Zinc finger protein-associated gene signature serves as a potential predictor for prognosis and therapeutic response in lung adenocarcinoma Authors: Mi Zou, Guangda Zheng, Yanju Bao Published: 2026 ABSTRACT Zinc finger proteins (ZNFs), characterized by zinc ion-binding domains, participate in cell proliferation, differentiation, and metastasis in lung adenocarcinoma (LUAD). However, associations between ZNFs-related genes and clinical outcomes, immune cell infiltration, and immunotherapy remain unclear. To explore feasibility of using ZNFs-related genes as prognostic tools for LUAD risk stratification. Retrospective analyses were conducted utilizing data from TCGA and GSE26939. After screening differentially expressed ZNFs, regression analyses were performed to construct prognostic signature. Enrichment analysis identified biological processes and pathways involved in signature genes, while immune landscape was examined by multiple algorithms. The drug sensitivity analysis identified potential candidate drugs related to the signature genes. Cell experiments indicated the function of the key risk gene CTCFL in promoting the malignant behavior of LUAD cells. A prognostic signature comprising 12 ZNFs-related genes (CBFA2T3, CTCFL, GFI1B, IGF2BP1, RIMS2, TRIM29, TRIML2, ZIC2, ZNF208, KLF10, ZNF750, and ZNF257) stratified LUAD patients into two risk groups, demonstrating robust performance in predicting clinical outcomes. These genes were significantly enriched in epidermal development, intermediate filament cytoskeleton, endopeptidase inhibitor activity, hormone activity, and neuroactive ligand-receptor interactions. Low-risk patients exhibited higher levels of immune cell infiltration (e.g., DCs, B cells, and neutrophils) and superior responses to immunotherapy (anti-CTLA-4 and PD-1/CTLA-4 dual blockade). Possible therapeutic compounds for LUAD patients included SHP-099, Dimethylfasudil, EMD-534085, and PF-2771. The expression of CTCFL enhanced the malignant cellular behavior in LUAD. ZNFs-related gene signature provides predictive insights into LUAD patient survival, immune cell infiltration, and immune checkpoint blockade therapy, serving as a valuable tool to guide clinical decision-making.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s00726-026-03528-y. Introduction Lung adenocarcinoma (LUAD), originating from glandular tissues in lungs, is characterized by tumor cells growing along alveolar walls, forming papillary or glandular structures (Willner et al. 2024). A global study on histological subtypes of lung cancer reported approximately 1 million new LUAD cases worldwide in 2020, accounting for nearly half of all new lung cancer cases (Zhang et al. 2023b). Typically diagnosed at advanced stages, LUAD exhibits a 5-year survival rate below 20% even after conventional treatments, molecular targeted therapies, and immunotherapy (Li et al. 2024; Xi et al. 2024). Many patients develop treatment resistance, leading to tumor recurrence (Hurtado et al. 2024). Thus, more accurate and reliable prognostic biomarkers are needed to guide personalized treatment strategies and improve patient survival. Transcriptomic alterations are common features of human cancers (Rhie et al. 2018). In many cases, transcriptional dysregulation results from changes in transcription factor expression or activity (Rhie et al. 2016; Yao et al. 2015). As the largest family of transcriptional regulators in human genome, zinc finger proteins (ZNFs) play roles in telomere and DNA maintenance/repair, chromatin remodeling, apoptosis, and autophagy (Vilas et al. 2018; Zhang et al. 2021b). ZNFs are critically involved in tumorigenesis, progression, and metastasis (Jen and Wang 2016). Li et al. (2023a) reported that ZNF687 overexpression correlates with unfavorable LUAD prognosis, exerting oncogenic effects by activating PI3K/AKT to enhance epithelial-mesenchymal transition and modulate cell cycle progression. ZNF768 induces LUAD cell proliferation by upregulating cell cycle-related genes (e.g., CDK1, PLK1, and PCNA) (Poirier et al. 2021). ZNF280A acts as a LUAD driver by fostering progression through EIF3C expression modulation (Liu et al. 2021). ZNF143 is a likely therapeutic target across multiple cancers, including LUAD (Mei et al. 2023). ZNFs may function as oncogenes in LUAD development and progression. However, existing studies primarily focus on individual ZNFs, leaving collective impact of ZNFs-related gene signatures on LUAD patient survival poorly understood. Therefore, this study established a predictive model using ZNFs-related genes influencing LUAD prognosis. The model stratified LUAD patients into two risk groups to compare survival duration, immune infiltration status, and responses to immune checkpoint blockade therapy. Mutation profiles of signature genes and their drug sensitivity implications were evaluated. Materials and methods Data collection A dataset comprising 59 LUAD and 541 adjacent normal samples was downloaded from TCGA (https://portal.gdc.cancer.gov/) and served as a training set. The GSE26939 and GSE72094 datasets, downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/), were employed as validation cohorts. GSE26939 comprises 112 LUAD samples, while GSE72094 includes 398 LUAD samples. Additionally, a single-cell RNA sequencing (scRNA-seq) dataset for LUAD was obtained from previously published work (Bischoff et al. 2021), consisting of 10 normal tissue samples and 10 LUAD tissue samples. ZNF-related genes were retrieved from UniProt (https://www.uniprot.org/), yielding 4,188 genes (Table S1). Analysis of differentially expressed ZNF-related genes (DEZNFs) and PPI network construction Differential analysis between LUAD (n = 59) and normal (n = 541) samples from TCGA was completed utilizing "edgeR" package in R. Genes with FDR < 0.05 and |log₂FC|> 1 were considered significantly expressed. Intersection of differentially expressed genes (DEGs) and ZNF-related genes yielded DEZNFs. STRING (https://cn.string-db.org/) was employed to construct, visualize, and analyze PPI network of DEZNFs. Development and validation of prognostic model DEZNFs (confidence score > 0.4) were subjected to univariate Cox regression and LASSO regression analyses utilizing "survival" and "glmnet" packages, respectively, to screen for prognosis-associated DEZNFs. Multivariate Cox regression was done to identify signature genes and their corresponding coefficients. Riskscore was calculated by formula:Riskscore=∑ Coefficient(gene)*Expressionvalue(gene) Patients were stratified into high- and low-risk groups with median riskscore. Kaplan–Meier curves were plotted utilizing "survival" package to compare survival differences between two groups. Time-dependent ROC curves generated by "timeROC" package assessed predictive accuracy, which was validated utilizing GSE26939. Riskscore distribution, survival status plots, and a heatmap of signature gene expression were generated. Construction and validation of the nomogram Univariate Cox regression was used to estimate HRs and 95% CIs for clinical features (age, sex, T, N, M, and tumor stage) and riskscores, followed by multivariate Cox regression to identify independent prognostic factors. A nomogram was constructed using the “rms” R package by integrating risk scores with clinical characteristics, enabling the visual prediction of 1-, 3-, and 5-year overall survival probabilities for patients. The predictive performance of the nomogram was subsequently evaluated using ROC curve analysis. Functional enrichment analysis of DEGs in high- and low-risk groups To elucidate functional pathways associated with DEGs (Risk-DEGs) between risk groups, KEGG and GO analyses were done utilizing "clusterProfiler" R package. Predictive analysis of immune microenvironment and immune therapy response sensitivity The "estimate" package was utilized to compute immune, stromal, and ESTIMATE scores for each sample in both risk groups. The ESTIMATE algorithm infers the proportion of stromal and immune cells in tumor samples using gene expression signatures, generating three scores: Immune Score (reflecting immune cell infiltration), Stromal Score (reflecting stromal cell content), and ESTIMATE Score (the sum of the two, reflecting the overall non-tumor cell content in the tumor microenvironment). ssGSEA was applied to assess 29 immune signatures. Tumor Immune Dysfunction and Exclusion (TIDE, http://tide.dfci.harvard.edu/) scores and Immunophenoscore (IPS, based on TCIA database: https://tcia.at/) were calculated to evaluate sensitivity of two risk groups to immune checkpoint blockade therapy. The immunophenotype score (IPS) was derived following the standard pipeline of The Cancer Immunome Atlas (TCIA), which integrates expression data from four categories of immune-related genes: major histocompatibility complex (MHC) molecules, immunomodulators, effector immune cells, and immunosuppressive cells. In general, a higher IPS reflects a tumor phenotype with enhanced immunogenicity and greater immune cell infiltration, suggesting a more favorable likelihood of response to immune checkpoint blockade therapy. Meanwhile, the TIDE score is primarily applied to infer potential mechanisms underlying tumor immune evasion, with higher scores being positively correlated with an increased risk of resistance to immunotherapy. Mutation analysis of signature genes R package "maftools" was employed to estimate and visualize mutation profiles of signature genes in two risk populations. Drug response prediction Utilizing data from CellMiner (https://discover.nci.nih.gov/cellminer/home.do), we computed correlations between signature gene expression and drug sensitivity. Cell culture Normal lung epithelial cells BEAS-2B and human LUAD cell lines A549 (CTCC-001–0036) and HCC78 (CTCC-001-0872) were purchased from MEISEN Cell (China). A549 cells were cultured in F-12 K medium (CTCC-002-007), while HCC78 cells were cultured in RPMI-1640 medium. Both media were supplemented with 10% FBS and 1% P/S. All cells were maintained in a humidified incubator at 37 ℃ with 5% CO₂. Cell transfection The vector sh-CTCFL and the corresponding negative control (sh-NC) were synthesized by GenePharma (China). Subsequently, they were respectively transfected into A549 cells using Lipofectamine™ 3000 transfection reagent (Thermo Fisher, USA). After 48-h transfection, the cells were collected for subsequent analysis. qRT-PCR Total RNA was extracted using TRIzol reagent (Invitrogen, USA), followed by concentration and purity measurement via a NanoDrop One micro-ultraviolet–visible spectrophotometer. Reverse transcription was performed using SuperScript II reverse transcriptase (Invitrogen, USA) to generate cDNA. qRT-PCR was conducted using TB Green® Premix Ex Taq™ II (Takara, Japan) and specific primers (Table 1) on an ABI 7500 PCR system (Applied Biosystems, USA). The relative mRNA expression levels were calculated by the 2−ΔΔCT method, with GAPDH as the internal reference gene.Table 1Primer sequences for qRT-PCRPrimer sequence (5’-3’)CTCFLGCCCAGGATGTTCACTGACTACTAGGGCTCCGATGGTCTTGAPDHAATGGGCAGCCGTTAGGAAAGCGCCCAATACGACCAAATC Table 1Primer sequences for qRT-PCRPrimer sequence (5’-3’)CTCFLGCCCAGGATGTTCACTGACTACTAGGGCTCCGATGGTCTTGAPDHAATGGGCAGCCGTTAGGAAAGCGCCCAATACGACCAAATC Primer sequences for qRT-PCR Western blot (WB) The cells were lysed in the RIPA lysis buffer (Beyotime Biotechnology), followed by adding 1% protease inhibitor PMSF and phosphatase inhibitor. The lysates were then incubated on ice for 30 min and centrifuged at 12,000 rpm for 15 min at 4 ℃. The supernatant was collected, and the protein concentration was determined using the Pierce™ BCA protein assay kit (Thermo Fisher Scientific). Equal protein aliquots (20–30 µg) were resolved by SDS-PAGE and transferred to a PVDF membrane (Millipore). Following a blocking step in 5% skim milk for 1 h, the membrane was subjected to overnight incubation at 4 ℃ with specific primary antibody, rabbit anti-human CTCFL antibody (39,851, Thermo Fisher Scientific), with Mouse anti-human GAPDH antibody (ab9485, Abcam) as the internal control. The following day, the membrane was washed with TBST at room temperature for 3 times, each for 10 min, and then cultured with the corresponding HRP-labeled goat anti-rabbit (ab6721, Abcam) secondary antibody at room temperature for 1 h. After thorough washing with TBST, the membrane was exposed on a chemiluminescence imaging system (Tanon) using the ECL substrate (Millipore). Ultimately, quantification was performed using ImageJ software (National Institutes of Health, USA). CCK8 The transfected cells were inoculated in 96-well plates (1 × 104 cells/well) and cultured in an incubator (37 ℃, 5% CO₂). The cells were washed twice with PBS, with the addition of 100 µL of cell culture medium. After replacing the culture medium, 100 µL of Cell Counting Kit-8 reagent (YEASEN, China) was introduced to each well at 0 h, 24 h, 48 h, and 72 h, and incubated at 37 ℃ for 2 h. Finally, the optical density (OD) of each well at 450 nm was measured using a microplate reader to assess cell proliferation. Colony formation assay Transfected cells from each group were digested with trypsin and resuspended to form single-cell suspensions. Following counting, cells were inoculated into 6-well plates (1 × 103 cells/well). The plates were cultured in an incubator (37 ℃, 5% CO₂) for 7–14 days to observe cell growth. When cell clones were visible, the culture medium was discarded to terminate the culture. Subsequently, the cells were stained with 0.1% crystal violet at room temperature for 20–30 min and washed thoroughly. After drying, samples were photographed and counted to determine cell growth and clone ability. Wound healing Transfected cells were seeded in 6-well plates (1 × 105 cells/well) and cultured at 37 ℃ with 5% CO₂ conditions for 24 h to reach 90% confluence. Subsequently, a scratch was made on the plate surface using a 200 µL pipette tip, followed by marking on the back of the plate to fix the microscope observation field. Following the creation of the scratch, the plate was washed three times with PBS to remove residual cells. After the creation of the scratch, the plate was washed three times with PBS to remove residual cells. Subsequently, serum-free medium was added to inhibit cell proliferation, thereby isolating the effects of cell migration. Images of the same field of view were taken under a microscope at 0 h and 24 h after scratch formation to evaluate cell migration ability. Transwell Chambers with a membrane pore size of 8.0 μm and containing Matrigel (356,234, BD Biosciences, USA) were applied for Transwell invasion assay. The upper chamber was filled with 50 μL of pre-cooled medium diluted 40 times with Matrigel and incubated at 37 ℃ for 30 min to obtain a matrix layer. The digested transfected cells were centrifuged (1000 rpm, 5 min), washed once with PBS, and resuspended in serum-free medium to adjust the cell density to 1 × 105/mL. 100 μL of the cell suspension was added to the upper chamber, while 600 μL of medium containing 10% FBS was added to the lower chamber. After incubating for 24 h at 37 ℃ with 5% CO₂, the chambers were subjected to three PBS washes, fixation with 4% PFA for 15 min, and further three PBS washes. Subsequently, the cells were stained with 0.1% crystal violet for 15 min. After washing three times with PBS, the non-migrated cells in the upper chamber were removed. Finally, images were captured under a microscope, and the migrated cells were counted to evaluate invasive capacity. Statistical analysis Each experimental group was set up with three replicates. The results were presented as the mean ± SD. Data analysis was performed using GraphPad Prism 10.1.2 software. Comparisons between the two groups were conducted using Student’s t-test. P < 0.05 denoted statistically significant. Results Identification and characterization of DEZNFs Differential analysis between LUAD and normal samples identified 5,576 DEGs (Table S2). Intersection of ZNF-related genes and DEGs yielded 288 DEZNFs (Fig. 1). Screening via a PPI network revealed 186 DEZNFs with strong interactions (Table S3). Fig. 1Venn diagram of DEGs and ZNF-related genes Venn diagram of DEGs and ZNF-related genes Construction and validation of the prognostic risk model To establish prognostic model, we did regression analyses on 186 DEZNFs, ultimately identifying 12 signature genes (Table S4, Fig. 2A–C). Riskscore for each patient was computed based on levels of these genes and their corresponding coefficients using formula: Fig. 2Prognostic model development. A, B LASSO regression analysis. C Multivariate Cox regression showing the association of 12 ZNF-related genes with patient prognosis \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \begin{aligned} Risk{\mathrm{score}} & = - 0.1730*CBFA2T3 + 0.0524*CTCFL - 0.0886*GFI1B + 0.0329 \\ & *IGF2BP1 + 0.0350*RIMS2 + 0.0611*TRIM29 + 0.0412*TRIML2 + 0.0466 \\ & *ZIC2 - 0.1083*ZNF208 + 0.3530*KLF10 - 0.0883*ZNF750 - 0.0932*ZNF257 \\ \end{aligned} $$\end{document} Prognostic model development. A, B LASSO regression analysis. C Multivariate Cox regression showing the association of 12 ZNF-related genes with patient prognosis Patients were divided into high-risk and low-risk groups based on the median riskscore. Kaplan‑Meier survival curves revealed that, across both the TCGA training cohort and the two GEO validation cohorts, patients in the high‑risk group exhibited significantly lower survival rates than those in the low‑risk group (Fig. 3A-C, P < 0.05). Model reliability in predicting 1-, 3-, and 5-year survival rates was assessed by AUC values. Training set exhibited an AUC of 0.73 for all time points (Fig. 3D), GSE26939 validation set showed AUCs of 0.74, 0.71, and 0.61, respectively (Fig. 3E), while GSE72094 validation set showed AUCs of 0.66, 0.69, and 0.9, respectively (Fig. 3F), supporting model robustness and effectiveness in predicting LUAD patient survival.Fig. 3Predictive performance of the prognostic model. A, D Survival and ROC curves for high- and low-risk groups in the training set. B, E Survival and ROC curves for high- and low-risk groups in the GSE26939 dataset. C, F Survival and ROC curves for high- and low-risk groups in the GSE72094 dataset. G–I Distribution of survival time (top panel), riskscores (middle panel), and signature gene expression (bottom panel) in the training set (G), GSE26939 dataset (H), and GSE72094 dataset (I) Predictive performance of the prognostic model. A, D Survival and ROC curves for high- and low-risk groups in the training set. B, E Survival and ROC curves for high- and low-risk groups in the GSE26939 dataset. C, F Survival and ROC curves for high- and low-risk groups in the GSE72094 dataset. G–I Distribution of survival time (top panel), riskscores (middle panel), and signature gene expression (bottom panel) in the training set (G), GSE26939 dataset (H), and GSE72094 dataset (I) High-risk group was associated with earlier mortality (Fig. 3G-I). Evaluation of signature gene expression patterns across three independent datasets showed that TRIML2, CTCFL, IGF2BP1, TRIM29, and ZIC2 were consistently highly expressed in the high-risk group, whereas ZNF750, CBFA2T3, GFI1B, and ZNF208 were consistently lowly expressed; the remaining three genes exhibited inter-dataset heterogeneity, possibly related to differences in sample sources and sequencing platforms (Fig. 3G-I). Construction and evaluation of the nomogram Univariate and multivariate Cox regression analyses were completed on age, sex, T, N, M, stage, and riskscore to identify independent prognostic factors in LUAD patients. Riskscore emerged as an independent prognostic factor (Fig. 4A–B). Subsequently, a nomogram integrating clinical characteristics with the riskscore was developed to predict patient survival at 1, 3, and 5 years (Fig. 4C). ROC curve analysis demonstrated that nomogram (AUC = 0.74) exhibited higher sensitivity and specificity than riskscore or clinical features alone (Fig. 4D). Fig. 4Nomogram construction and evaluation. A, B Univariate and multivariate Cox regression analyses of clinical features and riskscores in relation to LUAD patient survival. C, D Nomogram and ROC curves for survival prediction Nomogram construction and evaluation. A, B Univariate and multivariate Cox regression analyses of clinical features and riskscores in relation to LUAD patient survival. C, D Nomogram and ROC curves for survival prediction Functional enrichment analysis of risk-DEGs To elucidate underlying mechanisms of prognostic model, we conducted functional enrichment analysis. GO analysis revealed that Risk-DEGs were primarily involved in biological processes such as epidermis development, keratinocyte differentiation, and antimicrobial humoral response; cellular components including intermediate filament cytoskeleton and cornified envelope; and molecular functions such as receptor ligand activity, endopeptidase inhibitor/regulator activity, and hormone activity (Fig. 5A). KEGG analysis indicated enrichment in pathways like neuroactive ligand-receptor interaction, Staphylococcus aureus infection, and protein digestion and absorption (Fig. 5B). Fig. 5Functional enrichment analysis of Risk-DEGs. A GO enrichment analysis. B KEGG enrichment analysis. BP Biological Process, CC Cellular Component, MF Molecular Function Functional enrichment analysis of Risk-DEGs. A GO enrichment analysis. B KEGG enrichment analysis. BP Biological Process, CC Cellular Component, MF Molecular Function Immunological characteristics analysis The immune infiltration microenvironment plays a pivotal role in modulating tumor progression, therapeutic response, and patient prognosis by orchestrating the intricate interplay between immune cells and the tumor microenvironment (Liu et al. 2022; Mellman et al. 2023). To systematically explore the immune microenvironment characteristics associated with the ZNF-related risk score, we analyzed patterns of immune cell infiltration within the tumor microenvironment across patient groups stratified by risk level. Based on the ESTIMATE algorithm, it was found that low-risk group exhibited significantly higher immune and ESTIMATE scores (P < 0.05, Fig. 6A), indicating a stronger association with immune cell infiltration. Given differences in immune scores, we utilized ssGSEA to quantify their correlation with immune cell infiltration and immune function. Nine immune cell types (aDCs, B cells, DCs, iDCs, mast cells, neutrophils, pDCs, T helper cells, and TILs) were negatively correlated with riskscores (P < 0.05, Fig. 6B). Low-risk group showed elevated activity in HLA, T cell co-stimulation, and type II IFN response (Fig. 6C). Fig. 6Association between riskscores and the tumor microenvironment. A Differences in stromal, immune, and ESTIMATE scores between high- and low-risk groups. B, C Correlation between riskscores and immune cell infiltration/function. D, E Evaluation of immune checkpoint blockade therapy responses in both groups. F Differential expression analysis of immune checkpoint molecules between high- and low-risk groups in TCGA-LUAD. Significance: *P < 0.05; **P < 0.01; ***P < 0.001; ns, not significant Association between riskscores and the tumor microenvironment. A Differences in stromal, immune, and ESTIMATE scores between high- and low-risk groups. B, C Correlation between riskscores and immune cell infiltration/function. D, E Evaluation of immune checkpoint blockade therapy responses in both groups. F Differential expression analysis of immune checkpoint molecules between high- and low-risk groups in TCGA-LUAD. Significance: *P < 0.05; **P < 0.01; ***P < 0.001; ns, not significant Analysis of immune checkpoint inhibitor (PD-1 and CTLA-4) responsiveness demonstrated that low-risk group had significantly higher IPS values (P < 0.05, Fig. 6D), suggesting better responses to CTLA-4 inhibitors and PD-1/CTLA-4 dual blockade. TIDE algorithm supported these findings, indicating a lower likelihood of immune evasion in low-risk group (P < 0.05, Fig. 6E). It should be noted that both IPS and TIDE are computational models and do not directly serve as measures of treatment response; prospective clinical validation is required to confirm these associations. Collectively, these analyses indicate that low-risk patients are more likely to derive benefit from immunotherapy. Furthermore, we compared the differential expression profiles of immune checkpoint molecules between the high- and low-risk groups. The results revealed that, relative to the low-risk group, the high-risk group exhibited significantly downregulated expression of CTLA4, TIGIT, CD28, ICOS, BTLA, CD27, CD40LG, and CD96 (P < 0.05, Fig. 6F). Mutation analysis of signature genes Comparative mutation profiling revealed a higher overall mutation rate in high-risk patients than in low-risk patients (34.12% vs. 28.69%, Fig. 7A–B). Among 12 signature genes, RIMS2 was the most frequently mutated in both groups, followed by ZNF208, with higher mutation frequencies of these two genes observed in high-risk group. Fig. 7Mutation landscape of signature genes in high- and low-risk groups. A High-risk group. B Low-risk group Mutation landscape of signature genes in high- and low-risk groups. A High-risk group. B Low-risk group Drug sensitivity analysis Correlations between signature gene expression and drug IC50 values were predicted using CellMiner. Top 12 most significant candidate drugs (P < 0.001) are depicted in Fig. 8A. Positive correlations were observed between: KLF10 and IC50 values of four drugs (BLU-667, BMS-690514, VT-464, SHP-099), ZIC2 and IC50 value of Dimethylfasudil, CBFA2T3 and IC50 values of EMD-534085/PF-2771. Negative correlations were found between: TRIM29 and IC50 values of Imexon, VX-322, Chlorambucil, and Uracil mustard, CTCFL and IC50 value of NMS-E628. Fig. 8Drug sensitivity analysis. A Top 12 candidate drugs significantly correlated with signature gene expression. B Association between signature gene expression levels and drug sensitivity Drug sensitivity analysis. A Top 12 candidate drugs significantly correlated with signature gene expression. B Association between signature gene expression levels and drug sensitivity Compounds targeting KLF10, ZIC2, and CBFA2T3 were more effective in low-expression groups, whereas drugs associated with TRIM29 and CTCFL showed higher sensitivity in high-expression groups (Fig. 8B). Silent CTCFL inhibits the malignant behavior of LUAD Cells CTCFL (BORIS), a key DNA-binding protein involved in tumor regulation, is abnormally expressed in various cancers and closely related to tumor progression (Bergmaier et al. 2018; Soltanian and Dehghani 2018). In LUAD, CTCFL has been identified to serve as a biomarker linked to prognosis (Huang et al. 2023; Lopez-Sanchez et al. 2025). Prior studies have established that BORIS/CTCFL mechanistically governs the regulation of cancer-testis antigens (e.g., NY-ESO-1), as well as chromatin architecture and gene expression programs in malignant cells (Hong et al. 2005; Kang et al. 2007). Nevertheless, the functional mechanism of CTCFL in the context of LUAD remains to be comprehensively elucidated. Previously, our multivariate Cox regression analysis revealed that CTCFL, as one of the important risk factors in the LUAD prognosis model, is pronouncedly overexpressed in LUAD samples. Therefore, we further investigated the function of CTCFL in LUAD. In LUAD cell lines, detecting the expression of CTCFL, we found that CTCFL was markedly upregulated in multiple LUAD cell lines (A549, HCC78), with notably higher expression in the A549 cell line (p < 0.05) (Fig. 9A-B). Consequently, the A549 cell line was chosen for subsequent experiments. qRT-PCR and WB verification showed that a stable cell line with CTCFL knockdown expression was established (p < 0.05) (Fig. 9C). CCK-8 and clone formation results indicated that CTCFL knockdown remarkably attenuated the proliferation ability of A549 cells (p < 0.05) (Fig. 9D-E). As shown by wound healing and Transwell experiments, after knockdown of CTCFL, the migration and invasion abilities of A549 cells were significantly reduced (p < 0.05) (Fig. 9F-G). Collectively, CTCFL enhances the malignant behavior of LUAD cells. Fig. 9Functional exploration of the CTCFL gene. A, B qRT-PCR (A) and WB (B) detection of the transfection efficiency of CTCFL in different LUAD cell lines. C qRT-PCR and WB detection of the transfection efficiency of CTCFL in different LUAD cell lines. D The CCK-8 assay evaluation of the effect of CTCFL transfection on the viability of A549 cells. E The clone formation experiment of the effect of CTCFL transfection on the proliferation ability of A549 cells. F Wound healing of the effect of CTCFL transfection on the migration ability of A549 cells. G The Transwell assay of the effect of CTCFL transfection on the invasion ability of A549 cells. *, p < 0.05 Functional exploration of the CTCFL gene. A, B qRT-PCR (A) and WB (B) detection of the transfection efficiency of CTCFL in different LUAD cell lines. C qRT-PCR and WB detection of the transfection efficiency of CTCFL in different LUAD cell lines. D The CCK-8 assay evaluation of the effect of CTCFL transfection on the viability of A549 cells. E The clone formation experiment of the effect of CTCFL transfection on the proliferation ability of A549 cells. F Wound healing of the effect of CTCFL transfection on the migration ability of A549 cells. G The Transwell assay of the effect of CTCFL transfection on the invasion ability of A549 cells. *, p < 0.05 Discussion In this study, through the TCGA database, we evaluated the role of ZNFs-related genes in the prognosis of LUAD. The prognostic model based on ZNFs-related genes could serve as an independent prognostic indicator. The model was also associated with a unique pattern of immune cell infiltration and predicted patient sensitivity to immune checkpoint blockade therapy based on computational analysis. Additionally, we identified several potential candidate drugs related to signature genes. We screened out 12 ZNFs-related genes and constructed a reliable prognostic prediction model. Among these genes, CTCFL, IGF2BP1, RIMS2, TRIM29, TRIML2, ZIC2, and KLF10 were identified as important risk genes. As a zinc finger binding protein, CTCFL is widely expressed in various cell types and has been reported to be abnormally expressed in multiple cancers (Bergmaier et al. 2018). CTCFL holds an important regulatory role in various processes, including the self-renewal of cancer stem cells, chemotherapy resistance, tumorigenicity, sphere formation, and migration ability, thereby promoting tumorigenesis and progression (Soltanian and Dehghani 2018). In LUAD, CTCFL has been reported to be a potentially important prognostic biomarker (Huang et al. 2023; Lopez-Sanchez et al. 2025). Furthermore, prior studies have demonstrated that CTCFL is involved in regulating cancer-testis antigens, as well as modulating chromatin architecture and gene expression programs in malignant cells (Hong et al. 2005; Kang et al. 2007). Nevertheless, the mechanistic roles of CTCFL specifically within the LUAD context remain to be comprehensively elucidated. Notably, our functional experiments further confirmed the oncogenic effect of CTCFL in LUAD. We verified the high expression of CTCFL in LUAD cell lines (e.g., A549). Subsequently, in vitro experiments revealed that CTCFL knockdown markedly impaired the proliferation, migration, and invasion ability of LUAD cells. This finding was corroborated by the prediction results of the prognostic model, which clarified that CTCFL affects the prognosis by enhancing the malignant cellular behavior in LUAD. In addition to CTCFL, the RNA-binding protein IGF2BP1 accelerates various cancer progressions (Zhao et al. 2023). In LUAD, IGF2BP1 promotes the malignant phenotype of tumors (Wu et al. 2022) and interacts with CDCA1 to facilitate cell proliferation by regulating the PI3K/AKT signaling pathway (Feng et al. 2023). RIMS2 encodes a multi-domain presynaptic active region protein (Wang and Sudhof 2003). Multiple studies have shown that the dysregulated expression of RIMS2 may be closely linked to the tumorigenesis and progression of various cancers (Li et al. 2023c; Zhang et al. 2021c). However, while research on its role in LUAD remains insufficient, RIMS2 has been reported as a potential key prognostic biomarker for this cancer type (Ke et al. 2020). This is consistent with the conclusion of this study. TRIM29 is one of the members of the TRIM family and has complex functions with environmental dependence (Ai et al. 2014). Currently, the research on TRIM29 in LUAD is insufficient. In colorectal cancer, TRIM29 overexpression can promote cell invasion and metastasis, regulating epithelial-mesenchymal transition; while in breast cancer, the low expression of TRIM29 was found to enhance cell migration, invasion, and proliferation ability (Ai et al. 2014). As a paralog of the TRIM protein family, TRIML2 has been implicated in mediating the malignant tumor progression of gastric and oral cancers (Hayashi et al. 2019; Wang et al. 2025). However, its specific molecular mechanism remains unclear, particularly in the context of LUAD, where the gene is understudied. ZIC2 is an important regulatory factor of the Hedgehog signaling pathway and has been proven to promote tumor progression in various tumor types, including colorectal cancer (Tan et al. 2020), hepatocellular carcinoma (Lu et al. 2017), and epithelial ovarian cancer (Marchini et al. 2012). In LUAD, it has been shown that ZIC2 can enhance the stemness characteristics of cancer cells by upregulating OCT4 expression (Wei-Hua et al. 2020). KLF10 can be induced by the TGF-β signaling pathway and regulates tumorigenesis and development (Yoo et al. 2023). In LUAD, it may hold an oncogene role and the anti-proliferative effect of TGF-β signaling, demonstrating a bidirectional regulatory effect (Mishra et al. 2017). KLF10 can be induced by the TGF-β signaling and participates in tumorigenesis and development (Yoo et al. 2023). However, in LUAD, its specific function is not clear, exhibiting a bidirectional regulatory role. Its low expression is associated with poor survival, while it concurrently mediates the anti-proliferative function of TGF-β (Mishra et al. 2017). In summary, these key ZNFs-related genes may play an essential role in the tumorigenesis and progression of LUAD by modulating tumor proliferation, migration, stemness characteristics, and signaling pathway activity. Nevertheless, aside from CTCFL, the specific mechanism of action of other genes in lung adenocarcinoma still needs to be further explored. Functional enrichment analysis of Risk-DEGs in high- and low-risk groups revealed their involvement in biological processes such as epidermal development, intermediate filament cytoskeleton organization, endopeptidase inhibitor activity, hormone activity, and neuroactive ligand-receptor interactions. Epidermal development encompasses processes like differentiation, stratification, and terminal differentiation (Hu et al. 2018), with evidence suggesting its role in LUAD progression (Ding et al. 2022). Keratin 8, a key component of the intermediate filament cytoskeleton, is overexpressed in LUAD and drives cancer cell migration, invasion, and epithelial-mesenchymal transition via NF-κB signaling (Chen et al. 2022). Among endopeptidase inhibitors, SERPINB5 (a serine protease inhibitor) is upregulated in LUAD and enhances proliferation, invasion, and migration (He et al. 2023). Estrogen signaling through estrogen receptor β upregulates IL-6 and NUSAP1, contributing to LUAD progression (Huang et al. 2018; Zhang et al. 2024). Acetylcholine, via α5-nicotinic acetylcholine receptors, modulates FHIT expression and activates STAT3/DNMT1 pathway, facilitating LUAD cell migration and invasion (Jiao et al. 2023). Tumor immune microenvironment significantly influences LUAD patient outcomes (Sorin et al. 2023). Our data indicate that low-risk group exhibits higher infiltration of immune cells, including aDCs, B cells, DCs, iDCs, mast cells, neutrophils, pDCs, T helper cells, and TILs. Although mast cells and neutrophils can, under certain circumstances, exhibit pro-tumor or immunosuppressive activities, the concurrent reduction of these cell types alongside classical anti-tumor immune populations, including DCs, B cells, helper T cells, and TILs, within the high-risk group points to an immunosuppressive tumor microenvironment. This profile aligns with features of immune exclusion or immune desert phenotypes, which have been associated with unfavorable prognosis and diminished responsiveness to immunotherapy (Lei et al. 2020; Wang et al. 2024). Immune cell infiltration may also correlate with efficacy of immune checkpoint inhibitors (Herbst et al. 2014). CTLA-4 is an inhibitory molecule expressed on immune cells such as DCs, which regulates their function and reduces DC development and antigen-presenting activity (Laurent et al. 2010; Wang et al. 2011). Blocking CTLA-4 not only restores antigen-presenting function of DCs and T-cell activation but also represses cancer cell proliferation and drives their apoptosis (Chen et al. 2017). Research by Hollern et al. (2019) found that anti-CTLA4 therapy effectively activates B cells, leading to production of tumor-reactive antibodies and enhancing B-cell anti-tumor activity. Immunotherapy also influences neutrophil expression. One experiment showed that dual checkpoint inhibition of CTLA-4 and PD-1 increases neutrophil infiltration, which helps enhance the anti-tumor capacity of cytotoxic CD8 T cells, resulting in longer progression-free survival for cancer patients (Varaljai et al. 2023). Our IPS and TIDE scoring results predict that low-risk patients receiving anti-CTLA-4 therapy or combined PD-1 and CTLA-4 blockade are more likely to achieve clinical benefits. However, these findings are derived from computational predictions rather than direct clinical evidence, and future prospective clinical studies are warranted to validate the actual therapeutic benefits. Consistent with the suppressed immune microenvironment profile observed in the high-risk group, comparative analyses of immune checkpoint molecules revealed a concurrent downregulation of both inhibitory molecules (CTLA4, TIGIT, BTLA) and costimulatory molecules (CD28, ICOS, CD27, CD40LG), suggesting a paucity of TILs in high-risk tumors (Damei et al. 2023; Garcia-Barrientos et al. 2025; Mortezaee 2025). Specifically, CTLA4 and TIGIT are predominantly expressed on activated T cells and regulatory T cells (Tregs); their low expression in high-risk tumors more likely reflects reduced T cell infiltration rather than diminished immunosuppressive intensity (Li et al. 2023b; Long et al. 2025). Similarly, CD27 and CD40LG are key mediators of T cell–B cell cooperation and T cell survival, and their downregulation aligns with the phenotype of impaired adaptive immune responses in high-risk patients (Jacquot 2000; Rossetti et al. 2025). CD96, a recently identified inhibitory receptor on natural killer (NK) cells and T cells, was also significantly decreased in the high-risk group, further indicating reduced NK cell functional activity (Chai et al. 2025; Zhang et al. 2023a). Collectively, this immune checkpoint expression pattern confirms that high-risk LUAD patients exhibit a typical “cold tumor” immune microenvironment, characterized by insufficient immune cell infiltration and defective adaptive immunity. Such microenvironmental features may limit the therapeutic response to PD-1/PD-L1 blockade alone, suggesting that this patient subset may require combination strategies (e.g., with CTLA4 inhibitors or T cell activators) to improve clinical benefits. We also identified possible small-molecule compounds targeting signature genes, including SHP-099, Dimethylfasudil, EMD-534085, and PF-2771. SHP-099, an SHP2 inhibitor, synergizes with ceritinib (an ALK tyrosine kinase inhibitor) to suppress RAS/ERK1/2 reactivation and overcome resistance in NSCLC (Dardaei et al. 2018). Dimethylfasudil, a ROCK kinase inhibitor, selectively kills MYC-overexpressing cancer cells (Zhang et al. 2021a)—a relevant target given MYC amplification in > 30% of LUAD cases (Qu et al. 2021; Schaub et al. 2018). EMD-534085, a kinesin-5 inhibitor, induces mitotic arrest and death of cancer cells (Schiemann et al. 2010), though clinical trials showed limited efficacy in patients with advanced solid tumors or lymphomas (Hollebecque et al. 2013). CENPE, a mitotic kinesin, exhibits elevated expression in LUAD tissues; CENPE knockdown significantly hinders LUAD cell proliferation (Shan et al. 2019). Small-molecule inhibitor PF-2771 selectively targets CENPE, impairing growth of basal-type breast cancer cells through mechanisms involving genomic instability, DNA damage, and cell apoptosis (Kung et al. 2014). However, therapeutic potential of these drugs in LUAD requires validation through preclinical animal and clinical studies. Our study developed a ZNF-related gene signature that stratified LUAD patients into distinct risk groups with divergent survival outcomes, immune infiltration patterns, and immunotherapy responses. Although we have preliminarily verified, through in vitro experiments, the role of the key risk gene CTCFL in promoting malignant phenotypes of LUAD cells, thereby providing experimental support for the prognostic model, the functional regulation of other critical genes (e.g., RIMS2) and the combined effects of multiple genes (such as CTCFL and RIMS2) remains insufficiently elucidated. Furthermore, while the reliability of the risk model has been preliminarily substantiated through validation across multiple datasets, several limitations should be acknowledged. First, the validation cohorts employed in this study are derived predominantly from the GEO database, without covering LUAD samples from other public repositories such as cBioPortal. Consequently, the generalizability of the model across different ethnicities and clinical subgroups requires further assessment. Second, the present analyses are based on tissue-level bulk transcriptomic data, which preclude precise characterization of the expression patterns and functional roles of key genes within distinct cellular subpopulations of the tumor microenvironment. Therefore, future efforts should prioritize the integration of single-cell RNA sequencing datasets from lung cancer, enabling in-depth investigation of the regulatory mechanisms underlying core genes at single-cell resolution, thereby providing a more robust theoretical foundation for subsequent clinical translation. In addition, other clinical risk factors (e.g., smoking, air pollution, chronic respiratory diseases) were not incorporated due to data limitations; their integration may refine model predictive power. Supplementary Information Below is the link to the electronic supplementary material. Table S1: 4,188 ZNF-related genes Table S1: 4,188 ZNF-related genes Table S2: DEGs between LUAD and normal samples Table S2: DEGs between LUAD and normal samples Table S3: 186 strongly interacting genes from the PPI network Table S3: 186 strongly interacting genes from the PPI network Table S4: 12 ZNF signature genes used for riskscore modeling Table S4: 12 ZNF signature genes used for riskscore modeling Table S5: Expression patterns of the 12 signature genes in the TCGA training set, as well as the GSE26939 and GSE72094 validation sets Table S5: Expression patterns of the 12 signature genes in the TCGA training set, as well as the GSE26939 and GSE72094 validation sets Supplementary Material 6 Supplementary Material 6