Multiomics integration prioritizes potential drug targets for lung cancer Authors: Peize Meng, Zhiping Zhang, Zheng Ruan, Congcong Xu Published: 2026 ABSTRACT BackgroundTo provide more treatment options for patients with advanced lung cancer, this study aims to identify new therapeutic drug targets via the integration and analysis of multi-omics datasets.MethodsIn this study, we performed MR and co-localization analysis to identify candidate drug targets associated with lung cancer by leveraging gene expression quantitative trait loci (eQTLs) data from the eQTLGen consortium, and validated these findings orthogonally using protein expression trait loci (pQTL) data. In addition, we performed survival analysis, genetic diagnostic analysis, and differential expression analysis to further test their correlation with lung cancer. Summary-level data from ILCCO and FinnGen R9 were used as discovery datasets, and the TRICL consortium was used as an independent replication cohort. Together, we applied these methods to assess the causal relationship between these putative gene targets and lung cancer risk, and to determine whether they are associated with patient morbidity and prognosis.FindingAmong 2,645 candidate genes with robust cis-eQTL instruments, we prioritized four genes (SERPING1, THBS3, FLT4 and CHEK1) that showed strong causal associations with lung cancer risk and high posterior probabilities of co-localization (PP.H4 > 0.80) between eQTL and lung cancer GWAS signals. Genetically higher expression of SERPING1, THBS3 and FLT4 was associated with an increased risk of lung cancer, whereas higher CHEK1 expression was associated with a reduced risk. All four putative gene targets were consistently validated across independent cohorts.InterpretationThis multi-omics MR framework identifies genes whose genetically predicted expression is associated with increased or decreased lung cancer risk and may help prioritize candidate targets for future lung cancer drug development.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05415-0. Introduction Lung cancer is one of the most commonly diagnosed malignancies worldwide and remains the leading cause of cancer-related death. GLOBOCAN estimates suggest that lung cancer accounts for about 2.2–2.5 million new cases and roughly 1.8 million deaths globally, corresponding to nearly one in five cancer deaths [1, 2]. Non-small-cell lung cancer (NSCLC) accounts for roughly 85% of all lung cancer cases, while small-cell lung cancer (SCLC) accounts for approximately 15% [3, 4]. Although advances in surgery, radiotherapy, chemotherapy, targeted therapy and immunotherapy have improved outcomes for selected patients, many individuals still present with advanced disease, recurrence is common, and the 5-year survival rate for advanced-stage lung cancer remains unsatisfactory [5–7]. There is therefore a continued need to identify and validate novel therapeutic targets to expand treatment options and improve prognosis. Drug target discovery in lung cancer has traditionally relied on tumor genomic profiling and functional studies. Recurrent alterations in EGFR, ALK, ROS1, RET, MET, BRAF, KRAS and other genes have defined key oncogenic drivers and enabled the development of effective targeted therapies, especially in NSCLC [6, 8]. In parallel, large-scale sequencing and CRISPR-based functional genomic screens have uncovered additional cancer dependencies and vulnerabilities [9, 10]. More recently, multi-omics approaches integrating transcriptomic, epigenomic and proteomic data have been used to refine molecular subtypes and nominate candidate pathways in lung cancer [8, 11]. However, many proposed targets still lack robust support from human germline genetics, and systematic pipelines that link the druggable genome to lung cancer risk through causal inference remain scarce. Mendelian randomization (MR) has emerged as a powerful framework for drug target prioritization in complex diseases. By using genetic variants as instrumental variables for molecular traits, MR can strengthen causal inference and has been increasingly applied to validate and prioritize targets in cardiometabolic and neurological disorders, where genetic support has been associated with higher success rates in drug development [12–14]. The human “druggable genome” defined by Finan et al. provides a systematic catalogue of genes amenable to pharmacological modulation, which can be leveraged in such analyses [11]. Nevertheless, comprehensive MR-based screens of druggable genes for lung cancer, integrating expression quantitative trait loci (eQTL), protein quantitative trait loci (pQTL) and genome-wide association study (GWAS) data, are still limited. In this context, we aimed to systematically evaluate the causal relevance of druggable genes for lung cancer by combining cis-eQTL, cis-pQTL and lung cancer GWAS data in a two-sample MR framework, complemented by Bayesian co-localization, proteome-wide MR, differential expression analyses, survival analyses and ROC-based diagnostic evaluation, to prioritize druggable genes as potential therapeutic targets. Methods MR relies on three key assumptions: relevance, independence and exclusion restriction. First, the instrumental variables are strongly associated with the exposure trait (relevance assumption). Second, the instruments are not associated with any confounders of the exposure–outcome relationship (independence assumption). Third, the instruments affect the outcome only through the exposure and not via alternative pathways (exclusion restriction assumption). This study adheres to the Strengthening the Reporting of Observational Studies in Epidemiology using Mendelian Randomization (STROBE-MR) recommendations. Following Finan et al., we used the updated human druggable genome as our starting point. In that work, the authors defined 4479 druggable genes encoding targets of licensed drugs, proteins in clinical development, or members of key drug target families. In line with subsequent druggable genome-wide MR studies, we restricted this set to 4302 druggable genes located on autosomal chromosomes [1–22] and annotated with HGNC gene symbols [15]. Considering that cis-eQTL are usually more proximal to the gene of interest for drug-development MR studies, we obtained statistically significant cis-eQTL (false discovery rate < 0.05 within a ± 1 Mb window around each probe) from the eQTLGen consortium, a meta-analysis of whole-blood eQTL in 31,684 individuals [16, 17]. We intersected the 4302 autosomal druggable genes with the eQTLGen cis-eQTL results and, for each overlapping gene, extracted cis-eQTL located within ± 100 kb of each gene’s genomic position. Independent instruments were obtained by LD clumping at r² < 0.10. We required each SNP to be strongly associated with gene expression (P-value < 5E-8 and F-statistics > 10). After these steps, cis-eQTL instruments were available for 2645 druggable genes, which we refer to as ‘druggable genes with robust cis-eQTL instruments’ and carried forward as the exposure set in our MR analyses. Data resource of lung cancer Summary-level GWAS data for lung cancer were obtained from three large consortia: the Transdisciplinary Research in Cancer of the Lung/International Lung Cancer Consortium (TRICL/ILCCO) and the FinnGen R9 study. The ILCCO dataset provides detailed GWAS results for NSCLC and its main histological subtypes, including LUAD and LUSC, but it does not include SCLC. In our analyses, ILCCO therefore served as the primary discovery dataset for NSCLC, LUAD and LUSC. The FinnGen R9 dataset contains GWAS summary statistics for several lung cancer phenotypes, including small-cell lung cancer (SCLC), for which ILCCO has no data. Accordingly, FinnGen R9 was used as the discovery dataset for SCLC. The TRICL consortium includes GWAS data for both NSCLC and SCLC. In this study, TRICL was used as an independent replication dataset for all histological subtypes for which TRICL provided summary-level data (NSCLC, LUAD, LUSC, SCLC) [18]. The pooled lung cancer data in this study are publicly available from the Integrative Epidemiology Unit (IEU) (https://gwas.mrcieu.ac.uk/) and FinnGen (https://r9.finngen.fi/) without the need for a platform application. Lung cancer patients are classified according to The International Classification of Diseases (ICD). Data sources used for lung cancer research in this study can be found in Table S7. Lung diseases To minimize confounding factors, we chose six respiratory diseases that could potentially be linked to lung cancer, including pulmonary tuberculosis, pulmonary fibrosis and chronic obstructive pulmonary disease. (Table S8). Genetic data on lung diseases are all from European populations. Two-sample MR analysis was used to clarify causality. Summary-data-based MR analysis Our eQTL data were obtained from the eQTLGen Consortium’s 37 datasets, pQTL data were obtained from deCODE genetics and Suhre et al.’s study, and NSCLC data were obtained from ILCCO. These datasets were derived from independent cohort studies, and, to the best of our knowledge, do not include overlapping participants. We used the TwoSampleMR package to perform two-sample MR analysis on the 2,645 druggable genes with robust cis-eQTL instruments. For each SNP-exposure association, we first calculated the proportion of variance in the exposure explained by the SNP (R2) as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}^{2}=\:\frac{2\times\:EAF\times\:\:(1-EAF)\times\:{\beta\:}^{2}}{2\times\:EAF\times\:\left(1-EAF\right)\times\:{\beta\:}^{2}+2\times\:EAF\times\:\:(1-EAF)\times\:SE\times\:N\times\:{\beta\:}^{2}},$$\end{document} where EAF is the effect allele frequency, β is the estimated effect of the SNP on gene expression (or protein level), SE is its standard error and N is the exposure sample size. The F-statistic for each SNP was then derived from r² using\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F=\frac{{r}^{2}\times\:\:(N-2)}{1-{r}^{2}},$$\end{document} SNPs with F < 10 were considered weak instruments and were excluded from the MR analysis. For each gene, the total instrument strength was obtained by summing r² across all selected SNPs [19, 20]. The primary MR analysis used the Wald ratio method for genes with a single instrument and an inverse-variance-weighted (IVW) meta-analysis for genes with multiple instruments to derive overall causal effect estimates [21, 22]. To validate the robustness of the results, we applied the following complementary methods: (i) MR-Egger, which provides consistent causal effect estimates when instrumental variables exhibit pleiotropic effects, but the results are susceptible to outlier SNPs [23]. (ii) The weighted median method can provide a consistent causal estimate even if up to 50% of the total weight comes from invalid instruments, by taking the median of the SNP-specific ratio estimates [24]. (iii) The simple median method, which takes the median of the SNP-specific causal estimates and is robust to outliers when at least half of the total weight comes from valid instruments. In this study, we used IVW as the primary method and considered results from MR-Egger, weighted median and simple median as sensitivity analyses [25]. To control for multiple comparisons, we performed Bonferroni correction, and defined p-values < 1.89E-5 (p = 0.05/2645)as significant for the discovery set. For the validation set, replication of significant targets was considered if the p-value < 0.05/N, where N represents the number of significant targets in the discovery set. For sensitivity analysis, we used several complementary approaches to assess potential violations of the MR assumptions. First, we used Cochran’s Q statistic to test for heterogeneity across SNP-specific causal estimates [26]. We then applied the MR-Egger intercept test to evaluate directional horizontal pleiotropy, whereby a significant intercept (P < 0.05) would indicate that the genetic instruments may violate the core MR assumptions [27]. In our analysis, all instrumental variables were valid, and no pleiotropic effects were observed. All analyses in this MR study were conducted using RStudio (version 4.2.1). Co-localization analysis and power calculations We performed co-localization analysis to determine whether lung cancer and eQTLs share causal genetic variations. We utilized the Coloc R package to test whether the SNP associations between potentially drug-targetable genes and lung cancer are driven by linkage disequilibrium. For each genetic locus in the eQTL dataset, the coloc framework evaluates five mutually exclusive hypotheses: (i) the locus is associated with neither the gene nor lung cancer; (ii) it is associated only with the gene (cis-eQTL); (iii) it is associated only with lung cancer; (iv) it is associated with both traits, but through distinct causal variants; and (v) it is associated with both traits through the same causal variant. We considered a posterior probability for hypothesis 5 (PP.H4) > 0.80 as strong evidence of co-localization [28]. Proteome-wide MR analysis based on protein expression To further validate the causal relationship between genes and lung cancer, we performed proteome-wide MR analyses on genes that passed an initial co-localization standard (PP.H4 > 0.80). We utilized two independent cis-pQTL datasets: one from Ferkingstad et al. (N = 35,559 Icelanders) and another from Suhre et al. (approximately 1,000 Europeans) [29, 30]. For each gene, instrumental variables were selected from the corresponding protein’s cis-pQTLs at a genome-wide significant threshold (P < 5E-8). We prioritized the Ferkingstad et al. dataset; if no suitable instrumental variables were identified, we then screened the Suhre et al. dataset. The pQTL data were sourced from deCODE Genetics (Ferkingstad et al.) and the IEU Open GWAS platform (Suhre et al.), respectively (see Table S7 for details). Differential expression analysis and Survival analysis To characterize differences in gene expression between tumor and normal lung tissues and to evaluate the prognostic relevance of the prioritized genes, we performed differential expression and survival analyses using public transcriptomic resources. The UALCAN database (https://ualcan.path.uab.edu) is a web portal based on TCGA RNA-seq data that facilitates comparisons of gene expression between primary tumors and matched normal tissues and provides basic survival analyses [31]. We used UALCAN to compare the expression levels of the selected genes between lung cancer subtypes (LUAD and LUSC) and normal lung tissues. GEPIA is an interactive tool that integrates RNA-seq data from 9736 tumor samples and 8587 normal samples from the TCGA and GTEx projects. We used GEPIA to perform Kaplan-Meier survival analyses for OS and DFS according to high versus low gene expression in NSCLC and its subtypes. These analyses were intended as supportive functional evidence to complement the primary genetic findings from MR and co-localization. Protein-protein interactions and assessing druggability To determine whether proteins encoded by the four genes interact with each other, we constructed a protein-protein interactions (PPI) network through the STRING database (https://string-db.org/). We explored potential therapeutic targets by querying these protein interactions in the DrugBank database. The DrugBank database documents drug names along with their interactions with identified proteins, offering valuable insights into the drug development process. We further prioritized potential druggable targets by integrating relevant information, including drug–gene interactions and gene functions [32]. Genetic diagnostic analysis To further evaluate the diagnostic potential of the prioritized genes, we performed receiver operating characteristic (ROC)-based diagnostic analyses. Based on the results of two-sample MR, co-localization and survival analyses, we focused on four genes that showed the strongest and most consistent associations with lung cancer risk: THBS3 [ENSG00000169231.13], SERPING1 [ENSG00000149131.15], CHEK1 [ENSG00000149554.13] and FLT4 [ENSG00000037280.16]. For each gene, we assessed its ability to discriminate between tumor and normal lung tissues within the relevant histological subtype (NSCLC overall, LUSC or LUAD). RNA-seq data for the TCGA-LUSC (lung squamous cell carcinoma) and TCGA-LUAD (lung adenocarcinoma) projects were downloaded from the TCGA data portal (https://portal.gdc.cancer.gov). Gene-level expression values in fragments per kilobase of transcript per million mapped reads (FPKM) were extracted and transformed using log₂ (FPKM + 1) prior to analysis. For each gene-subtype pair, we plotted ROC curves to evaluate the diagnostic performance of gene expression in distinguishing tumor from normal tissues. The area under the ROC curve (AUC) and its 95% confidence interval were calculated using the DeLong method implemented in the pROC R package (version 1.18.0), and optimal cut-off values were determined by maximizing the Youden index. All statistical analyses and visualizations were performed in R (version 4.2.1) using pROC and ggplot2 (version 3.3.6), and a two-sided α = 0.05 was considered statistically significant. Results To identify therapeutic gene targets associated with lung cancer, we began with 4,302 autosomal druggable genes defined according to Finan et al. and then intersected them with eQTLGen cis-eQTL data, yielding 2,645 druggable genes with robust cis-eQTL instruments for further analysis (Fig. 1, Table S1). Using two-sample MR analysis, we assessed the relationship between the expression of these genes and lung cancer risk. Our investigation leveraged lung cancer data from TRICL, ILCCO, and FinnGen Consortium R9 release, in addition to cis-eQTL data from the eQTL Consortium, and cis-pQTL data from Ferkinstad et al. [10] and Suhre, et al. [29]. All aforementioned data are publicly available. We adjusted the threshold using the Bonferroni correction method. For MR results that underwent multiple testing and heterogeneity checks, validated in the replication set, we performed co-localization analysis to ascertain whether instrumental variables share common genetic variations associated with lung cancer risk. Moreover, we conducted supplementary causal analysis at the protein level, differential expression analysis, diagnostic analysis, and survival analysis to further validate and bolster the credibility of our findings. Fig. 1Overview of the study design. DEG, differentially expressed genes; GWAS, genome-wide association study; IVW, inverse-variance weighted; NSCLC, non-small-cell lung cancer; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; SCLC, small-cell lung cancer; PPI, Protein-protein interactions; pQTL, protein quantitative trait loci; ILCCO, International Lung Cancer Consortium; TRICL, Transdisciplinary Research in Cancer of the Lung; IV, Instrumental Variable; FDR, false discovery rate; N, sample size Overview of the study design. DEG, differentially expressed genes; GWAS, genome-wide association study; IVW, inverse-variance weighted; NSCLC, non-small-cell lung cancer; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; SCLC, small-cell lung cancer; PPI, Protein-protein interactions; pQTL, protein quantitative trait loci; ILCCO, International Lung Cancer Consortium; TRICL, Transdisciplinary Research in Cancer of the Lung; IV, Instrumental Variable; FDR, false discovery rate; N, sample size MR analysis After considering multiple testing corrections, the discovery cohort identified 89, 94, 85 and 36 candidate drug targets associated with lung cancer risk in NSCLC, LUAD, LUSC, SCLC groups, respectively (P < 0.05/2645 [IVW], Bonferroni correction). To validate these findings, we replicated the genes from the discovery cohort using data from the TRICL database, obtaining 41, 33, 40 and 8 candidate drug targets corresponding to NSCLC, LUAD, LUSC, SCLC, respectively. All significant SNPs are listed in Table S2. Co-localization analysis and power calculations We performed co-localization analysis to determine whether SNPs associated with lung cancer and candidate drug targets share the same causal genetic variants. The analysis suggested co-localization between several genes and certain lung cancer subtypes: SERPING1, THBS3, HIST1H4H with NSCLC; FLT4 with LUAD; CHEK1 with LUSC, and PSMA4 with SCLC. In total, six genes met the criteria for co-localization. Specifically, SERPING1 (PP.H4 = 0.897902828, Fig. 2A), THBS3 (PP.H4 = 0.956411758, Fig. 2B), CHEK1 (PP.H4 = 0.815339571, Fig. 2D), and FLT4 (PP.H4 = 0.812321356, Fig. 2E) exhibited primary co-localization support (PP.H4 > 0.8) and were identified as level 1 target genes. Meanwhile, HIST1H4H (PP.H4 = 0.75427073, Fig. 2C) and PSMA4 (PP.H4 = 0.679008889, Fig. 2F) demonstrated secondary co-localization support (0.5 < PP.H4 < 0.8) and were classified as level 2 target genes (Table S3). Power calculations indicated adequate statistical power for all six genes (Table S3). Fig. 2Regional Manhattan plots evaluating co-localization between eQTL signals from eQTLGen and lung cancer GWAS associations Regional Manhattan plots evaluating co-localization between eQTL signals from eQTLGen and lung cancer GWAS associations The upper panel displays the -log10 (P) values of SNP-gene expression associations from the eQTLGen dataset, and the middle panel presents the -log10 (P) values of SNP-lung cancer risk associations from the corresponding GWAS. The x-axis indicates genomic coordinates. SNPs are colored according to their linkage disequilibrium (LD, r²) with the lead SNP, which is highlighted by a purple diamond. The P-values shown in the plots are directly derived from the eQTL and GWAS summary statistics. A Bonferroni-corrected significance threshold was applied during instrument selection for the Mendelian randomization analyses; however, this threshold is not used in co-localization analysis and is therefore not displayed in the figure. Panels correspond to the lead eQTL variants used as genetic proxies for gene expression: rs28362951 for SERPING1, rs2974937 for THBS3, rs3999544 for HIST1H4H, rs76045215 for CHEK1, rs2387281 for FLT4, and rs931794 for PSMA4. MR results of six significantly lung cancer-related genes For the replication phase, employing either the Wald ratio or IVW method, six genes associated with lung cancer were successfully validated in the TRICL dataset, meeting the co-localization criteria (Table S4/5 and Fig. 3). We then performed a meta-analysis encompassing two datasets and found that five lung cancer-related genes exhibited significant correlations. Specifically, for NSCLC, we observed that a one-standard-deviation increase in the predictive genetic expression level resulted in odds ratios (OR) with 95% confidence intervals (95% CI) of 1.12 (1.05, 1.13) for THBS3, 1.24 (1.04, 1.49) for SERPING1, and 1.20 (1.15, 1.24) for HIST1H4H. Consistent with the MR results, genetically higher CHEK1 expression was associated with a lower risk of LUSC, and CHEK1 also showed a protective trend in proteome-wide MR analyses. In the context of SCLC, PSMA4 demonstrated an OR of 2.96 (2.36, 3.70). However, for LUAD, the post-META analysis did not reveal a statistically significant effect (p = 1.38E-01), with the confidence interval inclusive of 1; the OR 95% CI stood at 1.39 (0.90, 2.16). Beyond lung cancer, we then performed MR analysis to examine whether these level 1 genes (PP.H4 > 0.8) were causally linked to other pulmonary disorders. The analysis revealed that THBS3 is a risk factor for chronic bronchitis, while SERPING1 is correlated with bronchiectasis and appears to have a protective effect. Conversely, FLT4 and CHEK1 showed no association with any of the six evaluated pulmonary diseases. The MR results for these gene–disease pairs are shown in Figure S1. Fig. 3Datasets: Aggregated lung cancer GWAS data from ILCCO and FinnGen, with replication data from TRICL. Results of MR analysis for six candidate drug targets. ILCCO, International Lung Cancer Consortium; TRICL, Transdisciplinary Research in Cancer of the Lung; NSCLC, non-small-cell lung cancer; SCLC, small-cell lung cancer; IVW, inverse-variance weighted; OR, odds ratio; 95% CI, 95% confidence interval Datasets: Aggregated lung cancer GWAS data from ILCCO and FinnGen, with replication data from TRICL. Results of MR analysis for six candidate drug targets. ILCCO, International Lung Cancer Consortium; TRICL, Transdisciplinary Research in Cancer of the Lung; NSCLC, non-small-cell lung cancer; SCLC, small-cell lung cancer; IVW, inverse-variance weighted; OR, odds ratio; 95% CI, 95% confidence interval Detection of heterogeneity and directional pleiotropy MR-Egger intercept analysis did not indicate directional pleiotropy in any of the analyses (Table S5). The P-values were 0.519, 0.568, 0.753, 0.446 and 0.808 for THBS3, SERPING1, HIST1H4H, FLT4 and PSMA4, respectively. Because fewer than three SNPs were available for CHEK1, MR-Egger analysis could not be performed for this gene. None of the MR-Egger intercept P-values were statistically significant, supporting the absence of horizontal pleiotropy. Heterogeneity was assessed using Cochran’s Q statistic, which provided no evidence of heterogeneity. Steiger directionality tests did not suggest reverse causation between exposures and outcomes. Together, these sensitivity analyses support the robustness of our MR results. Proteome-wide MR analysis of protein expression THBS3 gene encodes Thrombospondin 3. Four SNPs were selected as cis-pQTL instruments at the genome-wide significance threshold (P < 5E-8). MR analysis for NSCLC indicated that higher genetically predicted thrombospondin 3 levels were associated with increased NSCLC risk (OR = 1.40, P = 4.05E-02), consistent with cis-eQTL-based results. For SERPING1, which encodes plasma protease C1 inhibitor, no eligible instruments remained after LD clumping at P < 5E-8. For FLT4, encoding Vascular Endothelial Growth Factor Receptor-3, 23 SNPs passed screening; MR analysis for LUAD showed a positive association between FLT4 and LUAD (OR = 1.060, P = 4.78E-02), again concordant with the cis-eQTL findings. For CHEK1, encoding checkpoint kinase 1 (CHK1), no genome-wide significant cis-pQTL remained after LD filtering in the protein dataset of Ferkingstad et al. [10]. Therefore, we performed a secondary screening by utilizing data from Suhre K et al. [29], identifying only one valid SNP. Subsequent MR analysis with LUSC as the outcome indicated that CHK1 may function as a protective factor for LUSC (OR = 0.80, P = 9.39E-02), aligning with the gene expression-based findings. (Table S6) Differential expression analysis and survival analysis We then performed differential expression analyses to compare expression levels of the candidate genes between lung cancer subtypes and normal lung tissues (Figure S2). In both LUAD and LUSC, the expression of THBS3 (Figure S2A/B) and HIST1H4H (Figure S2I/J) was higher in tumors than in normal tissues, whereas SERPING1 expression (Figure S2C/D) was lower in tumors. In LUAD, FLT4 expression (Figure S2G/H) was lower in tumors than in normal tissues, while in LUSC, CHEK1 expression (Figure S2E/F) was higher in tumors. The UALCAN database does not provide expression data for SCLC, so differential expression analyses were restricted to NSCLC subtypes. We used GEPIA for survival analysis and, for each candidate gene, stratified patients by expression level (high vs. low). In NSCLC overall, THBS3 expression was not associated with differences in overall survival (OS). However, among LUSC patients, those in the low-THBS3 expression group had longer disease-free survival (DFS). In contrast, LUSC patients with high SERPING1 expression had shorter OS. We observed no significant differences in survival between NSCLC patients with high versus low HIST1H4H or FLT4 expression. For LUSC patients stratified by CHEK1 expression, the high-expression group had better OS than the low-expression group. GEPIA, which is based on TCGA and GTEx data, does not contain survival data for SCLC patients (Figure S3). Protein-protein interactions and assessment of druggability PPI analysis revealed that proteins encoded by the potentially druggable genes under investigation do not exhibit interactions with each other, demonstrating their independence (Figure S4). We conducted pharmacological assessments on the co-localized target genes associated with lung cancer. Our analysis revealed that targeted drug development has been pursued for five of these genes (CHEK1, FLT4, SERPING1, THBS3, PSMA4). However, the protein encoded by HIST1H4H has not been incorporated into the STRING database. Notably, the protein encoded by FLT4 is linked to lymphatic vessel generation. Currently, drugs targeting the VEGF pathway, such as sunitinib and nintedanib, have been utilized in the treatment of NSCLC. For CHEK1, most targeted drugs function as inhibitors, and there is no ongoing research on gene activators or analogs. In addition, only a limited number of drugs or components act on plasma protease C1 inhibitor, as summarized in Supplementary Table S9. Genetic diagnostic analysis To determine whether the gene expression levels of the previously screened lung cancer-related genes (PP.H4 > 0.8) could be used as markers for lung cancer diagnosis, we assessed the predictability of the genes for lung cancer by using the receiver operating characteristic (ROC) curve and their respective areas under the curve (AUC). Gene expression levels of the tested genes can be found in Table S10. As shown in Fig. 4, all four genes had AUC values above 0.5. CHEK1 had the highest AUC (0.997) for LUSC, indicating excellent discriminative ability. Secondly, the FLT4 gene expression level was also adequate for the prediction of LUAD, with an AUC = 0.867. Furthermore, for the diagnostic prediction of NSCLC, the AUCs of SERPING1 and THBS3 were 0.967 and 0.697, respectively, which indicated that the diagnostic prediction of NSCLC by SERPING1 was better than that of THBS3. Fig. 4Diagnostic performance of prioritized level 1 target genes for lung cancer. A ROC curve for THBS3 to predict NSCLC. B ROC curve for SERPING1 to predict NSCLC. C ROC curve for CHEK1 to predict LUSC. D ROC curve for FLT4 to predict LUAD. AUC, area under curve; CI, confidence interval; FPR, false positive rate; TPR, true positive rate; ROC, receiver operating characteristic. ROC, receiver operating characteristic; AUC, area under the ROC curve; NSCLC, non-small-cell lung cancer; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; TPR, true-positive rate; FPR, false-positive rate; CI, confidence interval Diagnostic performance of prioritized level 1 target genes for lung cancer. A ROC curve for THBS3 to predict NSCLC. B ROC curve for SERPING1 to predict NSCLC. C ROC curve for CHEK1 to predict LUSC. D ROC curve for FLT4 to predict LUAD. AUC, area under curve; CI, confidence interval; FPR, false positive rate; TPR, true positive rate; ROC, receiver operating characteristic. ROC, receiver operating characteristic; AUC, area under the ROC curve; NSCLC, non-small-cell lung cancer; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; TPR, true-positive rate; FPR, false-positive rate; CI, confidence interval Discussion In our study, we identified four druggable genes (THBS3, SERPING1, CHEK1, and FLT4) strongly associated with lung cancer, with significant evidence of genetic co-localization. THBS3 and SERPING1 were identified as risk factors for NSCLC. The THBS3 gene encodes thrombospondin-3 (TSP-3) [33], which is involved in cell-cell and cell-matrix interactions, as well as associated with platelet function and angiogenesis [34]. Previous studies have shown that other thrombospondins, such as TSP-1 and TSP-4, are significantly associated with tumors and may serve as potential therapeutic targets. In particular, TSP-4 is known to promote angiogenesis [35] and is associated with tumor cell invasive properties [36]. In our previous study, the results of multiple analyses showed that elevated THBS3 expression was associated with poor overall survival in clear cell renal cell carcinoma [37]. However, the specific functions and regulatory mechanisms of TSP-3 remain unclear, and further exploration of its role in disease onset and progression is needed. Notably, no targeted drugs against THBS3 are listed in DrugBank database, highlighting a lack of approved interventions for this protein. Thus, further preclinical exploration is needed to develop potential therapies targeting THBS3. The SERPING1 gene, encoding plasma protease C1 inhibitor, is also strongly associated with NSCLC. Plasma protease C1 inhibitor can affect the angiogenic process by regulating the activity of angiogenic factors, and also promotes vascular-cell permeability, thus enhancing the nutrition of tumor cells, promoting tumor blood supply and growth [38]. Additionally, it has been implicated in promoting epithelial-mesenchymal transition in NSCLC cells [39]. Previous studies have also noted aberrant SERPING1 expression in lung cancer, with elevated levels observed in recurrent and metastatic NSCLC patients, correlating with poor NSCLC prognosis [40]. However, these studies did not explicitly establish a causal relationship between SERPING1 expression and NSCLC risk. Our work expands upon these findings by identifying SNPs associated with SERPING1 expression, revealing a positive causal correlation with NSCLC risk. In addition, the AUC of SERPING1 in the ROC curve for NSCLC in the diagnostic analysis was 0.967, indicating that SERPING1 may serve as a highly promising diagnostic and potential therapeutic biomarker for NSCLC. Furthermore, our systematic search of the DrugBank database uncovered investigational drugs such as PPL-100, which inhibit SERPING1 and are primarily studied for HIV infection and treatment. However, there is limited research on the therapeutic effects of PPL-100 for NSCLC, and the potential benefits of PPL-100 for NSCLC remain to be validated [41]. CHEK1 was found to be a protective factor for LUSC, encoding checkpoint kinase 1 (CHK1) of the serine/threonine protein kinase family [42]. CHK1, a critical protein kinase in the DNA damage response (DDR), plays a pivotal role in maintaining genetic stability within cells [43]. Previous studies have proposed CHEK1 as a tumor suppressor gene. Research by Andrés J. López-Contrera suggests that CHK1 protects the organism’s genome from mutations, thereby resisting cancer [44]. However, other studies suggest that CHK1 may promote tumor growth. Overexpression of CHEK1 has been observed in various tumors, showing positive correlations with tumor grade and disease recurrence in malignancies such as breast cancer, liver cancer, and cervical cancer [45–47]. This suggests that CHEK1 may not function strictly as a canonical tumor suppressor gene (Figure S5). Our findings revealed a negative correlation between the genetically predicted expression levels of the CHEK1 gene and LUSC, aligning with Andrés J. López-Contrera’s findings. Importantly, MR analysis using pQTLs showed results consistent with those of eQTLs, further providing genetic evidence that the CHEK1 gene is druggable. Currently, most clinical drugs targeting genes like CHEK1 are inhibitors, as developing gene-activating drugs is more challenging. Drugs that inhibit CHEK1 can be found in the DrugBank database, such as PF-477,736 and LY-2,608,204 [48]. However, direct targeted activation drugs for the CHEK1 gene are currently unavailable and present substantial development challenges. FLT4, a gene strongly associated with LUAD identified in this study, encodes VEGFR3. Previous studies have shown that VEGFR-3 activation through binding VEGF-C promotes lymphangiogenesis [49, 50], enhances cancer invasiveness, and induces epithelial-mesenchymal transition in NSCLC cells [51]. Our MR study further clarifies the causal relationship between FLT4 and NSCLC. Notably, in our TCGA/GTEx-based analysis, FLT4 expression appeared lower in LUAD tumors than in normal tissues, whereas previous studies have reported increased FLT4 expression in NSCLC [52]. This discrepancy may reflect differences in histological composition, sample size, or analytical pipelines across cohorts and warrants further experimental validation. In this study, we also identified several FDA-approved drugs targeting VEGFR-3 in the DrugBank database, including Tivozanib, Sunitinib and Sorafenib. While these drugs are primarily used for treating digestive tract tumors and renal cell carcinoma, their efficacy in lung cancer treatment has yet to be fully validated. Our findings provide a rationale for clinical trials to explore the potential of VEGFR-3 inhibitors for LUAD treatment. One major strength of this study is the use of MR and co-localization analyses based on human genetic variation to estimate the causal effects of genes on lung cancer, thereby minimizing bias from confounding and reverse causation. Previous MR studies have mainly focused on validating drug targets for cardiovascular diseases, where genetic evidence for target proteins has been linked to higher success rates in clinical drug development [12–15]. In oncology, several MR analyses have explored the causal relevance of circulating proteins, risk factors or pathways for specific cancers, but few have systematically screened the druggable genome for lung cancer while integrating multi-omics evidence. Our work extends this literature by combining cis-eQTL and cis-pQTL instruments for druggable genes with lung cancer GWAS data, and by incorporating Bayesian co-localization, proteome-wide MR, differential expression analyses and survival analyses in a single framework. This multi-layered approach provides convergent human genetic and functional support for the prioritization of lung cancer drug targets, complementing previous single-layer MR studies in other disease areas. Secondly, many circulating proteins related to gene expression did not meet the threshold required for MR analysis, failing to yield enough effective instrumental variables to represent proteins. Therefore, our study fills this gap. Thirdly, our diagnostic analyses resulted in high predictive accuracy with AUCs greater than 0.8 for three genes. There are several limitations worth noting in this study. First, the lung cancer data from the ILCCO cohort lacks stratification based on smoking, which hinders further analysis. Second, our study population primarily comprises individuals of European descent. Differences in allele frequencies, linkage disequilibrium structures and environmental exposures across ancestries may influence both instrument strength and causal effect estimates. Therefore, future studies should aim to replicate and extend our analyses in non-European and multi-ethnic cohorts to assess the generalizability of the prioritized targets. Third, our differential expression and survival analyses rely on TCGA/GTEx-based resources accessed via UALCAN and GEPIA, and we did not re-analyse raw RNA-seq count data using methods such as DESeq2 or limma-voom in independent cohorts. These expression analyses should therefore be viewed as supportive evidence rather than definitive validation, and future studies with additional transcriptomic datasets are warranted. Finally, among the identified four risk genes strongly associated with lung cancer susceptibility, some present challenges in drug development. For instance, THBS3 currently lacks specific targeted medications, while CHEK1 only has inhibitors, lacking necessary activators and active analogs. Although agents targeting FLT4 and the VEGF pathway exist, they predominantly inhibit angiogenesis, whereas specific therapies directed at VEGFR-3-mediated lymphangiogenesis remain limited. Additionally, although Bonferroni correction effectively controls the family-wise error rate in the presence of multiple testing, its stringency may also reduce power and lead to the omission of some true associations. Finally, the statistical power for CHEK1 (77%) was slightly below the conventional threshold of 80%, which may be related to the limited number of SNPs available as instruments for this gene. Therefore, we advocate for further foundational drug research and collaborative efforts among researchers to develop novel drugs for the high-risk genes uncovered in this study. Such initiatives may ultimately lead to significant advancements in the management and treatment of advanced-stage lung cancer. Conclusion This study integrated MR and co-localization analyses to identify several potentially druggable genes associated with lung cancer risk. By incorporating evidence from multiple drug databases, we further prioritized these genes as candidate therapeutic targets. Consistent MR estimates and supporting ancillary analyses, including co-localization, indicated that elevated FLT4 expression increases LUAD risk, whereas higher CHEK1 expression exerts a protective effect in LUSC. In addition, increased SERPING1 and THBS3 expression appeared to promote the progression of NSCLC. These genes may serve as candidate biomarkers for lung cancer and could provide guidance for future studies on diagnosis and treatment. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2.