The TCGA database (20) provided bulk tumor transcriptomic data for 497 LUAD patients, along with their clinical information, transcriptomic profiles, CNV (Copy Number Variation) and SNV (Single Nucleotide Variants) data. CNV data was analyzed using GISTIC 2.0 (21), while SNV data was processed with the maftools R package (22). External validation was performed using datasets GSE72094, GSE31210, GSE68465, GSE41271, GSE42127, GSE13213 and GSE8894 obtained from GEO (23). The selection of these GEO datasets was based on the following criteria: 1) the inclusion of comprehensive prognostic data for LUAD patients; 2) a sufficiently large sample size; 3) data derived from well-characterized cohorts with consistent inclusion criteria; 4) the availability of detailed molecular and clinical data relevant to our analysis; 5) rigorous quality control measures applied to the dataset and 6) patients didn’t received chemotherapy of other treatments before. Detailed information for these datasets can be found in the supplementary tables ( Supplementary Table S1 ). For genes with multiple corresponding probes, the average expression value of these probes was used to represent the gene expression level. The 27 Regulatory B cell (Breg) associated genes were obtained from the study by Zhou et al. (19).

The single-cell RNA-sequencing dataset GSE131907 was acquired from the GEO database. We selected 11 primary tumor samples and 10 brain metastasis samples for analyse. Batch effects were mitigated using the Harmony R package (24). Using the Harmony and Seurat (25) packages, we meticulously performed all necessary procedures, including NormalizeData, FindVariableFeatures, ScaleData, RunPCA, FindNeighbors,RunHarmony, FindClusters, and RunUMAP, ensuring the accuracy of each step. This comprehensive analysis resulted in the acquisition of 52,832 single cells. The AUCell package was utilized to assess gene set activity in single-cell RNA-seq data, identifying cells with active gene sets by calculating the area under the curve (AUC) (26).

After employing the robust ConsensusClusterPlus package (27) for clustering analysis, we successfully delineated two distinct subgroups among LUAD patients, characterized by the genetic profiles of 27 Breg-related genes. These subgroups exhibited notable disparities in TME compositions.

Differential expression analysis between patients at different Breg-related clusters was conducted using the Limma package, defining differentially expressed genes (DEGs) based on criteria of adjusted P-value (Padj) < 0.001 and absolute log2(Fold-change) > 0.5. To identify key genes among these DEGs, we established the following criteria:

To establish a reliable and accurate gene signature associated with regulatory B cells for predicting the prognosis of LUAD patients, we developed an advanced artificial intelligence (AI) framework integrating 32 algorithms from traditional regression, machine learning, and deep learning. By systematically exploring 274 algorithmic combinations, we aimed to achieve both predictive accuracy and model stability.

Our methodological approach encompassed traditional regression models, including stepwise Cox regression, LASSO, Ridge, Elastic Net (Enet), and the Cox Proportional Hazards Model (Coxph). We incorporated ensemble learning techniques such as Random Survival Forest (RSF), Conditional Random Forest (CForest), Oblique Random Survival Forest (obliqueRSF), and Boruta, alongside tree-based algorithms like Recursive Partitioning and Regression Trees (Rpart), Conditional Inference Trees (Ctree), and Ranger.

For boosting-based methods, we implemented Gradient Boosting Machine (GBM), LightGBM, Extreme Gradient Boosting (XGBoost), Gradient Boosting with Component-Wise Linear Models (GLMBoost), Gradient Boosting with Regression Trees (BlackBoost), Generalized Additive Model Boosting (GamBoost), and CoxBoost. Additionally, we integrated Partial Least Squares Regression for Cox (plsRcox), Prediction Analysis for Microarrays (Pamr), and Linear Discriminant Analysis (LDA) for feature selection and dimensionality reduction.

Deep learning-based survival models included DeepHit, DeepSurv, Cox-Time Survival Neural Network (CoxTime), Logistic-Hazard Survival Neural Network (Logistic-Hazard), and PC-Hazard Survival Neural Network (PC-Hazard). We further incorporated Survival Support Vector Machine (Survival-SVM) and parametric survival models such as Survreg and Akritas Conditional Non-Parametric Survival Estimator (Akritas). To enhance variable selection and stability, we employed the Variable Selection-Oriented LASSO Bagging Algorithm (VSOLassoBag).

This comprehensive AI-driven approach provides a robust foundation for constructing an effective gene signature for LUAD prognosis, integrating a diverse range of statistical, machine learning, and deep learning methodologies.

To uncover the biological pathways linked to specific genes, we performed comprehensive enrichment analyses using the R package ClusterProfiler (28). This included Gene Ontology (GO) annotation, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis, and Gene Set Enrichment Analysis (GSEA).

The “GSVA” R package (29) was employed to perform single-sample gene set enrichment analysis (ssGSEA), enabling the quantification of enrichment scores for individual gene sets across different samples. In addition, we utilized the ‘IOBR’ package (30) to assess immune cell infiltration levels in LUAD patients by integrating six distinct computational algorithms: quanTIseq, TIMER, EPIC, MCP-counter, ESTIMATE, and xCell. Information on the activation stages of the seven-phase Cancer Immunity Cycle was obtained from the Tracking Tumor Immunophenotype (TIP) database (31). Furthermore, we explored the association between immune-related molecule expression and BREGI.

The TIDE (Tumor Immune Dysfunction and Exclusion) score, a computational predictor of immune checkpoint blockade (ICB) therapy response in LUAD patients, was obtained from the TIDE website (32), with lower scores indicating higher sensitivity to ICB treatment. Additionally, the Immunophenoscore (IPS) for LUAD patients were retrieved from the TCIA database (33), where a higher IPS score reflects greater responsiveness to ICB therapy. Five immunotherapy cohorts—GSE91061, phs000452, PRJEB23709, phs000452, and PRJNA482620—were obtained from the TIGER database (34). Additionally, another immunotherapy cohort for NSCLC patients (PMID: 37024582) was sourced from the study by Ravi et al. (35). We calculated the BREGI across these cohorts and analyzed the prognostic differences between patients stratified into different BREGI groups.

To evaluate the therapeutic response of high-BREGI LUAD patients, we integrated data from the GDSC1 database (36) with analyses conducted using the ‘oncoPredict’ package (37). Drug sensitivity was assessed based on IC50 values, where a lower IC50 indicates greater responsiveness to the treatment.

The human lung adenocarcinoma cell line A549 was procured from the Cell Bank of the Chinese Academy of Sciences. Cells were maintained at 37°C in a humidified atmosphere with 5% CO2, using DMEM medium (Bioscience, China) supplemented with 10% fetal bovine serum (FBS; Gibco, USA). Small hairpin RNAs targeting TBRG4 (shTBRG4) and non-targeting control shRNA were acquired from Hanheng Biology (Shanghai, China). Transfection of HBLV vector-encoded shTBRG4 and control shRNA was conducted over 24 hours using polybrene reagent, following standard operational procedures. The nucleotide sequences were as follows: negative control (NC): TTCTCCGAACGTGTCACGTAA; shTBRG4: TCAAGCAGCAATGGTACTTAT.​Post-transfection, total RNA was extracted from A549 cells using an RNA isolation kit (Vazyme, China) according to the manufacturer’s protocols. Reverse transcription and PCR amplification were performed with kits obtained from TransGen Biotech (China), adhering to the supplier’s instructions. GAPDH served as the internal reference gene, and relative gene expression was calculated using the 2⁻ΔΔCT method. Primer sequences utilized were: TBRG4 forward 5′-TTCAACAGCCGAAGCAAGGA-3′ and reverse 5′-GGGAGTAGATGCTCGTTCCTTC-3′; GAPDH forward 5′-GGTGTGAACCATGAGAAGTATGA-3′ and reverse 5′-GAGTCCTTCCACGATACCAAAG-3′.

Cells were plated in 96-well plates at a density of 3,000 cells per well, with 200 μl DMEM medium per well. Culture medium was replenished daily. At each time point, 10 μl CCK-8 solution was added to each well, followed by a 2-hour incubation. Absorbance at 450 nm was subsequently determined using a microplate reader.

Transfected A549 cells were seeded in 6-well plates at 1×10⁵ cells per well. After 24 hours of culture, when cell confluence reached approximately 100%, linear wounds were created in the monolayer using a 10-μl pipette tip. Detached cells were gently removed by PBS washing, and reference marks were made on the plate bottom. Wound areas were imaged at 0 and 24 hours, with quantitative analysis of wound closure performed using ImageJ.

Migration and invasion assays were executed using Transwell chambers (Corning Inc., USA). For invasion assays, chamber inserts were pre-treated with 10 μg Matrigel. A total of 10,000 cells suspended in 200 μl FBS-free DMEM were added to the upper chamber, while 600 μl DMEM containing 10% FBS was placed in the lower chamber. Following 24 hours of incubation at 37°C, cells adhering to the membrane were fixed in paraformaldehyde and stained with hematoxylin. Migrated or invaded cells in the lower chamber were visualized and photographed under a high-power microscope.

Statistical analyses were conducted using R (version 4.1.3). The appropriate test, either the Wilcoxon test or t-test, was used to compare two groups, based on data distribution. Spearman correlation was employed for correlation analyses. Kaplan-Meier survival analysis was performed to assess overall survival differences between groups, and the log-rank test was used to determine statistical significance. To evaluate the prognostic value of BREGI alongside clinicopathological factors, both univariate and multivariate Cox regression analyses were carried out. The CompareC package was used to compare the C-index of BREGI with various models and clinical factors. Statistical significance was defined as *p < 0.05, **p < 0.01, and ***p < 0.001, with “Ns” indicating a p-value ≥ 0.05.

First, we examined the chromosomal locations of the 27 Breg-related genes, as illustrated in Figure 1A . Next, we investigated the differential expression of these genes between normal lung tissues and LUAD tissues, revealing that the majority exhibited significant expression differences ( Figure 1B ). We then analyzed the SNV and CNV alterations of these 27 genes in LUAD. Somatic mutations in these genes were detected in 127 LUAD patients (22.8%), with missense mutations being the most prevalent. Among them, FLNA exhibited the highest mutation frequency ( Figure 1C ). Additionally, CNV analysis ( Figure 1D ) indicated that MYC had the highest frequency of copy number gain.

To further explore the immune landscape, we applied the ssGSEA algorithm to assess immune cell infiltration levels in LUAD patients. Moreover, the ESTIMATE algorithm was used to calculate StromalScore, ImmuneScore, and ESTIMATEScore. Correlation analysis demonstrated that Breg-related genes were significantly positively associated with most immune cell populations ( Figure 1E ) as well as immune-related scores ( Figure 1F ).

Finally, Gene Ontology (GO) analysis of these 27 genes revealed their involvement in multiple immune-suppressive biological processes, such as “negative regulation of immune system process,” “negative regulation of lymphocyte activation,” and “negative regulation of cytokine production” ( Figure 1G ).

We performed molecular subtyping of LUAD patients based on 27 Breg-related genes. Patients were classified into k groups (k = 2–9), and the clustering heatmap ( Figure 2A ) and PAC plot (Proportion of Ambiguous Clustering, Figure 2B ) indicated that k = 2 provided the optimal clustering solution. The expression patterns of these 27 genes in the two clusters are shown in Figure 2C , revealing that the majority of them were highly expressed in Cluster C1.

Next, we compared the immune characteristics between the two clusters. ssGSEA analysis demonstrated that all immune cell types ( Figure 2D ) and immune-related activities ( Figure 2E ) were significantly higher in C1 than in C2. Furthermore, six distinct algorithms (EPIC, ESTIMATE, TIMER, MCPcounter, QuantiSeq, and xCell) were employed to evaluate immune cell infiltration levels, showing that most immune cell populations exhibited higher infiltration in C1 ( Figure 2F ). These findings suggest that Cluster C1 exhibits a “hot immune” phenotype.

We then examined the relationship between the clusters and clinical characteristics of LUAD patients. Cluster C2 was significantly associated with unfavorable OS events ( Figure 2G ), more advanced T stage ( Figure 2H ), N stage ( Figure 2I ), and clinical stage ( Figure 2J ), as well as shorter overall survival ( Figure 2K ), indicating poorer clinical classification and prognosis. Therefore, patients in Cluster C2 exhibited a “cold immune” phenotype, which may contribute to disease progression.

To further investigate the immunological differences between the two Breg-related clusters, we first analyzed the expression of immune-related molecules. The heatmap ( Figure 3A ) demonstrated that all immune-related molecules were upregulated in Cluster C1 compared to Cluster C2. Specifically, PD-1 and PD-L1, two critical immune checkpoint molecules, were significantly higher in C1 ( Figure 3B ).

Next, we explored T-cell receptor (TCR) diversity between the clusters. TCR richness ( Figure 3C ) and TCR Shannon diversity ( Figure 3D ) were both significantly elevated in Cluster C1, indicating a more diverse and active TCR repertoire. Additionally, we assessed tumor mutational burden (TMB) between the clusters; however, no significant difference was observed ( Figure 3E ).

Furthermore, we evaluated the immunogenic potential of the clusters using the Immunophenoscore (IPS), a composite metric predictive of response to immune checkpoint inhibitors. IPS scores were significantly higher in Cluster C1 ( Figure 3F ), suggesting greater immune responsiveness in this subgroup. Finally, our investigation into cancer progression demonstrated that most essential steps, including cancer antigen presentation, priming and activation, exhibited increased activity in the C1 group ( Figure 3G ), reinforcing the notion that Cluster C1 possesses a more active immune microenvironment.

These findings further support the classification of Cluster C1 as a “hot immune” phenotype and Cluster C2 as a “cold immune” phenotype, highlighting their potential implications for immunotherapy responsiveness and disease progression.

To further elucidate the differences between the two Breg-related clusters, we performed a differential expression analysis. We defined the differentially expressed genes (DEGs) as those with an adjusted P-value (Padj) < 0.001 and an absolute |log2(Fold-change)| > 0.5 ( Supplementary Table S2 ). The heatmap presents the top 50 DEGs, with the top 25 genes significantly upregulated in Cluster C1 and the top 25 genes significantly upregulated in Cluster C2 ( Figure 4A ). Next, GO analysis of these DEGs revealed their involvement in immune-related biological functions, including “immune receptor activity,” “T cell receptor binding,” and “negative regulation of immune system process” ( Figure 4B ). KEGG analysis further demonstrated that these genes were enriched in immune-related signaling pathways ( Figure 4C ). GSEA showed that genes highly expressed in Cluster C1 were associated with immune-related functions such as “immune effector process,” “immune response,” and “T cell activation” ( Figure 4D ), whereas genes highly expressed in Cluster C2 were enriched in biological processes such as “muscle tissue development,” “microtubule-based process,” and “heart morphogenesis” ( Figure 4E ).

To gain deeper insight into the cellular composition of the two clusters, we performed single-cell RNA sequencing (scRNA-seq) analysis using the GSE131907 dataset. The UMAP plot illustrates the clustering of single-cell transcriptomic data ( Figure 4F ), while Figure 4G shows the marker genes used to define each cell type. We then applied the AUCell algorithm to compute gene set activity scores based on the top 50 upregulated genes in Cluster C1 and Cluster C2, generating two scores: C1AUC and C2AUC. The violin plots depict the distribution of C1AUC ( Figure 4H ) and C2AUC ( Figure 4I ) across different cell types. Subsequently, we stratified all cells into four groups based on the median values of C1AUC and C2AUC: high-C1AUC (HC1), low-C1AUC (LC1), high-C2AUC (HC2), and low-C2AUC (LC2). We hypothesized that cells classified as high-C1AUC & low-C2AUC (HC1LC2) best represented the characteristics of Cluster C1, while those classified as low-C1AUC & high-C2AUC (LC1HC2) were most representative of Cluster C2. The UMAP distribution of HC1LC2 and LC1HC2 cells is shown in Figure 4J .

We then compared the cellular composition between HC1LC2 and LC1HC2 groups ( Figure 4K ). The results revealed that immune cells known to play critical roles in anti-tumor immunity, including CD4+ T cells, CD8+ T cells, B cells, myeloid cells, and NK cells, were more abundant in the HC1LC2 group. In contrast, mast cells, plasma cells, and endothelial cells were more prevalent in the LC1HC2 group. Notably, all epithelial (malignant) cells were exclusively enriched in the LC1HC2 group.

These findings further reinforce the notion that Cluster C1 exhibits a “hot immune” phenotype characterized by active anti-tumor immunity, whereas Cluster C2 displays a “cold immune” phenotype potentially facilitating tumor progression.

To select key genes from the DEGs for constructing the Breg-related gene signature, we applied the following criteria:

This process led to the identification of 32 genes for model construction. We then tested 274 algorithmic combinations on the TCGA dataset and calculated the C-index for each across the validation cohorts. The combination of RSF and ridge integration produced the highest average C-index of 0.668 across the validation cohorts. Based on this, we selected this combination to construct the Regulatory B Cell-related Index (BREGI, Figure 5A ). The BREGI is composed of 18 Breg-related genes. The Ridge algorithm assigned a coefficient to each of these genes ( Figure 5B , Supplementary Table S2 ). By multiplying the expression level of each gene by its respective coefficient and summing the results, the BREGI score is calculated. Furthermore, we analyzed the correlation between BREGI, the 18 genes composing BREGI, and clinical factors of LUAD patients ( Supplementary Figure S1A ). Higher BREGI, as well as elevated expression of LAMC2, PCDH7, STEAP1, and TBRG4, were associated with more aggressive clinical features. The differential expression of these 18 Breg-related genes between normal lung tissue and LUAD tissue is shown in Supplementary Figure S1B . Next, we examined the relationship between these 18 genes and immune cells and found significant correlations with various immune cell populations ( Supplementary Figure S1C ). GO analysis revealed that these 18 genes were involved in multiple immune-related functions, such as “negative regulation of immune effector process,” “negative regulation of complement activation,” and “negative regulation of leukocyte-mediated immunity” ( Supplementary Figure S1D , indicating that these genes are significantly associated with Breg cells.

Then we stratified LUAD patients into high- and low-BREGI groups based on the optimal BREGI threshold. The high-BREGI group showed significantly worse prognosis compared to the low-BREGI group across all cohorts ( Figure 5C ). Furthermore, time-dependent ROC analysis revealed that BREGI demonstrated high AUC values when predicting patient prognosis over 1–5 years, confirming its predictive accuracy and reliability ( Figure 5D ). To test whether BREGI can serve as an independent prognostic factor for LUAD patients, we conducted both univariate ( Supplementary Figure S2A ) and multivariate ( Supplementary Figure S2B ) Cox regression analyses across all cohorts. Except for the multivariate Cox regression analysis in the GSE31210 cohort, where the p-value for BREGI was 0.069, BREGI demonstrated significant prognostic value in all other cohorts. These analyses indicate the robust predictive performance of BREGI and its potential as an independent prognostic factor for LUAD patients.

To further demonstrate the superiority of BREGI, we compared it with clinical features. Specifically, we assessed the AUC values ( Figure 6a ) and concordance indices (C-index, Figure 6b ) of BREGI and clinical factors across eight cohorts, revealing that BREGI outperformed most clinical variables. Furthermore, we compiled 105 previously published gene signatures associated with various phenotypes and compared their C-index with that of BREGI across the eight cohorts ( Figure 6c ). The results indicated that BREGI ranked relatively lower in the TCGA cohort, likely due to overfitting, as most studies used TCGA as a training set. However, in the validation cohorts, except for GSE31210, BREGI demonstrated superior C-index performance compared to most models. These findings further highlight the robustness and superiority of BREGI.

Given the prognostic significance of age, gender, and clinical stage in LUAD patients, we proposed that incorporating these variables with BREGI could enhance predictive accuracy. To test this hypothesis, we constructed a nomogram integrating BREGI with age, gender, and clinical stage ( Supplementary Figure S3A ). In the TCGA cohort, this model achieved a C-Index of 0.715, demonstrating strong predictive performance. Further validation through calibration curves ( Supplementary Figure S3B ) and decision curve analysis (DCA) ( Supplementary Figure S3C ) affirmed its reliability in prognostic estimation. Notably, ROC analyses ( Supplementary Figures S3D-F ) indicated that the nomogram outperformed individual predictors, including BREGI, age, gender, and clinical stage, in forecasting patient outcomes. These findings highlight the potential of this integrated nomogram as a valuable tool for clinical decision-making and risk stratification in LUAD.

To further explore the relationship between BREGI and the TME, we first examined its association with Breg-related clusters. Our analysis revealed that patients in Cluster C1 exhibited a “hot immune” phenotype. Notably, the proportion of Cluster C1 patients was significantly higher in the low-BREGI group than in the high-BREGI group, suggesting a stronger association between low BREGI and the hot immune phenotype ( Figure 7A ). ssGSEA analysis further demonstrated that the majority of immune cell infiltration levels ( Figure 7F ) and immune-related activities ( Figure 7G ) were elevated in the low-BREGI group. Using six algorithms, we consistently observed higher immune cell infiltration in the low-BREGI group ( Figure 7H ). Correlation analysis showed that BREGI was negatively associated with most immune cells ( Supplementary Figure S4C ) but positively correlated with other TME components, such as keratinocytes and epithelial cells.

To validate these findings at the single-cell level, we calculated BREGI in the GSE131907 scRNA-seq dataset ( Figure 7I ) and stratified all cells into high- and low-BREGI groups based on the median BREGI value ( Figure 7J ). The low-BREGI group exhibited a higher proportion of CD8+ T cells, CD4+ T cells, B cells, NK cells, and myeloid cells, whereas epithelial and endothelial cell proportions were lower ( Figure 7K ), resembling the HC1LC2 group ( Figure 4K ). Based on these findings, we hypothesize that similar to Cluster C1, the low-BREGI group exhibits a “hot immune” phenotype, while the high-BREGI group is characterized by a “cold immune” profile.

To assess the relevance of BREGI to immunotherapy, we analyzed its association with TCR Shannon diversity ( Figure 7B ), TCR richness ( Figure 7C ), TIDE score ( Figure 7D ), and IPS score ( Figure 7E ). The low-BREGI group exhibited significantly higher TCR Shannon diversity, TCR richness, and IPS scores, indicating greater sensitivity to immunotherapy. Conversely, the high-BREGI group had higher TIDE scores, which predict resistance to immunotherapy, further supporting the link between low BREGI expression and improved immunotherapy response.

Additionally, we compared the expression of various immune-related molecules between the BREGI subgroups ( Supplementary Figure S4A ) and found that most were upregulated in the low-BREGI group, with BREGI expression negatively correlated with their levels ( Supplementary Figure S4B ). Finally, we calculated BREGI expression in six real-world immunotherapy cohorts and assessed patient outcomes following immune checkpoint blockade. The results demonstrated significantly worse prognosis in high-BREGI patients, suggesting that low-BREGI patients derive greater clinical benefit from immunotherapy ( Figure 7L ).

Furthermore, we performed a comprehensive genomic comparison between the BREGI-high and BREGI-low groups within the TCGA dataset. Initially, we focused on the top 20 genes with the highest mutation frequencies. Our analysis revealed that patients in the BREGI-high group exhibited a significantly higher overall mutation frequency (96.3%, Supplementary Figure S5A ) compared to those in the BREGI-low group (90.53%, Supplementary Figure S5B ). Further analysis demonstrated that the top 30 genes with the most pronounced differences in mutation frequency were consistently more frequently mutated in the BREGI-high group ( Supplementary Figure S5C ), with significant co-mutation relationships observed among these genes ( Supplementary Figure S5D ).

Moreover, BREGI expression was found to be significantly positively correlated with multiple genomic features, including silent mutations, non-silent mutations, aneuploidy score, fraction of altered genome (FGA), fraction of genome gained (FGG), and fraction of genome lost (FGL) ( Supplementary Figure S5E ). Additionally, our findings revealed substantial differences in CNV events between the two groups. Patients in the BREGI-high group exhibited heightened genomic instability, characterized by a greater frequency and complexity of CNV events ( Supplementary Figure S5F ), whereas those in the BREGI-low group displayed lower levels of genomic instability ( Supplementary Figure S5G ). ChromPlots further demonstrated that patients in the BREGI-high group had significantly higher G-scores ( Supplementary Figure S5H ) compared to those in the BREGI-low group ( Supplementary Figure S5I ).

These findings suggest that LUAD patients with high BREGI expression are more likely to exhibit aggressive and malignant genomic characteristics.

Given the significantly worse prognosis, increased malignancy, and immunotherapy resistance observed in high-BREGI patients, there is an urgent need to identify potential therapeutic options targeting this subgroup. To address this, we compared the IC50 values of various drugs between the high- and low-BREGI groups, where lower IC50 values indicate higher sensitivity. The results revealed that multiple chemotherapeutic agents, including paclitaxel, docetaxel, and vinorelbine, as well as targeted therapies such as gefitinib and cetuximab, exhibited lower IC50 values in the high-BREGI group ( Supplementary Figure S6A ). Furthermore, correlation analysis demonstrated a significant negative association between BREGI expression and the IC50 values of these drugs ( Supplementary Figure S6B ), suggesting that they may serve as potential treatment options for high-BREGI patients.

Among the 18 genes incorporated in the construction of BREGI, the role of TBRG4 in lung adenocarcinoma remained relatively less unexplored., prompting us to further investigate its functional role. Bioinformatics analyses first identified TBRG4 as a significant risk factor, with elevated TBRG4 expression correlating with poorer prognosis in LUAD patients ( Figure 8A ). To validate its biological relevance, we subsequently performed in vitro experiments to dissect the potential phenotypic roles of TBRG4 in LUAD progression. Efficient knockdown of TBRG4 in human LUAD A549 cells was achieved using shRNA, as confirmed by significant reduction in TBRG4 mRNA expression ( Figure 8B ). The CCK-8 assays revealed that TBRG4 depletion exerted a profound inhibitory effect on A549 cell proliferation ( Figure 8C ). Consistently, wound healing assays demonstrated impaired migratory capacity in TBRG4-knockdown A549 cells ( Figure 8D ). Furthermore, transwell assays validated that TBRG4 silencing suppressed both migratory and invasive properties of A549 cells ( Figure 8E ).

Collectively, these findings establish that TBRG4 knockdown robustly attenuates the malignant phenotypes of LUAD cells, including proliferation, migration, and invasion. This functional evidence, coupled with its identification as a risk factor via bioinformatics, underscores TBRG4’s pathogenic relevance in LUAD progression.