RNA sequencing data and corresponding clinical information for LUAD were obtained from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/) using the TCGAbiolinks package [14], comprising 513 LUAD and 58 normal lung tissue samples. External validation cohorts were derived from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo), specifically from GSE50081 (127 LUAD samples) [15] and GSE26939 (116 LUAD samples)[16]. Additionally, two GEO datasets containing paired LUAD tumor and adjacent normal tissues, GSE32863 (58 paired samples)[17] and GSE75037(83 paired samples)[18], were used for external validation of FGD3 expression differences. The GEO datasets were normalized using the limma package [19]. Annotation GPL570, GPL9053 and GPL6884 were used for ID conversion. Genes with missing values were excluded, and for those with one-to-many mapping relationships, the mean expression value was retained. Genes belonging to the zinc finger protein family (Supplementary Table S1) were retrieved from the Universal Protein Resource (UniProt) database (https://www.uniprot.org/).

Differential expression analysis of the TCGA count data from LUAD tumors and adjacent normal tissue samples was performed using the limma package in R software (version 4.4.0) [19]. Genes met the criteria of log2|fold change| values > 1 and adjusted p-values < 0.05 were selected. By intersecting these genes with the previously obtained ZNF gene set, differentially expressed ZNFs (DE-ZNFs) were identified. Heatmaps were generated using the pheatmap package, and volcano plots were constructed with the ggplot2 package.

Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Gene Set Enrichment Analysis (GSEA) were performed using the clusterProfiler package[20]. A p-value < 0.05 was set as the threshold for statistical significance. The results of the enrichment analyses were visualized using the ggplot2 and ggnewscale packages.

Using TCGA as the training set, univariate Cox regression analysis was performed on DE-ZNFs with the survival package, using overall survival (OS) as the evaluation criterion. Genes significantly associated with OS (p < 0.05) were retained. Subsequently, least absolute shrinkage and selection operator (LASSO) regression was conducted using the glmnet and survival packages, and the results were visualized with the ggplot2 and forestplot packages. To avoid overfitting, multivariate Cox regression analysis was then applied to identify prognostic genes. The risk score for the model was calculated as follows, where β represents the regression coefficient, E denotes the expression level of each prognostic gene, and i is the number of genes. LUAD patients were stratified into high- and low-risk groups based on the median risk score. Kaplan–Meier survival curves were generated, and log-rank tests were conducted to evaluate survival differences between groups. The model’s predictive performance was assessed using receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC) representing its discriminatory ability. To further validate the predictive power of the model, its performance was evaluated in two independent datasets, GSE50081 [15] and GSE26939 [16].

CIBERSORT analysis was performed using the IOBR package to compare the relative abundances of 22 immune cell types between the high- and low-risk groups[21]. Single-sample gene set enrichment analysis (ssGSEA) was performed with the GSVA package to evaluate differences in 28 immune cell types[22]. Additionally, xCell analysis was conducted using the xCell package to quantify the abundances of 64 immune cell types[23]. Correlation heatmaps between the eight model genes and the 22 CIBERSORT-derived immune cell types were generated using the corrplot and ggcorrplot packages. The estimate package was applied to calculate stromal scores, immune scores, estimate scores (sum of stromal and immune scores), and tumor purity. Pearson’s product–moment correlation was used to assess associations between the risk score and stromal score, immune score, estimate score, and tumor purity.

LUAD data were downloaded from the TIDE database (http://tide.dfci.harvard.edu/) for tumor immune dysfunction and exclusion (TIDE) analysis[24]. A higher TIDE score indicates a poorer response to immunotherapy. The immunophenoscore (IPS) for each patient was obtained from the Cancer Immunome Atlas (https://tcia.at/) and compared between different risk groups. Unlike the TIDE score, a higher IPS is positively associated with a better predicted response to immunotherapies targeting PD-L1 and CTLA-4[25]. Additionally, immunotherapy data from the urothelial carcinoma IMvigor210 dataset were used to predict responses to immunotherapy in each risk group[26].

The tumor mutation burden (TMB) data for the TCGA cohort were downloaded from the cBioPortal database (https://www.cbioportal.org/) and calculated using the Maftools package[27]. Waterfall plots were generated to visualize the top 15 genes with the highest mutation frequency in the high- and low-risk groups. The ggpubr package was used to create violin and bar plots, and chi-square tests were applied for group comparisons. The rtracklayer, RCircos, and magrittr packages were employed to plot copy number variation (CNV) landscapes, enabling visualization of changes in DE-ZNFs model genes across human chromosomes.

Differentially expressed genes (DEGs) between the two groups were identified using the DESeq2 package with the Wald test. Genes with an absolute log₂ (fold change) > 1 and p < 0.05 were selected. GO, KEGG, and GSEA of the DEGs were performed using the clusterProfiler package. Gene set variation analysis (GSVA) was conducted using the GSVA package[22].

Drug sensitivity data were obtained from two major pharmacogenomic databases: the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Therapeutics Response Portal (CTRP), which contain large-scale screening results of numerous anti-cancer drugs across hundreds of cell lines. GDSC1, GDSC2 and CTRP2 drug response data were obtained from the GDSC website (https://www.cancerrxgene.org/) and the CTRP website (http://portals.broadinstitute.org/ctrp.v2.1/). The oncoPredict package was applied to predict the sensitivity of high- and low-risk LUAD patients, as defined by the prognostic model, to commonly used chemotherapeutic agents and molecularly targeted drugs[28]. The half-maximal inhibitory concentration (IC₅₀) represents the drug concentration required to inhibit 50% of tumor cell viability. A lower IC₅₀ value indicates greater drug sensitivity. IC₅₀ values were expressed in micromolar (µM) units, and the natural logarithm of the IC₅₀ was used for downstream drug response analyses.

To evaluate the association between clinical characteristics and risk grouping, univariate and multivariate Cox regression analyses were performed to determine whether the clinical variables and the risk score were independent prognostic factors for LUAD patients. Clinical data for LUAD patients in the TCGA cohort were obtained from the UCSC Xena platform (https://xenabrowser.net/datapages/). Fisher’s exact test was applied to compare clinical characteristics between high- and low-risk groups using the tableone package. A nomogram was then constructed to enhance the prognostic accuracy and predictive power of the model[29]. The risk score, T stage, and N stage were identified as independent prognostic variables to estimate the probability of overall survival (OS) at 1, 2, and 3 years. Calibration curves were generated to assess the agreement between predicted and observed outcomes, thereby evaluating the nomogram’s predictive performance.

The expression levels of the prognostic genes were validated using the TCGA LUAD dataset. Bar plots were generated with the ggpubr and ggsci packages, and group comparisons were conducted using the Wilcoxon test. In addition, immunohistochemical (IHC) data for the prognostic genes in LUAD tissues were obtained from the Human Protein Atlas (HPA) (https://www.proteinatlas.org/). Receiver operating characteristic (ROC) curves for the prognostic model genes were constructed using the pROC package to evaluate their discriminatory performance.

LUAD cell lines (H3122, H2228, H1650, H1993, and PC9) were purchased from the Cell Bank of the Chinese Academy of Sciences (Shanghai, China). Human bronchial epithelial (HBE) cells were kindly provided by the Heilongjiang Cancer Institute (Harbin, China). All cell lines were authenticated by short tandem repeat (STR) analysis prior to experimentation. H3122, H1650, H2228, and PC9 cells were cultured in RPMI 1640 medium (HyClone, USA) supplemented with 10% fetal bovine serum (Pan-Biotech, Germany) and 1% penicillin–streptomycin (HyClone, USA). H1993 cells were maintained in DMEM medium (HyClone, USA). All cells were incubated at 37°C in a humidified atmosphere containing 5% CO₂. A total of 12 paired fresh-frozen LUAD tumor tissues and adjacent normal tissues were collected from patients at Harbin Medical University Cancer Hospital between 2024 and 2025 for protein extraction. Additionally, 14 paired paraffin-embedded LUAD and adjacent normal tissue sections were obtained from the same institution between 2020 and 2022 for immunohistochemistry. None of the patients had received preoperative therapy.

Cells and tissues were harvested and lysed on ice for 30 min using lysis buffer (Boster, Wuhan, China; #AR0102-100) supplemented with 1% protease inhibitor cocktail (Boster, Wuhan, China #AR1192). The lysates were centrifuged at 12,000 rpm for 15 min at 4°C, and the supernatants were collected. Protein concentration was determined using a protein quantification kit (Beyotime, Shanghai, China; #P0012S). Equal amounts of protein (25 µg) were loaded per lane, separated by SDS–PAGE, and transferred to PVDF membranes. The membranes were blocked with 5% nonfat milk overnight at 4°C and incubated with primary antibodies for 24 h at 4°C. The following day, membranes were washed and incubated with HRP-conjugated secondary antibodies for 50 min at room temperature. Immunoreactivity was detected using an ECL kit (Thermo Fisher Scientific, Waltham, MA, USA). The primary antibodies were as follows: rabbit anti-FGD3 (Proteintech, Wuhan, China; #20347-1-AP; 1:1000) and mouse anti-GAPDH (Abcam, Cambridge, UK; #ab8245; 1:50,000).

Lorlatinib was purchased from MedChemExpress (Monmouth Junction, NJ, USA; #HY-15728). A total of 5,000 cells were seeded into each well of 96-well plates and cultured in 100 µl RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS) for 12 hours. After cell adhesion, the culture medium was replaced with 100 µl of complete medium containing varying concentrations of lorlatinib. Following 48 hours of drug treatment, the medium was removed, and 100 µl of fresh medium containing 10 µl of CCK-8 reagent (Seven, Shanghai, China; #SC11901) was added to each well. The plate was incubated for 1 hour at 37°C, and absorbance at 450 nm was measured using a multifunctional microplate reader. Dose-response curves were generated based on the absorbance values at each drug concentration, and the IC₅₀ values for each cell group were calculated.

Parental H3122 cells were initially exposed to a low concentration of lorlatinib (0.04 µM) until they achieved full proliferation. The cells were then passaged and treated with a doubled concentration of lorlatinib. This process was repeated incrementally until the lorlatinib concentration reached 1 µM. The resulting cells exhibited an IC₅₀ value more than 50 times higher than that of the parental cells. This resistant cell line was designated as H3122 lorlatinib-resistant cells (H3122LR). H3122LR cells were subsequently expanded and continuously maintained in medium containing the same lorlatinib concentration during each passage.

Immunohistochemistry was performed using the high sensitive and rapid immunohistochemical kit (pH 9.0) (Elabscience, Wuhan, China; #E-IR-R220). Briefly, paraffin-embedded sections were baked at 65°C for 60 minutes, then placed in antigen retrieval and deparaffinization solution and microwaved at low power for 30 minutes, followed by natural cooling to room temperature. After washing, sections were incubated with peroxidase blocking buffer for 15 minutes, washed again, and incubated overnight at 4°C with primary antibody against FGD3 (Proteintech, Wuhan, China; #20347-1-AP; 1:100). On the following day, sections were incubated with secondary antibody at room temperature for 30 minutes, followed by DAB staining for 2 minutes and hematoxylin counterstaining for 30 seconds. The sections were then dehydrated through graded ethanol series, cleared with xylene, mounted with neutral resin, and air-dried. FGD3 immunostaining was independently evaluated by two investigators blinded to patient information. The percentage of positive cells was scored as 0 (0%), 1 (1–25%), 2 (26–50%), 3 (51–75%), and 4 (76–100%). Staining intensity was scored as 0 (negative), 1 (light yellow), 2 (dark yellow), and 3 (brown). The final immunoreactive score (IRS) was calculated by summing the percentage and intensity scores, ranging from 0 to 7.

FGD3 was overexpressed in H3122LR cells either transiently or stably. For transient overexpression, cells at ~ 70% confluence were transfected with an FGD3 plasmid or empty vector using Lipofectamine 3000 (Invitrogen, USA). After 6 h, the medium was replaced, and cells were cultured for 24–48 h before downstream assays. For stable overexpression, cells were transfected as above and selected with puromycin until resistant colonies were established. Transfection efficiency and FGD3 expression were confirmed by Western blot. Plasmid sequences are provided in Supplementary Table S2.

H3122LR cells transfected with an FGD3 overexpression plasmid or empty vector were seeded into 6-well plates and cultured in complete medium. After 10 days, colonies were fixed with 4% paraformaldehyde, stained with crystal violet, and counted. Experiments were performed in triplicate.

Cell viability was assessed using the Cell Counting Kit-8 (CCK-8, Sevenbio, Beijing, China) according to the manufacturer’s instructions. Cells were seeded in 96-well plates and cultured for 1–4 days. CCK-8 solution was then added to each well at a 1:10 dilution and incubated at 37°C for 1 h. Absorbance at 450 nm was measured using a microplate reader (Multiskan FC, Thermo, China).

H3122LR cells transfected with an FGD3 overexpression plasmid or empty vector were seeded in 6-well plates and cultured for 24 h. A scratch was made through the center of each well using a 10 µL pipette tip, and serum-free medium was added. Wound closure was quantified as the percentage of wound area reduction over the healing period. All experiments were performed in triplicate.

For the transwell assay, 105 cells were seeded in the upper chamber (8-µm pore size, Corning, USA) in serum-free medium, either uncoated (migration) or coated with Matrigel (BD Biosciences, invasion). The lower chamber contained 600 µL of medium with 10% FBS. After 24 h (migration) or 48 h (invasion), cells on the lower surface of the membrane were fixed with 4% paraformaldehyde and stained with crystal violet. Cell numbers were quantified as the average number of cells per field of view.

For variables with normal distribution between two groups, unpaired Student’s t-tests were used to assess statistical significance. When data were not normally distributed, Mann–Whitney U tests were applied. For categorical clinical variables with small sample sizes, Fisher’s exact test was used to evaluate group differences. Pearson’s or Spearman’s correlation analyses were performed to examine linear relationships between variables. Kaplan–Meier survival analysis was conducted to compare overall survival between high- and low-risk groups, with differences assessed by the log-rank test. Receiver operating characteristic (ROC) curves and area under the curve (AUC) values were used to evaluate the predictive performance of the risk score. All analyses were performed using R software (version 4.4.0) and GraphPad Prism (version 9.5). Data are presented as means, with statistical significance defined as p < 0.05.

The workflow of this study was illustrated in Fig. 1. A total of 19,934 RNA-seq genes and clinical information for 513 LUAD patients and 58 control samples were downloaded. Among them, 1,834 ZNFs genes were obtained. After differential analysis, 270 DE-ZNFs were identified, including 83 downregulated and 187 upregulated genes (Fig. 2A, B). To evaluate the potential functions of DE-ZNFs in LUAD, GO enrichment analysis was performed. Sixteen biological process (BP) pathways, 5 cellular component (CC) pathways, and 23 molecular function (MF) pathways were identified, with the top five displayed. DE-ZNFs were mainly involved in the protein polyubiquitination process (Fig. 2C, D). KEGG enrichment analysis revealed a total of 111 enriched pathways, including key biological processes such as ubiquitin-mediated proteolysis, lysine degradation, endocytosis, and chromatin remodeling (Fig. 2E). Notable signaling pathways identified by GSEA and KEGG enrichment included TNFA signaling via NF-kappa B, KRAS signaling, and the E2F transcription factor pathway (Fig. 2E, F).

Univariate Cox regression analysis identified 63 DE-ZNFs significantly associated with OS in the TCGA training set (Fig. 3A). After excluding samples with missing values and OS = 0, 500 cancer samples remained. LASSO regression was then applied to further select 19 DE-ZNFs (Fig. 3B, C). Finally, multivariate Cox regression was used to establish the predictive model and identify prognostic-related genes: TRIM6, TRIM29, CTCFL, FGD3, GATA4, CASZ1, TRAF2, and ZNF322 (Fig. 3D). The risk score was calculated as followsLUAD samples were then divided into high- and low-risk groups based on the median risk score. Risk curves and scatter plots illustrated that the high-risk group exhibited higher risk scores and mortality rates compared to the low-risk group (Fig. 3E, F). The heatmap showed that TRIM29, GATA4, TRIM6, CTCFL, and TRAF2 were highly expressed in the high-risk group, whereas FGD3, CASZ1, and ZNF322 were highly expressed in the low-risk group (Fig. 3G). Kaplan–Meier analysis demonstrated that patients in the high-risk group had significantly poorer OS than those in the low-risk group (Fig. 4A). Time-dependent ROC analysis yielded AUC values of 0.756, 0.676, and 0.683 for 1-, 2-, and 3-year survival, respectively (Fig. 4B). Two independent GEO validation datasets also confirmed shorter OS in the high-risk groups (Fig. 4C, D). For the GSE50081 validation dataset, 1-, 2-, and 3-year AUC values were 0.714, 0.735, and 0.758, respectively; for GSE26939, the corresponding AUCs were 0.721, 0.792, and 0.792 (Fig. 4E, F).

According to the results from CIBERSORT, ssGSEA, and xCell analyses, the low-risk group exhibited a higher proportion of immune cell infiltration (Fig. 5), suggesting stronger tumor immunity compared with the high-risk LUAD patients. The prognostic-related gene FGD3 showed positive correlations with the infiltration of memory B cells, CD8 + T cells, and M1 macrophages, while ZNF322 and CASZ1 were associated with increased resting CD4 memory T cells (Fig. 6A). Patients in the high-risk group had lower immune scores, lower stromal scores, and higher tumor purity based on the ESTIMATE algorithm (Fig. 6B). The heatmap demonstrated that FGD3 was significantly positively correlated with immune and stromal scores, but negatively correlated with tumor purity. In contrast, CASZ1 and TRAF2 exhibited opposite correlation patterns compared to FGD3 (Fig. 6C). This prognostic model represents an innovative and effective biomarker for selecting anti-tumor immunotherapies. The TIDE score indicated that the high-risk group exhibited poorer immunotherapy efficacy (Fig. 6D). Validation using the IMvigor210 dataset confirmed that patients in the high-risk group had worse survival outcomes (Fig. 6E). To explore the relationship between risk score and immune checkpoint blockade (ICB) biomarkers, we found that expression levels of PD-L1, PD-L2, and CTLA4 were lower in the high-risk group, suggesting a reduced response to therapies targeting these molecules (Fig. 6F). CD80 and CD86, key ligands for T cell activation, were also downregulated in tumor cells from the high-risk group, which may contribute to immune escape (Fig. 6F). Furthermore, based on the IPS score, the low-risk group might derive greater benefit from anti-CTLA4 therapy (Fig. 6G). This comprehensive analysis elucidates the complex interactions between zinc finger proteins and immune therapy, confirming the prognostic model related to ZNFs as a valuable tool for guiding immunotherapy in LUAD.

To investigate the molecular consequences of genetic differences between the high- and low-risk groups and to better understand the biological processes associated with poor survival in the high-risk group, we analyzed the genomic variability of the ZNFs model within the TCGA LUAD cohort. Oncoprint visualization revealed that somatic mutation frequency was higher in high-risk patients (95.92%) compared to the low-risk group (88.21%). The top 15 mutated genes were identified, with missense mutations and multi-hit mutations being the most common types. Notably, mutation rates of TP53, TTN, MUC16, and KRAS exceeded expectations and were significantly elevated in the high-risk group relative to the low-risk group (Fig. 7A, B). Consistently, tumor mutation burden (TMB) was significantly higher in the high-risk group (Fig. 7C, D). However, stratification of patients based on the median TMB showed that those with higher TMB had longer overall survival (Figure S1). The chromosomal locations of the 8 prognosis-related ZNFs genes are displayed (Fig. 7E). GO functional enrichment analysis of DE-ZNFs between risk groups indicated predominant localization in the collagen-containing extracellular matrix, cornified envelope, and axon terminus. Molecular functions enriched in these genes included hormone activity, bile acid binding, and heparin binding (Fig. 8A). After ID conversion, 311 downregulated and 554 upregulated genes were retained for further analysis. KEGG pathway analysis revealed significant differences between risk groups in pathways such as chemical carcinogenesis–DNA adducts, neuroactive ligand–receptor interaction, steroid hormone biosynthesis, and drug metabolism via cytochrome P450 (Fig. 8B). GSEA identified 229 significantly enriched pathways (p < 0.05), with the top 10 pathways predominantly enriched in the high-risk group. These pathways were associated with cell division, cell cycle regulation, DNA replication, transcription and translation, chromosome maintenance, DNA methylation, and keratinization (Fig. 8C), potentially explaining the shorter overall survival observed in the high-risk group. GSVA enrichment analysis further demonstrated that poor prognosis in the high-risk group was associated with activation of the MYC signaling pathway, glycolysis, mTOR signaling, DNA repair, E2F transcription factors, unfolded protein response, and oxidative phosphorylation (Fig. 8D). These findings suggest potential therapeutic targets for LUAD patients involving ZNFs. To identify potential chemotherapeutic and targeted agents for high-risk LUAD patients, we performed drug sensitivity prediction using available pharmacogenomic datasets. The high-risk group exhibited potential resistance to carboplatin (Fig. 9A), gefitinib (Fig. 9B), and crizotinib (Fig. 9C), whereas increased sensitivity was predicted for paclitaxel (Fig. 9D), docetaxel (Fig. 9E), erlotinib (Fig. 9F), vinorelbine (Fig. 9G), and trametinib (Fig. 9H).

Clinical information for the high- and low-risk groups was downloaded, and inter-group comparisons were performed using Fisher’s exact test, as summarized in Table 1. The high-risk group exhibited a higher proportion of male patients, advanced T stage, increased lymph node metastasis, and higher overall stage. Univariate and multivariate Cox regression analyses were conducted on clinicopathological factors—including gender, age, stage, TNM classification, and risk score—to identify independent prognostic factors for LUAD. Variables with p-values less than 0.05 in both analyses were deemed independent risk factors; accordingly, T stage, N stage, and risk score were selected (Fig. 10A, B). These independent prognostic factors were then integrated to construct a nomogram, facilitating the prediction of 1-, 2-, and 3-year overall survival (OS) probabilities in the TCGA-LUAD cohort. Due to the limited sample size in the T4 and N3 categories, singular fit issues arose during Cox regression modeling; hence, T3 and T4 were combined, as were N2 and N3 groups (Fig. 10C). Calibration plots for 1-, 2-, and 3-year OS demonstrated good agreement between predicted and observed survival outcomes (Fig. 10D).

Our study validated the expression differences of the eight ZNF prognostic model genes between patients and normal samples in the TCGA-LUAD database. We validated the differential expression of the eight ZNF prognostic model genes between LUAD patients and normal samples in the TCGA-LUAD dataset. The results showed that TRIM6, GATA4, TRAF2, and ZNF322 were significantly upregulated in LUAD patients, whereas TRIM29, FGD3, and CASZ1 were downregulated (Fig. 11). Protein expression levels of these eight prognostic genes in LUAD and normal tissues were further analyzed using data from the Human Protein Atlas (HPA) database. The protein expression patterns of GATA4, TRAF2, and ZNF322 were consistent with their RNA expression trends, showing elevated levels in LUAD patients. Similarly, FGD3 and CASZ1 exhibited lower protein expression in patients, consistent with RNA data. No significant differences were observed between tumor and normal tissues for CTCFL at either the protein or RNA level. Notably, TRIM6 protein was expressed at low levels in both tumor and normal tissues, and TRIM29 protein expression showed an inverse pattern relative to its RNA expression (Fig. 12). Kaplan–Meier survival analysis demonstrated that higher expression of TRIM6, TRIM29, and TRAF2, along with lower expression of FGD3, CASZ1, and ZNF322, were associated with poorer overall survival in LUAD patients. Expression levels of the remaining two genes did not show significant prognostic relevance (Fig. S2).

Among the eight prognostic-related genes, FGD3 exhibited the strongest correlation with ESTIMATE scores and a relatively high correlation coefficient within the prognostic model. Consequently, we prioritized FGD3 for downstream validation. Paired RNA-seq data from TCGA showed significantly lower FGD3 expression in LUAD tissues compared to normal samples (Fig. S3A), and consistent results were observed in two paired GEO datasets (Fig. S5). Protein expression of FGD3 was further validated in LUAD cell lines and normal bronchial epithelial (HBE) cells. Except for PC9, FGD3 protein levels in H3122, H2228, H1650, and H1993 cells were significantly lower than those in HBE cells (Fig. 13A). A similar expression pattern was observed in matched tumor and adjacent normal tissues from LUAD patients (Fig. 13B). The immunohistochemistry results showed that compared to LUAD tissues, FGD3 expression was significantly higher in normal tissues. Cellular localization revealed that FGD3 was expressed in both the cytoplasm and nucleus, with a greater expression in the cytoplasm. FGD3 was more highly expressed in the epithelial cells of the trachea and bronchi (Fig. 14). GO, KEGG, and GSEA enrichment analyses revealed that genes differentially expressed with FGD3 were mainly enriched in cilium movement and cell adhesion pathways in LUAD (Fig. S3B–D). Drug sensitivity prediction indicated that the low FGD3 expression group was less sensitive to the first-generation ALK inhibitor crizotinib (Fig. 13C). In lorlatinib-resistant H3122 cells (H3122LR) (Fig. S4), FGD3 expression was decreased compared to parental cells (Fig. 13A). Moreover, FGD3 expression showed a dose- and time-dependent decline in ALK-positive cell lines H3122 and H2228 with prolonged lorlatinib treatment and increasing drug concentration (Fig. 13D). The FGD3 plasmid was transfected into H3122LR cells, and transfection efficiency was confirmed by Western blot (Fig. 15A). Colony formation assays showed that FGD3 overexpression markedly reduced the number of colonies formed by H3122LR cells (Fig. 15B). Consistently, cell viability was significantly decreased upon FGD3 overexpression (Fig. 15C). Furthermore, transwell assays demonstrated that upregulation of FGD3 substantially impaired both migration and invasion capabilities (Fig. 15D). Wound-healing assays further indicated that FGD3 overexpression attenuated the cells’ wound closure ability (Fig. 15E). FGD3 overexpression can partially reverse lorlatinib resistance in H3122LR cells (Fig. 15F).