Bulk RNA-sequencing data and associated clinicopathological details for GC samples were sourced from TCGA official data portal (https://gdc portal.nci.nih.gov/). The mRNA and long noncoding RNA transcriptome profiles were obtained via the TCGAbiolinks package (https://bioconductor.org/packages/release/bioc/html/TCGAbiolinks.html) (17). Additionally, clinicopathological data and comprehensive expression data were obtained from three validated GC cohorts: GSE15459 (18), GSE84433 (19) and GSE84437 (19). Furthermore, two chemotherapy datasets were used to assess the contribution of the α-KG-related gene index (AKGI) to the prediction of therapy benefits: GSE26899 (20) and GSE13861 (21). Raw count data underwent transcripts per million normalization for standardization.

To construct the signatures list of α-KG-related genes, genes associated with eight distinct α-KG-related pathways were obtained from trusted scientific databases, including Gene Set Enrichment Analysis (GSEA) gene sets (https://www.gsea-msigdb.org/gsea/index.jsp), Kyoto Encyclopedia of Genes and Genomes (KEGG) (https://www.genome.jp/kegg/) and through manual compilation. After removing duplicates, a total of 508 α-KG-related genes were selected for subsequent analysis. Differential expression analysis was conducted using the ‘limma’ package to identify genes critical to α-KG-related functions (22). Differentially expressed α-KG-related genes between tumor and normal were identified by setting a cutoff value of P<0.05 and |Log2FoldChange (FC)|>1.0. Enrichment analysis was performed using the ‘clusterProfiler’ package (23).

To improve the comparability across different cohorts, all data were standardized by first performing Z-score transformation. Cox regression analysis was then performed to identify differentially expressed α-KG-related genes that were significantly associated with patient survival. Genes with a P-value of <0.05 were considered statistically significant and were included for further investigation. To construct an AKGI with high accuracy and generalizability, a combination of 10 machine learning algorithms were used, including supervised principal components (SuperPC), partial least squares regression for Cox (plsRcox), gradient boosting machine (GBM), stepwise Cox, least absolute shrinkage and selection operator (LASSO), Ridge, survival support vector machine (survival-SVM), CoxBoost and elastic net (Enet). By harnessing the distinct advantages of each algorithm, their integration enhanced the overall performance of the prognostic model in predicting STAD outcomes. Among the 10 machine learning algorithms employed in this model, the LASSO, StepCox and CoxBoost algorithms, which are capable of feature selection and dimensionality reduction, were combined with other machine learning methods to construct 115 prognostic signatures.

To identify the best algorithmic combination for constructing the optimal prognostic model, the C-index values was calculated for each model across the GSE15459, GSE84433, GSE84437 and TCGA cohorts. By comparing the average C-index values across these cohorts, the prognostic signature with the highest score was selected as the optimal model for further analysis. The risk score for each patient was then computed using the algorithmic combination selected for the optimal model. To divide patients into high- and low-risk groups, the ‘surv_cutpoint’ function in the ‘survminer’ package was used to determine the optimal cutoff point, which corresponded to the risk score that maximized the distinction in OS time between the two groups.

The optimal predictive model was derived from the training and validation cohorts, risk scores were calculated for each sample using this model, and samples were subsequently classified into high- and low-AKGI groups based on these scores. To evaluate the predictive performance of the constructed signature, Kaplan-Meier survival analysis was performed using the ‘survival’ and ‘survminer’ packages in both the training and validation cohorts. Decision curve analysis (DCA) images were drawn to reflect the clinical benefit of the predictive nomogram model using ‘stdca.R’ package. The discriminative ability of the model between the high- and low-risk groups was assessed by calculating the area under the curve (AUC) in the receiver operating characteristic (ROC) analysis. Furthermore, the ‘timeROC’ and ‘cmprsk’ packages were used to plot time-dependent ROC and decision curves to further gauge the predictive accuracy of the model in the TCGA-STAD cohort.

A prognostic nomogram was developed based on independent prognostic factors, including the risk score derived from the signature and key clinical features, using the ‘nomogramEx’ package in the TCGA-STAD cohort. To assess the ability of the nomogram to predict OS, several analyses were performed in the TCGA-STAD cohort. A calibration curve analysis was then performed using the ‘calibrate’ package, allowing for a comparison between the predicted survival probabilities and the actual survival outcomes observed in the cohort.

Several previously published signatures related to tumor microenvironment (TME) cell types, immunotherapy responses, immune suppression and immune exclusion were obtained using the IOBR package. A standardized approach was then used to calculate the enrichment score for each sample, facilitating a thorough analysis of the immunological differences between high- and low-AKGI patients (24). Subsequently, a correlation analysis between AKGI and the enrichment scores of several immune characteristics and biological functions was performed. Using P<0.05 and |r|>0.3 as the selection criteria, items significantly associated with AKGI were identified. Additionally, certain prominent indicators were focused on, such as the stemness (25) and m6A index (26). The distribution of tumor mutational burden (TMB), silent mutations and missense mutations were also compared between the two groups and patients were reclassified based on AKGI. For evaluating the immunotherapy response, the survival of patients was first assessed with delayed response to immunotherapy, followed by the use of the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm (http://tide.dfci.harvard.edu/). to estimate their likelihood of responding to treatment (27). In addition, the predictive role of AKGI was also evaluated using two chemotherapy treatment cohorts, GSE26899 and GSE13861.

To screen potential therapy agents for patients with high AKGI scores, expression data from human cancer cell lines sourced from the Broad Institute's Cancer Cell Line Encyclopedia (https://sites.broadinstitute.org/ccle) was utilized. This provides extensive molecular profiles across several cancer types and their corresponding cell lines, offering insights into genetic alterations and drug responses. To obtain drug sensitivity data, two comprehensive datasets we used: the Cancer Therapeutics Response Portal (CTRP) v.2.0 (https://portals.broadinstitute.org/ctrp) and the Profiling Relative Inhibition Simultaneously in Mixtures (PRISM) Repurposing datasets (19Q4; http://depmap.org/portal/prism/). The area under the dose-response curve (AUC) was used as a key measure of drug sensitivity, where lower AUC values indicated higher drug sensitivity. When evaluating the evidence for candidate drugs, Connectivity Map (CMap) identified potential therapeutic compounds for GC by revealing small molecules whose gene expression signatures inversely correlated with GC-specific transcriptional profiles (connectivity scores: −100 to 0), thereby nominating drug candidates for experimental validation, and PubMed (https://pubmed.ncbi.nlm.nih.gov/) was searched for literature related to the candidates, and all available clinical trials and descriptions involving these drugs were analyzed.

GC AGS and MKN74 cells, obtained from Procell Life Science & Technology, Co., Ltd., were used in the current study and cultured in a saturated humidity atmosphere (37°C and 5% CO₂). AGS cells were cultured in Ham's F-12 medium (cat. no. PM150810; Pricella^®; Elabscience Bionovation Inc.) containing 10% fetal bovine serum (FBS; cat. no. C04001; VivaCell, Shanghai, China) and 100 IU/ml penicillin/ streptomycin antibiotics (P/S; cat. no. 15070063, Thermo Fisher Scientific, Inc.). Meanwhile, RPMI-1640 medium (cat. no. PM150110; Pricella; Elabscience Bionovation Inc.) containing 10% FBS and 100 IU/ml P/S was used to culture MKN74 cells.

A Cell Counting Kit-8 (CCK-8) assay was applied to confirm the cytotoxicity effect of α-KG interventions against the GC progression. Briefly, AGS and MKN74 cells (5×10³/well) were plated in 96-well plates (Corning, Inc.) and cultured at 37°C for 24 h. For the α-KG interventions, 2, 4, 8, 10 and 20 mM α-KG was individually supplemented into the culture medium with the cells incubated at 37°C for a further 24 h. In addition, both GC cells treated with the culture medium supplemented with Dulbecco's phosphate-buffered saline (DPBS; cat. no. 14190250; Thermo Fisher Scientific, Inc.) were set as the negative control (NC) group. After the α-KG interventions, 100 µl CCK-8 reagents (cat. no. MA0218; Dalian Meilun Biotech Co., Ltd.) were added to the 96-well plates with the plates further incubated at 37°C for another 4 h. Lastly, the absorbance of 96-well plates at 450 nm was recorded using a microtiter plate reader (Multiskan™; Thermo Fisher Scientific, Inc.), and the inhibiting rates (IR) of α-KG interventions on the proliferation were calculated.

After the optimization of α-KG concentration based on the viability and 50% IR data, a wound healing assay was performed to analyze the effects of α-KG interventions on the migration potentials of GC cells (28). Accordingly, 5×10⁵ AGS or MKN74 cells were seeded in 6-well plates (Corning, Inc.) and further cultured at 37°C for 24 h. Prior to the α-KG intervention, cells were allowed to grow until reaching a high confluence (90–100 %) to ensure consistency across cell lines. A wound across the cell monolayer was prepared by scratching the plates with 200 µl plastic pipette tips, and cellular debris and non-adherent cells were removed by washing with DPBS solution. Cells were subsequently incubated in serum-free medium with α-KG or DPBS at 37°C for 24 h. After treatment, the wound closure was microscopically (TI-S; Nikon Corporation) recorded and further analyzed using ImageJ software (V1.8.0; National Institutes of Health). Furthermore, the effect of α-KG interventions on the invasion potentials of GC cells was assessed using a Transwell invasion assay (29). AGS or MKN74 cells (2×10³/well) were collected, resuspended in serum-free medium containing α-KG or DPBS solution and seeded into the upper Transwell chamber precoated with Matrigel at 37°C for 0.5 h (Corning, Inc.). Moreover, the lower chamber was supplemented with 600 µl culture medium containing 10% FBS (cat. no. C04001; Shanghai VivaCell Biosciences, Ltd.). After culturing at 37°C for 24 h, the GC cells in the upper chamber were removed using a cotton swab, and the Transwell chambers were fixed with methanol at room temperature for 10 min, stained with crystal violet solution (cat. no. G1062; Beijing Solarbio Science & Technology Co., Ltd.) at room temperature for 10 min, and microscopically (TI-S; Nikon Corporation) recorded. The number of GC cells that migrated through the Transwell system was quantified using ImageJ software (V1.8.0; National Institutes of Health).

The potential effect of α-KG treatments on the malignant characteristics of GC cells was assessed using a clonogenicity assay. A total of 1×10³ AGS or MKN74 cells were collected, seeded in a 6-well culture plate (Corning, Inc.) and cultured at 37°C for a further 7 days. The culture medium containing α-KG or DPBS solution was replaced daily. Once the visible colonies developed, the colonies were incubated with 4% paraformaldehyde solution (cat. no. P1110, Beijing Solarbio Science & Technology Co., Ltd.) at 37° for 15 min and stained with crystal violet solution at room temperature for 30 min. The number of colonies was further microscopically (TI-S; Nikon Corporation) assessed. Quantification was performed using ImageJ software (V1.8.0; National Institutes of Health), with the colony area threshold adjusted to identify clusters containing >50 cells.

The effect of α-KG treatments on inducing the cell cycle arrest of AGS or MKN74 cells was detected using a Cell Cycle and Apoptosis Analysis Kit (cat. no. MA0334; Dalian Meilun Biotech Co., Ltd.). As previously reported (3), after treatment with α-KG or DPBS, GC cells were collected and fixed with ice-cold 70% ethanol at −20°C. After fixation, cells were stained at room temperature for 15 min with a propidium iodide (PI) solution, which served as the analyte reporter by intercalating with DNA to allow quantification of DNA content. Flow cytometry analysis was performed using a FACSCalibur™ flow cytometer (BD Biosciences). Cell cycle distribution across the G₀/G₁, S, and G₂/M phases was analyzed using FlowJo software (version 10.8.1; BD Biosciences).

Annexin V-FITC/PI staining. The potential effect of α-KG treatments on triggering apoptosis in AGS or MKN74 cells was quantitatively assessed using Annexin V-FITC/PI staining. As per the manufacturer's manual, GC cells of both groups post-treatment were collected and incubated with Annexin V-FITC solution (cat. no. CA1020; Beijing Solarbio Science & Technology Co., Ltd.) in the dark at 37°C for 5 min. After PI solution staining, apoptosis was immediately analysed using a FACSCalibur™ flow cytometer (BD Biosciences, San Jose, CA, USA). Flow cytometric data were analysed using FlowJo software (version 10.8.1; BD Biosciences, http://www.flowjo.com/).

MitoTracker staining and JC-1 staining were utilized to assess α-KG treatment-induced mitochondrial dysfunctions. For MitoTracker staining to assess mitochondrial activity, GC cells of both groups post-treatment were incubated at 37°C with 200 nM MitoTracker staining regent (cat. no. C1049; Beyotime Institute of Biotechnology) for 30 min and further counterstained with DAPI solution at 37°C for 10 min. The MitoTracker staining of both groups was assessed under a microscope (TI-S; Nikon Corporation), and the MitoTracker staining intensity of both groups was further analyzed using ImageJ software (V1.8.0; National Institutes of Health). For assessing the mitochondrial membrane potential (ΔΨm), GC cells of both groups post-treatment were incubated with 10 µM JC-1 staining solution (cat. no. C2006; Beyotime Institute of Biotechnology) for 20 min, followed by an analysis of JC-1 staining intensity under a microscope (TI-S; Nikon Corporation) using ImageJ software (V1.8.0; National Institutes of Health).

Levels of oxidative stress and ferroptosis-related biomarkers [superoxide dismutase (SOD), malondialdehyde (MDA), reactive oxygen species (ROS) and iron] in GC cells of both groups post-treatment were measured using the following commercial kits according to the manufacturers' protocols: MDA (cat. no. S0131; Beyotime Institute of Biotechnology); ROS (cat. no. S0033; Beyotime Institute of Biotechnology); SOD (cat. no. BC5165; Beijing Solarbio Science & Technology Co., Ltd.); and iron (cat. no. BC5315; Beijing Solarbio Science & Technology Co., Ltd.).

The anticancer effects of α-KG treatments against GC progression were analyzed using RT-qPCR of apoptosis, oxidative stress, and ferroptosis-related genes. Accordingly, the mRNA of GC cells of both groups post-treatment was extracted using TRIzol™ reagent (cat. no. 12183555; Thermo Fisher Scientific, Inc.) with the synthesis of cDNA performed at 65°C for 10 min using the Prime Script™ RT reagent kit (cat. no. RR047A; Takara Biotechnology Co., Ltd.). For PCR amplification, specific primers targeting genes related to apoptosis (Bax and Bcl-2), oxidative stress [nuclear factor erythroid 2-related factor-2 (Nrf2) and Kelch-like ECH-associated protein 1 (Keap1)] and ferroptosis [glutathione peroxidase 4 (GPX4) and solute carrier family 7 member 11 (SLC7A11)] were designed using the National Center for Biotechnology Information website (https://blast.ncbi.nlm.nih.gov/Blast.cgi) and commercially synthesized by Invitrogen™ (Thermo Fisher Scientific, Inc.). RT-qPCR of each group was then performed using the PikoReal system (Thermo Fisher Scientific, Inc.) with a commercial kit (cat. no. RR820A; Takara Biotechnology Co., Ltd.). The thermocycling conditions used in this study were as follows: An initial denaturation at 9°C for 3 min; followed by 8 cycles of denaturation at 98°C for 15 sec, annealing at 60°C for 15 sec and extension at 72°C for 30 sec; with a final extension at 72°C for 5 min. After RT-qPCR, relative gene expression was calculated using the 2^−ΔΔCq method with the ubiquitously expressed GAPDH used as an internal control (30). Primer sequences were as follows: Bax, forward 5′-CCCGAGAGGTCTTTTTCCGAG-3′ and reverse 5′-CCAGCCCATGATGGTTCTGAT-3′; Bcl-2, forward 5′-GGTGGGGTCATGTGTGTGG-3′ and reverse 5′-CGGTTCAGGTACTCAGTCATCC-3′; Nrf2, forward 5′-TTCCCGGTCACATCGAGAG-3′ and reverse 5′-TCCTGTTGCATACCGTCTAAATC-3′; Keap1, forward 5′-GTGTCCATTGAGGGTATCCACC-3′ and reverse 5′-GCTCAGCGAAGTTGGCGAT-3′; GPX4, forward 5′-GAGGCAAGACCGAAGTAAACTAC-3′ and reverse 5′-CCGAACTGGTTACACGGGAA-3′; SLC7A11, forward 5′-TCTCCAAAGGAGGTTACCTGC-3′ and reverse 5′-AGACTCCCCTCAGTAAAGTGAC-3′; and GAPDH, forward 5′-CTGGGCTACACTGAGCACC-3′ and reverse 5′-AAGTGGTCGTTGAGGGCAATG-3′.

To further investigate the regulatory mechanism by which α-KG treatment suppresses GC progression, total RNA was extracted from GC cells in both treated and control groups using TRIzol^® reagent (cat. no. 15596018; Invitrogen; Thermo Fisher Scientific, Inc.) according to the manufacturer's instructions. For transcriptomic analysis, polyadenylated mRNA was enriched using oligo(dT) magnetic beads, fragmented into short sequences and reverse transcribed into cDNA using the SuperScript III First-Strand Synthesis System (cat. no. 18080051; Invitrogen; Thermo Fisher Scientific, Inc.) with dNTPs (cat. no. R0192; Thermo Fisher Scientific, Inc.) and random hexamer primers. The reverse transcription temperature protocol was as follows: 2°C for 10 min, 50°C for 50 min and 70°C for 15 min. Double-stranded cDNA was synthesized, end-repaired, A-tailed and ligated to sequencing adapters using the NEBNext Ultra II DNA Library Prep Kit for Illumina (cat. no. E7645S; New England Biolabs, Inc.). The ligation products were purified using AMPure XP beads and amplified by PCR to construct the cDNA library. DNA library quality was assessed using an Agilent 2100 Bioanalyzer with the High Sensitivity DNA Kit (cat. no. 5067-4626; Agilent Technologies, Inc.) to verify integrity, and concentrations were quantified using Qubit 4 Fluorometer (Thermo Fisher Scientific, Inc.). Libraries were sequenced using an Illumina NovaSeq 6000 platform with 150-bp paired-end reads. The sequencing kit used was the NovaSeq 6000 S4 Reagent Kit (300 cycles) (cat. no. 20028312; Illumina, Inc.). The final library was loaded at a concentration of 300 pM, determined using qPCR with the KAPA Library Quantification Kit (cat. no. KK4824; Roche Diagnostics), reported in molar concentration. Raw sequencing reads were processed by trimming low-quality bases and adapter sequences using Trimmomatic (v0.39; http://github.com/usadellab/Trimmomatic). Clean reads were aligned to the human genome reference hg19 using STAR (v2.7.3a). Gene expression levels were quantified as fragments per kilobase of transcript per million mapped reads using featureCounts (v2.0.1). Differentially expressed genes were identified with a threshold of P<0.05 and |log₂(fold change)|≥1. Subsequent functional analyses, including Gene Ontology and KEGG pathway enrichment, were performed using the DAVID database (https://david.ncifcrf.gov/). GSEA was also conducted to identify significantly enriched signaling pathways related to α-KG treatment (31–33).

In the present study, all statistical analyses were performed in R v.4.1.0 (The R Foundation). Each experiment was conducted in triplicate, unless otherwise specified. Descriptive data were analysed using the Student's t-test and are presented as the mean±standard deviation, unless otherwise indicated. For comparisons, normally distributed variables were analyzed using unpaired Student's t-tests, whilst non-normally distributed variables were assessed using the Wilcoxon rank-sum test. A two-sided Fisher's exact test was performed for the contingency tables. Correlation analyses were performed using Spearman's rank correlation. P<0.05 was considered to indicate a statistically significant difference.

As shown in Fig. 1A, an in-depth reanalysis of several previously published cohorts was performed to train and validate the predictive model in the present study. This included four bulk RNA sequencing datasets (TCGA-STAD, GSE84437, GSE8433 and GSE15459) as well as two treatment-related cohorts (GSE26899 and GSE13861). In total, a concatenated set of 509 genes (Table SI) derived from eight KEGG pathways associated with α-KG synthesis, metabolism and regulation were analyzed.

Based on the TCGA-STAD cohort data, 105 α-KG-related genes were identified that exhibited significant differential expression between tumor and normal samples (adjusted P<0.05 and |log2FC|>1; Fig. 1B and S1). Subsequently, using patient survival data, univariate Cox analysis was performed and 26 prognostic genes were identified from the differentially expressed α-KG-related genes (Fig. 1C and D). These were then used to construct a predictive model.

Using the 26 α-KG-related candidate genes, the AKGI was developed using an ensemble machine learning algorithm. In the TCGA-STAD cohort, the Leave-One-Out Cross Validation strategy was applied to construct 115 predictive models and compute the C-index for each model across all validation cohorts (Fig. 2A). Among the four datasets, the random survival forest (RSF) model demonstrated the highest average C-index of 0.66, establishing it as the most robust AKGI model (Fig. 2A and B). By using RSF we scored the samples, and based on their risk scores, patients were classified into high- and low-AKGI groups according to the optimal cutoff. Using the optimal cutoff value, patients were stratified into high and low AKGI groups (Figs. 2C and S2). Survival analysis revealed that in all four datasets, patients with GC with low AKGI levels had significantly longer OS and disease-free survival compared with those with high AKGI levels (Figs. 2D and E, and S2). These findings highlight the potential of AKGI as a prognostic predictor for patients with GC.

To evaluate the independent predictive ability of AKGI, further assessments were performed using multiple approaches. ROC curves based on the TCGA-STAD cohort demonstrated that the AUC values for AKGI were 0.91, 0.93 and 0.92 for predicting 1-, 3- and 5-year survival, respectively (Fig. 2F). Calibration curves confirmed the accuracy of the AKGI prediction model for reflecting actual observations (Fig. 2G).

Subsequently, multivariate Cox regression analysis was performed to evaluate the impact of several clinical factors on the prognosis of patients with GC, including AKGI, sex, age, clinical stage and pathological tumor (T) stage. Based on this analysis, a nomogram was constructed to predict patient outcomes, which indicated that AKGI was a significant prognostic risk factor for patients with GC (P<0.05; Fig. S3), with statistical significance observed at 1-, 2- and 3-year intervals. Additionally, sex, age and pathological T stage were also identified as high-risk factors for GC prognosis.

Moreover, DCA revealed that AKGI provided significantly better clinical outcomes than other indicators for patients with GC, including age, T stage and sex (Fig. 2H and I). Finally, time-dependent C-index analysis further demonstrated that AKGI exhibited superior predictive performance compared with other clinical factors (Fig. 2J).

To assess the biological functions associated with the AKGI prognostic model score, 501 gene sets and functional scores were collected and a correlation analysis was performed with AKGI (Table SII). The analysis identified 34 gene sets significantly correlated with AKGI (|R|>0.25; P<0.05; Fig. S4). Among these, the mRNA stemness index (mRNAsi) exhibited the strongest negative correlation with AKGI (R=−0.433; Fig. 3A). Notably, another stemness index, the DNA methylation stemness index, demonstrated a weaker correlation with AKGI (R=−0.1; Fig. 3B).

In addition to mRNAsi, gene sets related to glycosylphosphatidylinositol, ferroptosis, cell cycle, TGF-β signaling, EMT and DNA replication demonstrated significant correlations with AKGI; however, notable m6A scores were not correlated with AKGI (Fig. 3). Furthermore, AKGI demonstrated associations with several immune-related gene sets, including chemokines and macrophage M2 regulation, suggesting that AKGI may be involved in immune modulation and shaping the immune microenvironment in GC (Fig. 3).

Given the correlation between AKGI and several immune-related markers in GC, the role of AKGI in the immune microenvironment was comprehensively evaluated. Results from multiple immune cell infiltration algorithms demonstrated significantly lower infiltration levels of tumor-associated immune cells, including T, natural killer (NK) and stromal cells, in the high AKGI group, suggesting an immunosuppressive state in this group (Fig. 4A). Additionally, molecular markers associated with immune suppression and rejection, such as those involved in the EMT pathway and TGF-β signaling, were predominantly enriched in the high AKGI group (Fig. 4B). This finding further supports that the high AKGI group exhibits characteristics consistent with an immunosuppressive state.

Gene sets related to malignant tumor treatment and immunotherapy markers were also assessed. The results revealed that markers associated with favorable immunotherapy outcomes were significantly enriched in the low AKGI group compared with the high AKGI group (Fig. 4C and D). TMB was compared between high and low AKGI groups. The results revealed that TMB, silent mutations and missense mutations were significantly higher in the low AKGI group compared with the low AKGI group (Fig. 4E-G), indicating greater immunogenicity in this subgroup.

Further joint analyses of several mutation metrics demonstrated that AKGI could synergize with TMB, silent mutations and missense mutations to predict patient prognosis more effectively. Notably, patients with GC with lower AKGI and higher TMB levels exhibited significantly improved survival outcomes than those with higher AKGI (Fig. 4H-J).

To comprehensively evaluate the potential application of AKGI in GC therapy, two treatment cohorts were analyzed, GSE26899 and GSE13861, which provided extensive prognostic and treatment-related data for this patient population. In both cohorts, the 26 genes used to construct the AKGI demonstrated significant differences in expression levels (Fig. 5A). After performing AKG scoring using the RSF model and determining the cut-point for high/low group division (Fig. S5). Prognostic analysis revealed that the high and low AKGI groups had markedly different survival outcomes, with the low AKGI group showing a significantly improved prognosis (Fig. 5B and C). These findings further indicate the utility of low AKGI in assessing therapy outcomes.

Additionally, the TIDE algorithm was applied to evaluate the responses of patients to immunotherapy. The results revealed that the low AKGI group exhibited a significantly improved response to immunotherapy compared with the high AKGI group (P=0.00179; Fig. 5D). In subgroup mapping analysis of patients with GC receiving immunotherapy, a high AKGI level was significantly associated with an improved response to cytotoxic T-lymphocyte associated protein 4 treatment in comparison with a low AKGI level (Bonferroni-corrected P=0.008; Fig. 5E).

Finally, the expression differences of AKGI-related genes between treatment responder and non-responder groups were assessed. A total of 11 genes demonstrated a significantly higher expression trend in the high AKGI group compared with in the low AKGI group (Fig. 5F), further confirming that AKGI and its associated genes may provide a valuable predictive insight into the efficacy of immunotherapy and chemotherapy for patients with GC.

Given the poor response to therapy in patients with high AKGI, the CTRP and PRISM databases were utilized to identify potential therapeutic drugs for this patient group (Fig. 6A). The algorithm produced meaningful insights through differential drug response analysis between high AKGI (top decile) and low AKGI (bottom decile) groups. Specifically, compounds with lower AUC values in the high AKGI group were identified (log2FC>0.10). To further refine the selection, Spearman correlation analysis was performed between AUC values and AKGI scores, focusing on compounds with a negative correlation coefficient (Spearman r<-0.20 for CTRP and PRISM). As a result, six potential compounds were identified from the CTRP database (Fig. 6B) and four from the PRISM database (Fig. 6C). All 10 compounds exhibited lower AUC values in the high AKGI group compared with in the low AKGI group, indicating a negative correlation with AKGI.

Although the 10 candidate compounds demonstrated higher drug sensitivity in patients with high AKGI scores, this alone does not confirm their therapeutic efficacy in GC. Therefore, to evaluate their therapeutic potential, four compounds with more comprehensive information were selected from the CMap and PubMed databases for further investigation. First, CMap was used to identify compounds with gene expression patterns opposite to those observed in GC-specific profiles. Specifically, compounds that could downregulate gene expression in tumor tissues after treatment were focused on, countering the elevated expression seen in untreated tumor tissues. A total of two of the selected compounds exhibited CMap scores of <-50, suggesting potential therapeutic effects in GC. Subsequently, an extensive literature review was performed using PubMed to identify clinical trials and supporting evidence regarding the use of the four candidate compounds for GC treatment. To further assess their potential, the fold change differences in mRNA expression levels of the target genes between tumor and normal tissues were calculated. A higher fold change indicated a greater potential for targeting GC (Fig. 6D).

Among the candidate compounds, the drug targets of dasatinib and YM-155 exhibited significant expression differences between tumor and normal tissues (Fig. 6E). Overall, the analysis identified dasatinib and YM-155 as strong candidates with promising therapeutic potential for patients with GC with high AKGI signatures.

The aforementioned results demonstrated the prognostic predictive value of AKGI for patients with GC as well as the biological functions of α-KG in GC. To validate these findings, the GC cell lines, MKN74 and AGS, were treated with α-KG and transcriptomic sequencing was performed on the treated cells. Phenotypic analysis revealed that compared with the NC group, α-KG treatment significantly inhibited AGS and MKN74 cell viability (Figs. 7A and S6A), migration (Figs. 7B and S6B), invasion (Figs. 7C and S6C) and colony formation (Figs. 7D and S6D), which is consistent with previous reports highlighting the suppressive effects of α-KG on malignant tumors (8,34,35).

Flow cytometry analysis further demonstrated that compared with the NC group, α-KG-treated cells exhibited cell cycle arrest in the G2 and S phases (Figs. 7E and S6E), accompanied by a significant increase in apoptosis levels (Figs. 7F and S6F). This observation aligns with the aforementioned conclusion that AKGI is closely associated with cell cycle regulation and ferroptosis.

Subsequently, the impact of α-KG treatment on AKGI-related gene expression was assessed. Comparative transcriptomic analysis between α-KG-treated cells and the control group revealed significant differential expression of genes within the AKGI model (Fig. 7G). As the RFS model did not yield conclusive gene regression coefficients, single-sample (ss)GSEA for KEGG enrichment analysis was applied on both the transcriptomic data from α-KG-treated cells and the expression profiles of high and low AKGI groups from TCGA. The results demonstrated that the transcriptomic expression patterns of α-KG-treated cells closely resembled those of the low AKGI group, with significant enrichment in pathways such as KRAS signaling, EMT, apoptosis and ROS signaling pathways (Fig. 7H).

Further validation of genes significantly enriched in both the AKGI model and transcriptomic sequencing data was performed using RT-qPCR analysis (Figs. 7J and S6H). These experiments demonstrated that, in α-KG-treated MKN74 and AGS cells, the expression levels of genes involved in apoptosis, ROS-related pathways and ferroptosis signaling were significantly disturbed in the α-KG treatment group with the NC group. In addition, the quantitative result of oxidative stress and ferroptosis-related biomarkers, as well as the RT-qPCR results of oxidative stress and ferroptosis-related genes, further indicated the potential effect of α-KG for inducing oxidative stress and ferroptosis in GC cells (Figs. 7I and S6G). Additionally, mitochondrial activity was assessed in α-KG-treated cells. Mitochondrial activity was significantly reduced following α-KG treatment, in comparison with negative controls, suggesting impaired mitochondrial function, possibly associated with apoptosis or oxidative stress (Fig. S7).

In summary, the aforementioned findings suggest that the inhibitory effects and mechanisms of α-KG on GC cell lines in vitro are closely aligned with the biological processes observed in the low AKGI group. This provides further evidence supporting the molecular mechanisms by which α-KG exerts its role in GC.