Prediction of prognosis, efficacy of lung adenocarcinoma by machine learning model based on immune and metabolic related genes Authors: Cong Xue, Yi-Zhi Dai, Gui-Long Li, Yi Zhang Published: 2024 ABSTRACT BackgroundThe aim of this study is to integrate immune and metabolism-related genes in order to construct a predictive model for predicting the prognosis and treatment response of LUAD(lung adenocarcinoma) patients, aiming to address the challenges posed by this highly lethal and heterogeneous disease.Material and methodsUsing TCGA-LUAD as the training subset, differential gene expression analysis, batch survival analysis, Lasso regression analysis, univariate and multivariate Cox regression analysis were performed to construct prognostic related gene models. GEO queue as validation subsets, is used to validate build Riskscore. Then, we explore the Riskscore and mutation status, immune cell infiltration, the relationship between immune therapy and chemotherapy, and build the model of the nomogram.ResultsThe Riskscore has been determined to be composed of seven gene. In the high-risk group defined by this score, both early-stage and advanced-stage LUAD patients exhibit a decreased overall survival rate. The mutation status of patients as well as immune cell infiltration show associations with the Riskscore value obtained from these genes’ expression levels. Furthermore, there exist variations in response to immunotherapy as well as sensitivity to commonly used chemotherapy drugs among different individuals. Lastly, when using a column line plot model based on the calculated Riskscore values, we obtain a concordance index (C-index) was 0 .716 (95% CI 0.671–0.762), and time-dependent ROC predicted probabilities of 1-, 3- and 5-year survival for LUAD patients were 0.752, 0.725 and 0.654, respectively.ConclusionIn conclusion, we have successfully developed a predictive model incorporating immune and metabolism-related genes, encompassing gene expression levels of CAT/CCL20/GPI/INSL4 NT5E/GSTA3/GNPNAT1. This comprehensive model not only enables the prognosis prediction for LUAD patients but also facilitates the prediction of their response to first-line chemotherapy drugs and immune checkpoint inhibitors, thus demonstrating its broad potential in clinical applications. However, our study still has limitations as it is based on TCGA and GEO databases with limited pathological characteristics of patients. Therefore, more practical and valuable factors are needed to predict efficacy. The crosstalk between metabolism and immunity remains to be explored. Finally, this study lacks experimental evidence for the underlying gene expression of prognosis and further research is required. Background Lung adenocarcinoma is the predominant histological subtype of lung cancer, resulting in millions of fatalities annually [1]. Despite the significant advancements in lung cancer survival achieved through chemotherapy and immune checkpoint inhibitors, not all patients derive benefits from immunotherapy due to the extensive heterogeneity observed in lung cancer genomes. Consequently, there exists an urgent imperative for novel methodologies to identify potential candidates who would benefit from personalized treatment approaches. Hence, the development of innovative genetic models holds immense significance for both diagnosis and prognosis of LUAD. Metabolism is a fundamental process of life, which produces energy, redox equivalents, and macromolecules and metabolites required for various life activities. Studies have found that tumors still require anaerobic glycolysis to obtain energy when metabolic substrates are sufficient, a process known as the Warburg Effect [2].Subsequently, it was found that metabolic abnormalities in tumors are not limited to the Warburg effect, but also abnormalities in central metabolic pathways. This heterogeneity is known as tumor metabolic reprogramming [3]. Recent studies have shown that metabolic reprogramming in tumors is not only a phenotypic feature, but also can be used to counteract the body's anti-tumor response [3, 4]. Immunity refers to the process of recognizing and eliminating “non-self” components. Various immune cell populations in the body work together to maintain the homeostasis of the immune system, and through this process, cancer cells are monitored and eliminated. The immune system can regulate its function according to different signals in the environment to start or turn off specific immune responses. Immunotherapy for cancer can kill tumor cells by activating the patient's anti-tumor immune response [5]. Such as PD-1 or PD-L1, Although immunotherapy has many advantages compared with traditional chemotherapy alone, the response to treatment in each patient will vary greatly due to tumor heterogeneity and different degrees of immune infiltration. Common biomarkers such as PD-L1 expression cannot accurately predict the sensitivity of patients to treatment [6, 7]. In the Checkmate-032 study, there was no significant difference in objective response rates between the PD-L1 positive (≥ 1%) subgroup and the PD-L1 negative (< 1%) subgroup (24.0% vs. 26.2%) [8]. Additionally, various factors such as tumor stage and biopsy site may influence the expression of PD-L1 [9]. Not only does the metabolic reprogramming of cancer confer the characteristics of rapid tumor proliferation, but also certain metabolites are involved in the regulation of anti-tumor immune response [10–12]. This regulation affects tumor progression and drug resistance by altering the intensity of immune response [3, 13, 14]. In addition, it seems that we can improve the sensitivity of tumors to drugs by regulating their abnormal metabolism [15, 16]. Many studies have shown that there is an interactive relationship between tumor metabolic reprogramming and immune response [15, 17, 18]. It is worth noting that this influence is not unidirectional, as immune checkpoint inhibitors can also affect the functional status of immune cells by altering the immune microenvironment [8]. The immune system and tumor metabolism are distinguishing features that differentiate tumors from normal tissues. There exists a mutual correlation between these two aspects. Many predictive models only consider one aspect, thus we boldly hypothesize that by integrating immune and metabolism-related genes, we can develop a risk model for predicting the prognosis of LUAD patients. but also evaluates the efficacy of chemotherapy and immune checkpoint inhibitors. Furthermore, we combined the prognostic model with nomogram based on clinicopathological factors (age, gender, T, N, M stage) and found that it performed well in 1-, 3-, and 5-year survival rates, suggesting that our prognostic model has high clinical value. Material and methods Data acquisition The transcriptome profiles, mutation data and corresponding clinical information of LUAD patients in the present study were retrieved and downloaded from the TCGA portal (https://portal.gdc.cancer.gov/, up to November 10, 2023) and GEO database (https://www.ncbi.nlm.nih.gov/geo/, up to November 10, 2023 Table 1). Some cases with survival of less than 3 months or incomplete clinical data were then filtered out. In addition, 483 immune-related genes were obtained from IMMPORT database (https://immport.niaid.nih.gov/home). Metabolically associated Genes were screened from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (https://www.genome.jp/kegg/). 977 Metabolically associated Genes were extracted from the KEGG database. (http://www.broadinstitute.org/gsea/msigdb).Table 1The baseline characteristic of the enlisting cohortGEO cohortpatients with LUADAgeSexStageStatusGSE11969n = 94 > 70 = 18 ≤ 70 = 76Female = 43Male = 51I–II = 67III–IV = 27Dead = 43Alive = 51GSE13213n = 117 > 70 = 20 ≤ 70 = 107Female = 57Male = 60I–II = 92III–IV = 25Dead = 49Alive = 68GSE41271n = 183 > 70 = 62 ≤ 70 = 121Female = 90Male = 93I–II = 130III–IV = 53Dead = 71Alive = 112GSE30219n = 85 > 70 = 17 ≤ 70 = 68Female = 19Male = 66I–II = 84III–IV = 1Dead = 45Alive = 40GSE31210n = 226 > 70 = 5 ≤ 70 = 221Female = 121Male = 105I–II = 226Dead = 35Alive = 191GSE37745n = 105 > 70 = 81 ≤ 70 = 24Female = 60Male = 45I–II = 88III = 17Dead = 76Alive = 29GSE42127n = 133 > 70 = 81 ≤ 70 = 52Female = 65Male = 68I–II = 111III = 22Dead = 43Alive = 90GSE50081n = 127 > 70 = 35 ≤ 70 = 52Female = 62Male = 65I–II = 127Dead = 51Alive = 76GSE72094n = 387 > 70 = 189 ≤ 70 = 198Female = 216Male = 171I–II = 315III–IV = 72Dead = 107Alive = 280GSE115002n = 52 > 70 = 5 ≤ 70 = 47Female = 26Male = 26/Dead = 38Alive = 14TCGA-LUADn = 465 > 70 = 146 ≤ 70 = 309Female = 26Male = 26I–II = 370III–IV = 95Dead = 291Alive = 174 Table 1The baseline characteristic of the enlisting cohortGEO cohortpatients with LUADAgeSexStageStatusGSE11969n = 94 > 70 = 18 ≤ 70 = 76Female = 43Male = 51I–II = 67III–IV = 27Dead = 43Alive = 51GSE13213n = 117 > 70 = 20 ≤ 70 = 107Female = 57Male = 60I–II = 92III–IV = 25Dead = 49Alive = 68GSE41271n = 183 > 70 = 62 ≤ 70 = 121Female = 90Male = 93I–II = 130III–IV = 53Dead = 71Alive = 112GSE30219n = 85 > 70 = 17 ≤ 70 = 68Female = 19Male = 66I–II = 84III–IV = 1Dead = 45Alive = 40GSE31210n = 226 > 70 = 5 ≤ 70 = 221Female = 121Male = 105I–II = 226Dead = 35Alive = 191GSE37745n = 105 > 70 = 81 ≤ 70 = 24Female = 60Male = 45I–II = 88III = 17Dead = 76Alive = 29GSE42127n = 133 > 70 = 81 ≤ 70 = 52Female = 65Male = 68I–II = 111III = 22Dead = 43Alive = 90GSE50081n = 127 > 70 = 35 ≤ 70 = 52Female = 62Male = 65I–II = 127Dead = 51Alive = 76GSE72094n = 387 > 70 = 189 ≤ 70 = 198Female = 216Male = 171I–II = 315III–IV = 72Dead = 107Alive = 280GSE115002n = 52 > 70 = 5 ≤ 70 = 47Female = 26Male = 26/Dead = 38Alive = 14TCGA-LUADn = 465 > 70 = 146 ≤ 70 = 309Female = 26Male = 26I–II = 370III–IV = 95Dead = 291Alive = 174 The baseline characteristic of the enlisting cohort > 70 = 18 ≤ 70 = 76 Female = 43 Male = 51 I–II = 67 III–IV = 27 Dead = 43 Alive = 51 > 70 = 20 ≤ 70 = 107 Female = 57 Male = 60 I–II = 92 III–IV = 25 Dead = 49 Alive = 68 > 70 = 62 ≤ 70 = 121 Female = 90 Male = 93 I–II = 130 III–IV = 53 Dead = 71 Alive = 112 > 70 = 17 ≤ 70 = 68 Female = 19 Male = 66 I–II = 84 III–IV = 1 Dead = 45 Alive = 40 > 70 = 5 ≤ 70 = 221 Female = 121 Male = 105 Dead = 35 Alive = 191 > 70 = 81 ≤ 70 = 24 Female = 60 Male = 45 I–II = 88 III = 17 Dead = 76 Alive = 29 > 70 = 81 ≤ 70 = 52 Female = 65 Male = 68 I–II = 111 III = 22 Dead = 43 Alive = 90 > 70 = 35 ≤ 70 = 52 Female = 62 Male = 65 Dead = 51 Alive = 76 > 70 = 189 ≤ 70 = 198 Female = 216 Male = 171 I–II = 315 III–IV = 72 Dead = 107 Alive = 280 > 70 = 5 ≤ 70 = 47 Female = 26 Male = 26 Dead = 38 Alive = 14 > 70 = 146 ≤ 70 = 309 Female = 26 Male = 26 I–II = 370 III–IV = 95 Dead = 291 Alive = 174 Data processing After the aforementioned transcriptome sequencing raw expression data from different platforms were combined and log2-transformed, all of which were implemented using the R package. To eliminate batch differences between two datasets, we applied the ComBat method from the R package “sva” for standardization. In addition, principal component analysis (PCA) was conducted to evaluate the performance of batch effect removal. NMF (Non-negative matrix factorization) consensus clustering analysis Non-negative matrix factorization (NMF) algorithm is a non-negative factorization of a matrix under the condition that all the elements of the matrix are non-negative, so as to find out the relationships and interactions between them. Elements with similar characteristics are grouped into one group and elements with different characteristics are grouped into another group. Prior to performing the NMF clustering algorithm, the “limma” package with the thresholds of p < 0.05 and |logFC| (fold change) > 1 was utilized to screen out the differentially expressed HP-related genes. Afterwards, on the basis of these genes, the NMF clustering method was implemented to sort TCGA. The number of clusters (K) from 2 to 8 were tested by running ten iterations per K, and the optimal K number was eventually determined in accordance with silhouette, consensus, as well as cophenetic. Then, Kaplan–Meier (K-M) curve was carried out to estimate the differences of OS between distinct subtypes. Moreover, the ESTIMATE algorithms were performed to elucidate the immune infiltration landscape of different clusters. Construction and validation of the risk model To quantify immune-related correlation patterns for individual tumors, we divided the patients into training and validation groups in a ratio of 7:3, and univariate Cox proportional hazard regression was conducted to identify immune-related prognostic markers. We further applied significant factors to Least Absolute Shrinkage and Selective Operator (LASSO) and univariate Cox proportional hazard regression analyses to construct the risk model. The risk score was calculated based on the coefficients of the candidate genes. According to the median risk score, patients were divided into low- and high-risk groups.,Then Kaplan–Meier (K-M) curve was carried out to estimate the differences of OS between distinct subgroup. To improve the accuracy and practicability of the clinical predictive model, we constructed a nomogram model that included the following parameters: risk score, TNM stage, age, and sex. A calibration curve of the nomogram model was established to assess the consistency between the predicted and observed results. TMB (Tumor mutational burden) calculation and prognostic analysis Somatic mutation information was extracted with a Perl script (23), and the TMB score of each sample was calculated through dividing the number of variants by exon length (38 million). For prognostic analysis and clinical correlation analysis, R was utilized to merge the patients’ TMB scores with corresponding clinical information. Kaplan–Meier (K-M) analysis was conducted to compare the difference in overall survival (OS) with the log-rank test for statistical significance. Prediction of ICB (Immune checkpoint blockade) therapy response Predict the effectiveness of patient immunotherapy using TIDE (http://tide.dfci.harvard.edu/). Obtain immune checkpoint scores for patients in the TCGA cohort using TICA (https://www.cancerimagingarchive.net/). Additionally, 298 urothelial carcinoma patients with both transcriptomic data and treatment response to immunotherapy from the IMvigor210 cohort were used for speculating the immunotherapy response of the signature. Drug response prediction By logging in to the Gene Set Cancer Analysis (GSCA) (http://bioinfo.life.hust.edu.cn/GSCA/#/drug) database, the association between the sensitivity of drugs derived from the Genomics of Drug Sensitivity in Cancer (GDSC) database and the mRNA expression of hub genes was explored [19]. At the same time, the chemical structural formulas of GDSC agents that exhibited a significant positive or negative correlation with the expression of these hub genes were collected by querying the MedChemExpress website (https://www.medchemexpress.cn/). Results Classification and characterization of lung adenocarcinoma The overview of the process used in this study was shown in Fig. 1. After extraction from KEGG and MMPORT databases, a total of 947 genes related to metabolic pathways and 483 genes related to immunity were obtained. In the TCGA LUAD cohort (n = 502), 343 pairs of gene expression differences were found in contrast to adjacent normal tissues. The GEO-LUAD cohort was merged using the combat package and principal component analysis (PCA) was applied to evaluate the merging effect (Fig. 2). Through GEO cohort validation, 272 genes were identified that were expressed in both GEO and TCGA cohorts (Fig. 3A). Using these 272 pairs of genes for Non-negative matrix factorization (NMF) algorithm, the TCGA cohort was divided into two subtypes (Fig. 3B). According to the results of survival analysis, they were named high-risk group and low-risk group, and it was found that there were statistically significant differences in disease-free survival and survival between these two types (Fig. 3C).Fig. 1Flow chart of this studyFig. 2Principal Component Analysis (PCA) dimensionality reduction before merging GEO-LUAD Cohorts (A). Principal Component Analysis (PCA) dimensionality reduction before merging GEO-LUAD Cohorts (B)Fig. 3Construction of a NMF subtype based on Immune and Metabolic Related Genes in the TCGA-LUAD cohort. A Identification of DEGs between tumor and normal. B NMF consensus clustering for k = 2. C Kaplan–Meier analysis of overall survival (OS) for Cluster C1 and C2. D Kaplan–Meier analysis of disease-free survival (DFS) for Cluster C1 and C2. E ESTIMATE for Cluster C1 and C2 Flow chart of this study Principal Component Analysis (PCA) dimensionality reduction before merging GEO-LUAD Cohorts (A). Principal Component Analysis (PCA) dimensionality reduction before merging GEO-LUAD Cohorts (B) Construction of a NMF subtype based on Immune and Metabolic Related Genes in the TCGA-LUAD cohort. A Identification of DEGs between tumor and normal. B NMF consensus clustering for k = 2. C Kaplan–Meier analysis of overall survival (OS) for Cluster C1 and C2. D Kaplan–Meier analysis of disease-free survival (DFS) for Cluster C1 and C2. E ESTIMATE for Cluster C1 and C2 We used the ESTIMATE algorithm to estimate the population abundance of immune cells and stromal cell populations infiltrating the tissues in both subtypes. The ESTIMATE evaluation results showed that C1 subtype had a higher score compared to C2 subtype, indicating a more significant level of immune cell infiltration in C1 subtype (Fig. 3E). Establishment and validation of prognostic models In order to further explore the value of differential genes, The TCGA-LUAD cohort was randomly divided into a training group and a testing group in a 7:3 ratio. we combined univariate Cox regression, LASSO regression (Fig. 4A, B) and multivariate Cox regression (Fig. 4C), used to select 7 genes to construct an immune prognostic risk model. Risk scores were obtained by combining expression levels and regression coefficients (Riskscore = CAT exp × -0.166751936312374 + CCL20 exp × 0.0563504470190809 + GPI exp × 0.133661988590366 + INSL4 exp * 0.126270665159998 + NT5Eexp * 0.105058801123867 + GSTA3exp × -0.064669483229661) + GNPNAT1 exp × 0.136377362257813). The training set and test set patients were divided into low-risk group and high-risk group based on median risk score, and the robustness of the prognostic risk model was evaluated. In both TCGA cohort (Fig. 5A, B, C) and GEO cohort (Figure D), patients in the high-risk group had shorter survival time than those in the low-risk group ROC curves showed that Riskscore had the largest area under AUC curve among all clinical features (Fig. 5E). In TCGA cohort, the areas under AUC curve at 1 year, 3 years, and 5 years were: 0.752, 0.755, and 0.654 respectively (Fig. 5F).Fig. 4Screening for prognostic genes in LUAD using the TCGA-LUAD cohort.Lasso regression (A, B), multivariate Cox regression analyses (C)Fig. 5We were performed to screen candidate genes. Kaplan–Meier curves demonstrated the overall survival of LUAD patients in the Riskscore-high and -low groups of TCGA-LUAD train cohort (A), TCGA-LUAD test cohort (B), TCGA-LUAD cohort (C) and e GEO-LUAD cohort (D). The ROC curves to validate the accuracy of the risk model in predicting the clinical outcomes of patients with LUAD (E, F). All of statistical tests were performed using Cox regression analysis Screening for prognostic genes in LUAD using the TCGA-LUAD cohort.Lasso regression (A, B), multivariate Cox regression analyses (C) We were performed to screen candidate genes. Kaplan–Meier curves demonstrated the overall survival of LUAD patients in the Riskscore-high and -low groups of TCGA-LUAD train cohort (A), TCGA-LUAD test cohort (B), TCGA-LUAD cohort (C) and e GEO-LUAD cohort (D). The ROC curves to validate the accuracy of the risk model in predicting the clinical outcomes of patients with LUAD (E, F). All of statistical tests were performed using Cox regression analysis According to the Riskscore, calculate the scores of each sample and then divide the GEO cohort and TCGA cohort into two groups based on the median risk score (Fig. 6A, B). Compare the gene expression patterns in different risk groups between the two cohort. The results show that regardless of whether it is in different risk groups in TCGA or GEO cohort similar abnormal gene expressions are observed in both (Fig. 6C, D). Additionally, there is a higher distribution of death cases in the high-risk group (Fig. 6E, F). In addition, after dividing the TCGA cohort into two groups based on patient staging (I–II stage and III–IV stage), we found that the high-risk group exhibited shorter survival (Fig. 6G, H). After further evaluation, we conducted a thorough analysis of the clinical characteristics of two patient groups. In the TCGA-LUAD cohort, significant differences were revealed in gender grading, stage classification, T staging, and N staging between the high-risk and low-risk subgroups(Fig. 6I).Fig. 6Prognostic value of Riskscore. According to the median risk score, TCGA-LUAD cohort was divided into high-risk group and low-risk group. A GEO-LUAD cohort was divided into high-risk group and low-risk group. B Gene expression profiles associated with high and low-risk groups in the TCGA-LUAD cohort prognostic model (C). Gene expression profiles associated with high and low-risk groups in the GEO-LUAD cohort prognostic model (D). The scatter plots are for the TCGA-LUAD cohort and the GEO-LUAD cohort, respectively. E, F Disparities in survival rates persist among high and low-risk subgroups within both the early-stage (G) and Advanced stage (H) cohorts. Differences in clinical characteristics between the high-risk and low-risk groups (I) Prognostic value of Riskscore. According to the median risk score, TCGA-LUAD cohort was divided into high-risk group and low-risk group. A GEO-LUAD cohort was divided into high-risk group and low-risk group. B Gene expression profiles associated with high and low-risk groups in the TCGA-LUAD cohort prognostic model (C). Gene expression profiles associated with high and low-risk groups in the GEO-LUAD cohort prognostic model (D). The scatter plots are for the TCGA-LUAD cohort and the GEO-LUAD cohort, respectively. E, F Disparities in survival rates persist among high and low-risk subgroups within both the early-stage (G) and Advanced stage (H) cohorts. Differences in clinical characteristics between the high-risk and low-risk groups (I) Construction of a Nomogram After univariate Cox regression and Multivariate Cox analysis, we found that the risk score is an independent risk factor and used it to construct a column chart for prediction (Fig. 7A, B). In the column chart, scores for various factors such as age, stage, and risk score were calculated, and the total score can be used as a predictive tool (Fig. 7C). At the same time, a calibration curve was plotted to measure the consistency between actual observed prognosis values and predicted values from the column chart (Fig. 7D). The performance of the column chart was evaluated using ROC curves with AUC values of 0.715 at 1-year survival period, 0.735 at 3-year survival period, and 0.739 at 5-year survival period (Fig. 7E).Fig. 7Diagnostic efficiency of the Nomogram model. A and B Forest plots of the A univariate and B multivariate Cox regression analyses in TCGA LUAD cohorts. A nomogram model constructed based on the TCGA-LUAD cohort was used to assess the 1-, 3-, and 5-year overall survival of LUAD patients. C The calibration plots for the internal validation of the nomogram predicting 1-, 3- and 5-year D. Time-dependent ROC curve demonstrated the ability of the model to predict overall survival at 1-, 3- and 5-year in LUAD patients (E) Diagnostic efficiency of the Nomogram model. A and B Forest plots of the A univariate and B multivariate Cox regression analyses in TCGA LUAD cohorts. A nomogram model constructed based on the TCGA-LUAD cohort was used to assess the 1-, 3-, and 5-year overall survival of LUAD patients. C The calibration plots for the internal validation of the nomogram predicting 1-, 3- and 5-year D. Time-dependent ROC curve demonstrated the ability of the model to predict overall survival at 1-, 3- and 5-year in LUAD patients (E) Clinical differences between high-risk and low-risk groups After analysis, it was found that there is a correlation between the genes used to construct the prognostic model and the risk scores (Fig. 8A), and there are also differences in gene expression between high-risk and low-risk groups (Fig. 8B–H). However, when dividing the high and low expression groups based on the median expression of each gene, we observed survival differences between the high and low expression groups of CAT\CCL20\GPI\GNPNAT1 in TCGA cohort (Fig. 9A–D).Fig. 8Model of gene expression and Riskscore relationships in the TCGA-LUAD cohort. A Differential expression of prognostic genes between high and low risk groups in TCGA-LUAD cohort (B–H)Fig. 9CAT\CCL20\GPI\GNPNAT1 expression and its relationship with overall survival in the TCGA-LUAD cohort (A–D) Model of gene expression and Riskscore relationships in the TCGA-LUAD cohort. A Differential expression of prognostic genes between high and low risk groups in TCGA-LUAD cohort (B–H) CAT\CCL20\GPI\GNPNAT1 expression and its relationship with overall survival in the TCGA-LUAD cohort (A–D) The analysis of the correlation between Riskscore and tumor mutation showed that risk score was positively correlated with tumor (Fig. 10A), and there were differences in TMB between high and low risk groups (Fig. 10B). Further investigation into the relationship between TMB and survival time found that the high mutation group had a shorter survival time (Fig. 10C). According to the division of risk groups, the TCGA cohort was divided into four groups: high-risk high-TMB group, high-risk low-TMB group, low-risk high-TMB group, and low-risk low-TMB group. The results showed that the low-risk high-TMB group had the longest survival time, followed by the low-risk low-TMB group and the high-risk high-TMB group (Fig. 10D). In terms of gene mutations, the analysis of the top 10 genes TP53, TTN, MUC16, CSMD3, RYR2, LRP1B, ZFHX4, USH2A, KRAS and IRP2 showed that the mutation rates of each gene in the high-risk group were significantly higher than those in the low-risk group (Fig. 10E, F).Fig. 10The tumor mutation burden characteristics in low- and high-risk group. Correlation analysis between Riskscore and TMB. (A) The difference in TMB between two groups. B Survival difference between high- and low-TMB. C Survival analysis of TMB along with Riskscore. D Mutational landscape in the high-risk group. E The difference in TMB between two groups. F TMB: Tumor mutational burden The tumor mutation burden characteristics in low- and high-risk group. Correlation analysis between Riskscore and TMB. (A) The difference in TMB between two groups. B Survival difference between high- and low-TMB. C Survival analysis of TMB along with Riskscore. D Mutational landscape in the high-risk group. E The difference in TMB between two groups. F TMB: Tumor mutational burden To further investigate immune infiltration, we employed the MCPcounter method. We calculated the high- and low-risk groups, risk scores, and the population abundance of stromal cell populations. The results showed a correlation between B-cell lineage, T-cell, myeloid cell, endothelial cell, and fibroblast with Riskscore (Fig. 11A). In the high- and low-risk groups, the abundance of myeloid cells, endothelial cells, fibroblasts, and neutrophils varied (Fig. 11B–E).Fig. 11Immuno-infiltration analysis and Riskscore relationships in the TCGA-LUAD cohort (A). The difference in myeloid cells, endothelial cells, fibroblasts, and neutrophils between low- and high-risk group in TCGA-LUAD (B-E) Immuno-infiltration analysis and Riskscore relationships in the TCGA-LUAD cohort (A). The difference in myeloid cells, endothelial cells, fibroblasts, and neutrophils between low- and high-risk group in TCGA-LUAD (B-E) Potential drug sensitivity analysis We used TIDE and TICA models to predict clinical differences between high-risk and low-risk groups, and used IMvigor210 as an external validation set. According to TIDE scores, high-risk groups showed higher TIDE score and Worse treatment response than low-risk groups (Fig. 12A, B). According to TICA scores, there was no statistically significant difference in AIPS scores in PD1 positive groups, but in PD1 negative groups, regardless of CALT4 expression, low-risk groups had higher scores (Fig. 12C). External validation on the IMvigor210 cohort showed that patients classified as high-risk groups had shorter survival and (Fig. 12D). The efficacy of immunotherapy was also assessed. The high-risk group demonstrated a higher likelihood of disease progression (PD) and disease stability (SD), whereas the low-risk group exhibited a greater probability of achieving complete remission (CR) and partial remission (PR) (Fig. 12E).Fig. 12Immunotherapy response of low- and high-risk TCGA-LUAD cohort. A, B Difference of the TIDE . C TCIA score of the low- and high-risk group in the of CTLA4- PD1-, CTLA4- PD1 + , CTLA + PD1−, and CTLA + PD1 + subgroups. D Kaplan–Meier curves for Prognostic Model in the IMvigor210. E Differences in Riskscore between patients with CR/PR and individuals with SD/PD after treatment with ICIs in the IMvigor210 Immunotherapy response of low- and high-risk TCGA-LUAD cohort. A, B Difference of the TIDE . C TCIA score of the low- and high-risk group in the of CTLA4- PD1-, CTLA4- PD1 + , CTLA + PD1−, and CTLA + PD1 + subgroups. D Kaplan–Meier curves for Prognostic Model in the IMvigor210. E Differences in Riskscore between patients with CR/PR and individuals with SD/PD after treatment with ICIs in the IMvigor210 The OncoPredict was used to conduct drug sensitivity analysis on high- and low-risk groups, specifically for common chemotherapy regimens of lung adenocarcinoma. The results showed that there were varying degrees of response differences among commonly used chemotherapy drugs such as paclitaxel (Fig. 13A), docetaxel (Fig. 13B), cisplatin (Fig. 13C), gemcitabine (Fig. 13D), and vinorelbine (Fig. 13E) in the high- and low-risk groups. These drugs are commonly used chemotherapy drugs for lung adenocarcinoma and showed higher scores in the low-risk group.Fig. 13Prediction of chemotherapeutic drugs in high-risk and low-risk groups. A–E The boxplot shows drugs with different IC50 values between the high and low risk groups Prediction of chemotherapeutic drugs in high-risk and low-risk groups. A–E The boxplot shows drugs with different IC50 values between the high and low risk groups Discussion In this study, we combined immune and metabolism-related genes and identified two molecular subtypes through MNF non-negative matrix factorization. Subsequently, we established gene prognosis models related to immunity or metabolism using univariate Cox regression analysis, multivariate Cox regression analysis, and LASSO Cox regression analysis. Univariate and multivariate Cox regression analyses were conducted on both the Riskscore and various clinical features, demonstrating that the Riskscore serves as an independent prognostic factor irrespective of other clinical features. Furthermore, the hazard ratios (HR) > 1, indicating that a high Riskscore is an unfavorable prognostic indicator. Patients were divided into high- and low-risk groups based on the median risk score, and differences in overall survival were observed among these groups as well as within each subtype. Based on these results, we further plotted a prognostic nomogram incorporating patient characteristics such as gender, age, T stage, N stage, and M stage which demonstrated good predictive efficacy for patient prognosis. Additionally, CAT, CCL20, GPI and GNPNAT1 showed a close correlation with survival time. These findings suggest that immune- and metabolism-related genes play important roles in the occurrence and development of LUAD. The CAT gene encodes catalas. Disrupting the ROS balance in tumor tissue can disturb its homeostasis and inhibit tumor growth. Meanwhile, delivering catalase to the lesion through different carriers can enhance the tumor treatment response [20, 21]. CCL20 is the ligand for CCR6 [20]. Activation of the CCL20/CCR6/ERK signaling pathway through autocrine or paracrine mechanisms promotes cell proliferation and migration [22]. Related studies have confirmed the high expression of CCL20 and CCR6 in lung cancer tissues, which may promote tumor cell proliferation and migration through autocrine or paracrine mechanisms [23] and its elevated levels may be associated with poor prognosis [24, 25]. GPI plays an important role in glycolysis and gluconeogenesis. Abnormal expression of GPI has been associated with metastasis and poor prognosis in colorectal cancer, renal cell carcinoma, breast cancer, and endometrial cancer [26, 27]. Additionally, decreased expression of GPI may lead to increased oxidative stress sensitivity, resulting in age-related death associated with oxidative stress [28]. GNPNAT1 is a key enzyme in the biosynthesis pathway of uridine diphosphate-N-acetylglucosamine.glutamine plays a crucial role in tumor tissues and its end product uridine diphosphate-N-acetylglucosamine has critical functions [29]. Enhancing the expression of this gene can promote malignant phenotypes of cancer cells and is associated with platinum resistance [30] and Kras mutations [31]. In vitro experiments have also confirmed the relationship between its expression levels in lung cancer cells and prognosis [32, 33]. INSL4 belongs to the insulin superfamily and was first identified in the human placenta [34]. It has been found to have a promoting effect on cancer [35], but due to its specificity as a gene in primates, there have not been any in vivo experiments conducted to validate its mechanism. NT5E encoding (CD73) catalyzes the dephosphorylation of nucleoside 5'-monophosphate, converting it into adenosine. Relevant studies have found that overexpression of this gene is associated with the collaborative action of immune cells in the immune microenvironment of lung cancer [36], affecting immune evasion mechanisms. The protein encoded by GSTA3 is GST. The ability of cells to transform toxic chemicals and endogenous substances is correlated with the expression level of GST [37]. We also examined the mutation status of TCGA samples and discovered that the high-risk group exhibited a greater frequency of mutations. Specifically, TP53 displayed an elevated mutation rate within this high-risk cohort, which is correlated with poorer prognosis [38, 39]. Within this same high-risk group, 10 genes demonstrated a mutation rate surpassing 30%, whereas only 5 genes did so within the low-risk group. Furthermore, survival analysis indicated that patients with increased tumor mutational burden (TMB) experienced prolonged overall survival time. This could potentially be attributed to the fact that heightened TMB encompasses more neoantigens and possesses enhanced immunogenicity, thereby leading to improved treatment response [40]. Chemotherapy and immunotherapy are important treatment strategies for LUAD. We used the oncoPredict R package to predict the responsiveness of high- and low-risk groups to common chemotherapy drugs. The results showed that compared to the high-risk group, the low-risk group performed better in terms of sensitivity scores for commonly used chemotherapy drugs such as vinorelbine, docetaxel, paclitaxel, and cisplatin. Additionally, we utilized the TCIA\TIDE model to predict the effectiveness of immunotherapy and found that the low-risk group exhibited a better immune response. By using the IMvigor210 immune therapy cohort as an external validation dataset, we obtained consistent results. Specifically, patients in the low-risk group had a higher response rate and longer survival when treated with PDL1 immune therapy. This study is based on the TCGA and GEO databases, and it classifies lung adenocarcinoma based on two key features: immunity and metabolism. A risk scoring system consisting of seven genes was identified. In the high-risk group, patients had shorter overall survival and poor response to chemotherapy and immunotherapy. This model takes into account both patient prognosis and efficacy prediction of common treatment options, making it convenient with broad application prospects and high research value. However, it should be noted that this model still has limitations. For example, there may be differences between different sequencing platforms which could lead to errors. Additionally, the pathway mechanisms involved in this study have not been experimentally validated, which may reduce the credibility of the results. Although this study is retrospective in design, it does establish new prognostic and treatment prediction models. In the future, we can further explore the interactions among genes composing the predictive model as well as downstream regulation to better understand the immune-metabolic crosstalk mechanism in LUAD. In conclusion, we have successfully developed a predictive model incorporating immune and metabolism-related genes, encompassing gene expression levels of CAT/CCL20/GPI/INSL4 NT5E/GSTA3/GNPNAT1. This comprehensive model not only enables the prognosis prediction for LUAD patients but also facilitates the prediction of their response to first-line chemotherapy drugs and immune checkpoint inhibitors, thus demonstrating its broad potential in clinical applications. Conclusion In conclusion, we have successfully developed a predictive model incorporating immune and metabolism-related genes, encompassing gene expression levels of CAT/CCL20/GPI/INSL4 NT5E/GSTA3/GNPNAT1. This comprehensive model not only enables the prognosis prediction for LUAD patients but also facilitates the prediction of their response to first-line chemotherapy drugs and immune checkpoint inhibitors, thus demonstrating its broad potential in clinical applications. However, our study still has limitations as it is based on TCGA and GEO databases with limited pathological characteristics of patients. Therefore, more practical and valuable factors are needed to predict efficacy. The crosstalk between metabolism and immunity remains to be explored. Finally, this study lacks experimental evidence for the underlying gene expression of prognosis and further research is required.