Next Article in Journal
Zinc Ferrite-Integrated Halloysite Nanotubes as a Platform for Folate-Mediated Targeted Cisplatin Delivery
Previous Article in Journal
Shared and Divergent Transcriptional Programs of Oligodendrocyte Differentiation Across Vertebrate Species Revealed by scRNA-seq Analysis
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
IJMS
Volume 27
Issue 10
10.3390/ijms27104282
ijms-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Machine Learning-Based Identification of Biomarkers for Early-Stage Non-Small Cell Lung Cancer Through Gene Expression Analysis

by
Zorka Szollár
1,†,
Fanni Dzsubák
1,†,
Ádám Ürmös
1,
Barbara N. Borsos
1,2,
Balázs Bende
1,
Zoltán G. Páhi
1,2,* and
Tibor Pankotai
1,2,*
1
Hungarian Centre of Excellence for Molecular Medicine (HCEMM), Genome Integrity and DNA Repair Core Group, 6728 Szeged, Hungary
2
Competence Centre of the Life Sciences Cluster of the Centre of Excellence for Interdisciplinary Research, Development and Innovation, University of Szeged, 6720 Szeged, Hungary
*
Authors to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Int. J. Mol. Sci. 2026, 27(10), 4282; https://doi.org/10.3390/ijms27104282 (registering DOI)
Submission received: 15 April 2026 / Revised: 5 May 2026 / Accepted: 6 May 2026 / Published: 11 May 2026
(This article belongs to the Section Molecular Oncology)
Browse Figures
Versions Notes

Abstract

Tumor progression is primarily driven by DNA mutations; however, this mechanism alone does not fully account for all aspects of tumor development. Beyond genetic alterations, epigenetic changes also significantly influence the mutational landscape and affect gene expression without altering the DNA sequence. To gain a more comprehensive understanding of these regulatory mechanisms, it is essential to analyze gene expression at the transcriptional level. In this study, we examined non-small cell lung cancer (NSCLC) samples to identify specific gene expression changes, particularly in early-stage tumors. We conducted a bioinformatic analysis of RNA-sequencing data, followed by validation using an independent dataset from The Cancer Genome Atlas. Our analysis revealed a set of differentially expressed genes, seven of which were validated in patient-derived samples. Among these genes, EFNA4 and TEDC2 were significantly upregulated, whereas CDC42EP2, STX11, THBD, TMEM88, and GPM6A were notably downregulated in tumor tissues compared with adjacent normal tissues. Our findings highlight a distinct gene expression signature that differentiates NSCLC samples from normal lung tissues at the transcriptional level. These results underscore the potential of transcriptomic profiling as a promising tool for early-stage cancer detection and biomarker discovery.

1. Introduction

Lung cancer is the leading cause of cancer-related deaths worldwide, with non-small cell lung cancer (NSCLC) accounting for 85% of cases [1]. Despite available interventions such as surgery or therapies, including radiotherapy, chemotherapy, and targeted therapy, the prognosis for many patients remains poor, with a five-year survival rate of only 23% [2]. Many early-stage patients who undergo surgery still face risks of recurrence or metastases, which often necessitate systemic treatment [3].
Advancements in DNA-based diagnostics have improved cancer treatment by focusing on mutation detection, though this represents only part of the disease’s complexity [4,5]. Testing for early and locally advanced NSCLC includes blood tests, medical imaging, pulmonary function tests, and biomarker testing. These evaluations often assess programmed death-ligand 1 (PD-L1) levels, epidermal growth factor receptor (EGFR) mutations, and anaplastic lymphoma kinase (ALK) rearrangements [6].
The National Comprehensive Cancer Network guidelines recommend broad-panel next-generation sequencing (NGS) as the primary method for identifying mutations associated with NSCLC. If no driver oncogene mutations are detected, RNA sequencing (RNA-seq) can be used to identify gene fusions [7]. EGFR mutations and ALK rearrangements can be detected with polymerase chain reaction (PCR) or NGS and fluorescence in situ hybridization or immunohistochemistry, respectively [8,9]. Besides EGFR and ALK rearrangements, NSCLC is driven by diverse mutations, each affecting prognosis and treatment. Among these, TP53 and Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations are associated with poor survival, whereas EGFR alterations are amenable to multiple targeted therapies [10,11]. Additionally, ALK and ROS proto-oncogene 1, receptor tyrosine kinase (ROS1) rearrangements primarily occur in younger, non-smoking patients; however, the available inhibitors often face resistance [12]. The amplification of the MET proto-oncogene, receptor tyrosine kinase (MET), which is linked to aggressive behavior, can be treated with specific drugs [13]. Phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha (PIK3CA) mutations are common in squamous carcinoma. B-Raf proto-oncogene, serine/threonine kinase (BRAF) V600E and human epidermal growth factor receptor 2 (HER2) alterations, though subtle, have significant clinical implications [14,15].
Lung cancer progression is driven by genetic mutations and/or epigenetic modifications, leading to intratumoral heterogeneity by promoting lineage plasticity. Epigenetic dysregulation contributes to tumor diversity and resistance to targeted therapies [16]. These alterations play a critical role in hallmark cancer processes, including uncontrolled proliferation, resistance to apoptosis, angiogenesis, and metastasis [17]. Lung tumors often exhibit global DNA hypomethylation, leading to genomic instability, along with hypermethylation of CpG islands, which silences tumor suppressor genes and promotes oncogenesis; consequently, this manifests in transcriptomic dysregulation [18]. Epigenetic changes significantly impact transcription activity and gene expression. RNA analysis may reveal these shifts in expression, providing insight into how epigenetic modifications affect cellular function and therapeutic response [19]. While tissue biopsies remain the gold standard for genotyping, plasma-based circulating tumor DNA (ctDNA) liquid biopsies complement tissue testing, providing a more comprehensive analysis [20,21]. Standard diagnostics and therapies offer essential benefits but also face limitations. By incorporating RNA analysis and liquid biopsies for ctDNA analysis, a minimally invasive alternative to traditional tissue biopsies can be developed. Integrating ctDNA analysis with multi-omics approaches holds promise for improved cancer management [21,22].
In this study, we investigate the role of transcriptomic changes in NSCLC, with a particular focus on early-stage disease. By leveraging RNA-seq data and bioinformatics analysis, we have identified differentially expressed genes (DEGs) that can serve as potential biomarkers for early detection of NSCLC. Our results have been validated using an independent dataset from The Cancer Genome Atlas (TCGA) and confirmed in patient samples with RT–qPCR analysis. Our findings reveal a distinct gene expression signature that distinguishes NSCLC from normal tissue, underscoring the importance of transcriptional profiling in understanding tumor development. Notably, we have identified seven genes with significant expression alterations: two upregulated genes—ephrin A4 (EFNA4) and tubulin epsilon and delta complex 2 (TEDC2)—and five downregulated genes—CDC42 effector protein 2 (CDC42EP2), syntaxin 11 (STX11), thrombomodulin (THBD), transmembrane protein 88 (TMEM88), and glycoprotein M6A (GPM6A). These genes may play crucial roles in the progression of NSCLC and can serve as biomarkers for early-stage detection.

2. Results

Jeong-Sun Seo et al. conducted the first large-scale RNA-seq study of lung adenocarcinoma, providing valuable insights into its molecular profile. They comprehensively profiled lung adenocarcinoma across all stages, identifying key driver somatic mutations, such as EGFR, KRAS, neuroblastoma rat sarcoma viral oncogene (NRAS), BRAF, PIK3CA, MET, and catenin beta 1 (CTNNB1). Moreover, they discovered potential new drivers, including lemur tyrosine kinase 2 (LMTK2), AT-rich interacting domain 1A (ARID1A), neurogenic locus notch homolog protein 2 (NOTCH2), and SWI/SNF-related, matrix-associated, actin-dependent regulator of chromatin, subfamily a, member 4 (SMARCA4). In contrast, our study focused specifically on the transcriptomic profile during the early stages of this malignancy [23] (Figure 1).
We analyzed transcriptomic data from 31 stage 1A and 24 stage 1B tumor samples, along with their matched adjacent normal pairs. All specimens were obtained as fresh surgical samples. Among the 30 male and 25 female participants, 27 individuals were smokers. Eight tumor samples lacked matched adjacent normal samples and were therefore excluded from this study. Principal Component Analysis (PCA) was performed to assess potential outliers among the samples (Figure 2A). During this preliminary quality control, two sample pairs were excluded due to their outlier nature.
Using RNA-seq data from samples that met the inclusion criteria, we utilized the edgeR package, which is suitable for paired RNA-seq experiments, to identify genes with altered expression between tumor and adjacent normal samples. Our differential gene expression analysis identified 3421 DEGs, out of 19,886 analyzed genes. Significance was defined by an adjusted p-value of <0.05 and an absolute log2 fold change > 1 (Figure 2B,C).
To confirm the biological relevance of these identified genes in relation to NSCLC, multiple pathway analyses, including the KEGG, GO, and Reactome databases, were performed. The GSEA using the KEGG database highlighted that the most significantly activated pathways were those associated with DNA replication, homologous recombination, and cell cycle (Figure 2D, Supplementary Figure S1A–E). These pathways are involved in various cellular processes such as maintaining genomic integrity, coordinating accurate DNA synthesis, and regulating checkpoint control. These processes are often disrupted in various tumors, leading to the development of targeted therapies that specifically inhibit proteins involved in these pathways. The most suppressed pathways are the calcium and the MAPK signaling pathways. The downregulation of these pathways contributes to the regulation of proliferation, differentiation, and apoptosis. Several targeted therapies have been developed to selectively inhibit the proteins involved in these suppressed circuits. Over-representation analysis (ORA) based on the GO database revealed significant over-representation of pathways associated with angiogenesis, extracellular matrix organization, and epithelial cell proliferation. Dysregulation of these processes promotes tumor progression by enhancing extracellular matrix remodeling, facilitating invasion and metastasis, and supporting neovascularization required for sustained tumor growth. (Figure 2E, Supplementary Figure S1F–H).
Transcripts per million (TPM) values were calculated for sample normalization. TPM was chosen for its ability to normalize both gene length and sequencing depth, enabling accurate comparisons across samples. Utilizing the 3424 genes that were significantly differentially expressed and their TPM values, we built a Random Forest model to identify a gene expression pattern that effectively distinguishes tumor samples from normal samples with high accuracy. Random Forest was selected for this purpose due to its superior performance compared to other supervised learning models we evaluated.
In a 10-iteration loop, we randomly split the dataset into training (80%) and testing (20%) sets. Furthermore, a 10-fold cross-validation was implemented. By running the model multiple times, we aimed to minimize the impact of randomization and enhance its generalizability. The optimal number of trees (ntree) was set to 200, which proved to be the best-performing configuration. The number of variables randomly selected at each node was determined using a common practice: the square root of 3424 and the total number of genes in the model. Across 10 iterations, the model achieved an average accuracy of 97.8%, indicating it can classify a sample from the testing set as a tumor or a normal sample with 97.8% accuracy. To identify the genes that contributed most to the classification, we used the importance function. From the ranked list of genes provided by this function, we selected the top 10 genes for further analysis: STX11, RADIL, THBD, TEDC2, SERTM1, CDC42EP2, TMEM88, RASAL1, GPM6A, and EFNA4 (Figure 3A,B).
To validate our results, we constructed an independent Random Forest model using the top 10 genes and their TPM values as the training set. We employed a TCGA dataset as a testing set to determine whether the gene expression patterns of these 10 genes were identified in an independent secondary dataset. By using the same model parameters as in the first model, this new model achieved an average accuracy of 95.52%, demonstrating strong performance in distinguishing tumors from normal samples (Figure 3C). The biological relevance of these 10 genes and the pathways in which they are involved are illustrated in Figure 3D. Collectively, this finding further confirms the potential of our gene expression pattern as a robust and reliable diagnostic marker for early-stage NSCLC. The high level of accordance between our internal dataset and the TCGA data strongly supports the validity and generalizability of our findings (Figure 4, Supplementary Figure S2).
To further validate our results, we performed RT–qPCR measurements on tumor and adjacent normal tissue samples from human subjects. From the genes identified in the NGS analysis, we selected 7 genes (i.e., CDC42EP2, STX11, THBD, TMEM88, GPM6A, EFNA4, and TEDC2) based on differential expression between normal and tumor samples, prioritizing those with the most significant expression changes. To ensure the robustness of our findings, we established a well-defined patient cohort, comprising 24 individuals diagnosed with NSCLC (Table 1). Frozen tissue samples were collected from both the tumor and adjacent normal tissues from the same patients. An expert pathologist evaluated the tumor samples to maintain sample integrity and ensure an accurate representation of tumor biology. Only samples with at least 30% tumor content were included in the analysis.
We quantified RNA expression levels across these samples using RT–qPCR, calculated the mean expression values, and visualized the average expression levels (Figure 5). The RT–qPCR results demonstrated a strong correlation with the NGS data, reinforcing the reliability of our transcriptomic analysis. Specifically, five genes—CDC42EP2, STX11, THBD, TMEM88, and GPM6A—showed marked downregulation in tumor tissue compared with matched adjacent normal samples (Figure 5A). This downregulation may indicate a potential tumor suppressor function, as reduced expression levels can contribute to impaired cellular homeostasis, loss of cell adhesion, or enhanced metastatic potential (Figure 3D). In contrast, two genes, EFNA4 and TEDC2, exhibited significantly elevated mRNA levels in tumor samples (Figure 5B), suggesting their potential role in tumor progression or oncogenic signaling pathways.
To assess the discrimination potential of the combined expression of seven genes (CDC42EP2, EFNA4, GPM6A, STX11, TEDC2, THBD, and TMEM88), we applied a generalized linear model (GLM)-based approach similar to that implemented in the combiROC R package. Our results show that the combined Receiver Operating Characteristic (ROC) curve derived from sequencing data achieved a higher area under the curve (AUC) than that from RT–qPCR data. Importantly, two independent validation datasets were used, including the RT–qPCR dataset, which serves as biological validation. Although differences in performance metrics across datasets were expected, the combined ROC analysis of seven genes consistently yielded a well-separated classification model across all three cohorts (i.e., ENA, TCGA, and RT–qPCR).
These findings align with existing literature, further supporting the involvement of these genes in lung cancer biology. The observed gene expression patterns across all three independent cohorts suggest that this gene set could serve as a potential diagnostic or prognostic biomarker panel for NSCLC. However, further functional validation, mechanistic exploration of these gene candidates, and assessment of their potential utility in clinical applications, such as patient stratification and therapeutic targeting, still need to be elucidated.

3. Discussion

Tumor progression is a highly complex process that primarily involves transcriptomic reprogramming in cells [16]. Epigenetic and transcriptomic changes are increasingly recognized as early indicators of tumor development. To explore whether tumor and normal tissues in early-stage NSCLC can be robustly distinguished based on these changes, we conducted integrated bioinformatic and experimental analyses. Our findings reveal a seven-gene expression signature that significantly differentiates early-stage NSCLC tumors from adjacent normal tissues. This signature was validated across three independent cohorts and confirmed with RT–qPCR in a Hungarian cohort, underscoring its reliability and translational potential. Recent studies further support the concept that transcriptomic dysregulation frequently precedes overt genomic instability during early lung tumorigenesis. While mutation-based diagnostics remain central in NSCLC management, increasing evidence indicates that epigenetic alterations, including DNA methylation changes, chromatin remodeling, and histone modification patterns, can drive transcriptional reprogramming at very early stages of malignant transformation. Baumann et al. have recently described these alterations as “epigenomic echoes,” emphasizing that early-stage NSCLC often exhibits stable transcriptomic disturbances even before the accumulation of dominant driver mutations [19]. Similarly, Marei highlighted that epigenetic regulators shape therapeutic resistance and lineage plasticity by maintaining aberrant transcriptional states rather than solely inducing DNA mutations [16].
This concept strongly supports our analytical strategy, which focuses on transcript-level alterations rather than predefined mutational panels. Unlike DNA-based testing that relies on detecting specific driver alterations such as EGFR, ALK, or KRAS, transcriptomic profiling captures the functional output of both genetic and epigenetic dysregulation. This process is particularly relevant in early-stage tumors, where mutational burden may remain low, and ctDNA sensitivity is limited. The observed dysregulation of genes involved in angiogenesis, extracellular matrix remodeling, and Wnt/β-catenin signaling in our dataset likely reflects these upstream epigenetic disturbances. Therefore, our seven-gene signature may represent not only diagnostic markers but also transcriptional readouts of early oncogenic reprogramming.
Our study offers several improvements over existing diagnostic modalities for early-stage NSCLC. Currently, low-dose CT screening remains the gold standard for early detection, yet it suffers from high false-positive rates and lacks molecular resolution [4]. Tissue- or plasma-based genomic profiling—such as testing for EGFR mutations, ALK/ROS1 fusions, and PD-L1 expression—is routinely employed for therapeutic guidance, but not for early detection. Although ctDNA is a promising minimally invasive tool, its efficacy is often limited in early-stage disease due to low tumor burden. In contrast, our RNA-seq-based transcriptomic profiling captures gene expression shifts that may occur before significant genomic instability and detectable mutations arise. This method provides a complementary and potentially more sensitive approach for early diagnosis. Importantly, our pipeline does not rely on predefined mutation panels; instead, it enables data-driven discovery of dysregulated gene signatures that reflect early malignant transformation.
To identify molecular features that distinguish tumors from normal tissue, we applied a machine learning-based feature selection pipeline. Using a Random Forest classifier, we ranked genes based on their contribution to classification accuracy. Subsequently, we validated the top candidates based on their biological relevance and expression patterns. Among the most informative DEGs, EFNA4 and TEDC2 were significantly upregulated in tumor tissues. EFNA4, a member of the ephrin family involved in cell signaling, has been linked to tumor progression and angiogenesis in various cancers, including lung cancer [24]. EFNA4 upregulation may reflect enhanced tumor cell communication, migration, and neovascularization in early tumorigenesis, and it also promotes lymph node metastasis in later stages [24,25]. TEDC2 has recently been shown to be implicated in cell cycle regulation, and higher expression levels have been associated with a worse prognosis [26]. On the other hand, CDC42EP2, STX11, THBD, TMEM88, and GPM6A were markedly downregulated in tumor samples. These genes have diverse biological roles, including vesicular transport (STX11) and immune signaling (THBD) [27,28]. Their suppression in early-stage tumors may indicate loss of normal tissue architecture, immune evasion, or altered differentiation pathways—hallmarks of malignant transformation.
Recent studies further support the importance of transcriptional regulation and epithelial remodeling as central mechanisms in early lung tumorigenesis. Beyond classical driver mutations, disruption of transcriptional control and RNA-processing pathways can promote epithelial–mesenchymal transition, immune escape, and metastatic competence through coordinated reprogramming of gene expression [29,30]. RNA-binding proteins have emerged as critical regulators of these processes. For example, recent work has demonstrated that RBMX functions as an important regulator of transcriptional stability and epithelial plasticity, contributing to tumor-associated signaling and progression in lung cancer models [31]. Similarly, studies investigating infection-associated lung disease have reported that host transcriptional regulators and RNA-binding proteins coordinate epithelial remodeling and inflammatory signaling, thereby providing mechanistic links between transcriptional dysregulation and disease progression [32]. These findings are highly relevant to our results, particularly for genes such as TMEM88, CDC42EP2, and THBD, whose downregulation may reflect disruption of epithelial integrity and altered signaling pathways rather than isolated gene-specific effects. Therefore, the identified seven-gene signature may represent a broader transcriptional reprogramming event associated with early malignant transformation rather than simply a collection of independent biomarkers. For instance, TMEM88 is known to negatively regulate the Wnt/β-catenin signaling pathway, a key driver of NSCLC progression; its downregulation may facilitate tumorigenesis and metastasis [33]. Importantly, while some of these genes have been implicated in lung cancer, others—such as TEDC2 and CDC42EP2—remain poorly characterized in this context, suggesting that our approach has revealed previously unrecognized components of early tumor biology.
Because of its exclusive focus on early-stage (i.e., stage 1) disease and its incorporation of machine learning-based feature selection across multiple datasets, our study complements prior RNA-seq efforts, such as the comprehensive lung adenocarcinoma transcriptome reported by Seo and colleagues [23]. While Seo’s dataset spanned all tumor stages and emphasized mutation discovery, our approach directly addresses the need for molecular stratification at early stages using gene expression as the central feature. Our Random Forest model achieved a classification accuracy of over 95% in both internal and TCGA-derived testing datasets, indicating the diagnostic discriminative capacity of the identified gene set. Furthermore, by validating these findings across ethnically diverse cohorts and Hungarian clinical samples, we support the generalizability of the signature.
From a translational perspective, the reproducibility of this gene set across populations with different nationalities suggests it may serve as a ubiquitous molecular marker, given the well-documented genomic variability of NSCLC worldwide. Previous studies have emphasized substantial heterogeneity across NSCLC subtypes and populations, with significant differences in the prevalence of driver mutations, gene fusions, and transcriptomic profiles—largely influenced by ancestry, sex, and smoking status [34,35]. Although the observed consistency of this transcriptomic pattern is promising, the broader applicability requires larger, prospective validation. Moreover, transcriptomic consistency often correlates with core regulatory functions, suggesting that these genes might also serve as early-stage therapeutic targets, either alone or in combination with existing pathway inhibitors. However, these data remain exploratory and require further functional validation.
An important future direction is to integrate this transcriptomic signature with circulating biomarkers and broader multi-omics platforms. Recent studies have demonstrated that isolated single-omics approaches often fail to capture the biological complexity of NSCLC, whereas integrative strategies that combine genomics, transcriptomics, proteomics, and liquid biopsy-derived biomarkers significantly improve diagnostic precision [5]. Chen et al. showed that combining ctDNA with RNA and protein biomarkers enhances sensitivity for early-stage cancer detection and improves patient stratification compared with mutation analysis alone [22]. Similarly, plasma-based assays incorporating exosomal RNA, circulating proteins, and methylation patterns have shown superior predictive performance compared to conventional ctDNA-only approaches.
Our current findings fit well into this emerging framework. The seven-gene panel identified here can be translated into minimally invasive platforms using circulating exosomal RNA or plasma-derived transcriptomic signatures. The use of these signatures would be particularly valuable for stage 1 NSCLC patients, in whom tissue access is often limited, and ctDNA levels are frequently below reliable detection thresholds. Moreover, integration with proteomic validation can further prioritize biologically actionable targets, particularly for genes such as EFNA4 and TEDC2, where transcriptional upregulation may directly reflect functional pathway activation [24,26,33]. Such multi-layer validation would substantially increase the clinical applicability of these biomarkers and improve their potential for future diagnostic implementation.
Nonetheless, our study also has limitations. Several genes within the expression signature remain poorly characterized, necessitating functional validation via in vitro and in vivo assays. Moreover, although the RT–qPCR cohort included 20 patients, future studies should expand validation to larger and more diverse populations, including underrepresented ethnicities and patients at later stages of disease. Another important limitation of our study is the relatively small size of both the validation and clinical cohorts, which may influence the robustness and generalizability of the identified biomarker panel. Although the transcriptomic signature was consistently validated across the ENA dataset, the TCGA cohort, and our independent Hungarian patient samples, the absolute number of early-stage matched tumor–normal pairs remained limited, particularly in the RT–qPCR validation cohort. Small cohort sizes increase the risk of statistical overfitting, especially in machine learning-based feature selection approaches such as Random Forest analysis, where model performance may be artificially inflated if biological heterogeneity is underrepresented. In NSCLC, substantial variability exists across histological subtypes, smoking status, sex, ancestry, and molecular driver composition, all of which may significantly influence transcriptomic profiles. Our present cohort mainly consisted of early-stage adenocarcinoma cases; therefore, it may not fully represent the broader biological diversity of NSCLC, particularly squamous cell carcinoma or rare histological subtypes. Furthermore, prospective multicenter validation in larger, clinically heterogeneous cohorts will be necessary to determine the true diagnostic sensitivity, specificity, and reproducibility of this seven-gene panel under routine clinical conditions. Such validation will be particularly important before considering implementation in screening programs or liquid biopsy-based early detection platforms. It will also be important to assess whether the observed gene expression shifts are lung-specific or detectable in circulating compartments, such as exosomes. If validated, such expression markers can form the basis for minimally invasive, liquid biopsy-based diagnostics—especially valuable in settings where tissue sampling is limited or impossible.

4. Conclusions

In conclusion, our study identifies a seven-gene transcriptomic signature that effectively distinguishes early-stage NSCLC from adjacent normal tissue across three independent datasets. This gene panel holds promise not only as a diagnostic tool but also as a gateway to understanding the mechanisms underlying early lung cancer biology by combining RNA-seq-based expression profiling with machine learning-guided biomarker discovery and clinical sample validation. Our results provide a foundation for future studies aimed at supporting precision diagnostics in NSCLC. However, these tissue-based findings have to be translated into non- or minimally invasive modalities. Further research should investigate the utility of these markers in liquid biopsy formats, their specificity across different cancer types, and their potential as therapeutic targets in early intervention strategies.

5. Materials and Methods

5.1. Bioinformatic Analysis

In this study, the RNA-seq data analyzed were obtained from the European Nucleotide Archive (ENA) under Gene Expression Omnibus (GEO) accession GSE40419, corresponding to the ENA project PRJEB3132. Our inclusion criteria required matched tumor and normal samples from the same patient. We focused on early-stage samples, including 31 stage 1A and 24 stage 1B tumor samples, along with their matched adjacent normal samples. Eight tumor samples were excluded from the study cohort because they lacked accurately matched adjacent normal samples. Additionally, two sample pairs were excluded as outliers during preliminary quality control.
The initial analysis was performed in a Linux environment (Debian) using Bioconda-distributed tools. Quality assessment was performed with FastQC (version 0.12.1), followed by trimming with Trimmomatic (version 0.38) (SLIDINGWINDOW:4:20, MINLEN:20). After a second quality check with FastQC, alignment was performed with HISAT2 (version 2.2.1) using the GRCh38 reference genome. For quantification, FeatureCounts (version 0.12.1), a highly efficient read summarization tool, was employed [36,37,38].
Gene expression analysis was performed in R (version 4.3.0). The packages used for visualization, data handling, and utility included readr (version 2.1.5), dpylr (version 1.1.4), ggplot2 (version 3.5.1), and ggrepel (version 0.9.5). A pre-filtering step was implemented to identify and remove outliers. PCA was conducted using the FactoMineR (version 2.11) and factoextra (version 1.0.7) packages to visualize sample distances, leading to the exclusion of two additional sample pairs because the normal samples deviated significantly from the rest. Differential gene expression analysis was performed from raw counts using the edgeR package (version 3.42.4), identifying significant DEGs with adjusted p-values < 0.05 and absolute log2-fold changes > 1. Heatmaps were generated with the ComplexHeatmap package (version 2.16.0) to visualize the expression values of these genes [39,40,41,42,43,44].
For feature selection, a supervised machine learning model was built using the caret (version 6.0-94) and RandomForest (version 4.7-1.1) packages. The dataset used for the Random Forest model comprised 90 samples—45 tumors and 45 matched normals—along with normalized gene expression data. For the model, genes meeting the thresholds of |logFC| > 1 and nominal p-value < 0.05 were selected, resulting in 3424 genes. Since the primary goal of the model was classification, genes were filtered based on nominal p-values rather than adjusted p-values. This approach allowed the retention of additional genes that could provide useful discriminatory information, thereby improving predictive performance. Data were normalized using TPM [45].
The dataset was divided into training (80% of the data, 72 samples) and testing (20% of the data, 18 samples) sets while preserving the tumor–normal ratio. To ensure reproducibility, a fixed random seed was applied. R with 200 trees, and the mtry parameter was optimized using the square root of 3424, the number of input variables. All other settings were left at default values. Hyperparameter optimization was performed by 10-fold cross-validation. To evaluate model robustness, the entire workflow, including the data partitioning step, was repeated across 10 iterations, and the mean performance across these runs was reported.
Importance values were calculated using Mean Decrease in Accuracy, a permutation-based importance metric that is generally preferred for interpretation because it is directly tied to model prediction performance. No additional normalization or scaling was applied. Importance scores were extracted using the varImp() function with no scaling, meaning that raw importance values were used. To ensure robustness, the model employed 10-fold cross-validation, and the entire process was repeated across 10 independent iterations, where randomized data partitioning occurred each time with controlled random seeds. For each run, the decrease in accuracy was measured for each gene, and mean importance values were subsequently calculated by averaging these scores across runs. Genes were then ranked by their mean importance values, enabling the identification of those with high potential in distinguishing between classes.
For external validation, early-stage RNA-sequenced tumor–normal pairs were obtained from the TCGA–Lung Adenocarcinoma Dataset (LUAD) using the TCGAbiolinks package (version 2.30.0). This dataset included 29 stage 1 tumor samples with matched normal counterparts. The model was constructed using the same modeling parameters but with only the 10 highest-ranking genes, as identified by the first model’s average Mean Decrease in Accuracy (importance). In this case, the entire dataset was used for training, with an independent external TCGA dataset serving as the validation set.
Pathway enrichment analyses were also performed using the Gene Ontology (GO), Reactome, and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. Gene annotation was carried out utilizing the bioMart (version 2.58.0), AnnotationDbi (version 1.64.1), and org.HS.eg.db (version 3.17.0) packages. Gene set enrichment analysis (GSEA) and ORA were performed using clusterProfiler (version 4.8.3). The findings from the GSEA and ORA were visualized with enrichplot (version 1.20.3) [46,47,48,49].
To assess the joint discriminatory performance of the seven-gene panel, a combined ROC analysis was performed using a GLM-based approach, conceptually similar to the strategy implemented in the combiROC (version 0.3.4) R package. All seven genes (CDC42EP2, EFNA4, GPM6A, STX11, TEDC2, THBD, and TMEM88) were included as predictors in a logistic regression framework. To ensure robust estimation of model performance, a repeated random subsampling cross-validation procedure was applied. In each iteration, the dataset was randomly partitioned into training (80%) and testing (20%) subsets using stratified sampling to preserve the class distribution between tumor and normal samples. This process was repeated 20 times with a fixed random seed. For each iteration, the model was trained on the training subset and evaluated on the independent testing subset. Predicted probabilities were obtained for the testing set and used to construct ROC curves. The final ROC curve was generated by aggregating predictions across all cross-validation iterations. The analyses were performed in R. ROC analyses were conducted using the pROC package (version 1.19.0.1), data import was performed using readr (version 2.1.5), figures were generated using svglite (version 2.2.2), and results were exported using openxlsx (version 4.2.8.1) [2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52].
Statistical analyses of RT–qPCR data were conducted using IBM SPSS Statistics 29.0. The statistical significance of differences in expression values between normal and NSCLC samples was assessed using the independent-samples t-test.
The study was conducted in accordance with the Declaration of Helsinki and approved by the Scientific and Research Ethics Commission (TUKEB) of the Hungarian Medical Research Council (protocol code: BM/32453-3/2024; date of approval: 18 January 2025).

5.2. Cohort Characteristics

In the initial analysis, RNA-seq data from early-stage tumor samples of 55 patients were obtained from the ENA. Of these, 8 patients were excluded because no matched adjacent normal lung tissue samples were available. During PCA, an additional 2 sample pairs were removed because their corresponding normal samples clustered separately from the main group of normal tissues and were therefore considered outliers. Consequently, a total of 45 tumor samples and 45 matched adjacent normal RNA-seq samples were retained for downstream analyses, including DEG identification and Random Forest model development.
For validation, RNA-seq data from the TCGA–Lung Adenocarcinoma (LUAD) cohort were used. Specifically, 29 stage 1 tumor samples with their corresponding adjacent normal lung tissue RNA-seq samples were found and included as an independent validation dataset.
For RT–qPCR validation, a total of 24 patients, including 15 females and 9 males, diagnosed with lung cancer, were included in the study. The participants’ ages ranged from 50 to 85 years, with a mean age of 72 years. The cohort enrolled patients diagnosed with pT1a (n = 5), pT1b (n = 3), and pT1c (n = 16). Histopathological evaluation identified invasive solid adenocarcinoma as the predominant subtype (n = 9), invasive papillary adenocarcinoma (n = 5), invasive adenocarcinoma (n = 4), and mixed adenocarcinoma (n = 2). Less frequently observed subtypes included adenosquamous carcinoma (n = 1), invasive acinar adenocarcinoma (n = 1), mixed-type lung cancer (n = 1), and papillary adenocarcinoma (n = 1) (Table 1). However, due to extremely low RNA concentrations, four patients’ samples were excluded from further measurements and analyses. Therefore, the final RT–qPCR measurements were performed on 20 NSCLC and 20 matched adjacent healthy lung tissue samples derived from 20 patients.

5.3. RNA Isolation

From each patient sample, five and ten 5-micrometer slices were collected from the tumor and the corresponding normal tissue, respectively. Total RNA was isolated using the ReliaPrep RNA Cell Miniprep System Kit (Promega, Madison, WI, USA) according to the manufacturer’s instructions. The RNA concentrations were measured with a NanoDrop OneC spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA). Reverse transcription was carried out using TaqMan Reverse Transcription Reagents (Thermo Fisher Scientific, Waltham, MA, USA), according to the manufacturer’s instructions, with the following thermal profile: 25 °C for 10 min, 37 °C for 60 min, and 95 °C for 5 min. In each RT–qPCR reaction, an equal amount of cDNA was utilized.

5.4. RT–qPCR

For RT–qPCR evaluation, 20 patient samples were selected for further analysis. RT–qPCR reactions were conducted in a final volume of 10 µL using GoTaq qPCR Master Mix (Promega, Madison, WI, USA) with a QIAGEN Rotor-GeneQ 5-plex HRM qPCR System (Qiagen, Hilden, Germany). All RT–qPCR amplifications followed the same thermal profile: 95 °C for 7 min, 45 cycles of 95 °C for 15 s, and 60 °C for 30 s, followed by a melting curve analysis. The cycling conditions were optimized to achieve efficient amplification of the target gene while minimizing non-specific amplification. The primers, listed in Table 2, were designed using the Primer3 software (https://primer3.ut.ee/, accessed on 02-03-2025). The specificity of the primers was verified using NCBI BLAST (http://www.ncbi.nlm.nih.gov/tools/primer-blast/, version: BLAST+ 2.16.0, accessed on 02-03-2025). The primers were also tested to determine the most suitable concentrations for RT–qPCRs. All RT–qPCR reactions were performed in duplicate, and expression levels were determined based on the mean value of the duplicate measurements. For each primer pair, a no-template control (NTC) was included to detect any DNA contamination. The relative quantification method was applied in each case. In accordance with MIQE guidelines, Ct values of each sample were normalized to the geometric mean of two validated reference genes (TBP and Cyclophilin B). Relative expression levels were calculated using the ΔΔCt method, enabling accurate comparison across samples.

6. Patents

This work is submitted for patenting under the following accession number: European Patent Application No. 25181645.0 (June 2025).

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/ijms27104282/s1.

Author Contributions

Z.S. carried out the investigation, formal analysis, and visualization and contributed to writing the original draft as well as reviewing and editing the manuscript. F.D. performed the investigation, formal analysis, and visualization and contributed to writing the original draft and reviewing and editing the manuscript. Á.Ü. was involved in the investigation, visualization, and formal analysis. B.N.B. performed the investigation and visualization and was involved in reviewing and editing the manuscript. B.B. contributed to conceptualization and methodology. Z.G.P. was involved in conceptualization, methodology, and formal analysis and contributed to writing the original draft and reviewing and editing the manuscript. T.P. carried out conceptualization, methodology, formal analysis, writing of the original draft, and reviewing and editing of the manuscript and was responsible for funding acquisition. All authors have read and agreed to the published version of the manuscript.

Funding

The project was supported by the EU’s Horizon 2020 Research and Innovation Program [No. 739593 to TP]; the Ministry of Culture and Innovation of Hungary under the National Research, Development and Innovation Fund’s TKP2021-EGA funding scheme [TKP-2021-EGA-05 to TP]; and the Ministry of Culture and Innovation of Hungary under the National Research, Development and Innovation Fund’s 2022-2.1.1-NL funding scheme [2022-2.1.1-NL-2022-00005 to TP] and the University of Szeged Open Access Fund [Grant Nr. 8490 to TP]. This research work was conducted with the support of the National Academy of Scientist Education Program of the National Biomedical Foundation under the sponsorship of the Hungarian Ministry of Culture and Innovation [to TP]. The funders had no role in the study design, data collection, analysis, interpretation, or manuscript writing.

Institutional Review Board Statement

The study was conducted in accordance with the Declaration of Helsinki and approved by the Scientific and Research Ethics Commission (TUKEB) of the Hungarian Medical Research Council (protocol code: BM/32453-3/2024; date of approval: 18 January 2025).

Informed Consent Statement

Informed consent was obtained from all subjects involved in the study.

Data Availability Statement

The original contributions presented in this study are included in the article/Supplementary Materials. Further inquiries can be directed to the corresponding author(s).

Acknowledgments

We thank Manuéla Katona, who provided support in RT–qPCR measurements.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
ALKanaplastic lymphoma kinase
BRAFB-Raf proto-oncogene, serine/threonine kinase
CDC42EP2CDC42 effector protein 2
ctDNAcirculating tumor DNA
EFNA4ephrin A4
EGFRepidermal growth factor receptor
ENAEuropean Nucleotide Archive
GOGene Ontology
GPM6Aglycoprotein M6A
GSEAGene set enrichment analysis
HER2human epidermal growth factor receptor 2
KEGGKyoto Encyclopedia of Genes and Genomes
KRASKirsten rat sarcoma viral oncogene homolog
NGSnext-generation sequencing
NSCLCNon-small cell lung cancer
ORAover-representation analysis
PCAPrincipal Component Analysis
PCRpolymerase chain reaction
PD-L1programmed death-ligand 1
RNA-seq
ROC		RNA sequencing
Receiver Operating Characteristic
ROS1ROS proto-oncogene 1, receptor tyrosine kinase
RT–qPCR Reverse Transcription Quantitative Polymerase Chain Reaction
STX11syntaxin 11
TCGAThe Cancer Genome Atlas
TEDC2tubulin epsilon and delta complex 2
THBDthrombomodulin
TMEM88transmembrane protein 88
TPMtranscripts per million

References

  1. Tang, F.H.; Wong, H.Y.T.; Tsang, P.S.W.; Yau, M.; Tam, S.Y.; Law, L.; Yau, K.; Wong, J.; Farah, F.H.M.; Wong, J. Recent Advancements in Lung Cancer Research: A Narrative Review. Transl. Lung Cancer Res. 2025, 14, 975–990. [Google Scholar] [CrossRef]
  2. Lahiri, A.; Maji, A.; Potdar, P.D.; Singh, N.; Parikh, P.; Bisht, B.; Mukherjee, A.; Paul, M.K. Lung Cancer Immunotherapy: Progress, Pitfalls, and Promises. Mol. Cancer 2023, 22, 40. [Google Scholar] [CrossRef]
  3. Hendriksen, B.S.; Ortiz, B.A.; Wigle, D.A.; Cassivi, S.D.; Saddoughi, S.A.; Puig, C.A.; Shen, K.R.; Reisenauer, J.S.; Tapias, L.F. Risk Factors and Survival Associated with Lung Cancer Recurrence after Curative-Intent Surgery: Beyond TNM Pathological Staging. J. Thorac. Dis. 2025, 17, 10285–10297. [Google Scholar] [CrossRef] [PubMed]
  4. Pulumati, A.; Pulumati, A.; Dwarakanath, B.S.; Verma, A.; Papineni, R.V.L. Technological Advancements in Cancer Diagnostics: Improvements and Limitations. Cancer Rep. 2023, 6, e1764. [Google Scholar] [CrossRef] [PubMed]
  5. Pankotai-Bodó, G.; Oláh-Németh, O.; Sükösd, F.; Pankotai, T. Routine Molecular Applications and Recent Advances in Breast Cancer Diagnostics. J. Biotechnol. 2024, 380, 20–28. [Google Scholar] [CrossRef] [PubMed]
  6. Šutić, M.; Vukić, A.; Baranašić, J.; Försti, A.; Džubur, F.; Samaržija, M.; Jakopović, M.; Brčić, L.; Knežević, J. Diagnostic, Predictive, and Prognostic Biomarkers in Non-Small Cell Lung Cancer (NSCLC) Management. J. Pers. Med. 2021, 11, 1102. [Google Scholar] [CrossRef]
  7. Liu, S.V.; Nagasaka, M.; Atz, J.; Solca, F.; Müllauer, L. Oncogenic Gene Fusions in Cancer: From Biology to Therapy. Signal Transduct. Target. Ther. 2025, 10, 111. [Google Scholar] [CrossRef]
  8. Mayer, C.; Ofek, E.; Fridrich, D.E.; Molchanov, Y.; Yacobi, R.; Gazy, I.; Hayun, I.; Zalach, J.; Paz-Yaacov, N.; Barshack, I. Direct Identification of ALK and ROS1 Fusions in Non-Small Cell Lung Cancer from Hematoxylin and Eosin-Stained Slides Using Deep Learning Algorithms. Mod. Pathol. 2022, 35, 1882–1887. [Google Scholar] [CrossRef]
  9. Dalurzo, M.L.; Avilés-Salas, A.; Soares, F.A.; Hou, Y.; Li, Y.; Stroganova, A.; Öz, B.; Abdillah, A.; Wan, H.; Choi, Y.-L. Testing for EGFR Mutations and ALK Rearrangements in Advanced Non-Small-Cell Lung Cancer: Considerations for Countries in Emerging Markets. OncoTargets Ther. 2021, 14, 4671–4692. [Google Scholar] [CrossRef]
  10. Rubio-Perez, J.; Hernandez, R.; Santolaya, C.; Martin-Soberon, M.C.; Zazo, S.; Carvajal, N.; Rojo, F. New Therapeutic Approaches for EGFR Mutated Non-Small Cell Lung Cancer on Osimertinib Era. Cancer Treat. Res. Commun. 2025, 44, 100945. [Google Scholar] [CrossRef]
  11. Huang, L.; Guo, Z.; Wang, F.; Fu, L. KRAS Mutation: From Undruggable to Druggable in Cancer. Signal Transduct. Target. Ther. 2021, 6, 386. [Google Scholar] [CrossRef]
  12. Song, Z.; Zheng, Y.; Wang, X.; Su, H.; Zhang, Y.; Song, Y. ALK and ROS1 Rearrangements, Coexistence and Treatment in Epidermal Growth Factor Receptor-Wild Type Lung Adenocarcinoma: A Multicenter Study of 732 Cases. J. Thorac. Dis. 2017, 9, 3919–3926. [Google Scholar] [CrossRef] [PubMed]
  13. Lennerz, J.K.; Kwak, E.L.; Ackerman, A.; Michael, M.; Fox, S.B.; Bergethon, K.; Lauwers, G.Y.; Christensen, J.G.; Wilner, K.D.; Haber, D.A.; et al. MET Amplification Identifies a Small and Aggressive Subgroup of Esophagogastric Adenocarcinoma with Evidence of Responsiveness to Crizotinib. J. Clin. Oncol. 2011, 29, 4803–4810. [Google Scholar] [CrossRef] [PubMed]
  14. Yan, N.; Guo, S.; Zhang, H.; Zhang, Z.; Shen, S.; Li, X. BRAF-Mutated Non-Small Cell Lung Cancer: Current Treatment Status and Future Perspective. Front. Oncol. 2022, 12, 863043. [Google Scholar] [CrossRef] [PubMed]
  15. Scheffler, M.; Bos, M.; Gardizi, M.; König, K.; Michels, S.; Fassunke, J.; Heydt, C.; Künstlinger, H.; Ihle, M.; Ueckeroth, F.; et al. PIK3CA Mutations in Non-Small Cell Lung Cancer (NSCLC): Genetic Heterogeneity, Prognostic Impact and Incidence of Prior Malignancies. Oncotarget 2015, 6, 1315–1326. [Google Scholar] [CrossRef] [PubMed]
  16. Marei, H.E. Epigenetic Regulators in Cancer Therapy and Progression. npj Precis. Oncol. 2025, 9, 206. [Google Scholar] [CrossRef]
  17. Tufail, M.; Hu, J.-J.; Liang, J.; He, C.-Y.; Wan, W.-D.; Huang, Y.-Q.; Jiang, C.-H.; Wu, H.; Li, N. Hallmarks of Cancer Resistance. iScience 2024, 27, 109979. [Google Scholar] [CrossRef]
  18. Ehrlich, M. Dna Hypomethylation in Cancer Cells. Epigenomics 2009, 1, 239–259. [Google Scholar] [CrossRef]
  19. Baumann, A.A.; Buribayev, Z.; Wolkenhauer, O.; Salybekov, A.A.; Wolfien, M. Epigenomic Echoes—Decoding Genomic and Epigenetic Instability to Distinguish Lung Cancer Types and Predict Relapse. Epigenomes 2025, 9, 5. [Google Scholar] [CrossRef]
  20. Raez, L.E.; Brice, K.; Dumais, K.; Lopez-Cohen, A.; Wietecha, D.; Izquierdo, P.A.; Santos, E.S.; Powery, H.W. Liquid Biopsy Versus Tissue Biopsy to Determine Front Line Therapy in Metastatic Non-Small Cell Lung Cancer (NSCLC). Clin. Lung Cancer 2023, 24, 120–129. [Google Scholar] [CrossRef]
  21. García-Pardo, M.; Czarnecka-Kujawa, K.; Law, J.H.; Salvarrey, A.M.; Fernandes, R.; Fan, Z.J.; Waddell, T.K.; Yasufuku, K.; Liu, G.; Donahoe, L.L.; et al. Association of Circulating Tumor DNA Testing Before Tissue Diagnosis with Time to Treatment Among Patients with Suspected Advanced Lung Cancer. JAMA Netw. Open 2023, 6, e2325332. [Google Scholar] [CrossRef] [PubMed]
  22. Chen, G.; Zhang, J.; Fu, Q.; Taly, V.; Tan, F. Integrative Analysis of Multi-Omics Data for Liquid Biopsy. Br. J. Cancer 2023, 128, 505–518. [Google Scholar] [CrossRef]
  23. Seo, J.-S.; Ju, Y.S.; Lee, W.-C.; Shin, J.-Y.; Lee, J.K.; Bleazard, T.; Lee, J.; Jung, Y.J.; Kim, J.-O.; Shin, J.-Y.; et al. The Transcriptional Landscape and Mutational Profile of Lung Adenocarcinoma. Genome Res. 2012, 22, 2109–2119. [Google Scholar] [CrossRef] [PubMed]
  24. Zhao, X.; Chen, Y.; Sun, X.; He, Z.; Wu, T.; Wu, C.; Chen, J.; Wang, J.; Diao, K.; Liu, X.-S. Oncogenic EFNA4 Amplification Promotes Lung Adenocarcinoma Lymph Node Metastasis. Cancers 2022, 14, 4226. [Google Scholar] [CrossRef] [PubMed]
  25. Saintigny, P.; Peng, S.; Zhang, L.; Sen, B.; Wistuba, I.I.; Lippman, S.M.; Girard, L.; Minna, J.D.; Heymach, J.V.; Johnson, F.M. Global Evaluation of Eph Receptors and Ephrins in Lung Adenocarcinomas Identifies EphA4 as an Inhibitor of Cell Migration and Invasion. Mol. Cancer Ther. 2012, 11, 2021–2032. [Google Scholar] [CrossRef]
  26. Fang, L.; Yu, W.; Zhu, P.; Yu, G.; Ye, B. TEDC2 Correlated with Prognosis and Immune Microenvironment in Lung Adenocarcinoma. Sci. Rep. 2023, 13, 5006. [Google Scholar] [CrossRef]
  27. Morser, J. Thrombomodulin Links Coagulation to Inflammation and Immunity. Curr. Drug Targets 2012, 13, 421–431. [Google Scholar] [CrossRef]
  28. Kögl, T.; Chang, H.-F.; Staniek, J.; Chiang, S.C.C.; Thoulass, G.; Lao, J.; Weißert, K.; Dettmer-Monaco, V.; Geiger, K.; Manna, P.T.; et al. Patients and Mice with Deficiency in the SNARE Protein SYNTAXIN-11 Have a Secondary B Cell Defect. J. Exp. Med. 2024, 221, e20221122. [Google Scholar] [CrossRef]
  29. Yang, J.; Zhou, F.; Luo, X.; Fang, Y.; Wang, X.; Liu, X.; Xiao, R.; Jiang, D.; Tang, Y.; Yang, G.; et al. Enhancer Reprogramming: Critical Roles in Cancer and Promising Therapeutic Strategies. Cell Death Discov. 2025, 11, 84. [Google Scholar] [CrossRef]
  30. Spitschak, A.; Casalegno Garduño, R.; Pützer, B.M. Cancer in Transition: Discovery of Tumor-Intrinsic Transcriptional Programs Shaping the Immune and Microenvironmental Landscape. Biomark. Res. 2026, 14, 45. [Google Scholar] [CrossRef]
  31. Wang, C.; Peng, Y.; Yang, H.; Jiang, Y.; Khalid, A.K.; Zhang, K.; Xie, S.; Bermudez, L.; Yang, Y.; Zhang, L.; et al. RBMX2 Links Mycobacterium Bovis Infection to Epithelial–Mesenchymal Transition and Lung Cancer Progression. Elife 2025, 14, RP107132. [Google Scholar] [CrossRef]
  32. Dey, S.; Lu, W.; Pathinayake, P.S.; Waters, M.; Haug, G.; Larby, J.; Weber, H.C.; Wark, P.A.B.; Eapen, M.S.; Sohal, S.S. Epithelial-to-Mesenchymal Transition Is an Active Process in the Large Airways of Patients with Asthma-COPD Overlap and Partially Abrogated by Inhaled Corticosteroid Treatment: A Bronchoscopy Endobronchial Biopsy Study. Front. Immunol. 2025, 16, 1531279. [Google Scholar] [CrossRef]
  33. Zhang, X.; Yu, X.; Jiang, G.; Miao, Y.; Wang, L.; Zhang, Y.; Liu, Y.; Fan, C.; Lin, X.; Dong, Q.; et al. Cytosolic TMEM88 Promotes Invasion and Metastasis in Lung Cancer Cells by Binding DVLS. Cancer Res. 2015, 75, 4527–4537. [Google Scholar] [CrossRef] [PubMed]
  34. Miura, K.; Shukuya, T.; Greenstein, R.; Kaplan, B.; Wakelee, H.; Kurokawa, K.; Furuta, K.; Kato, S.; Suh, J.; Sivakumar, S.; et al. Ancestry-, Sex-, and Age-Based Differences of Gene Alterations in NSCLC: From the Real-World Data of Cancer Genomic Profiling Tests. J. Natl. Compr. Cancer Netw. 2024, 22, e247021. [Google Scholar] [CrossRef] [PubMed]
  35. Laguna, J.C.; García-Pardo, M.; Alessi, J.; Barrios, C.; Singh, N.; Al-Shamsi, H.O.; Loong, H.; Ferriol, M.; Recondo, G.; Mezquita, L. Geographic Differences in Lung Cancer: Focus on Carcinogens, Genetic Predisposition, and Molecular Epidemiology. Ther. Adv. Med. Oncol. 2024, 16, 17588359241231260. [Google Scholar] [CrossRef] [PubMed]
  36. Kim, D.; Paggi, J.M.; Park, C.; Bennett, C.; Salzberg, S.L. Graph-Based Genome Alignment and Genotyping with HISAT2 and HISAT-Genotype. Nat. Biotechnol. 2019, 37, 907–915. [Google Scholar] [CrossRef]
  37. Bolger, A.M.; Lohse, M.; Usadel, B. Trimmomatic: A Flexible Trimmer for Illumina Sequence Data. Bioinformatics 2014, 30, 2114–2120. [Google Scholar] [CrossRef]
  38. Liao, Y.; Smyth, G.K.; Shi, W. FeatureCounts: An Efficient General Purpose Program for Assigning Sequence Reads to Genomic Features. Bioinformatics 2014, 30, 923–930. [Google Scholar] [CrossRef]
  39. Gu, Z.; Eils, R.; Schlesner, M. Complex Heatmaps Reveal Patterns and Correlations in Multidimensional Genomic Data. Bioinformatics 2016, 32, 2847–2849. [Google Scholar] [CrossRef]
  40. Chen, Y.; Chen, L.; Lun, A.T.L.; Baldoni, P.L.; Smyth, G.K. EdgeR v4: Powerful Differential Analysis of Sequencing Data with Expanded Functionality and Improved Support for Small Counts and Larger Datasets. Nucleic Acids Res. 2025, 53, gkaf018. [Google Scholar] [CrossRef]
  41. Lê, S.; Josse, J.; Husson, F. FactoMineR: An R Package for Multivariate Analysis. J. Stat. Softw. 2008, 25, 1–18. [Google Scholar] [CrossRef]
  42. Wickham, H.; François, R.; Henry, L.; Müller, K.; Vaughan, D. Dplyr: A Grammar of Data Manipulation. CRAN: Contributed Packages 2014. Available online: https://cran.r-project.org/package=dplyr (accessed on 5 May 2026).
  43. Wickham, H.; Hester, J.; Bryan, J. Readr: Read Rectangular Text Data. CRAN: Contributed Packages 2015. Available online: https://cran.r-project.org/package=readr (accessed on 5 May 2026).
  44. Slowikowski, K. Ggrepel: Automatically Position Non-Overlapping Text Labels with “Ggplot2.” CRAN: Contributed Packages 2016. Available online: https://cran.r-project.org/web/packages/ggrepel/ggrepel.pdf (accessed on 5 May 2026).
  45. Kuhn, M. Building Predictive Models in R Using the Caret Package. J. Stat. Softw. 2008, 28, 1–26. [Google Scholar] [CrossRef]
  46. Wu, T.; Hu, E.; Xu, S.; Chen, M.; Guo, P.; Dai, Z.; Feng, T.; Zhou, L.; Tang, W.; Zhan, L.; et al. ClusterProfiler 4.0: A Universal Enrichment Tool for Interpreting Omics Data. Innovation 2021, 2, 100141. [Google Scholar] [CrossRef]
  47. Durinck, S.; Moreau, Y.; Kasprzyk, A.; Davis, S.; De Moor, B.; Brazma, A.; Huber, W. BioMart and Bioconductor: A Powerful Link between Biological Databases and Microarray Data Analysis. Bioinformatics 2005, 21, 3439–3440. [Google Scholar] [CrossRef]
  48. Bioconductor-AnnotationDbi. Available online: https://bioconductor.org/packages/release/bioc/html/AnnotationDbi.html (accessed on 5 May 2026).
  49. Bioconductor-Enrichplot. Available online: https://bioconductor.org/packages/release/bioc/html/enrichplot.html (accessed on 5 May 2026).
  50. Schauberger, P.; Walker, A. Openxlsx: Read, Write and Edit Xlsx Files. CRAN: Contributed Packages 2014. Available online: https://cran.r-project.org/web/packages/openxlsx/openxlsx.pdf (accessed on 5 May 2026).
  51. ReAD: Powerful Replicability Analysis of Genome-Wide Association Studies. CRAN: Contributed Packages 2023. Available online: https://cran.r-project.org/web/packages/ReAD/ReAD.pdf (accessed on 5 May 2026).
  52. Robin, X.; Turck, N.; Hainard, A.; Tiberti, N.; Lisacek, F.; Sanchez, J.-C.; Müller, M. PROC: An Open-Source Package for R and S+ to Analyze and Compare ROC Curves. BMC Bioinform. 2011, 12, 77. [Google Scholar] [CrossRef]
Ijms 27 04282 g001
Figure 1. Schematic overview of the experimental design and analytical workflow.
Figure 1. Schematic overview of the experimental design and analytical workflow.
Ijms 27 04282 g001
Ijms 27 04282 g002
Figure 2. Multi-layered transcriptomic analysis of tumor versus normal samples. (A) Principal Component Analysis (PCA) plot illustrating the distribution of samples from the ENA dataset. Samples highlighted were excluded due to abnormal clustering; specifically, a tumor sample exhibited expression patterns that were too similar to those of normal tissues, while a normal sample did not sufficiently cluster with other normal controls. (B) Heatmap displaying the normalized expression levels (transcripts per million [TPM], log2-transformed) of significantly differentially expressed genes (DEGs) between tumor and normal samples. (C) Volcano plot depicting log2 fold change values of DEGs. Genes with log2 fold change beyond ±1 and adjusted p-value < 0.05 are highlighted, and the most significant DEGs are annotated with their gene names. (D) Dot plot displaying pathway enrichment results for the DEGs. Enrichment analysis was conducted using the clusterProfiler R package, with 10,000 permutations for increased accuracy. A fixed random seed was used to ensure reproducibility. (E) Dot plot showing the result of the over-representation analysis (ORA), which was conducted using enrichGO from clusterProfiler, applying adjusted p-value < 0.10 thresholds to identify significantly enriched terms. (F) Combined Receiver Operating Characteristic (ROC) analysis of the seven genes on the ENA dataset.
Figure 2. Multi-layered transcriptomic analysis of tumor versus normal samples. (A) Principal Component Analysis (PCA) plot illustrating the distribution of samples from the ENA dataset. Samples highlighted were excluded due to abnormal clustering; specifically, a tumor sample exhibited expression patterns that were too similar to those of normal tissues, while a normal sample did not sufficiently cluster with other normal controls. (B) Heatmap displaying the normalized expression levels (transcripts per million [TPM], log2-transformed) of significantly differentially expressed genes (DEGs) between tumor and normal samples. (C) Volcano plot depicting log2 fold change values of DEGs. Genes with log2 fold change beyond ±1 and adjusted p-value < 0.05 are highlighted, and the most significant DEGs are annotated with their gene names. (D) Dot plot displaying pathway enrichment results for the DEGs. Enrichment analysis was conducted using the clusterProfiler R package, with 10,000 permutations for increased accuracy. A fixed random seed was used to ensure reproducibility. (E) Dot plot showing the result of the over-representation analysis (ORA), which was conducted using enrichGO from clusterProfiler, applying adjusted p-value < 0.10 thresholds to identify significantly enriched terms. (F) Combined Receiver Operating Characteristic (ROC) analysis of the seven genes on the ENA dataset.
Ijms 27 04282 g002
Ijms 27 04282 g003
Figure 3. Validation and functional relevance of top classifier genes identified by Random Forest analysis. (A) Bar plot illustrating the top 10 genes with the highest importance scores as identified by the Random Forest classifier. (B) Heatmap displaying the normalized expression levels (TPM, log2-transformed) of the 10 most informative genes—EFNA4, TEDC2, RASAL1, RADIL, SERTM1, GPM6A, CDC42EP2, STX11, THBD, and TMEM88—in the ENA dataset. These genes were selected based on their classification importance in the Random Forest model. (C) Heatmap displaying the expression of these genes—RADIL, SERTM1, GPM6A, THBD, TMEM88, STX11, EFNA4, RASAL1, CDC42EP2, and TEDC2—in TCGA samples, which are also represented in the ENA dataset. The replication of expression patterns across both datasets supports the robustness of the classification model. Expression values were normalized using TPM and log2 transformation. (D) Heatmap summarizing the biological pathways most affected by the differential expression of the top 10 genes. Enriched pathways include those related to cytokine response, cell junction assembly, cell–cell signaling, and signaling receptor activity.
Figure 3. Validation and functional relevance of top classifier genes identified by Random Forest analysis. (A) Bar plot illustrating the top 10 genes with the highest importance scores as identified by the Random Forest classifier. (B) Heatmap displaying the normalized expression levels (TPM, log2-transformed) of the 10 most informative genes—EFNA4, TEDC2, RASAL1, RADIL, SERTM1, GPM6A, CDC42EP2, STX11, THBD, and TMEM88—in the ENA dataset. These genes were selected based on their classification importance in the Random Forest model. (C) Heatmap displaying the expression of these genes—RADIL, SERTM1, GPM6A, THBD, TMEM88, STX11, EFNA4, RASAL1, CDC42EP2, and TEDC2—in TCGA samples, which are also represented in the ENA dataset. The replication of expression patterns across both datasets supports the robustness of the classification model. Expression values were normalized using TPM and log2 transformation. (D) Heatmap summarizing the biological pathways most affected by the differential expression of the top 10 genes. Enriched pathways include those related to cytokine response, cell junction assembly, cell–cell signaling, and signaling receptor activity.
Ijms 27 04282 g003
Ijms 27 04282 g004
Figure 4. Multi-layered transcriptomic analysis of tumor versus normal samples in the TCGA dataset. (A) Heatmap displaying the normalized expression levels (TPM, log2-transformed) of significant DEGs in tumor samples compared with normal samples from the TCGA dataset. Significance was determined by an adjusted p-value of <0.05 and log2 fold change beyond ±1. (B) Volcano plot illustrating the log2 fold change values of DEGs from TCGA. Genes that exhibit a log2 fold change beyond ±1 and an adjusted p-value of <0.05 are highlighted, and the top DEGs are annotated. (C,D) Dot plots demonstrating GSEA and ORA for DEGs derived from TCGA. These analyses were performed using the clusterProfiler package with the parameters consistent with those in previous analyses (Figure 2D,E), ensuring both consistency and reproducibility. (E) Combined ROC analysis of the seven genes on the TCGA dataset.
Figure 4. Multi-layered transcriptomic analysis of tumor versus normal samples in the TCGA dataset. (A) Heatmap displaying the normalized expression levels (TPM, log2-transformed) of significant DEGs in tumor samples compared with normal samples from the TCGA dataset. Significance was determined by an adjusted p-value of <0.05 and log2 fold change beyond ±1. (B) Volcano plot illustrating the log2 fold change values of DEGs from TCGA. Genes that exhibit a log2 fold change beyond ±1 and an adjusted p-value of <0.05 are highlighted, and the top DEGs are annotated. (C,D) Dot plots demonstrating GSEA and ORA for DEGs derived from TCGA. These analyses were performed using the clusterProfiler package with the parameters consistent with those in previous analyses (Figure 2D,E), ensuring both consistency and reproducibility. (E) Combined ROC analysis of the seven genes on the TCGA dataset.
Ijms 27 04282 g004
Ijms 27 04282 g005
Figure 5. Validation of DEGs in NSCLC using RT–qPCR in a Hungarian patient cohort. RT–qPCR analysis was performed on tumor and adjacent normal tissue samples from 20 NSCLC patients to validate the differential expression of selected genes identified in the NGS dataset. Expression levels were log2-transformed for visualization. (A) Boxplot displaying the expression levels of downregulated genes in NSCLC tumor samples compared with matched normal tissues, including CDC42EP2, STX11, THBD, TMEM88, and GPM6A. Statistical significance was determined using the independent-samples t-test: * p-value < 0.05; *** p-value < 0.001. (B) Boxplot displaying the expression levels of upregulated genes EFNA4 and TEDC2, which were significantly higher in tumor samples. Statistical significance was determined using the independent-samples t-test: * p-value < 0.05; *** p-value < 0.001. (C) Paired boxplot visualization comparing gene expression between tumor and matched normal samples on a per-patient basis, highlighting individual variation. The red line indicates the median expression level of the control group. (D) Combined ROC analysis of the 7-gene RT–qPCR panel.
Figure 5. Validation of DEGs in NSCLC using RT–qPCR in a Hungarian patient cohort. RT–qPCR analysis was performed on tumor and adjacent normal tissue samples from 20 NSCLC patients to validate the differential expression of selected genes identified in the NGS dataset. Expression levels were log2-transformed for visualization. (A) Boxplot displaying the expression levels of downregulated genes in NSCLC tumor samples compared with matched normal tissues, including CDC42EP2, STX11, THBD, TMEM88, and GPM6A. Statistical significance was determined using the independent-samples t-test: * p-value < 0.05; *** p-value < 0.001. (B) Boxplot displaying the expression levels of upregulated genes EFNA4 and TEDC2, which were significantly higher in tumor samples. Statistical significance was determined using the independent-samples t-test: * p-value < 0.05; *** p-value < 0.001. (C) Paired boxplot visualization comparing gene expression between tumor and matched normal samples on a per-patient basis, highlighting individual variation. The red line indicates the median expression level of the control group. (D) Combined ROC analysis of the 7-gene RT–qPCR panel.
Ijms 27 04282 g005
Table 1. Cohort characteristics. (A) Characteristics of participants included in the initial sequencing analysis obtained from the ENA dataset. (B) Validation cohort derived from the TCGA dataset. (C) Characteristics of patients included in the Hungarian population-based RT–qPCR analysis.
Table 1. Cohort characteristics. (A) Characteristics of participants included in the initial sequencing analysis obtained from the ENA dataset. (B) Validation cohort derived from the TCGA dataset. (C) Characteristics of patients included in the Hungarian population-based RT–qPCR analysis.
(A) Characteristics of Participants (ENA)
Enrolled patients with lung cancern = 45Smoking status
Age interval38–72smokern = 19
Mean age64current smokern = 2
Gender never smokedn = 23
Femalen = 22smoking status: NAn = 1
Malen = 23
Stage
1An = 25
1Bn = 20
(B) Characteristics of participants (TCGA)
Enrolled patients with lung cancern = 29Histological subtype
Age interval51–86Adenocarcinoma, NOSn = 23
Mean age67Bronchiolo-alveolar carcinoma, non-mucinousn = 1
Gender Mucinous adenocarcinoman = 1
Femalen = 18Adenocarcinoma with mixed subtypesn = 1
Malen = 11Papillary adenocarcinoman = 1
(C) Characteristics of participants (RT–qPCR)
Enrolled patients with lung cancern = 24Histological subtype
Age interval50–85Invasive solid adenocarcinoman = 9
Mean age72Invasive papillary adenocarcinoman = 5
Gender Invasive adenocarcinoman = 4
Femalen = 15Mixed adenocarcinoman = 2
Malen = 9Adenosquamous carcinoman = 1
Tumor size (T) Invasive acinar adenocarcinoman = 1
pT1an = 5Mixed-type lung cancern = 1
pT1bn = 3Papillary adenocarcinoman = 1
pT1cn = 16
Table 2. Primer sequences for the genes of interest and the housekeeping genes used in RT–qPCR.
Table 2. Primer sequences for the genes of interest and the housekeeping genes used in RT–qPCR.
GeneForward Primer (5’–3’)Reverse Primer (5’–3’)
TEDC2CTCAAGGAGAAGGGGCACCTGGTCTGTGTGGAACTGAGC
TMEM88CACTCTCAGTTCCTGCGCTCCGATAAAGGGCTCGGCTGTA
GPM6ATGCGAATCTACTGAGCTGAACCCAGTTGGCAGACAGAACCA
STX11AAGTGGGACGTGTTTTCCGACTCTCGATCTCGTTGAGGGC
THBDCCTAATGACAGTGCGCTCCTCTGGTGTTGTTGTCTCCCGT
CDC42EP2TATCTGAAGCGTGGCAGTCGATGAATGGTGTGGCGGAAGT
EFNA4TTACTACTACATCTCGGTGCCCGAAGACGAAGAATCAGAAGCAG
CYCBCTTCCCCGATGAGAACTTCAAACTCACCTCCATGCCCTCTAGAACTTT
TBPACTCCACTGTATCCCTCCCCTATATTCGGCGTTTCGGGCA
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Szollár, Z.; Dzsubák, F.; Ürmös, Á.; Borsos, B.N.; Bende, B.; Páhi, Z.G.; Pankotai, T. Machine Learning-Based Identification of Biomarkers for Early-Stage Non-Small Cell Lung Cancer Through Gene Expression Analysis. Int. J. Mol. Sci. 2026, 27, 4282. https://doi.org/10.3390/ijms27104282

AMA Style

Szollár Z, Dzsubák F, Ürmös Á, Borsos BN, Bende B, Páhi ZG, Pankotai T. Machine Learning-Based Identification of Biomarkers for Early-Stage Non-Small Cell Lung Cancer Through Gene Expression Analysis. International Journal of Molecular Sciences. 2026; 27(10):4282. https://doi.org/10.3390/ijms27104282

Chicago/Turabian Style

Szollár, Zorka, Fanni Dzsubák, Ádám Ürmös, Barbara N. Borsos, Balázs Bende, Zoltán G. Páhi, and Tibor Pankotai. 2026. "Machine Learning-Based Identification of Biomarkers for Early-Stage Non-Small Cell Lung Cancer Through Gene Expression Analysis" International Journal of Molecular Sciences 27, no. 10: 4282. https://doi.org/10.3390/ijms27104282

APA Style

Szollár, Z., Dzsubák, F., Ürmös, Á., Borsos, B. N., Bende, B., Páhi, Z. G., & Pankotai, T. (2026). Machine Learning-Based Identification of Biomarkers for Early-Stage Non-Small Cell Lung Cancer Through Gene Expression Analysis. International Journal of Molecular Sciences, 27(10), 4282. https://doi.org/10.3390/ijms27104282

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop