Data from 420 LUAD and 59 non-neoplastic lung tissues, including RNA sequencing, whole exome sequencing (WES), and patient clinical information, were retrieved from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/) in June 2020. The “maftools” R package was used to analyze WES data for somatic variants. Immunohistochemical staining (IHC) corresponding to TCGA-LUAD patients was acquired from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/) and the Clinical Proteomic Tumor Analysis Consortium (CPTAC) database (https://proteomics.cancer.gov/programs/cptac/). Microarray profiles of mRNA expression and clinical information of 353 patients were combined from the Gene Expression Omnibus (GEO) datasets GSE31210 and GSE50081. Another database containing data on 443 patients with LUAD was obtained from the GEO dataset GSE68465. We also obtained information from 24 non-small cell lung cancer (NSCLC) patients treated with PD-1 blockade combined with chemotherapy from the GEO GSE207422 dataset.

A clinical study on patients with advanced or recurrent stage III/IV NSCLC who were treated with cytotoxic chemotherapeutic reagents in combination with ICI therapy (pembrolizumab or atezolizumab) was performed from 2020 to 2022 at Kanagawa Cancer Center (KCC, Yokohama, Japan) and Kurume University Hospital (KU, Fukuoka, Japan). The treatment response of the patients was determined according to the Response Evaluation Criteria in Solid Tumors (RECIST) 1.1. From the cohort, 20 patients whose formalin-fixed paraffin-embedded (FFPE) tumor tissues were available for RNA sequencing, were involved in the present study. We named this cohort the KCC-ICI. For these patients, the concentrations of PFAAs in the peripheral venous blood before the start of treatment were examined. OS was defined as the period from the date of treatment to the date of death from any cause.

To comprehensively evaluate the role of amino acid metabolism, we identified genes related to amino acid metabolism. Briefly, all human genes related to 14 amino acid metabolic pathways were listed in the Kyoto Encyclopedia of Genes and Genomes (KEGG) (https://www.genome.jp/kegg/). The resultant 293 genes were designated as AAMGs. The entire list of genes by gene symbols is provided in Supplementary Table 1 .

The “DESeq2” Bioconductor R package (v1.28.1) was used to normalize the process and identify the DEGs between the two groups from the RNA sequencing data (fold change > 2, P _adj < 0.05). Gene set enrichment analysis (GSEA) was performed based on the Gene Ontology (GO), KEGG, and Reactome Pathway (REACTOME) databases. The R packages “clusterProfiler” and “ReactomePA” were used for enrichment analysis and visualization of the results. Statistical significance was set at a normalized enrichment score |(NES)| > 1 and FDR < 0.05. Gene set variation analysis (GSVA) was performed to identify significantly correlated pathways using a reference gene set “c2.cp.kegg.v7.4. symbols.gmt” was downloaded from the GSEA website (https://www.gsea-msigdb.org/gsea/downloads.jsp), with an FDR < 0.05. The abundance of immune cell infiltration, stroma score, and immune score were determined by the ESTIMATE algorithm using RNA sequencing data.

Drug response data for cytotoxic drugs in human cancer cell lines and the corresponding genomic markers were downloaded from the Genomics of Drug Sensitivity in Cancer (GDSC) website (http://www.cancerrxgene.org) as the GDSC v2 dataset. To predict the drug response profiles for patients involved in TCGA-LUAD cohort, analysis using the R package “oncoPredict” was applied to the GDSC v2 and RNA sequencing data of each LUAD tumor of the corresponding patient. The half-maximal inhibitory concentration (IC₅₀) values of drugs were calculated for each patient, which were used to speculate how a drug inhibits certain biological or metabolic processes (21). Response to ICIs was predicted using the tumor immune dysfunction and exclusion (TIDE) algorithm based on the gene expression related to T cell dysfunction and exclusion, obtained from the TIDE website (http://tide.dfci.harvard.edu/) (22).

For total RNA isolation from FFPE tissues of patients, we first marked the area of the tumor on the hematoxylin-eosin-stained section of each tissue block. Then, total RNA was extracted from the corresponding tumor area on the unstained serially sliced sections with macrodissection using RNeasy FFPE Kit (Qiagen, Hilden, Germany), according to the manufacturer’s instructions. The amount of RNA was measured using a NanoDrop1000 (Thermo Scientific, Wilmington, DE, USA) and RNA integrity was assessed using the RNA Nano 6000 Assay Kit of the Agilent Bioanalyzer 2100 system (Agilent Technologies, Santa Clara, CA). RNA samples with 30% or greater DV200 values were subjected to RNA sequencing. Construction of cDNA libraries followed by RNA sequencing was performed by Takara Bio Inc. (Shiga, Japan) as a contract analysis using a SMART-Seq Stranded mRNA Kit (Clontech, Palo Alto, CA, USA) and a NovaSeq sequencing system (Illumina, San Diego, CA, USA) according to the manufacturer’s instructions. After confirming the read quality using FastQC, the sequence data were aligned to the human genome GRCh37 using STAR-2.5.2a (https://github.com/alexdobin/STAR/releases/tag/2.5.2a), and the mapped read count of each sample was calculated using Python 2.7. Transcripts per million (TPM) were calculated for transcription quantification using Salmon-1.1.0 (https://combine-lab.github.io/salmon/) with GRCh38.v99 as the reference index. Salmon estimated TPM data were summarized to gene expression levels using the “tximport” package(v1.22.0) for correlation analysis. Mapping to the reference by STAR-2.5.2a and a quality check of reads for each sample were performed using the Genomon 2 analysis pipeline (https://github.com/Genomon-Project).

Peripheral blood samples from the KCC-ICI cohort were collected in the morning in an overnight fasting state from the antecubital vein into tubes containing EDTA-2Na and immediately placed on ice. Plasma was separated via centrifugation at 3000 rpm for 15 min at 4°C and stored at −80°C until analysis. After thawing, plasma samples were deproteinized using acetonitrile at a final concentration of 50% before measuring amino acid concentrations using high-performance liquid chromatography–electrospray ionization mass spectrometry via precolumn derivatization, as described previously (20). Concentrations of the following 21 PFAAs were measured: alanine, arginine, asparagine, citrulline, α-aminobutyric acid (AABA), glutamate, glutamine, glycine, histidine, isoleucine, leucine, lysine, methionine, ornithine, phenylalanine, proline, serine, threonine, tryptophan, tyrosine, and valine. The amino acid concentration was determined by summing the concentrations of each of the 21 PFAAs.

For all statistical analyses, R version 4.1.0 was used, unless otherwise noted. Heatmaps were drawn by the “ComplexHeatmap” package with Spearman as a distance indicator; for the generation of Kaplan–Meier (KM) curves and calculation of Cox proportional hazard ratios, packages of the “survplot” and the “survminer” were used. Hazard ratios (HRs) and 95% confidence intervals (CIs) were calculated using the “Coxph” function. For the Concordance Index (C-index) and time-dependent AUC of receiver operating characteristic (ROC), the packages of the “dplyr” and the “survivalROC” were used. Univariate Cox regression and least absolute shrinkage and selection operator (LASSO) Cox regression analyses were performed by using the “glmnet” package in R. Wilcoxon rank sum test was used for differential significance test between the two groups. Categorical variables between groups were compared using Fisher’s exact chi-square test. A schematic overview of the study design is provided in Supplementary Figure 1 .

We compared RNA gene expression between 420 TCGA-LUAD tumor samples and 59 non-neoplastic lung samples, and first identified 4543 significant DEGs (fold change > 2, P_adj < 0.05). We further identified 76 DEGs that appeared in the list of AAMGs and designated them as tumor-specific AAMGs (TAAMGs). Among the 76 TAAMGs, 61 genes were upregulated and 15 were downregulated in tumor tissues compared to those in non-neoplastic lung tissues ( Supplementary Figure 2A ). To gain insight into the function of TAAMGs, enrichment analysis with KEGG pathways was carried out; cysteine and methionine metabolism; arginine and proline metabolism; tyrosine metabolism; glycine, serine, and threonine metabolism; alanine, aspartate, and glutamate metabolism; biosynthesis of amino acids; and tryptophan metabolism were the most enriched terms ( Supplementary Figure 2B ).

To identify TAAMGs related to patient prognosis, we first conducted a univariate Cox proportional hazards regression analysis using survival data of TCGA-LUAD patients and identified 16 candidate genes associated with OS with a P < 0.05 (these prognosis-related genes were designated as “PTAAMGs”) ( Supplementary Table 2 ). Next, we established a prognostic signature composed of nine key PTAAMGs (KYNU, PSPH, PPAT, MIF, GCLC, ACAD8, TYRP1, ALDH2, and HDC) by selecting the mostly marked genes with the optimal value of tuning parameter (λ) by ten-time cross-validation using minimum criteria in LASSO Cox regression analysis. In addition, we performed a collinearity test to check the independence of the key PTAAMGs. The results showed that the multicollinearity assumption was not violated when the variance inflation factor (VIF) less than two. We named this novel prognostic signature model as “PTAAMG-Sig”, and the risk score of each patient was calculated using the expression values (EV) of optimized genes and their multivariate Cox regression correlation coefficients. The PTAAMG-Sig formulation was as follows: 0.0717*EV(PPAT) + 0.0649*EV(MIF) + 0.0095*EV(GCLC) + 0.1195*EV(PSPH) + 0.1262EV(KYNU) – 0.1394EV(ALDH2) – 0.0422EV(ACAD8) – 0.1782 EV(HDC) – 0.1274 EV(TYRP1). Multivariate Cox regression analysis of the key PTAAMGs comprising the signature showed that the expression of KYNU, PSPH, PPAT, MIF, and GCLC contributed to poorer OS with HRs > 1, whereas those of ACAD8, TYRP2, ALDH2, and HDC were associated with better OS with HRs < 1. Only KYNU had a P-value less than 0.05 (P = 0.022), indicating that it was a risk factor of this prognosis-predicting model ( Figure 1A ). The KM curves for each of the nine key PTAAMG expression levels revealed a significant association with OS ( Supplementary Figure 3 ). When PTAAMG-Sig was applied to dichotomize TCGA-LUAD dataset with the risk score, 52 patients were classified into the high-risk group, with a value of −0.366 as the risk score cutoff. The high-risk group had significantly shorter OS than the low-risk group. Thus, PTAAMG-Sig efficiently stratified patient OS (C-index = 0.641, HR = 2.718, 95% CI 1.937–3.815, Log-rank P < 0.0001) ( Figure 1B ). We evaluated the signature using the ROC curve and calculated the area under the curve (AUC) at different time points (1, 3, and 5 years). The maximum AUC (AUC_max) of PTAAMG-Sig was 0.730 at the 1-year OS time point ( Figure 1C ). We found that patients who died showed a significantly higher risk score than those who survived (Wilcoxon rank sum test, P < 0.01), indicating the effect of the risk score by PTAAMG-Sig on survival status ( Figure 1D ). These data suggest that the model efficiently stratifies the OS of LUAD patients. The expression patterns of the key PTAAMGs genes used in PTAAMG-Sig showed that the expression levels of KYNU, PSPH, PPAT, MIF, and GCLC were significantly higher in the high-risk group, whereas those of ACAD8, TYRP2, ALDH2, and HDC were significantly higher in the low-risk group (Wilcoxon rank sum test, P < 0.0001) ( Figure 1E ).

Based on the analysis of the HPA and CPTAC databases, we found that the protein expression levels of PPAT, KYNU, PSPH, ALDH2, and MIF evaluated by IHC in LUAD tumor tissues were consistent with the mRNA expression levels of each molecule; that is, high expression of PPAT, KYNU, PSPH, and MIF, as well as low ALDH2 expression, was observed in tumor tissues compared with those in the adjacent non-neoplastic lung tissues ( Supplementary Figures 4A, B ). In addition, we checked the mutation landscape of the nine genes in TCGA-LUAD tumor tissues. In total, 9.6% of patients had nonsynonymous gene mutations, such as missense, nonsense, frameshift insertion, or splice site mutations ( Supplementary Figure 5A ). KYNU had the highest mutation rate (3%), followed by HDC (2%). PPAT and MIF showed no mutations in TCGA-LUAD samples. The results of the multi-omics analyses revealed that the nine key genes exhibited consistent expression patterns at the protein and mRNA levels and had low mutation rates.

To validate the predictive capability of PTAAMG-Sig, we investigated two independent lung cancer cohorts 2 public GEO cohorts of 353 patients (GSE31210 and GSE50081) and another GEO cohort of 443 patients (GSE68465). The combined GSE31210 and GSE50081 cohorts did not enroll patients at stage III and more advanced stages, whereas the GSE68465 cohort included stage III and stage IV patients with patients at the earlier stages. Detailed clinical information for TCGA and GEO cohorts is presented in Table 1 .

After removing the batch effect of the gene expression files, we performed Cox regression survival analysis in each cohort. In the GEO cohorts, PTAAMG-Sig stratified patients into high-risk (N = 95) and low-risk (N = 258) groups with a significant correlation to OS (HR = 2.434, 95% CI 1.176–4.497, Log-rank P = 0.012) ( Figure 2A ). Time-dependent ROC analysis showed a maximum predictive accuracy of 0.734 at the 2-year time point ( Figure 2B ). Compared with the alive patients, the dead patients showed significantly higher risk scores (Wilcoxon rank sum test, P < 0.001) ( Figure 2C ). Further validation of the signature with another GEO cohort, GSE68465, was performed because the cohort included stages I–IV, and the composition of patient stages was closer to that of TCGA patients. PTAAMG-Sig effectively stratified patients in the GSE68465 into high-risk (N = 179) and low-risk (N = 264) groups and revealed a significant association with OS (HR = 1.635, 95% CI 1.261–2.119, log-rank P < 0.001) ( Figure 2D ). Time-dependent ROC analysis showed a peak predictive accuracy of 0.695 at one year ( Figure 2E ). Patients who died had significantly higher risk scores than those who survived (Wilcoxon rank sum test, P < 0.0001) ( Figure 2F ).

TCGA cohort, and the GEO cohorts used for the validation of the PTAAMG-Sig patients were those before introduction of ICI to clinics. Therefore, we interrogated the efficacy of the PTAAMG-Sig in our original cohort enrolled 20 patients with NSCLC at advanced stages III/IV or with recurrence who received combined cytotoxic reagents and ICI therapy (KCC-ICI cohort). Univariate Cox regression analysis revealed no significant association between OS and sex, age, or stage (P > 0.05) ( Supplementary Table 3 ). The patients were divided into high-risk (N = 7) and low-risk (N = 13) groups based on PTAAMG-Sig, and a good stratification ability of OS was confirmed in this cohort (HR = 4.976, 95% CI 2.265–13.379, Log-rank P = 0.0087) ( Figure 3A ). The maximum predictive accuracy of this signature was 0.842 at the 330-day time point ( Figure 3B ). Surviving patients showed substantially lower risk scores than those who died owing to PTAAMG-Sig (Wilcoxon rank sum test, P < 0.05) ( Figure 3C ).

To partly strengthen the results obtained from this small cohort, we analyzed the GEO dataset GSE207422, which enrolled 24 patients with NSCLC who received neoadjuvant PD-1 blockade in combination with chemotherapy followed by surgical resection of the tumors ( Table 1 ). This dataset did not include information on OS but considered information on pathological responses to treatment after resection of the tumor. Nine patients were categorized as responders (with a major pathologic response defined as less than or equal to 10% viable tumor cells identified), and 15 were categorized as non-responders (non-major pathologic response). The responders had significantly lower PTAAMG-Sig risk scores than the non-responders (Wilcoxon rank sum test, P < 0.05) ( Supplementary Figure 6A ). Of patients with the PTAAMG-Sig high-risk scores, 28.6% were responders to ICI treatment, whereas all patients with low-risk scores were responders, and the difference was significant (chi-square test, P < 0.01) ( Supplementary Figure 6B ). We further combined the KCC_ICI and GSE207422 datasets and analyzed them as a cohort of 44 patients. The responders had significantly lower PTAAMG-Sig risk scores than the non-responders (Wilcoxon rank sum test, P < 0.05) ( Figure 3D ). Of the patients with PTAAMG-Sig high-risk scores, 51.9% were non-responders to ICI treatment, whereas 17.6% of the patients with low-risk scores were non-responders, and the difference was significant (chi-square test, P < 0.01) ( Figure 3E ).

Further evaluation of PTAAMG-Sig was performed in subgroups stratified by PD-L1 expression level in tumors. High PD-L1 expression with a tumor proportion score (TPS) ≥ 50% (n = 8) and PD-L1 positive expression with a TPS ≥ 1% (n = 13) were not significantly associated with OS ( Supplementary Figure 7A, B ). Combined with PTAAMA-Sig, we found that the high-risk patients in the low PD-L1 group had the shortest OS (log-rank P < 0.05) ( Supplementary Figure 7C ).

The somatic non-synonymous mutation profiles of PTAAMG-Sig high- and low-risk groups in TCGA-LUAD were analyzed. Among the top 20 genes in terms of mutation frequency in each risk group, 13 genes, including TP53, TTN, CSMD3, RYR2, ZFHX4, LRP1B, USH2A, MUC17, MUC16, SPTA1, NAV3, FLG, and XIRP2 were shared between the two risk groups ( Figure 4A ). The rate of TP53 (71% vs. 47%, P < 0.001), MUC17 (40% vs.18%, P < 0.0001), TTN (69% vs. 43%, P < 0.0001), and PCLO (33% vs. 16%, P < 0.001) mutations were significantly higher in the high-risk group, whereas the difference in mutations of KRAS (23% vs. 26%), TNR (23% vs. 15%), PCDH11X (23% vs. 13%), and ANK2 (23% vs. 18%) were not significant ( Figure 4B ). We further analyzed the relationship between the significantly differentially mutated genes and OS using univariate Cox regression analysis in the high- and low-risk groups. The results showed that KRAS, MUC17, TTN, TNR, PCDH11X, and PCLO mutations were positively correlated with long OS, whereas TP53 mutation was negatively correlated with long OS in the high-risk group (HR: 3.740, Log-rank P = 0.0046) ( Figure 4C ). In the low-risk group, ANK2 mutations were markedly associated with OS prognosis, but not with other factors ( Supplementary Table 4 ). Notably, no significant differences occurred between the two groups in the mutation rates of the nine key PTAAMGs genes used in the signature ( Supplementary Figure 5B ).

Although the tumor mutation burden (TMB) was considered a biomarker for the prognosis of various cancers, TMB did not predict OS prognosis in the TCGA training cohort. That is, the OS of the low- and high TMB groups (divided with a cutoff value of –0.167, calculated as log2TMB/Mb, based on the optimal AUC cutoff) was not significantly different (HR = 2.503, 95% CI 0.061–10.334, log-rank P = 0.190) ( Supplementary Figure 7D ). However, we found that patients classified as high-risk had a greater TMB than those in the low-risk group (Wilcoxon rank sum test, P < 0.0001) ( Supplementary Figure 7E ). PTAAMG-Sig and TMB effectively stratified patients into high-risk with high TMB levels (N = 52), low-risk with high TMB levels (N = 323), and low-risk with low TMB levels (N = 41). Compared with low-risk with lower TMB levels, high-risk patients with higher TMB levels had poorer OS rates and showed a significant association with OS (HR = 8.302, 95% CI 1.917–35.952, log-rank P < 0.0001) ( Supplementary Figure 7F ). These results indicate that TMB levels were one of the factors influencing the differentiation between the high and low PTAAMG-Sig risk groups and that PTAAMA-Sig was useful for stratifying prognosis in the subgroup with higher TMB levels.

To explore the relationship between PTAAMG-Sig and clinical information, we combined its signature with common clinical factors from TCGA-LUAD dataset. In the univariate Cox regression analysis, sex, age, and metastasis status (M) were not significantly associated with OS. In contrast, PTAAMG-Sig, clinical stage, T, and N were significantly associated with OS (P < 0.05) ( Table 2 ). Multivariate Cox regression analysis revealed that PTAAMG-Sig (HR: 2.297, 95% CI 1.600–3.296, P < 0.001) and clinical stage (HR: 1.329, 95% CI 1.065–1.659, P = 0.0118) were independent predictors of OS ( Figure 5A ). Smoking is an important risk factor for lung cancer. We divided the patients into non-smokers (less than 100 cigarettes smoked in their lifetime) and smokers (including current smokers and current reformed smokers). Tobacco smoking history did not predict the OS prognosis. OS was not significantly different between smokers and non-smokers (HR = 0.832, 95% CI: 0.544–1.174, log-rank P = 0.400) ( Supplementary Figure 8A ). Smokers and non-smokers showed no significant difference in the PTAAMG-Sig risk distribution (chi-square test, NS: P > 0.05) ( Supplementary Figure 8B ). In combination with PTAAMA-Sig, we found that in smokers, high-risk patients showed a significantly worse OS outcome than low-risk patients (HR = 3.399, 95% CI: 2.278–5.073, log-rank P < 0.0001) but not in non-smokers (HR = 2.893, 95% CI: 0.844–9.920, log-rank P = 0.0771) ( Supplementary Figures 8C, D ).

As PTAAMG-Sig could predict the prognosis of the LUAD patients in three independent cohorts with different treatment modalities, we further investigated the relationship between PTAAMG-Sig and the prediction of response to chemotherapy and ICI therapy. The drug sensitivity prediction of TCGA-LUAD patients was performed on the “oncoPredict” algorithm utilizing the data from the GDSC v2 database. Spearman rank correlation analysis demonstrated that the predicted IC₅₀ value of doramapimod (p38 MAPK inhibitor), BMS-754807 (IGF-1R/InsR inhibitor), GSK269962A (ROCK inhibitor), PF-4708671 (cell-permeable S6K1 inhibitor), JQ1 (BET bromodomain inhibitor), SB216763 (ATP-competitive GSK-3 inhibitor), uprosertib (AKT inhibitor), and axitinib (multi-target inhibitor for VEGFR1, VEGFR2, VEGFR3, and PDGFRβ) were positively correlated with the PTAAMG-Sig risk score, indicating that the PTAAMG-Sig high-risk group had a relatively higher resistance to these therapeutics than the low-risk group. Conversely, the PTAAMG-Sig high-risk group showed higher sensitivity to docetaxel (microtubule depolymerization inhibitor, docetaxel^a and docetaxel^b), savolitinib (c-MET inhibitor), BI-2536 (PLK1 and BPD4 inhibitor), paclitaxel (microtubule stabilizer), AZD7762 (ATP-competitive Chk inhibitor), AZD6738 (ATR kinase inhibitor), and MK-1775 (Wee1 inhibitor) than the low-risk group ( Figures 5B, C ).

Because genomic mutations can alter tumor immune profiles and response to immunotherapy, we further evaluated the predictive capability of PTAAMG-Sig for response to ICI therapy in TCGA-LUAD patients using the TIDE algorithm. TIDE predicted that 84.6% of TCGA-LUAD patients with PTAAMG-Sig high-risk scores were non-responders to ICI treatment, whereas 61.4% of LUAD patients with low-risk scores were non-responders, and the difference was significant (chi-square test, P < 0.01) ( Figure 5D ). This indicates that immunotherapy was more likely to be successful in patients in the PTAAMG-Sig low-risk group. In fact, TIDE-predicted non-responders showed significantly higher PTAAMG-Sig risk scores than responders (Wilcoxon rank sum test, P < 0.05) ( Figure 5E ).

To understand the functional significance of the PTAAMG-Sig signature, stromal and immune scores were analyzed using the ESTIMATE algorithm. The PTAAMG-Sig low-risk group showed significantly higher stromal and immune scores than the high-risk group, indicating that the tumors in the low-risk group had a TME with high stromal and immune activity ( Supplementary Figure 9A ). The DEGs between the high- and low-risk groups identified using PTAAMG-Sig were subjected to GSEA for the GO dataset. We identified that the “Metabolism of amino acids and derivatives” pathway was positively enriched in the PTAAMG-Sig high-risk group (NES = −1.691, FDR < 0.001) ( Supplementary Figure 9B ). Furthermore, in the high-risk group, pathways related to “DNA replication”, “cell cycle checkpoints”, and “apoptosis” were significantly positively enriched. In the low-risk group, the “SLC-mediated transmembrane transport” pathway was significantly enriched, which is possibly related to altered amino acid transmembrane transport and metabolism. “PD-1 signaling”, “Phosphorylation of CD3 and TCR zeta chains”, and “Cytokine-cytokine receptor interaction” pathways correlated with immune activity and ICIs treatment process were also significantly highly enriched in the low-risk group ( Supplementary Figure 9C , Supplementary Table 5 ). These results demonstrate the predictive capability of this signature for ICIs therapy.

The relationship between the PTAAMG-Sig risk scores and OS in the KCC-ICI cohort treated with combined chemotherapy and ICIs therapy is shown in Figure 3 . We further estimated the ability of this signature to predict patient responses to this therapy. The responders in the treatment group had significantly lower PTAAMG-Sig risk scores than non-responders (Wilcoxon rank sum test, P < 0.05) ( Figure 6A ).

Transcriptome analysis of the tumor FFPE specimens in this cohort demonstrated that the PTAAMG-Sig low-risk group exhibited a significantly higher immune score than the high-risk group. However, the stromal scores did not differ, indicating elevated immune activity in the low-risk group (Wilcoxon rank sum test, P < 0.05) ( Figure 6B ). To address the apparent differences in immune function between them, we performed GSVA using the Biocarta, KEGG, and Reactome databases. The PTAAMG-Sig low-risk group showed several significantly enriched pathways related to immune cell regulation, including T cell activation, inflamed status, and some other immune-related pathways, such as “Classic pathway”, “Graft versus host disease”, “Allograft rejection”, and “Diseases associated with surfactant metabolism” ( Figure 6C , left). The heatmap exhibits Spearman’s rank correlations between the PTAAMG-Sig risk scores and the enrichment scores of the significantly altered pathways ( Figure 6C , right). Among the pathways, the scatter plot analysis identified the “NO2-IL12 pathway” (R = −0.525, P = 0.0175), “T cytotoxic pathway” (R = −0.516, P = 0.0199), and “T helper pathway” (R = −0.478, P = 0.0330) as significantly negatively correlated pathways with the PTAAMG-Sig risk scores ( Figures 6C, D ). Subsequently, we investigated the differences in T, NK, and NK-T cells between the PTAAMG-Sig low- and high-risk groups by determining the infiltrating immune cell population in the TME, deconvoluted from the results of RNA sequencing. Activated CD8 T, central memory CD4 T, central memory CD8 T, NK, and NK-T cells were significantly increased in the low-risk group (Wilcoxon rank sum test, P < 0.05) ( Figure 6E ).

The PTAAMG-Sig model was developed from the expression profiles of genes related to amino acid metabolism pathways, based on their significance in tumor biology and immunity. We examined the relationship between PFAA concentration and PTAAMG-Sig risk scores in a KCC-ICI cohort treated with combined chemotherapy and ICI therapy. AABA was the only amino acid whose concentration differed significantly between the high- and low-risk PTAAG-Sig groups, with a lower concentration in the high-risk group (Wilcoxon rank sum test, P < 0.05) ( Figure 6F , Supplementary Figure 10A ). To address the target factors, we performed Spearman rank correlation analysis, and the results showed that the concentration of AABA was significantly negatively correlated with MIF gene expression levels in TME (R=-0.459, P=0.0419) ( Supplementary Figures 10B, C ).