Machine learning reveals CAT gene as a novel potential diagnostic and prognostic biomarker in non-small cell lung cancer Authors: Yi Tian, Wen-ya Zhao, Yi-ru Liu, Wen-wen Song, Qiao-xin Lin, Yan-na Gong, Yi-ting Deng, Dian-na Gu, Ling Tian Published: 2024 ABSTRACT AbstractBackgroundNon-small cell lung cancer (NSCLC) represents one of the most prevalent forms of lung cancer, with a five-year survival rate of 21.7%. There is an urgent need to identify pertinent biomarkers to inform the diagnosis and prognosis of tumors, particularly those that can be applied to different age groups. Herein, we would apply machine learning methods to specifically analyze the issue of biomarker applicability across different age groups in NSCLC.MethodsStudies have shown a higher incidence of NSCLC in people over 40 years of age, and due to the limitations of data set, studies of individuals under 40 years of age were not included in this study. To simulate the human aging model as closely as possible, we gathered corresponding non-small cell lung cancer (NSCLC) samples from the UCSC Xena database based on patient age information. These samples were then categorized into three groups: 40–60, 60–80, and over 80 years old. Subsequently, we employed four machine learning methods—Random Forest, LASSO regression analysis, XGBoost, and GBM—to identify gene sets with significant diagnostic value for each age group. By taking the intersection of these sets, we identified the optimal gene and assessed its prognostic significance in NSCLC. Then, the diagnostic value of CAT gene was validated using global public databases, including the GSE32863, GSE43458, GSE68571, GSE10072, and GSE63459 datasets from the Americas, the GSE30219 and GSE102511 datasets from Europe, and the GSE31210 and GSE19804 datasets from Asia. Furthermore, immunohistochemical staining was performed in an independent cohort from a tissue microarray. Additionally, cell culture and RT-qPCR were employed for external validation.ResultsThrough the implementation of machine learning methods, we successfully identified the catalase (CAT) gene. Our analysis revealed that individuals with high expression of the CAT gene experienced improved survival rates. Additionally, these individuals exhibited elevated immune scores. We further discovered that the CAT gene synergizes with multiple components of neutrophils, including TLRs, FcRn, and the selective GEF of Rho-family GTPases. In addition, we identified a potential immune checkpoint, TNFSF15, which is applicable to the human aging model. Finally, we validated the CAT gene's diagnostic value using databases encompassing the Americas, Europe, and Asia regions. Through external RT-qPCR validation, we verified that CAT expression in BEAS-2B was higher than that of A549. In an independent human cohort, we also verified that CAT is lowly expressed in lung cancer tissues. In addition, higher CAT levels were associated with improved survival in the 40–60 and 60–80 age groups.ConclusionsIn our analysis of the NSCLC database, we pinpointed the CAT gene, which holds promise for potential diagnostic and prognostic applications in the context of human aging. Furthermore, it may offer insights into addressing age-related heterogeneity of NSCLC.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-024-01670-1.Keywords: Machine learning, Catalase gene, Biomarker, Aging-related genes, Age heterogeneity, Non-small cell lung cancer Introduction Lung cancer stands as the second most prevalent cancer worldwide, with an estimated 2.2 million new cases and 1.8 million deaths [1]. Moreover, it is also the primary cause of cancer-related mortality. Consequently, it represents a significant and alarming public health concern on a global scale. Non-small cell lung cancer (NSCLC), one of the primary types of lung cancer, encompasses 85% of cases [2]. However, the survival rate for lung cancer patients stands at a mere 21.7% after five years post-diagnosis [3]. Thus, there is a pressing need to prioritize the advancement of research in the realm of NSCLC. Notably, the five-year survival rate for patients diagnosed with stage I NSCLC is approximately 80%, whereas patients with stage II and III disease have a five-year survival rate ranging from 13 to 60% [4]. Therefore, early diagnosis of lung cancer is crucial. Currently, the widespread emergence of gene sequencing has led to the identification of an increasing number of genetic markers, which have subsequently been entered into clinical studies, improving patient diagnosis and survival rates [5, 6]. Hence, the discovery of biomarker genes specific to NSCLC holds particular significance. With advancements in modern medicine, the average lifespan of human beings is gradually increasing, leading to a gradual increase in the aging of the population. Recently, aging has been identified by 12 hallmarks, with chronic inflammation being included for the first time [7]. Studies have indicated a bidirectional relationship between chronic inflammation and the development of age-related diseases such as cardiovascular disease, neurodegenerative disease, as well as cancer. Crosstalk between chronic inflammation and other features of aging leads to a vicious cycle that exacerbates the decline in cellular function and promotes aging. Thus, the drivers associated with chronic inflammation may be an ideal target for addressing aging [8]. However, no studies currently exist on the combination of genes related to aging and chronic inflammation in NSCLC. Therefore, it is imperative to investigate the relationship among aging, chronic inflammation and NSCLC. Current research on aging models has certain limitations. Firstly, the majority of aging-related cancer research centered on constructing animal models, resulting in a substantial disparity between the role of aging in cancer research and its manifestation in real human beings [9]. Secondly, evidences suggested that interventions targeting younger individuals might not yield the same efficacy in the elderly. For example, differences in response to VEGF-targeted therapy were observed in melanoma species in young versus older patients, with little benefit in the latter [10]. Similarly, another study was in young and older patients with triple-negative breast cancer, where the former had poorer survival and should be treated with different strategies [11]. Another study has also demonstrated variations in the validity of prognostic models across different age groups, indicating that models developed without considering age may not sufficiently meet the needs of all populations [12]. Based on these observations, we would model real-world disparities using a large public sequencing database of humans, and attempt to investigate heterogeneity among different age groups of NSCLC patients. In the context of the information time, the medical field has extensively embraced machine learning technology to analyze, model and train substantial medical data via computers, which has significantly contributed to disease diagnosis and enhanced diagnostic accuracy [13]. Machine learning is becoming more prevalent in clinical oncology, specifically in the diagnosis of cancer and the prediction of patient prognosis [14]. Hence, we herein apply machine learning techniques to identify genes that play a significant role in lung cancer diagnosis. In particular, we focus on age heterogeneity in NSCLC, and would investigate potential diagnostic and prognostic biomarker genes based on the common biological characteristics of chronic inflammation, aging and cancer in NSCLC. Of note, a higher incidence of NSCLC has been identified in the age group over 40 years [15]. And, due to data set limitations, this study did not include studies of individuals under 40 years of age. Furthermore, to corroborate the significance of these marker genes, we incorporate multiple databases from diverse global regions to expanding the scope of our study. Finally, we would experimentally validate the marker gene expression by using tissue array slides in an independent cohort. Materials and methods Construction of a human aging model based on UCSC Xena database In order to further model real human aging, we collected 980 NSCLC specimens from UCSC Xena database (https://xenabrowser.net/datapages/) for ages > 40 years [16]. We obtained corresponding tissue samples based on the age information provided, including 223 tumor tissue specimens and 26 normal ones from individuals aged 40 to 60, 699 tumor tissue specimens and 70 normal ones from individuals aged 60 to 80, and 58 tumor tissue specimens and 14 normal ones from individuals over 80 years. We collected 1977 aging-related genes (ARGs) and 519 chronic inflammation-related genes (CIRGs) from the GeneCards website (https://www.genecards.org/) [17]. The information on these genes is presented in the Supplementary Data. We accessed the UCSC Xena database, GeneCards website, Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and the Gene Expression Omnibus (GEO) database in January 2024. Establishment of candidate genes The "limma" R package was employed to conduct differential gene analysis in tumor tissue specimens and normal ones spanning various age groups [18]. Subsequently, a threshold of an adjusted p-value (adj. p) of < 0.05 and |log2 fold change (FC)|> 1.5 was set to identify differentially expressed genes (DEGs). Notably, a volcano plot was generated using the "ggplot2" R package, which was based on genes that were either up-regulated or down-regulated. Then, the "VennDiagram" R package was utilized to generate a Venn diagram illustrating the candidate genes among the datasets of DEGs, ARGs, and CIRGs. Finally, in order to gain a more comprehensive understanding of biological functions associated with the candidate genes, a functional analysis was conducted using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) methodologies [19, 20]. Machine learning approach identifies genes of high diagnostic value To improve the identification of genes with diagnostic relevance for diseases, we implemented four machine learning methodologies. As a preliminary step, we utilized the random forest algorithm to sift through differential genes. Specifically, we established 1000 seeds to locate optimal points with minimum error rates, resulting in the selection of the foremost 30 genes for disease diagnosis [21, 22]. To further refine our gene set, we employed least absolute shrinkage and selection operator (LASSO) logistic regression analysis techniques, commonly used for feature selection and binary classification. Specifically, we chose a set of genes consisting of both potential and candidate genes using LASSO feature selection criteria. Next, we identified genes with non-zero coefficients from the optimal LASSO regression model [23, 24]. Subsequently, we employed extreme gradient boosting (XGBoost) algorithm for feature selection, and used the XGBoost Tree (XGBTree) algorithm to identify genes with importance scores greater than zero [25, 26]. Furthermore, we used the gradient-boosted machine (GBM) method to screen characteristic genes, identifying those with importance scores greater than zero [27]. In our analysis, we utilized a tenfold cross-validation method to determine the optimal value for machine learning. We assessed the model's accuracy using the receiver operating characteristic (ROC) curve and corresponding area under the curve (AUC). Finally, we intersected the characteristic genes of the four machine learning methods to obtain high-value genes. The same methodology was applied to individuals aged 40–60, 60–80, and over 80. Assessment of prognostic value of core genes To address the heterogeneity of human aging models, we took the intersection of genes from different age groups to find the core genes that were significant for each age group. We further evaluated its differential expression in tumor and normal tissue samples across age groups, and then assessed the survival prognosis of high and low expression by the "survival" R package. To further evaluate its impact on the tumor microenvironment, we performed an immune infiltration score by the "estimate" R package. We then used the "Cibersort" R package to assess the difference between high and low expression of the core gene in 22 immune cells [28]. Then we assessed the correlation between it and the 22 immune cells by Pearson correlation analysis [29]. Finally, we analyzed the correlation between our core gene and neutrophil markers using the Pearson correlation method, focusing on six types of markers: chemoattractants, chemokines, Toll-like receptors (TLRs), neonatal Fc receptors for IgG (FcRn), selective guanine nucleotide exchange factors (GEFs) for Ras homologous (Rho) family GTPases, and adhesion molecules [30]. Evaluate the responsiveness of aging populations to immunotherapy and identify immune checkpoints applicable to aging populations The Immunophenotype Score (IPS) has demonstrated its utility in assessing the responsiveness of immunotherapies targeting CTLA-4 and PD-1 [31]. Subsequently, we applied the IPS to evaluate the potential immunotherapeutic efficacy of high and low CAT groups in different age groups. In light of the limited applicability of current anti-PD-1 and CTLA-4 immunotherapies, we performed a correlation analysis investigating 47 immune checkpoints in different age groups to identify immune checkpoints that may be more applicable to the aging population. Validation based on external datasets from different regions of the globe To further validate the diagnostic value of the core genes, we collected nine publicly available NSCLC datasets according to globally representative regions where possible: GSE32863, GSE43458, GSE68571, GSE10072 and GSE63459 were from America, GSE30219 and GSE102511 from Europe, as well as GSE31210 and GSE19804 from Asia. All datasets were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) for validation needs [32]. "Limma" R package was used to eliminate the batch effect on the data. Log2 transformation was performed for normalization. Cell culture and RT-qPCR for external validation The human lung adenocarcinoma cell line (A549) and normal pulmonary epithelial cells (BEAS-2B) were obtained from the cell bank at the Shanghai Institutes for Biological Sciences (Shanghai, China) (https://www.cellbank.org.cn/). Both A549 and BEAS-2B cell lines were authenticated via short tandem repeat (STR) analysis. Cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM; Gibco, USA) supplemented with 10% fetal bovine serum (FBS; Gibco, USA) at 37 °C in a 5% CO₂ incubator. To prevent bacterial contamination, penicillin–streptomycin solutions (Ncmbio, China) were added to the culture media. Additionally, Mycoplasma Elimination Reagent (Biosharp, China) was used during cell culture to prevent mycoplasma infection. Regular testing with the MycoBlue Kit (FuHeng, China) was conducted to ensure absence of mycoplasma contamination. Total RNA (1 μg) was extracted from both cell lines using RNA-easy Isolation Reagent (Vazyme, Nanjing, China). After reverse transcription of mRNA to cDNA, real-time quantitative polymerase chain reaction (RT-qPCR) was performed using Applied Biosystems ViiA 7 (Life Technologies, USA). Gene expression was analyzed using the 2–ΔΔCT method, with primer sequences obtained from Sangon (Shanghai, China) (listed in Supplementary Data). Immunohistochemistry (IHC) validation in an independent cohort Lung cancer tissue array slides (No. AF-LucSur2201) were purchased from AiFang biological (Hunan, China). All patients signed an informed consent for the use of their tissues. The tissue array slides contained 80 lung cancer tissue spots and 80 adjacent normal tissue spots (Supplementary Figure S1), and the location of all tissue spots and the corresponding patient information were shown in the Supplementary Data. Endogenous peroxidase activity was blocked by incubation with hydrogen peroxide, then the tissues were covered with drops of 3% BSA (G5001, Servicebio) and closed for 30 min at room temperature, the closure solution was gently shaken off, primary antibody (G1209, Servicebio) was added to the sections by drops with a proportion of PBS, and the sections were flatly placed in a humidified chamber and incubated at 4 °C overnight. The secondary antibody (AFIHCO01, AiFang) was added dropwise to the tissues and incubated at room temperature for 50 min, followed by DAB (AFIHCO04, AiFang) chromatography, hematoxylin (AFIHCO05, AiFang) staining, and the staining was examined under a microscope, and the images were collected and analyzed. Finally, H-Score = ∑(pi × i) = (percentage of weak intensity cells × 1) + (percentage of moderate intensity cells × 2) + (percentage of strong intensity cells × 3). Here, i represents the grade of positive cells: negative without coloring, 0 points; weakly positive in yellowish tint, 1 point; moderately positive in tan color, 2 points; strongly positive in tan color, 3 points. The value pi denotes the percentage of positive cells. The H-score was determined by two pathologists with over 5 years of experience. Staining was classified as high CAT expression and low one according to the median value, and the significance of CAT on patient prognosis was verified by plotting survival curves. Statistical analysis All statistical analyses were carried out using R version 4.2.1 software. T-test was utilized for two-group comparison. The Pearson correlation coefficient method was used to analyze the correlation. Significance was defined as a P-value less than 0.05, with *, **, and *** indicating P-values less than 0.05, 0.01, and 0.001, respectively. In our study, we implemented false discovery rate (FDR) control to enhance the confidence of our results. Specifically, we utilized the Benjamini–Hochberg method to manage FDR. This involved ranking the p-values from all tests and adjusting them based on the significance level to determine significant results. Results The establishment of human aging model on NSCLC The samples of NSCLC used in our study were categorized based on age into three groups: 40–60, 60–80, and over 80 years old. We collected tumor samples from 223 individuals and normal samples from 26 individuals in the 40 to 60-year-old group. Through our analysis, we identified 3387 differentially expressed genes, which were further visualized using a volcano plot to display the up-regulated and down-regulated genes. By intersecting these differentially expressed genes with those associated with aging and chronic inflammation using a Venn diagram, we obtained 104 candidate genes (Fig. 1A). Similarly, in the 60–80 age group, we analyzed 669 tumor samples and 70 normal ones, resulting in the identification of 3367 genes with differential expression. These genes were also represented in a volcano plot to highlight their up-regulation and down-regulation patterns. By intersecting these differentially expressed genes with those related to aging and chronic inflammation, we found 103 candidate genes (Fig. 1B). In the over 80 age group, we examined the differential expression of 58 tumor samples and 14 normal ones. Our analysis yielded 2764 differential genes, which were visualized using a volcano plot. Through the intersection of these genes with those associated with aging and chronic inflammation, we identified 90 potential candidate genes (Fig. 1C). These findings provide valuable insights into the genetic factors underlying aging and chronic inflammation in different age groups of NSCLC. Establishment of candidate genes in human aging model of NSCLC. A Samples aged 40–60 years, including 223 tumor tissue specimens and 26 normal tissue specimens, were differentially analyzed to find 3387 differential genes, and then 104 candidate genes were obtained after taking intersections with CIRGs and ARGs; B Samples aged 60–80 years, including 699 tumor tissue specimens and 70 normal tissue specimens, were differentially analyzed to find 3367 differential genes, and then 103 candidate genes were obtained after taking intersections with CIRGs and ARGs; C Samples aged over 80 years, including 58 tumor tissue specimens and 14 normal tissue specimens, were differentially analyzed to find 2764 differential genes, and then 90 candidate genes were obtained after taking intersections with CIRGs and ARGs Further, by GO enrichment analysis, it was observed that in individuals aged 40–60, 60–80, and over 80 years of NSCLC, the enrichment was primarily focused on leukocyte migration, leukocyte chemotaxis, cell chemotaxis, regulation of inflammatory response, cytokine-mediated signaling pathway, CXCR chemokine receptor and others (Supplementary Figure S2). Additionally, KEGG enrichment analysis highlighted that the main enrichment was in cytokine-cytokine receptor interaction, IL-17 signaling pathway, chemokine signaling pathway, TNF signaling pathway and so on (Supplementary Figure S3). Overall, the candidate genes appear to regulate tumor growth by modulating inflammation-related pathways. Machine learning techniques identify candidate genes of diagnostic value in NSCLC So, we applied machine learning technology to identify candidate genes with diagnostic relevance for NSCLC. Firstly, in the 40–60 age group, a total of 30 diagnostic genes were screened using the random forest method (Fig. 2A, B). Subsequently, 15 diagnostic genes were identified through the LASSO method (Fig. 2C), 13 genes through XGboost algorithm (Fig. 2D), and 20 genes through GBM algorithm (Fig. 2E). The training set was constructed utilizing the UCSC Xena database, while the GSE68465 set was employed as the validation set. The corresponding tenfold cross-validation AUC values for Random Forest in the training and validation sets are 1.000, and 1.000 (Fig. 3A) respectively, and for LASSO are also 1.000, and 1.000 (Fig. 3B), XGBoost are 0.992, and 1.000 (Fig. 3C), and GBM are 0.998, and 1.000 (Fig. 3D) respectively. By taking the intersection of the genes identified by these four machine learning algorithms, 3 genes, namely PECAM1, CAT, and FEN1, were identified (Fig. 3E). Furthermore, the respective AUC values for these three genes were determined as 0.998 for PECAM1 (Fig. 3F), 0.999 for CAT (Fig. 3G), and 0.978 for FEN1 (Fig. 3H). High diagnostic value genes screened by machine learning technology in the 40–60 age group of NSCLC. A The impact of the number of decision trees on the error rate; B 30 potential diagnostic genes were screened using the Random Forest method; C 15 diagnostic genes were identified through the Lasso method; D 13 diagnostic genes were identified through XGboost algorithm; E 20 diagnostic genes were identified via GBM algorithm Accuracy assessment of diagnostic genes screened by machine learning technology in the 40–60 age group of NSCLC. A–D The tenfold cross-validation AUC value in UCSC Xena database as the training set and in GSE68465 as the validation set with Random Forest method (A), LASSO method (B), XGboost algorithm (C), and GBM algorithm (D), respectively; EVenn diagrams for four machine learning methods to take intersections; (F–H) The AUC values of the intersecting genes, PECAM1 (F), CAT (G), and FEN1 (H), were taken Moving on to the 60–80 age group, a total of 30 diagnostic genes were screened using the random forest method (Supplementary Figure S4A, & S3B). This was followed by the identification of 21 diagnostic genes via the LASSO method (Supplementary Figure S4C), 16 genes through XGboost algorithm (Supplementary Figure S4D), and 20 genes through GBM algorithm (Supplementary Figure S4E). The corresponding AUC values for random forest, LASSO, XGboost, and GBM were 1.000 (Supplementary Figure S5A), 1.000 (Supplementary Figure S5B), 0.998 (Supplementary Figure S5C), and 0.998 (Supplementary Figure S5D), respectively. The intersection of genes identified by these four machine learning algorithms revealed three genes namely PECAM1, CAT, and ADRB2 (Supplementary Figure S5E). Additionally, the AUC values for these three genes were determined as 0.998 for PECAM1 (Supplementary Figure S5F), 0.993 for CAT (Supplementary Figure S5G), and 0.993 for ADRB2 (Supplementary Figure S5H). Lastly, in the over 80 age group, a total of 30 diagnostic genes were screened using the random forest method (Supplementary Figure S6A, & S5B). Following this, 11 diagnostic genes were identified through the LASSO method (Supplementary Figure S6C), 7 genes through XGboost algorithm (Supplementary Figure S6D), and 26 through GBM algorithm (Supplementary Figure S6E). The corresponding AUC values for random forest, LASSO, XGboost, and GBM were 1.000 (Supplementary Figure S7A), 1.000 (Supplementary Figure S7B), 0.974 (Supplementary Figure S7C), and 1.000 (Supplementary Figure S7D), respectively. The intersection of the genes identified by these four machine learning algorithms resulted in identification of CAT gene (Supplementary Figure S7E). Moreover, the AUC value for CAT gene was determined as 0.991 (Supplementary Figure S7F). In summary, four different machine learning techniques were used to screen genes with high diagnostic value in our human aging models across three different age groups, and the accuracy of these algorithms was verified by calculating AUC values, indicating their potential as early diagnostic markers for NSCLC. These findings provided valuable insights for future analyses. CAT gene expression is associated with clinical prognosis in NSCLC We further investigated the clinical significance of the expression of candidate gene CAT. By integrating high-value diagnostic genes identified across varying age groups of NSCLC, we successfully pinpointed CAT gene (Fig. 4A). Notably, we observed a higher expression level of CAT gene in the 60–80 age group compared to the 40–60 age group (Fig. 4B). Furthermore, our analysis revealed that CAT gene exhibited elevated expression levels in normal tissues but lower expression levels in tumor tissues (Fig. 4C, E, G). Interestingly, individuals with high CAT gene expression demonstrated improved survival rates compared to those with low expression levels (Fig. 4D, F, H). These data suggested that CAT played an important role in the aging model, and high CAT gene expression was associated with a higher survival rate in NSCLC. Clinical prognostic assessment of CAT gene in NSCLC. A CAT gene was obtained by taking the intersection of the three age groups in NSCLC; B The difference of CAT gene expression in the three age groups; C The differential expression of CAT gene in tumor tissues and normal tissues in the 40–60 age group; D The difference of the survival rate between high and low expression of CAT gene in the 40–60 age group; E The differential expression of CAT gene in tumor tissues and normal tissues in the 60–80 age group; F The difference of the survival rate between high and low expression of CAT gene in the 60–80 age group; G The differential expression of CAT gene in tumor tissues and normal tissues in the group of over 80 years old; H The difference of the survival rate between high and low expression of CAT gene in the group of over 80 years old We further explored the differences of CAT gene expression towards tumor microenvironment across distinct age groups. After assessing the StromalScore, ImmuneScore, and EstimateScore with “ESTIMATE” algorithm, we found that in the 40–60 age group, these scores of the high CAT gene expression group were higher than those of the low one. This indicated that there was a more active immune response in the tumor microenvironment, which may have a positive impact on the treatment and prognosis of the tumor. Further analysis revealed that CAT was positively correlated with monocytes and dendritic cells (Fig. 5A), which were thought to play an important role in immune surveillance of tumors and anti-tumor immune responses. In the 60–80 age group, the high-expression CAT gene group continued to exhibit higher StromalScore, ImmuneScore, and EstimateScore values compared to the low one. Additionally, CAT was also found to be positively correlated with monocytes and dendritic cells (Fig. 5B). In the over 80 age group, although StromalScore, ImmuneScore and EstimateScore did not reach statistical significance, the median ImmuneScore remained higher in the high CAT population than in the low one. And CAT also remained positively correlated with resting monocytes and dendritic cells (Fig. 5C). In short, these results indicated that patients with high CAT exhibited a stronger immune infiltration, and coordinated dendritic cells and monocytes to exert anti-tumor mechanisms across different age groups. In other words, NSCLC patients with higher CAT expression might have a better survival prognosis due to a favorable tumor immune microenvironment. Tumor microenvironment analysis of CAT gene in NSCLC. A–C The differences in StromalScore, ImmuneScore, and EstimateScore of CAT gene, and the correlation of 22 immune cells with CAT gene in the 40–60 age group (A), the 60–80 age group (B) and the group over 80 years old (C), respectively CAT gene coordinates neutrophils involving in antitumor response Now that CAT gene is associated with chronic inflammation, we thus would like to investigate its interconnection with neutrophils, the predominant type of leukocytes involved in chronic inflammatory responses. In the 40–60 age group of NSCLC, correlation analysis showed that formyl peptide receptor 1 (FPR1), a G protein-coupled receptor mediating the chemoattraction of macrophages and phagocytosis, was positively and statistically different from CAT (R = 0.32, P = 1e-06) among chemoattractants (Fig. 6A). To chemokines, CXCR1 (C-X-C chemokine receptor 1) (R = 0.3, P = 5.5e-06) and CXCR2 (R = 0.33, P = 3.5e-07) were positively correlated with CAT (Fig. 6B). For TLRs, TLR2 (R = 0.33, P = 5.7e-07) and TLR4 (R = 0.35, P = 5.2e-08) were positively correlated with CAT (Fig. 6C). In Fc receptors, FcRn was positively correlated with CAT (R = 0.42, P = 7e-11) (Fig. 6D). As for the selective GEFs for Rho family GTPases, ARHGEF6 (Rho guanine nucleotide exchange factor 6) was positively correlated with CAT (R = 0.36, P = 4.5e-08) (Fig. 6E). With regard to adhesion molecules, ITGA4 (the alpha4 integrin subunit gene) (R = 0.34, P = 1.5e-07) and P-selectin (SELP) (R = 0.37, P = 7.5e-09) were positively correlated with CAT (Fig. 6F). CAT gene coordinates neutrophils in the 40–60 age group of NSCLC. A–F The correlation of CAT gene with different components of neutrophils, including chemoattractants (A), chemokines (B), TLRs (C), Fc receptors (D), selective GEFs for Rho Family GTPases (E), and adhesion molecules (F) However, in the 60–80 age group of NSCLC, we found that LTB4R (leukotriene B4 receptor) (R = −0.34, P < 2.2e-16) and LTB4R2 (R = −0.3, P = 6.1e-16) were negatively associated with CAT among chemoattractants (Supplementary Figure S8A). As to TLRs, TLR2 (R = 0.36, P < 2.2e-16) and TLR4 (R = 0.35, P < 2.2e-16) were positively correlated with CAT (Supplementary Figure S8B). In Fc receptors, FcRn was positively correlated with CAT (R = 0.49, P < 2.2e-16) (Supplementary Figure S8C). As for the selective GEFs for Rho family GTPases, ARHGEF6 (R = 0.36, P < 2.2e-16), SPATA13 (R = 0.31, P < 2.2e-16), DDCK8 (R = 0.31, P < 2.2e-16), and FGD4 (R = 0.31, P < 2.2e-16) were positively correlated with CAT (Supplementary Figure S8D). With respect to adhesion molecules, ICAM2 (intercellular adhesion molecule 2) (R = 0.33, P < 2.2e-16) and SELP (R = 0.33, P < 2.2e-16) were positively correlated with CAT (Supplementary Figure S8E). However, no correlation was found in chemokines. Finally, in the over 80 age group, correlation analysis revealed that LTB4R (R = −0.41, P = 0.0012) and LTB4R2 (R = −0.38, P = 0.0034) were negatively correlated with CAT among chemoattractants (Supplementary Figure S9A). In TLRs, TLR4 (R = 0.3, P = 0.023) was positively correlated with CAT (Supplementary Figure S9B). As to Fc receptors, FcRn was positively correlated with CAT (R = 0.58, P = 1.6e-06) (Supplementary Figure S9C). AS for the selective GEFs for Rho family GTPases, ARHGEF6 (R = 0.35, P = 0.0074), SPATA13 (spermatogenesis-associated protein 13) (R = 0.34, P = 0.008), DOCK8 (dedicator of cytokinesis 8) (R = 0.33, P = 0.012), and FGD4 (FYVE, RhoGEF, and PH domain-containing 4) (R = 0.44, P = 0.00061) were positively associated with CAT Correlation (Supplementary Figure S9D). However, no correlation was found with regard to adhesion molecules and chemokines. Hence, we suggested that CAT primarily synergized with Toll-like receptors (TLRs), FcRn, and the selective GEF of Rho-family GTPases by taking intersections of three age groups. And the corresponding three key molecular proteins, TLR4, FcRn and ARHGEF6, were also identified. These results indicated that CAT might orchestrate multiple components of neutrophils to participate in antitumor responses. CAT gene helps evaluate immune checkpoint blockades and immunotherapeutic potentials in NSCLC Further, we wondered whether CAT gene could help evaluate immune checkpoint blockades and immunotherapeutic potentials in NSCLC. Then, we assessed the correlation of 47 immune checkpoints with the CAT gene, and found that TNFSF15 (tumor necrosis family superfamily member 15) and CD200R1(CD200 Receptor 1) were correlated with CAT in the 40–60 age group (Fig. 7A). In addition, TNFSF15 and CD40LG (CD40 ligand) were found to be correlated with CAT in the 60–80 age group (Fig. 7B), and TNFSF15 and HHLA2 (human endogenous retrovirus-H long terminal repeat-associating protein 2) were associated with CAT gene in the over 80 age group (Fig. 7C). The intersection of these analyses suggested that CAT might show a significant correlation with TNFSF15 immune checkpoints, indicating that TNFSF15 could be a potential immunotherapeutic target for NSCLC in different age groups. Evaluation of the potential immune checkpoint blockades and immunotherapeutic efficacy of CAT gene in NSCLC. A TNFSF15 and CD200R1 were found to be correlated with CAT gene in the 40–60 age group; B TNFSF15 and CD40LG were found to be correlated with CAT gene in the 60–80 age group; C TNFSF15 and HHLA2 were found to be correlated with the CAT gene in the group aged over 80; D IPS scores for PD1 and CTLA4 treatment in the 40–60 age group; E IPS scores for PD1 and CTLA4 treatment in the 60–80 age group; F IPS scores for PD1 and CTLA4 treatment in the over 80 age group We further assessed the potential anti-CTLA-4 and anit-PD-1immunotherapy efficacy by IPS score, and found that high expression of CAT gene in the 60–80 age group may be more effective for immunotherapy (Fig. 7E). However, no difference was found in the 40–60 age group (Fig. 7D) and the over 80 age group (Fig. 7F). In combination with the results of the previous correlation analysis of the 47 immune checkpoints, no correlation was found between CAT and PDCD1 and CTLA4. These results suggested that common anti-PD1 and anti-CTLA4 might not be suitable for all age groups. Intriguingly, TNFSF15 might be a potential immunotherapeutic target for NSCLC in all age groups. Prognosis value of CAT gene in all patients We attempted to merge all the data to explore the differences between the combined results and the three separate groups in the same way, and found that high expression of the CAT gene still had a superior survival rate and was statistically significant in the early stages of cancer. Although we did not find statistical significance in the intermediate and advanced stages of cancer, we still found a trend in this survival curve (Supplementary Figure S10A-D). Similarly, we found that high CAT expression still had higher immune infiltration scores, and there was still a significant correlation with dendritic cells and monocytes (Supplementary Figure S10E-F). The CAT gene also had the same immune checkpoints, and again, high expression of the CAT gene had superior immunotherapeutic efficacy (Supplementary Figure S10G-H). It should be noted that anti-CTLA-4 and anit-PD-1immunotherapy efficacy of CAT gene pertained only to the 60–80 age group and not to all populations. Nevertheless, the TNFSF15 immune checkpoint might be applicable to all populations, which opens up new possibilities for age-heterogeneous immunotherapy research. External datasets validation based on globally representative regions We wondered whether the prognostic value of CAT gene was universal. So, for reference, we collected and analyzed several databases from three globally representative regions, the United States, Europe and Asia. We found that AUC values of the CAT gene corresponding to GSE32863, GSE43458, GSE68571, GSE10072, and GSE63459 from the American region were 0.940, 0.983, 0.967, 0.993, and 0.868, respectively (Fig. 8A). And AUC values of CAT gene corresponding to GSE30219 from France, and GSE102511 from Lebanon were 0.948, 0.829, respectively (Fig. 8B, C). Moreover, AUC values of CAT gene corresponding to GSE31210 from Japan, and GSE19804 from China were 0.828, 0.913, respectively (Fig. 8D, E). These results suggested that the early diagnostic value of CAT was broad without regional restrictions, which may provide a novel biomarker to address regional heterogeneity. External datasets validation of CAT gene based on globally representative regions. A Distribution diagram of the verified GEO datasets around the world—AUC values of CAT gene corresponding to GSE32863, GSE43458, GSE68571, GSE10072, and GSE63459 from America; B AUC values of CAT gene corresponding to GSE30219 from France; C AUC values of CAT gene corresponding to GSE102511 from Lebanon; D AUC values of CAT gene corresponding to GSE19804 from China; E AUC values of CAT gene corresponding to GSE31210 from Japan External validation from an independent cohort We further validated diagnostic and prognostic values of CAT gene in an independent cohort using tissue microarray. We found that CAT expression was higher in lung normal tissues than in tumor tissues (Fig. 9A–D). And external RT-qPCR validation experiments also showed that the mRNA expression of CAT in BEAS-2B was higher than that in A549 (Fig. 9E). At the same time, we verified that patient with higher CAT expression had a better survival prognosis in the 40–60 and 60–80 age groups (Fig. 9F, G). In patients with stage I or II lung cancer, patients with higher CAT expression were revealed to have a better survival prognosis, while there did not appear to be the same trend in patients with stage III lung cancer, which is consistent with our previous findings in public databases (Fig. 9H, & Supplementary Figure S10I). These findings revealed that CAT could be applied as an independent biomarker in human aging models, mainly in the early diagnosis and prognostic value of NSCLC. External validation of CAT gene in an independent cohort based on tissue microarray. A–D In lung cancer tissue microarray, CAT expression was found to be higher in normal tissue compared to tumor tissue for the following patients: Patient ID 03385 (A), Patient ID 01086 (B), Patient ID 02699 (C), and Patient ID 00599 (D); E The CAT expression in BEAS-2B and A549 by RT-qPCR results. F–H High CAT expression had a better survival prognosis in the 40–60 (F) and 60–80 (G) age groups; H High CAT expression had a better survival prognosis in patients of stage I or II Discussion Aging is the inevitable decline of all physiological functions over time, driven by genetic programming and associated with an increased risk of cardiovascular and cerebrovascular disease, and cancer [33]. Screening for biomarkers that can clearly reflect aging and its regulatory processes has been proposed in recent studies as an urgent need in the current field of aging research [34]. With the advancement of computer technology in recent years, machine learning has grown explosively. Notably, the application of machine learning in oncology has also been gradually increasing, including risk assessment, early diagnosis, patient prognosis and treatment selection. It was reported that machine learning was very effective in predicting various types of cancers, including breast, brain, lung, liver and prostate cancers, by using computers to learn from data [35]. Herein, we utilized multiple machine learning methods to find biomarkers with high diagnostic value in NSCLC. In clinical practice, it was observed that patients of different age groups might have different responses to the same treatment regimen, such as in metastatic pancreatic ductal adenocarcinoma [36] and in aggressive B cell lymphoma [37]. It remains unknown whether the heterogeneity in patient age is also present in NSCLC and affects its diagnosis and prognosis. We herein established a human aging model of NSCLC based on the patient age groups, 40–60, 60–80, and over 80 years old, for the first time. We then applied four different machine learning approaches, random forest, LASSO, XGboost, and GBM algorithms, to identify candidate genes with diagnostic relevance for NSCLC based on this model by combining aging-related genes with chronic inflammation-related genes. We found that these candidate genes were predominantly enriched in inflammation-associated pathways, such as the IL-17 signaling pathway and the CXCR chemokine receptor. IL-17 was supposed to support cancer development by promoting chronic tissue inflammation, and an increase in IL-17 signature genes was also found in several human malignancies [38]. It was also shown that chemokines produced by the tumor itself, cancer-associated fibroblasts and infiltrating leukocytes directly promoted cancer cell proliferation by binding to chemokine receptors expressed by the tumor cells, whilst the CXCR chemokine receptor was also a potential target for immunotherapy [39]. Finally, we screened CAT genes with high diagnostic value in different age groups, and observed that CAT genes were lowly expressed in tumor tissues and highly expressed in normal tissues. We also found that high expression of CAT genes had higher survival rate by survival analysis. We further scored the tumor microenvironment for high and low expression of CAT genes, and found that highly expressed CAT genes had higher immune cell infiltration. It was reported that high Immunesocre might have a higher survival rate in NSCLC [40]. More specifically, we speculated that CAT was mainly positively correlated with dendritic cells and monocytes, indicating CAT gene might synergize with these immune cells to participate in the anti-tumor process, such as neo-antigen presentation and cytokine-mediated cell death and phagocytosis [41], et al. Furthermore, we evaluated the potential mechanisms involving in antitumor response by which CAT gene coordinated with neutrophils. We found that CAT gene synergized with TLR4, FcRn, and ARHGEF6. TLRs could trigger innate and adaptive responses that modulate anti-tumor activity [42]. Whereas recent studies showed that FcRn was involved in antigen presentation, humoral immunity and anti-tumor immunity in lung cancer, and the group positive for FcRn displayed a notably higher rate of disease-free survival and a tendency towards increased survival for NSCLC of tumor-node-metastasis stage I [43]. Some study suggested that ARHGEF6 acted as a tumor suppressor in LUAD and might serve as a new prognostic marker and potential therapeutic target [44]. Particularly, we found that immunotherapy with high expression of CAT gene may be more effective in the 60–80 age group according to the IPS score, but no significant differences were found in the other groups. Intriguingly, a correlation analysis combining the results of 47 immune checkpoints also did not reveal a correlation between CAT and PDCD1 and CTLA4. These data indicated that commonly used anti-PD1 and anti-CTLA4 therapy might not be suitable for all age groups. Nevertheless, TNFSF15 was found to be positively correlated with CAT gene in all age groups. TNFSF15 had been shown to be a potential therapeutic target for immune checkpoint inhibitors [45]. This might provide a new insight to address age heterogeneity of immune checkpoints as well. Finally, external validation was conducted using publicly available databases from America, Europe and Asia. The results reinforced the high diagnostic value of CAT gene in NSCLC, suggesting its potential as a biomarker to address regional heterogeneity. Moreover, an independent cohort based on tissue microarray also confirmed that CAT gene was expressed at lower levels in lung cancer tissues and at higher levels in normal tissues. And high CAT expression was linked to improved survival prognosis in the 40–60 and 60–80 age groups, as well as in patients with clinical stages I and II. Although our study revealed the high diagnostic and prognostic value of CAT gene in NSCLC through machine learning, we need to recognize that the study still has some limitations. Firstly, the validation of the external dataset included remains insufficient, necessitating a larger dataset to validate our findings further. Additionally, our exploration was limited to NSCLC, highlighting the need for future studies to encompass other solid tumors. Conclusion In conclusion, we discovered CAT gene with dual characteristics of aging and chronic inflammation based on human genome sequencing data through machine learning methods, which has high diagnostic value in NSCLC, and we also confirmed its high diagnostic value in multiple regional databases around the world. CAT gene is not only effective in different age groups, but also addresses the heterogeneity between age groups, which means that it has the potential to be one of the effective human aging biomarkers associated with NSCLC (Fig. 10). The article works summary chart Supplementary Information Supplementary material 1: Table 1. List of selected Aging-related genes. Table 2. List of selected chronic inflammation-related genes. Table 3. Primer sequences for qPCR. Table 4. The location of all tissue spots on the tissue microarray. Table 5. The corresponding patient information in the tissue microarray. Table 6. R packages and corresponding versions used in this study. Supplementary material 2: Figure S1. The overall view of tissue microarray. The tissue array slide contained 80 spots of lung cancer tissue and 80 spots of adjacent normal tissue. Supplementary material 3: Figure S2. GO analysis results. (A) GO analysis of 104 candidate gene sets in the 40-60 age group; (B) GO analysis of 103 candidate gene sets in the 60-80 age group; (C) GO analysis of 90 candidate gene sets in the group over 80 years old. Supplementary material 4: Figure S3. KEGG analysis results. (A) KEGG analysis of 104 candidate gene sets in the 40-60 age group; (B) KEGG analysis of 103 candidate gene sets in the 60-80 age group; (C) KEGG analysis of 90 candidate gene sets in the group over 80 years old. Supplementary material 5: Figure S4. High diagnostic value genes screened by machine learning technology in the 60-80 age group of NSCLC. (A) The impact of the number of decision trees on the error rate; (B) 30 potential diagnostic genes were screened using the Random Forest method; (C) 21 diagnostic genes were identified through the Lasso method; (D) 16 diagnostic genes were identified through XGboost algorithm; (E) 20 diagnostic genes were identified via GBM algorithm. Supplementary material 6: Figure S5. Accuracy assessment of diagnostic genes screened by machine learning technology in the 60-80 age group of NSCLC. (A) The AUC value of Random Forest method; (B) The AUC value of Lasso method; (C) The AUC value of XGboost algorithm; (D) The AUC value of GBM algorithm; (E) Venn diagrams for four machine learning methods to take intersections; (F-H) The AUC values of the intersecting genes, PECAM1 (F), CAT (G), and ADRB2 (H), were taken. Supplementary material 7: Figure S6. High diagnostic value genes screened by machine learning technology in the group aged over 80 of NSCLC. (A) The impact of the number of decision trees on the error rate; (B) 30 potential diagnostic genes were screened using the Random Forest method; (C) 11 diagnostic genes were identified through the Lasso method; (D) 7 diagnostic genes were identified through XGboost algorithm; (E) 26 diagnostic genes were identified via GBM algorithm. Supplementary material 8: Figure S7. Accuracy assessment of diagnostic genes screened by machine learning technology in the group aged over 80 of NSCLC. (A) The AUC value of Random Forest method; (B) The AUC value of Lasso method; (C) The AUC value of XGboost algorithm; (D) The AUC value of GBM algorithm; (E) Venn diagrams for four machine learnings to take intersections; (F) The AUC value of the intersecting CAT gene was taken. Supplementary material 9: Figure S8. CAT gene coordinates neutrophils in the 60-80 age group. (A-E) The correlation of CAT gene with different components of neutrophils, including chemoattractants (A), TLRs (B), Fc receptors (C), selective GEFs for Rho family GTPases (D), and adhesion molecules (E). Supplementary material 10: Figure S9. CAT gene coordinates neutrophils in the group over 80 years old. (A-D) The correlation of CAT gene with different components of neutrophils, including chemoattractants (A), TLRs (B), Fc receptors (C), selective GEFs for Rho family GTPases (D). Supplementary material 11: Figure S10. CAT gene re-analysis after pooling of the three age groups in NSCLC. (A) Survival differences of CAT gene in all patients; (B) Prognosis of CAT gene in the stage I group; (C) Prognosis of CAT gene in the stage II group; (D) Prognosis of the CAT gene in the stage III and IV group; (E) Tumor microenvironment scores of CAT gene; (F) Correlation of CAT gene with 22 immune cells; (G) Immune checkpoints associated with CAT gene;(H) Assessment of the potential immunotherapeutic efficacy of CAT gene. (I) No trend was found in patients of Stage III on tissue microarray cohort. Acknowledgements Not applicable. Author contributions TY performed majority of the data analysis and wrote the manuscript. TL wrote and revised the manuscript. ZW, LY, SW, LQ, GY and DY assisted with the data analysis. TL and GD conceived the study and supervised the research. All authors have read and agreed to the published version of the manuscript. Funding This study was supported by the funding of the National Natural Science Foundation of China (No. 82073203). Data availability The data supporting the findings of this study are given in the main manuscript and supplementary files. The publicly available non-small cell lung cancer dataset utilized in the manuscript was obtained from https://xenabrowser.net/datapages/. All data analyzed during the present study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This study was performed in accordance with the Declaration of Helsinki. And this study using the tissue microarray was approved by the Life Sciences Ethics Committee of Changsha Yaxiang Biotechnology Co., LTD. The Ethics report is available online at yxswll.ccrl.cn. The query code is Csyayj2024011. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Dian-na Gu, Email: yinuo801@126.com. Ling Tian, Email: TL09168@hotmail.com. References Associated Data Supplementary Materials Supplementary material 1: Table 1. List of selected Aging-related genes. Table 2. List of selected chronic inflammation-related genes. Table 3. Primer sequences for qPCR. Table 4. The location of all tissue spots on the tissue microarray. Table 5. The corresponding patient information in the tissue microarray. Table 6. R packages and corresponding versions used in this study. Supplementary material 2: Figure S1. The overall view of tissue microarray. The tissue array slide contained 80 spots of lung cancer tissue and 80 spots of adjacent normal tissue. Supplementary material 3: Figure S2. GO analysis results. (A) GO analysis of 104 candidate gene sets in the 40-60 age group; (B) GO analysis of 103 candidate gene sets in the 60-80 age group; (C) GO analysis of 90 candidate gene sets in the group over 80 years old. Supplementary material 4: Figure S3. KEGG analysis results. (A) KEGG analysis of 104 candidate gene sets in the 40-60 age group; (B) KEGG analysis of 103 candidate gene sets in the 60-80 age group; (C) KEGG analysis of 90 candidate gene sets in the group over 80 years old. Supplementary material 5: Figure S4. High diagnostic value genes screened by machine learning technology in the 60-80 age group of NSCLC. (A) The impact of the number of decision trees on the error rate; (B) 30 potential diagnostic genes were screened using the Random Forest method; (C) 21 diagnostic genes were identified through the Lasso method; (D) 16 diagnostic genes were identified through XGboost algorithm; (E) 20 diagnostic genes were identified via GBM algorithm. Supplementary material 6: Figure S5. Accuracy assessment of diagnostic genes screened by machine learning technology in the 60-80 age group of NSCLC. (A) The AUC value of Random Forest method; (B) The AUC value of Lasso method; (C) The AUC value of XGboost algorithm; (D) The AUC value of GBM algorithm; (E) Venn diagrams for four machine learning methods to take intersections; (F-H) The AUC values of the intersecting genes, PECAM1 (F), CAT (G), and ADRB2 (H), were taken. Supplementary material 7: Figure S6. High diagnostic value genes screened by machine learning technology in the group aged over 80 of NSCLC. (A) The impact of the number of decision trees on the error rate; (B) 30 potential diagnostic genes were screened using the Random Forest method; (C) 11 diagnostic genes were identified through the Lasso method; (D) 7 diagnostic genes were identified through XGboost algorithm; (E) 26 diagnostic genes were identified via GBM algorithm. Supplementary material 8: Figure S7. Accuracy assessment of diagnostic genes screened by machine learning technology in the group aged over 80 of NSCLC. (A) The AUC value of Random Forest method; (B) The AUC value of Lasso method; (C) The AUC value of XGboost algorithm; (D) The AUC value of GBM algorithm; (E) Venn diagrams for four machine learnings to take intersections; (F) The AUC value of the intersecting CAT gene was taken. Supplementary material 9: Figure S8. CAT gene coordinates neutrophils in the 60-80 age group. (A-E) The correlation of CAT gene with different components of neutrophils, including chemoattractants (A), TLRs (B), Fc receptors (C), selective GEFs for Rho family GTPases (D), and adhesion molecules (E). Supplementary material 10: Figure S9. CAT gene coordinates neutrophils in the group over 80 years old. (A-D) The correlation of CAT gene with different components of neutrophils, including chemoattractants (A), TLRs (B), Fc receptors (C), selective GEFs for Rho family GTPases (D). Supplementary material 11: Figure S10. CAT gene re-analysis after pooling of the three age groups in NSCLC. (A) Survival differences of CAT gene in all patients; (B) Prognosis of CAT gene in the stage I group; (C) Prognosis of CAT gene in the stage II group; (D) Prognosis of the CAT gene in the stage III and IV group; (E) Tumor microenvironment scores of CAT gene; (F) Correlation of CAT gene with 22 immune cells; (G) Immune checkpoints associated with CAT gene;(H) Assessment of the potential immunotherapeutic efficacy of CAT gene. (I) No trend was found in patients of Stage III on tissue microarray cohort. Data Availability Statement The data supporting the findings of this study are given in the main manuscript and supplementary files. The publicly available non-small cell lung cancer dataset utilized in the manuscript was obtained from https://xenabrowser.net/datapages/. All data analyzed during the present study are available from the corresponding author on reasonable request.