Comprehensive analysis of transcriptomics and radiomics revealed the potential of TEDC2 as a diagnostic marker for lung adenocarcinoma

Selection and processing of transcriptomics cohorts

The TCGA-LUAD cohort was selected from the TCGA database (https://portal.gdc.cancer.gov) to acquire transcriptomics data and clinicopathological information in FPKM form, and a total of 572 samples (including 513 tumor samples and 59 para-cancerous tissue samples) were included in this cohort. In this research, the RNA-sequencing data from TCGA was transformed into transcripts per kilobase million (TPM) values, facilitating better comparability between TCGA samples and microarray datasets (Wagner, Kin & Lynch, 2012). Transcriptome data of LUAD samples were obtained from the Gene Expression Omnibus (GEO, https://portal.gdc.cancer.gov) database using the search numbers GSE31210 and GSE30219. The dataset numbered GSE31210 included 20 normal samples and 226 tumor samples. The cohort with search number GSE30219 included 14 normal samples and 293 tumor samples. The downloaded data in the GEO database were processed by the R package oligo (Carvalho & Irizarry, 2010) according to the uniform data preprocessing routine.

Construction of co-expression networks

Weighted gene co-expression network analysis (WGCNA) identifies gene modules that are significantly associated with characterized phenotypes by constructing gene co-expression networks, which helps us to identify a set of potential candidate genes that are most relevant to LUAD (Langfelder & Horvath, 2008). Thus, in this study, based on the characteristic that gene co-expression analysis is sensitive to abnormal values, the Median Absolute Deviation (MAD) of all protein coding genes in the whole genome was calculated, and the genes with MAD greater than the top 70% were submitted to the “WGCNA” package (Langfelder & Horvath, 2008) for weighted gene co-expression network development. The distance-based adjacency index of the sample was calculated and the default parameters are defined to generate the module. After cluster analysis of modules, a heatmap of correlation between modules and traits was constructed.

Pathway annotation analysis

The data of the two queues downloaded from the GEO database were merged, and the RNA data of LUAD samples and normal samples were submitted to the R package “limma” (Ritchie et al., 2015) for difference analysis. The threshold of differentially expressed genes (DEGs) was defined as adj. p < 0.05 and | log2FC | > 1. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis was performed by importing the DEGs into the “clusterProfiler” package (Wu et al., 2021). The complete expression profile was loaded into clusterProfiler package for gene set enrichment analysis (GSEA). The results of GO and KEGG analysis were visualized as bars by ggplot2 package. GSEA enrichment maps were generated by gseaplot2.

Machine learning analysis

Machine learning analysis was performed to select diagnostic markers for LUAD, including Random Forest (RF), Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression and Support Vector Machines Recursive Feature Elimination (SVM-RFE). RF is an integrated learning method based on decision trees that efficiently handles high-dimensional data and provides an importance score for each feature. With the importance score of RF, we can filter out the genes with the most discriminative ability in the classification task, reduce the spatial dimensionality of features, and provide the generalization ability of the model (Hu & Szymczak, 2023). In this study, the ‘randomForest’ package in R was used to grow a forest of 610 trees using the default settings (Alderden et al., 2018). The glmnet function of “glmnet” package was used to perform the LASSO Cox regression model analysis, in which the parameters were family = “binomial”, alpha = 1, and nlambda = 100 (Engebretsen & Bohlin, 2019). The LASSO algorithm is able to select a small number of important features while avoiding overfit and to further improve the stability and prediction ability of the model (Jiang & Jiang, 2023). In addition, in the SVM-RFE, the SVM classifier was constructed using R package e1071, which the parameters were kernel = “linear”, and “cost = 1”. This is due to the ability of the SVM-RFE method to finely select features through a recursive process to ensure that the selected features have the maximum contribution to the model performance and thus optimize the feature set (Sanz et al., 2018). The genes jointly selected by the above machine learning algorithms were regarded as diagnostic markers of LUAD.

Analysis of immune system characteristics

We used the CIBERSORT and ESTIMATE algorithms, respectively, to further evaluate the association between the screened diagnostic markers and the tumor microenvironment. The documents required for CIBERSORT analysis were prepared in advance, including the official LM22, gene expression matrix and CIBERSORT code (Newman et al., 2015). The results of CIBERSORT were analyzed by Pearson correlation analysis together with the LUAD markers. The gene expression matrix of LUAD was converted into a GCT format file and read into ESTIMATE (Yoshihara et al., 2013). The immune, stromal and ESTIMATE scores of the sample were calculated by the estimateScore() function. Pearson correlation analysis was also performed to calculate the correlation coefficient between each score and each diagnostic marker screened.

Extraction of radiomics feature

Three regions of interest (ROIs) of gross tumor volume from TCGA-LUAD were manually segmented using ITK-SNAP (version 3.8.0) by a radiologist with more than 5 years of experience. A total of 107 radiomic features: including 14 shape features, 18 firstorder statistical features, and 75 texture features (including 16 Gray level size zone matrix features, 14 Gray level dependence matrix features, 24 Gray level cooccurrence matrix features, 16 Gray level run length matrix features, and five neighbouring gray tone difference matrix features), were collected from the ROIs based on CT images applying PyRadiomics (version 3.0) (van Griethuysen et al., 2017). It should be noted that the SD of original shape Flatness and original shape Least Axis Length was 0 and were not considered in the subsequent analysis. The Z-score algorithm was used to normalize the radiomic features for further analysis.

Construction of the radiomics model

Feature selection was realized applying LASSO regression analysis. To establish a robust binary classification model, we used the generalized linear model network (glmnet) with a binomial distribution for the logit link function. During the model training process, we set 1,000 different lambda values (nlambda = 1,000) and chose LASSO regression (alpha = 1) as the regularization method. To select the optimal regularization parameter lambda, we performed five-fold cross-validation (nfolds = 5) and used deviance as the evaluation metric. The cross-validation was carried out using the “cv.glmnet” function. Finally, we identified features related to TEDC2 expression from radiomic features and established the corresponding model. The formula for calculating radiomics score (Rad score) was obtained: $$Radscore=\sum _{i=1}^n{\beta }_i\times featur{e}_i+intercept$$

In the formula, β refers to the feature coefficient (β), and intercept within the radiomics signature was determined according to the average value of the included models.

Source of cells and RT-qPCR

Human LUAD cells A549 and human normal lung epithelial cells BEAS-2B were commercially purchased from The American Type Culture Collection (ATCC), thawed and cultured according to the supplied product instructions. Cells were cultivated at 37 °C in a DMEM (PM150210) medium that includes 10% fetal bovine serum (FBS), 1% glutamine, and a 1% solution of antibiotic/antifungal agents, with a controlled atmosphere of 5% CO₂. Total RNA was separated using E.Z.N.A. Total RNA Kit I (Omega Bio-Tek, Norcross, USA), followed by reverse transcription of total RNA into cDNA with the use of PrimeScript RT Master Mix (TaKaRa, Japan). qRT-PCR was carried out using the SYBR Green RT-PCR kit (Vazyme, Nanjing, China). Relative mRNA expression was quantified by the 2^−ΔΔCt method and normalized to GAPDH. Each primer sequence is described in Table 1.

Transwell assay

siRNA for TEDC2 was synthesized by GenePharma (Shanghai, China) and transfected into A549 cells seeded in six-well plates using Lipofectamine 2000 according to the prescribed protocol. After transfection, 200 μL of A549 cells were transferred to the upper chamber without matrigel and the upper chamber coated with Matrigel, respectively, and complete medium containing 10% FBS was added to the lower chamber for 24 h. Crystal violet was used for cell staining for 30 min after harvesting the cells in the lower chamber. Infiltrating cells were quantified under a microscope. Cells in four random fields were intercepted, and photographs were taken using an inverted microscope and the number counted.

Cell counting Kit-8 assay

Cell counting kit-8 (CCK8) assay for assessing cell viability of LUAD cell lines after silencing TEDC2. The A549 cells were seeded into a 96-well plate during the exponential growth phase at a density of 1 × 10⁴ cells per well and incubated at 37 °C and 5% CO₂ for 48 h. Following this incubation period, 10 μL of CCK8 was added to the cell culture and incubated at 37 °C for 2 h. The measurements of optical density (OD) were taken at a wavelength of 450 nm (OD 450) using a microplate reader manufactured by Bio-Rad Laboratories Inc. The data presented are the average results from three separate experiments.

Statistical analysis

Statistical tests were performed in R software (version 3.6.0). Cox regression analysis was run using the “survival” package and diagnostic efficiency was evaluated by generating receiver operating characteristic (ROC) curves using the “ROCR” package. Student’s t-test or Wilcoxon rank-sum test was used to analyze continuous variables, Pearson correlation analysis was used to assess the correlation between variables, and log-rank test was used to compare the differences in survival time between different groups of patients. A p value < 0.05 indicated statistical significance.

Identification and functional characterization of LUAD-related gene modules

WGCNA showed that stable average connectivity was achieved with a fit R² = 0.85 for the scale-free topological model, corresponding to a soft threshold of 7 (Figs. S1A–S1B). After executing WGCNA, 15 gene modules were obtained (Fig. S2C). The turquoise module has the largest pool of similarly expressed genes (Fig. S2D). The correlation between module expression profile and LUAD showed that blue was the module with the strongest correlation with LUAD, and correlation coefficient between module membership (MM) and gene significance (GS) reached 0.66, and 214 key genes in the blue module were identified using MM > 0.7 and GS > 0.4 as the screening criteria (Figs. 1A, 1B). These genes were significantly annotated in pathways, cellular components, biological processes, molecular functions associated with the cell cycle (Figs. 1C–1F). Therefore, the blue module may be the module that mediates the cell cycle of LUAD.

Cell cycle progression was hyperactive in LUAD

A total of 2,600 genes were dysregulated in LUAD samples relative to normal samples, including 1,593 significantly down-regulated genes and 1,007 significantly up-regulated genes (Figs. S2A, S2B). These two classes of DEGs that showed opposite expression patterns in LUAD were also involved in different functional regulation, and the up-regulated genes were significantly annotated to numerous pathways regulating the cell cycle, such as chromosome segregation, DNA conformation change, and DNA replication, etc. (Fig. S2C). The genes with downregulated expression were significantly annotated in kidney development, regulation of angiogenesis, tissue migration, regulation of epithelial cell proliferation and migration (Fig. S2D). GSEA analysis of LUAD expression profile also revealed significant activation of DNA repair and G2M checkpoint pathway, which are involved in cell cycle regulation in LUAD (Fig. S3). Therefore, based on these results, we suggest that hyperactivity of the cell cycle may be a major feature of LUAD.

Screening, validation and accuracy assessment of diagnostic markers

We found that 192 more of the 214 key genes identified in WGCNA-based were upregulated in LUAD patients (Fig. 2A). To further explore the impact of these key genes in LUAD, we first identified 18 characterized genes using the SVM-RFE algorithm (Fig. 2B). Subsequently, we identified six genes based on the RF algorithm (Fig. 2C) as well as screened 11 genes using the LASSO logistic regression method (Fig. 2D). We utilized a Wayne diagram in order to take the intersection of the genes screened by the three machine algorithms and ended up with four key genes, UBE2T, TEDC2, RCC1, and FAM136A for subsequent in-depth studies (Fig. 3A). As shown in Fig. 3D, we found that all four genes were significantly overexpressed in tumor samples compared to normal samples in the TCGA training cohort. The ROC curves of UBE2T, TEDC2, RCC1 and FAM136A showed their probability of being valuable biomarkers with the area under the ROC curves (AUC) value of 0.989, 0. 989, 989 and 0.987, respectively, which suggests that these four biomarkers have high predictive value (Fig. 3B). Finally, we combined these four biomarkers and found the AUC value of up to 0.963 based on a logit regression model, which again showed high diagnostic accuracy (Fig. 3C).

Subsequently, we further validated the screened diagnostic genes using the GSE30219 and GSE31210 datasets as validation. The AUC value of UBE2T, TEDC2, RCC1 and FAM136A as diagnostic markers in the GSE30219 cohort were 0.96, 0.95, 0.89 and 0.94, respectively (Fig. 4A). In GSE31210 cohort, the AUC value of UBE2T, TEDC2, RCC1 and FAM136A for the diagnosis of LUAD were 0.96, 0.94, 0.94 and 0.93, respectively (Fig. 4C). Additionally, all four diagnostic markers showed significantly higher expression in LUAD tumor samples than in normal samples in both cohorts (Figs. 4B, 4D). Therefore, the accuracy of the four genes as diagnostic markers in the GSE30219 and GSE31210 cohorts was also ideal.

Association of diagnostic markers with LUAD prognosis and TME

In addition to showing good performance for the diagnosis of LUAD, we also explored the relationship between these four genes and patient prognosis. Using univariate Cox regression analysis, we found that UBE2T, TEDC2, RCC1 and FAM136A were significantly associated with patients’ overall survival (Fig. S4A), whereas only UBE2T was significantly associated with patients’ progression free interval (PFI) (Fig. S4C). Additionally, multivariate Cox regression analysis showed that these four diagnostic genes were not independent factors for predicting the prognosis of LUAD (Figs. S4B, S4D).

Next, we further explored the relationship between these four key genes and tumor microenvironment (TME). Notably, we found that UBE2T, TEDC2, RCC1 and FAM136A were all significantly and positively associated with M1 Macrophages, activated memory CD4 T cells, and follicular helper T cells, whereas they were significantly and positively associated with M2 macrophage and resting memory CD4 T cells (Figs. 5A–5D). In addition, we found that each diagnostic gene was also negatively correlated with stromal, immune and ESTIMATE scores (Figs. 5E–5G). These results indicate that the four biomarkers we screened may act by inhibiting stromal and immune cells, which in turn promotes the malignant behavior of tumors.

Construction and discriminant ability evaluation of radiomics model

Since we did not identify features associated with the expression levels of UBE2T, RCC1 and FAM136A. Therefore, we only analyzed the mechanism of action of TEDC2 in a follow-up study. As shown in Figs. 6A, 6B, we selected the seven optimal features associated with TEDC2 expression, including “original shape Elongation”, “original firstorder Median”, “original firstorder Total Energy”, “original glrlm Run Length NonUniformity”, “original glszm Large Area Emphasis”, “original glszm Small Area Emphasis”, and “original glszm Small Area Low Gray Level Emphasis”, from 105 radiomics features by the LASSO algorithm. Therefore, the resulting radiomics model was: Rad score= −0.5888599–0.6561874* original shape Elongation + 2.6456259* original firstorder Median−1.1273691* original irstorder Total Energy + 0.9690754* original glrlm Run Length NonUniformity–1.4601943* original glszm Large Area Emphasis−1.4749247* original glszm Small Area Emphasis−0.9582481* original glszm Small Area Low Gray Level Emphasis.

To demonstrate the practicability of radiomics model in the diagnosis of LUAD, the sensitivity (SEN), accuracy (ACC), specificity (SPE), Positive Predictive Value (PPV) and Negative Predictive Value (NPV) scores of radiomics model were calculated by using 60% of the samples in TCGA as the training set and 40% as the verification set, which were 0.85, 0.9, 0.8, 0.8889 and 0.8182 respectively, and ROC-AUC was 0.96 (Fig. 6C). In the validation cohort, SEN, PPV, ACC, SPE, and NPV of the radiomics model were 0.8889, 1.0, 0.75, 1 and 0.8333, respectively. The ROC-AUC reached 0.9, and the Precision Recall (PR)-AUC reached 0.96 (Fig. 6D). The calibration curve and decision curve analysis (DCA) also showed that the radiomics model had an ideal discrimination power for LUAD (Figs. 6E, 6F).

Evaluation of prognostic ability of radiomics model

The prognosis of LUAD samples in TCGA was also evaluated according to the radiomics model, generating Kaplan-Meier curves and ROC curves. Radiomics model could significantly distinguish the prognosis of different LUAD patients, and had the best prediction effect on 5-year overall survival, with ROC-AUC of 0.88 (Figs. 7A, 7B). The rad scores of samples exhibiting high TEDC2 expression were significantly higher than those of samples exhibiting low TEDC2 expression (Figs. 7C, 7D) in both the training and validation sets.

Diagnostic markers were overexpressed in LUAD cells and promoted metastasis

We examined the expression of UBE2T, TEDC2, RCC1 and FAM136A in BEAS-2B and A549 cell lines, and found that they were all significantly overexpressed in A549 cells compared with BEAS-2B cells (Figs. 8A–8D). CCK-8 assays indicated that the proliferative capacity of A549 cells was significantly reduced after silencing TEDC2 expression (Fig. 8E). Compared with A549 cells without TEDC2 knockdown, the density of migrating and invading cells in either field was significantly loosed in A549 cells with TEDC2 knockdown, indicating that TEDC2 promotes the metastasis of LUAD cells (Figs. 8F, 8G).