase activity’. Significant CC terms included the ‘ubiquitin ligase complex’ and ‘nuclear envelope’, whereas enriched BP terms included ‘proteasome-mediated ubiquitin-dependent protein catabolic processes’, among others. KEGG pathway analysis further revealed significant enrichment in nine pathways, including ‘ubiquitin-mediated proteolysis’ (Fig. 2C and Table SXII). A PPI network was then constructed for the 97 candidate genes. Of these, 31 genes showed no interactions with other genes in the network, whereas the remaining genes formed 131 interactions, including an interaction between PSMD4 and DDI1 (Fig. 2D). No prominent hub gene was identified, as no node exhibited a markedly higher degree than the others.

Using the candidate genes, univariate Cox regression analysis in the TCGA-LUAD cohort was first performed and 14 genes were identified to be significantly associated with patient survival (P<0.05, HR≠1; Fig. 3A). The PH assumption was then evaluated and 11 potential prognostic genes that satisfied the assumption (P>0.05) were retained (Fig. S1). Subsequently, LASSO regression with 10-fold cross-validation was performed. The optimal regularization parameter λ (λ min=0.002767626) was selected as the value that minimized the mean cross-validated error. At this λ min=0.002767626, the regression coefficients of RNF180, MAP3K3, TRIM2, RNF125, RWDD3, RNF144B, KCTD5 and TRIM27 remained non-zero (Fig. 3B and C and Table SXIII). The LASSO-derived coefficients for these eight prognostic genes were as follows: RNF180 (−0.1427), MAP3K3 (0.0953), TRIM2 (0.1286), RNF125 (−0.0734), RWDD3 (0.1121), RNF144B (−0.0918), KCTD5 (0.1572) and TRIM27 (0.1039). The risk score for each patient was calculated using the linear combination formula ∑(β gene × expression gene): RiskScore = −0.1427×RNF180 + 0.0953×MAP3K3 + 0.1286×TRIM2 - 0.0734×RNF125 + 0.1121×RWDD3 - 0.0918×RNF144B + 0.1572×KCTD5 + 0.1039×TRIM27. These eight genes were therefore defined as the prognostic signature. Using the optimal risk-score cut-off of 1.183469, patients were stratified into an HRG (n=185) and LRG (n=330). Patients in the HRG had significantly poorer overall survival (P<0.0001) and a higher mortality rate (Fig. 3D-G). Time-dependent ROC analysis yielded AUC values >0.6 for 1-, 3- and 5-year survival (Fig. 3H). In the GSE31210 cohort (n=226), an optimal cut-off of 4.734456×10⁻⁵ was used to stratify 106 patients into the HRG and 120 into the LRG. The results obtained in GSE31210 were consistent with those from the TCGA-LUAD cohort, supporting the generalisability of the prognostic model to an independent dataset (Fig. 3I-M).

A multivariable prognostic model incorporating the risk score and six clinicopathological parameters was developed. Univariate Cox regression identified two significant predictors of overall survival: The risk score and M stage (M1) (HR≠1, P<0.05). Both variables satisfied the PH assumption (P>0.05) and remained statistically significant in the multivariate Cox regression (HR≠1, P<0.05), confirming their roles as independent prognostic factors (Fig. 4A and B, Fig. S2, Tables SXIV and SXV). A nomogram incorporating these independent prognostic factors was then constructed to estimate the probabilities of death at 1, 3 and 5 years in patients with LUAD. Higher total nomogram scores corresponded to higher predicted probabilities of death (Fig. 4C). The calibration curves showed close agreement between the predicted and observed outcomes at 1, 3 and 5 years, indicating good predictive performance (Fig. 4D). DCA further showed that, within a limited threshold probability range (e.g., 0.3–0.5), the nomogram provided a modest improvement in net clinical benefit compared with the risk score alone. For example, at a threshold probability of 0.3, the net benefit was 0.012 for the nomogram compared with 0.008 for the risk score alone. Although the improvement was incremental rather than substantial, these findings suggest potential clinical utility in selected settings (Fig. 4E). Thus, the nomogram may provide additional information for clinical decision-making, such as identifying patients who may warrant more intensive follow-up or consideration of treatment escalation, compared with the use of the risk score or M stage alone.

GSEA revealed 90 pathways significantly enriched in the HRG and 63 pathways significantly enriched in the LRG. Based on NES rankings, the top five pathways were ‘ribosome’, ‘cell cycle’, ‘oxidative phosphorylation’, ‘Parkinson's disease’ and ‘olfactory transduction’ (Fig. 5A and Table SXVI, Table SXVII, Table SXVIII, Table SXIX). Among the enriched pathways, the B-cell receptor signalling pathway was associated with immune function. To further investigate immune-related mechanisms in LUAD, additional analyses of the tumour immune microenvironment were performed. Compared with the LRG, the HRG exhibited significantly lower stromal and ESTIMATE scores, whereas immune scores were comparable between the two groups. By contrast, tumour purity was significantly higher in the HRG (Fig. 5B). In both risk groups, T lymphocytes, particularly CD8⁺ T cells, naïve CD4⁺ T cells and resting CD4⁺ memory T cells, accounted for a substantial proportion of the infiltrating immune cells (Fig. 5C). Significant differences in 11 immune cell populations were identified between the HRG and LRG (P<0.05), including activated CD4⁺ memory T cells and activated natural killer cells (Fig. 5D). Significant correlations were also observed between activated dendritic cells and M1 macrophages (cor=−0.46, P=2.572×10⁻²⁸) and between resting dendritic cells and monocytes (cor=0.39, P=1.379×10⁻²⁰) (Fig. 5E). TIDE scores were subsequently compared between the two groups. Violin plots showed significantly higher TIDE scores in the HRG than in the LRG (P=0.0037) (Fig. 5F). Collectively, these findings indicate that the HRG is characterized by a distinct immune microenvironment, including altered immune cell infiltration patterns, increased tumour purity and higher TIDE scores, which may be associated with poorer clinical outcomes and a reduced likelihood of response to immunotherapy.

The mutational landscapes of the two risk groups were systematically compared. Missense mutations represented the predominant mutation type in both groups. In the HRG, the most frequently mutated genes were TP53, TTN, CSMD3, MUC16 and RYR2 (Fig. 6A). A similar mutation pattern was observed in the LRG, in which TP53, TTN, MUC16, RYR2 and CSMD3 had the highest mutation frequencies (Fig. 6B). The mutation profiles of the prognostic genes were also examined separately in the HRG and LRG. TRIM2 showed the highest mutation frequency in the HRG, including nonsense and missense mutations, whereas RNF180 showed the highest mutation frequency in the LRG. Missense mutations were the predominant mutation type among the remaining prognostic genes in both groups (Fig. 6C and D). Drug sensitivity analysis identified 125 agents with significantly different IC₅₀ values between the HRG and LRG. To visualise the 10 drugs showing the most significant differences between the two groups, box plots were generated and ranked by P-value (P<0.0001). These drugs included axitinib, AZD1208, BMS-754807 and doramapimod (Fig. 6E, Table SXX). These findings highlight distinct molecular and pharmacological characteristics between the two risk groups and may provide potential clues for the development of precision treatment strategies in LUAD.

In the GeneMANIA network, the peripheral nodes represented genes co-expressed with the prognostic genes, whereas the central nodes represented the prognostic genes themselves. The differently coloured blocks indicated functional categories associated with these genes, such as ‘ubiquitin-like protein ligase activity’ (Fig. 7A). In the TCGA-LUAD cohort, expression analysis showed that KCTD5, RWDD3, TRIM2 and TRIM27 were significantly upregulated in LUAD tissues, whereas MAP3K3, RNF125, RNF144B and RNF180 were significantly downregulated (P<0.05; Fig. 7B). To validate the key prognostic genes, five representative genes (KCTD5, MAP3K3, RNF144B, TRIM2 and TRIM27) were selected for RT-qPCR validation based on their significant differential expression in the TCGA-LUAD cohort and their close association with LUAD progression reported in previous studies. The remaining three genes (RNF180, RNF125 and RWDD3) were not validated by RT-qPCR because of limited clinical sample availability and unsuccessful primer validation, although their differential expression was confirmed in the TCGA-LUAD cohort. RT-qPCR was subsequently performed using 20 paired clinical samples. The clinicopathological characteristics of the 20 patients are summarised in Table SII. Consistent with the bioinformatics findings, KCTD5, TRIM2 and TRIM27 were significantly upregulated in LUAD tissues, whereas MAP3K3 and RNF144B were significantly downregulated (P<0.05; Fig. 7C). Thus, the expression patterns observed in the clinical specimens were consistent with those identified in the TCGA-LUAD cohort. These findings further support the robustness of the prognostic gene signature and suggest that these genes may contribute to LUAD progression, providing a foundation for further investigation of the underlying molecular mechanisms and potential therapeutic targets.