Integrating MRI radiomics and transcriptomics to predict IDH mutation status and prognosis in glioma Authors: Rong-de Zhong, Qian Li, Heng Wang, Zeng-wei Kou, Shi-xiang Guo, Fan-fan Chen, Na Zhang, Guo-dong Huang, Yun-sheng Liu Published: 2026 ABSTRACT ObjectiveTo investigate MRI-based radiomic features in glioma and key genes related to IDH mutations, and to analyze their correlation.Methods61 MRI files from TCIA were divided into training and validation sets (7:3). An external validation set included 66 pathologically confirmed glioma patients. Regions of interest (ROIs) were manually delineated, extracting 1,037 radiomic features. Mann-Whitney U test and least absolute shrinkage and selection operator (LASSO) regression reduced and filtered features. RNA-seq data differential analysis identified key genes linked to glioma progression and IDH mutations. To ensure the comprehensiveness and robustness of our findings, we further validated the protein expression, tumor immune microenvironment, and cellular localization of these key genes utilizing Western blotting, immune infiltration analysis, single-cell RNA sequencing, and spatial transcriptomics. Correlation between radiomic features and key IDH mutation genes was analyzed.ResultsA model with nine IDH mutation-related radiomic features showed strong predictive performance. Univariate and multivariate Cox regression indicated these features predict glioma prognosis. Eight radiomic features differed significantly across MGMT methylation levels. Differential analysis identified eight key genes promoting glioma development and associated with IDH mutations. Crucially, Western blot experiments confirmed the differential expression of these proteins across IDH subtypes. Furthermore, single-cell and spatial transcriptomic analyses revealed that these genes are predominantly expressed in specific cellular subpopulations, such as astrocytes and microglia, and correlate with immune infiltrating cells, contributing to glioma malignancy. Among the nine selected radiomic features (RF1-RF9), RF3 and RF7 were significantly negatively correlated with key IDH mutation genes, while RF4 showed significant positive correlation. RF2 and RF6 exhibited significant negative correlations with MMP9.ConclusionMRI-based radiomic features and models are valuable for predicting IDH mutations in gliomas. Combined with bioinformatics analysis, they aid in identifying glioma pathological types and genetic characteristics.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12935-026-04384-2. Introduction Glioma is the most common and malignant form of central nervous system tumors, most of which arecharacterized by rapid progression, high recurrence, and high mortality rate [1, 2]. Currently, the standard of care for glioma is maximal resection of the tumor, followed by radiotherapy and chemotherapy with temozolomide. Despite these aggressive treatment strategies, the survival rates for glioma patients have not seen significant improvement. Emerging therapeutic approaches, including immunotherapy, targeted therapy, and gene editing, are gradually being integrated into glioma treatment. However, the challenge remains to enhance patient outcomes in the face of this aggressive disease [3]. Firstly, during the progression of glioma, the proliferation of blood vessels and the disruption of the blood-brain barrier can lead to the infiltration of tumor cells into adjacent brain tissue. This results in an unclear boundary between the tumor and normal cells, making complete surgical resection challenging [4]. Secondly, complex genetic mutations and epigenetic changes contribute to treatment resistance and recurrence, making successful outcomes difficult to achieve [5]. Thirdly, novel therapies that have shown effectiveness in other cancers often fail in clinical trials for glioma [6, 7]. Additionally, several clinical factors can influence the prognosis for glioma patients, including age at diagnosis, histopathologic type, and the ratio of tumor volume to the area of edema. Contrast-enhanced MRI is the preferred method for diagnosing and monitoring gliomas [8, 9]. However, the images produced offer only subjective and difficult-to-quantify information, limiting physicians’ ability to interpret the pathological and molecular characteristics of the lesions. Traditionally, obtaining this information requires invasive sampling and testing of tissue samples, which can be cumbersome, costly, and time-consuming. Thus, there is an urgent need for a reliable, non-invasive, low-risk assessment method. Radiomics can fulfill this need by analyzing medical images to provide convenient clinical diagnosis and prognosis predictions. When combined with transcriptomics, it can assist physicians in identifying tumor type, grade, and treatment response, facilitating the development of personalized treatment plans [10]. Gliomas undergo genetic re-editing to meet metabolic and environmental demands necessary for rapid growth [11]. One of the most common mutations is in the isocitrate dehydrogenase (IDH) enzyme, primarily involving the IDH1 and IDH2 typologies [12]. Mutated IDH leads to the production of 2-hydroxyglutarate (2HG), which can serve as a biomarker for assessing treatment efficacy and disease progression in glioma patients [13]. Additionally, epigenetic alterations caused by IDH mutations fundamentally reprogram the cellular state, primarily by maintaining progenitor-like characteristics and silencing differentiation genes. Despite driving early gliomagenesis, these IDH-mutant tumors exhibit less aggressive phenotypes and predict a favorable prognosis. They are highly prevalent in most adult lower-grade gliomas. Standard treatments significantly improve the prognosis for patients with these mutations, with studies indicating that the survival of IDH mutant patients is 2 to 3 times longer than that of those with wild-type IDH [14]. Therefore, preoperative knowledge of glioma molecular features and tumor subtypes is crucial for effective diagnosis, prognosis prediction, and treatment planning. While traditional transcriptomics provides valuable insights into the molecular characteristics of gliomas, it fails to capture the profound intra-tumoral heterogeneity and the complex cellular ecosystems within the tumor microenvironment. Recently, the advent of single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) has revolutionized our ability to dissect these microenvironments at a single-cell resolution and accurately map their spatial architectures. However, these advanced molecular techniques still rely on invasive surgical sampling. Bridging the gap between non-invasive, macroscopic imaging (radiomics) and microscopic spatial-cellular landscapes (scRNA-seq and ST) represents a critical and highly novel frontier. By combining radiomics with spatial and single-cell transcriptomics, we can non-invasively decode the spatial heterogeneity and specific cellular localizations of key genetic alterations, offering a truly comprehensive tool for evaluating IDH mutation status and exploring its underlying microenvironmental mechanisms. In this study, we performed imaging feature screening and modeling of different molecular subtypes of IDH by outlining regions of interest (ROIs) from The Cancer Imaging Archive (TCIA) and in-hospital MRIs, utilizing a machine learning approach. We obtained corresponding RNA-seq data from TCGA and conducted differential analysis to identify key genes associated with IDH mutations, followed by in-depth bioinformatic analyses incorporating single-cell and spatial transcriptomic evaluations. Additionally, we used Pearson correlation analysis to explore the relationships between imaging features and key genes linked to IDH mutations. Our results indicate that the constructed model can effectively predict IDH molecular subtypes, and several imaging features are significantly associated with prognosis and key IDH mutation-related genes. Materials and methods Research objects for this research This study involves a retrospective collection of glioma patients confirmed by pathology at the First Affiliated Hospital of Shenzhen University, covering the period from April 2022 to June 2024. Inclusion criteria included the availability of preoperative and postoperative MRI imagesand comprehensive clinical data, comprising pathology reports, age, gender, and immunohistochemical test outcomes. Exclusion criteria were based on poor image quality, indicated by a low signal-to-noise ratio and/or significant artifacts, and lesions considered too small for analysis (tumor diameter < 1.0 cm). A total of 66 cases were included in the study, comprising 36 males and 30 females, with a mean age of 50.2 ± 9.8 years. Open-access datasets containing RNA-seq and MRI data for 61 gliomas were obtained from TCGA (https://www.cancer.gov) and TCIA (https://www.cancerimagingarchive.net/), respectively, along with corresponding clinical information, which included 29 IDH wild-type and 32 IDH mutant gliomas. For the TCGA/TCIA dataset, IDH status was defined using the provided molecular and mRNA-seq data. In our hospital cohort, IDH status was determined by pathological IHC and sequencing results, incorporating both IDH1 and IDH2 mutations. All tumor classifications and molecular designations were performed in accordance with the 2021 WHO Classification of Tumors of the Central Nervous System. Additionally, data related to glioma immune infiltration, single-cell analysis, and spatial transcription were retrieved from LM2 (https://cibersortx.stanford.edu/), GSE17327 (https://www.ncbi.nlm.nih.gov/geo), and 10x Genomics (https://www.10xgenomics.com/), respectively. Data collection, feature extraction and screening The hospital’s MR images were retrieved from the PACS system and stored in DICOM format to facilitate the manual delineation of tumor regions of interest (ROIs). All MRI images were obtained using Siemens Verio or Magnetom Aera scanners, with the scanning sequences including T1WI, T2WI and T1CE, FLAIR and DWI (Supplementary Table 1). The online datasets were downloaded in DICOM format. Subsequently, 3D Slicer software version 5.6.1 (https://download.slicer.org/) was employed to systematically apply paint and erase techniques on the cross-sectional, coronal, and sagittal images from the MRI sequences, layer by layer along the tumor boundaries. The software automatically reconstructed the volume of interest (VOI) for the tumors. The ROI delineation was carried out independently by an experienced radiologist and a neurosurgeon for all 127 patients, both of whom were blind to the clinical and pathological information of the patients. The inter-observer feature stability analysis across all 127 patients demonstrated excellent reproducibility. Before feature extraction, the data underwent the following preprocessing steps: First, the dcm2nii software was used to convert the DICOM files into NIFTI format. Next, the FSL coordinate repositioning function standardized all patients’ T1CE images to the left posterior superior (LPS) coordinate system. Then, the FSL registration tool aligned the T1CE images with the MNI152 standard T1 modality atlas, ensuring that all brain images were consistent within a standard space. To ensure strict consistency and reproducibility across the multi-center data, identical extraction configurations were applied to both the TCIA datasets and the hospital external validation dataset. Specifically, following spatial normalization to the MNI152 template, all MR images were resampled to an isotropic voxel size of 1.0×1.0×1.0 mm³ using B-spline interpolation to achieve spatial uniformity and rotational invariance for texture analysis. Image intensity was discretized using a fixed bin width of 25 for the computation of higher-order texture matrices. To ensure that this normalization did not compromise the integrity of texture-based characteristics, a robustness analysis was performed on a subset of 30 randomly selected cases. Radiomic features were extracted from both native and MNI spaces, and the Intraclass Correlation Coefficient (ICC) was calculated to evaluate feature stability. Radiomic features were extracted using the open-source Python package, pyRadiomics (version 2.2.0; https://pyradiomics.readthedocs.io/). 1037 features were extracted in T1CE sequences from the tumor region: including 14 shape features describing the 3D geometric characteristics, 198 intensity features describing the first-order distribution of the intensities and 825 texture features were computed to describe the patterns, or the high-order intensity distributions with five classic texture matrix categories, including gray-level cooccurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray level size zone matrix (GLSZM), gray level dependence matrix (GLDM), and neighborhood gray-tone difference matrix (NGTDM). Before performing feature selection, patients of TCIA were randomly divided into training and test sets at a ratio of 7:3. The 66 in-house hospital cases were used as an independent external validation set. Firstly, the features were standardized using z score normalization to zero mean and unit variance for the training set, and the distribution of the features in the training set was used to normalize the test set and validation set. For feature pairs with correlation coefficients greater than 0.90, the feature with the better univariate predictive power (smaller Mann–Whitney U test p value) was retained. Finally, a least absolute shrinkage and selection operator (LASSO) regression model was applied in the training subset for further feature selection and signature construction. Ten-fold cross validation with minimum criteria was used to determine the optimal λ. In our analysis, the final value of λ that gave minimum cross validation error was 0.1098, yielding the optimal LASSO model. Nine features with nonzero coefficients remained. Construction, validation and efficacy analysis of predictive models The two-class random forest classification (IDH wild-type and IDH mutant-type of gliomas) was implemented by using the randomForest package V4.7–1.1 in R, in which each decision tree was built based on the classification and regression tree (CART). In this approach, radiomics-based model was trained in the training set based on the selected features. After assessing a set of numbers ranging from 50 to 800, the number of trees in the random-forest algorithm was set to 500 by balancing the accuracy and computational burden. A repeated cross-validation approach was used to optimize the models, which performed 15 repetitions of tenfold cross-validation in the training set. The diagnostic performance of each model was measured using the AUC, accuracy, precision, recall and F1-score values. In addition, the Mann-Whitney U test was used to compare the differences in radiomics features between IDH mutant and wild-type images. To evaluate the relative advantage of the Random Forest (RF) model, we implemented several state-of-the-art machine learning algorithms as baselines, including Logistic Regression (LR) and Support Vector Machine (SVM), using the same training and validation frameworks. Differential gene screening and validation The mRNA count data downloaded from the TCGA database was normalized, transformed, and duplicate samples were removed. The “DESeq2” package in R was then used to identify differentially expressed genes (DEGs) between gliomas and normal brain tissue, as well as between IDH mutant and IDH wild-type gliomas, with an adjusted p-value < 0.05 and |logFC| > 2 as the cutoff criteria for differential expression. Additionally, Venn analysis was performed to obtain the overlap genes. Subsequently, mRNA-seq data sourced from CGGA (https://www.cgga.org.cn/) was utilized to validate the differential expression levels of the overlap genes across different IDH molecular subtypes of gliomas. Concurrently, data from the TCGA was employed to confirm the expression of these overlap genes in various types of gliomas. To robustly identify the key genes driving glioma development and contributing to IDH mutations, we defined them as the overlap genes that consistently exhibited statistically significant differential expression across both validation databases, as determined by the Wilcoxon rank-sum test with a significance threshold of P < 0.05. Ultimately, we identified eight genes: MMP9, IGF2BP3, FGFBP2, EN1, HOXD9, HOXB2, HOXB7, and CD48. Prognostic prediction efficacy of key genes associated with IDH mutations in glioma We adopted a robust survival analysis algorithm to develop a glioma survival prediction model based on eight key genes. The nomogram was constructed using the “rms” package (6.8-2) to analyze the predicted probabilities of the outcome variable. Additionally, we generated clinical decision curves with the “dcurves” package (0.5.0) to evaluate the model’s practical value in clinical decision-making. Calibration curves were also created using the “rms” package (6.8-2) to assess the consistency between the predicted survival probabilities and actual observed outcomes. Finally, we utilized the “riskRegression” package (2023.12.21) to construct ROC curves, allowing us to evaluate the predictive efficacy of the risk score for three-year survival in glioma patients. Western blotting As described before [15], total protein was extracted by RIPA cell lysate (Cat. 89901; Thermo Fisher) and protein concentration was quantified by BCA Kit (Cat.P0012, Beyotime Biotechnology). Proteins were separated by SDS-PAGE electrophoresis, then transfer to the activated PVDF membrane. Primary antibodies were incubated after blocking with skim milk, overnight at 4˚C; Next day, after washed with TBST, the membrane were incubated with HRP-conjugated secondary antibodies for 1 h, after washing, positive signals were uncovered by ECL chemiluminescence. The primary antibodies of FGFBP2 (YP-Ab-03871), HOXD9 (YP-Ab-05661), HOXB2 (YP-Ab-15764), HOXB7 (YP-Ab-07810) were purchased from Upingbio; IGF2BP3(R24706), EN1 (R389169), CD48 (R382369) were obtained from Zenbio; GAPDH (60004-1-Ig) and MMP-9 (10375-2-AP) were purchased from Proteintech. Functional enrichment and immune infiltration analysis of key genes for IDH mutations The key genes associated with IDH mutations were enriched through GO and KEGG analyses using the “clusterProfiler” package in R software. The protocols were followed as described in the literature [16], and a threshold of P < 0.05 was considered statistically significant. The immune infiltration analysis utilized two databases: one was the RNA-seq data from TCGA, and the other was the “LM22.txt” data from the R package “CIBERSORT,” which recorded the gene expression of 22 types of immune cells. Using the CIBERSORT computational method, we obtained the abundance of these 22 immune cell types and employed the Spearman method to analyze the correlation between tumor-infiltrating immune cell types and the expression levels of key genes associated with IDH mutations. Single cell analysis and spatial transcriptomics The GSE173278 dataset, which comprised 99 samples from glioma patients, was utilized for single-cell analysis. We employed the R package “Seurat” (version 5.1.0) for data filtering, quality control, normalization, scaling, principal component analysis, and clustering annotation. Each cell cluster was labeled as a specific cell type, and the expression levels of IDH mutation-related genes across different cell types were visualized using bar scatter plots. Spatial transcriptomic data were downloaded from “Human Glioblastoma: Whole Transcriptome Analysis Mastering Biology to Advance Human Health” − 10x Genomics and analyzed using the Seurat package (v. 4.3.0). By integrating the single-cell data (GSE148842) with the RCTD algorithm, we obtained spatial cell annotations. The functions spatialDimPlot and SpatialFeaturePlot were used together to visualize cell expression levels in the spatial transcriptomics (ST) data. Statistical analysis Statistical analyses and plots were performed using R version 4.4.0 or Python. We analyzed correlations between radiomic features and age, sex, and key genes associated with IDH mutations, as well as correlations between key genes for IDH mutations and MGMT expression, sex, 1p/19q codeletion, IDH staging, andgliomagrading using Pearson correlation analysis. The association of radiomic features with gliomarecurrence, pathological type, P53, and 1p/19q codeletion, was evaluated using the Mann-Whitney U test. Single and multifactorial Cox regression analyses were conducted to assess the relationship between radiomic features and prognosis. The results of gene expression against radiomic features were presented in a heatmap correlation matrix. The interrater reliability of continuous radiomic features was assessed by calculating the ICC using a 2-way random-effects, absolute agreement, single-measures model (ICC(2,1)). Features yielding ICCs > 0.8 were considered excellent and were retained for subsequent analysis, while unstable features were excluded. Significance was defined as P < 0.05. Results Clinical characteristics This study included 127 patients with glioma from two cohorts, and the external validation set included 66 cases (from the First Affiliated Hospital of Shenzhen University, with 36 males and 30 females, and an average age of 50.23 years. Among the 66 glioma patients, glioblastoma multiforme (GBM) accounted for 43 cases (65.2%), low-grade glioma accounted for 23 cases, and the recurrence rate was 21.2%) (Fig. 1, Table 1). In the TCGA dataset used to screen for key genes associated with IDH mutations, there were 32 cases of IDH mutant patients and 29 cases of wild-type patients; LGG represented 27.9% of the cases, with a median survival time of 1.31 years, as detailed in Table 2. Fig. 1Flowchart of patient enrollment, exclusion criteria, and final cohort assignment in this study. Flowchart of patient enrollment, exclusion criteria, and final cohort assignment in this study. Table 1Demographic and clinical characteristics of patients from Shenzhen Second People’s HospitalCharacteristicsPatientsIDHStatisticP-value(n = 66)Mutant (n = 26)Wild-type (n = 40)Gender = Male (%)36 (54.5%)15 (57.69%)21 (52.5%)0.1710.679Age (mean ± SD)50.23 ± 9.8545.62 ± 8.7553.23 ± 9.44−3.2910.002Diagnosis = GBM (%)43 (65.2%)3 (11.54%)40 (100%)-< 0.001Recurrence (%)14 (21.2%)5 (19.23%)9 (22.5%)0.1010.751Location (%)11.410.022 Frontal lobe20 (30.3%)12 (46.15%)8 (20%) Parietal lobe5 (7.6%)0 (0%)5 (12.5%) Temporal lobe7 (10.6%)0 (0%)7 (17.5%) Multiple24 (36.4%)10 (38.46%)14 (35%) Other10 (15.2%)4 (15.38%)6 (15%)1p/19q codeletion= deletion (%)21 (31.8%)11 (42.31%)10 (25%)2.1760.14BRAF = Wild-type (%)57 (86.4%)23 (88.46%)34 (85%)0.0010.973TERT = Wild-type (%)32 (48.5%)17 (65.38%)15 (37.5%)4.9050.027Oligo2 (+) (%)63 (95.5%)25 (96.15%)38 (95%)01H3K27Me3 (+) (%)61 (92.4%)25 (96.15%)36 (90%)0.20.655Ki6711.5620.001 (+)27 (40.9%)4 (15.38%)23 (57.5%) (-)39 (59.1%)22 (84.62%)17 (42.5%)P530.1270.722 (+)44 (66.7%)18 (69.23%)26 (65%) (-)22 (33.3%)8 (30.77%)14 (35%)IDH mutation status includes both IDH1 and IDH2 variants, consistent with the latest WHO molecular diagnostic framework Table 1Demographic and clinical characteristics of patients from Shenzhen Second People’s HospitalCharacteristicsPatientsIDHStatisticP-value(n = 66)Mutant (n = 26)Wild-type (n = 40)Gender = Male (%)36 (54.5%)15 (57.69%)21 (52.5%)0.1710.679Age (mean ± SD)50.23 ± 9.8545.62 ± 8.7553.23 ± 9.44−3.2910.002Diagnosis = GBM (%)43 (65.2%)3 (11.54%)40 (100%)-< 0.001Recurrence (%)14 (21.2%)5 (19.23%)9 (22.5%)0.1010.751Location (%)11.410.022 Frontal lobe20 (30.3%)12 (46.15%)8 (20%) Parietal lobe5 (7.6%)0 (0%)5 (12.5%) Temporal lobe7 (10.6%)0 (0%)7 (17.5%) Multiple24 (36.4%)10 (38.46%)14 (35%) Other10 (15.2%)4 (15.38%)6 (15%)1p/19q codeletion= deletion (%)21 (31.8%)11 (42.31%)10 (25%)2.1760.14BRAF = Wild-type (%)57 (86.4%)23 (88.46%)34 (85%)0.0010.973TERT = Wild-type (%)32 (48.5%)17 (65.38%)15 (37.5%)4.9050.027Oligo2 (+) (%)63 (95.5%)25 (96.15%)38 (95%)01H3K27Me3 (+) (%)61 (92.4%)25 (96.15%)36 (90%)0.20.655Ki6711.5620.001 (+)27 (40.9%)4 (15.38%)23 (57.5%) (-)39 (59.1%)22 (84.62%)17 (42.5%)P530.1270.722 (+)44 (66.7%)18 (69.23%)26 (65%) (-)22 (33.3%)8 (30.77%)14 (35%)IDH mutation status includes both IDH1 and IDH2 variants, consistent with the latest WHO molecular diagnostic framework Demographic and clinical characteristics of patients from Shenzhen Second People’s Hospital IDH mutation status includes both IDH1 and IDH2 variants, consistent with the latest WHO molecular diagnostic framework Table 2Clinical characteristics of patients in the IDH-mutant and IDH-wild-typegroups from the TCIA databaseCharacteristicsTotal patient(n = 61)Mutant(n = 32)Wild-type(n = 29)StatisticP-valueAge (mean ± SD)51.23 ± 14.3544.03 ± 13.1559.17 ± 11.2−4.815< 0.001Gender = Male (%)33 (54.1%)14 (43.75%)19 (65.52%)2.9030.088 Low-grade glioma17 (27.9%)17 (53.13%)0 (0%) High-grade glioma44 (72.1%)15 (46.88%)29 (100%)1p/19q codeletion = Non-Codel (%)50 (83.3%)21 (65.62%)29 (100%)1.540.001MGMT = Unmethylated (%)14 (26.4%)1 (3.12%)13 (44.83%)26.009< 0.001OS (median [IQR])1.31 [0.93, 2.21]2.17 [1.30, 3.59]0.79 [0.61, 1.26]4.044< 0.001Status = alive (%)32 (52.50%)27 (84.38%)5 (17.24%)27.492< 0.001MMP9 (median [IQR])0.82 [0.27, 3.86]0.32 [0.15, 0.8]3.78 [1.66, 5.42]−4.549< 0.001IGF2BP3 (median [IQR])0.56 [0.16, 3.15]0.19 [0.1, 0.45]3.02 [1.69, 4.29]−4.463< 0.001FGFBP2 (median [IQR])0.84 [0.47, 1.47]0.7 [0.38, 1.04]1.3 [0.68, 2.14]−2.5270.011EN1 (median [IQR])0.27 [0.04, 2.67]0.06 [0.03, 0.39]2.6 [0.27, 3.45]−3.729< 0.001HOXD9 (median [IQR])0.50 [0.07, 2.20]0.11 [0.03, 0.56]1.35 [0.5, 3.01]−3.2460.001HOXB2 (median [IQR])0.54 [0.18, 2.17]0.27 [0.12, 0.58]1.55 [0.35, 3.4]−3.3360.001HOXB7 (median [IQR])0.67 [0.29, 2.08]0.43 [0.25, 0.8]1.75 [0.61, 3.16]−3.1770.001CD48 (median [IQR])0.74 [0.22, 1.94]0.29 [0.18, 0.66]1.65 [0.87, 2.19]−3.3790.001 Table 2Clinical characteristics of patients in the IDH-mutant and IDH-wild-typegroups from the TCIA databaseCharacteristicsTotal patient(n = 61)Mutant(n = 32)Wild-type(n = 29)StatisticP-valueAge (mean ± SD)51.23 ± 14.3544.03 ± 13.1559.17 ± 11.2−4.815< 0.001Gender = Male (%)33 (54.1%)14 (43.75%)19 (65.52%)2.9030.088 Low-grade glioma17 (27.9%)17 (53.13%)0 (0%) High-grade glioma44 (72.1%)15 (46.88%)29 (100%)1p/19q codeletion = Non-Codel (%)50 (83.3%)21 (65.62%)29 (100%)1.540.001MGMT = Unmethylated (%)14 (26.4%)1 (3.12%)13 (44.83%)26.009< 0.001OS (median [IQR])1.31 [0.93, 2.21]2.17 [1.30, 3.59]0.79 [0.61, 1.26]4.044< 0.001Status = alive (%)32 (52.50%)27 (84.38%)5 (17.24%)27.492< 0.001MMP9 (median [IQR])0.82 [0.27, 3.86]0.32 [0.15, 0.8]3.78 [1.66, 5.42]−4.549< 0.001IGF2BP3 (median [IQR])0.56 [0.16, 3.15]0.19 [0.1, 0.45]3.02 [1.69, 4.29]−4.463< 0.001FGFBP2 (median [IQR])0.84 [0.47, 1.47]0.7 [0.38, 1.04]1.3 [0.68, 2.14]−2.5270.011EN1 (median [IQR])0.27 [0.04, 2.67]0.06 [0.03, 0.39]2.6 [0.27, 3.45]−3.729< 0.001HOXD9 (median [IQR])0.50 [0.07, 2.20]0.11 [0.03, 0.56]1.35 [0.5, 3.01]−3.2460.001HOXB2 (median [IQR])0.54 [0.18, 2.17]0.27 [0.12, 0.58]1.55 [0.35, 3.4]−3.3360.001HOXB7 (median [IQR])0.67 [0.29, 2.08]0.43 [0.25, 0.8]1.75 [0.61, 3.16]−3.1770.001CD48 (median [IQR])0.74 [0.22, 1.94]0.29 [0.18, 0.66]1.65 [0.87, 2.19]−3.3790.001 Clinical characteristics of patients in the IDH-mutant and IDH-wild-typegroups from the TCIA database The excellent performance of radiomics model in distinguishing IDH mutations The workflow for radiomic ROI delineation, feature selection, modeling, and validation was illustrated in Fig. 2. We selected nine features (denoted as RF1–RF9) for analyzing (Supplementary Figure 1). Specifically, these include one shape metric detailing the lesion’s lateral extent, alongside eight texture parameters derived from matrices like GLSZM and GLCM. These features comprehensively quantify intratumoral heterogeneity, signal intensity variations, and spatial structural complexity (Supplementary Table 2). These features were reflected in tumor morphology and texture. Fig. 2Workflow for developing radiomics models using machine learning Workflow for developing radiomics models using machine learning Among them, features RF4, RF5, RF8, and RF9 were significantly downregulated in IDH mutant gliomas, while RF1, RF2, RF3, RF6, and RF7 were significantly upregulated in IDH mutant gliomas, as shown in Fig. 3. A random forest binary classification model was developed based on the nine radiomic features. In the test set, the AUC, accuracy, precision, recall, and F1-score for the IDH mutation model were 0.922, 0.842, 0.875, 0.778, and 0.824, respectively (Table 3). Fig. 3The cloud rain plot illustrates the quantitative comparison of radiomic features between IDH wild-type and IDH mutant variants. (A-I) Comparison of radiomic features of RF1-9 between IDH wild-type and IDH mutant variants The cloud rain plot illustrates the quantitative comparison of radiomic features between IDH wild-type and IDH mutant variants. (A-I) Comparison of radiomic features of RF1-9 between IDH wild-type and IDH mutant variants In the validation set, these metrics were 0.883, 0.803, 0.800, 0.900, and 0.847, respectively (Table 3). Furthermore, the performance of our proposed Random Forest (RF) model was compared with other baseline clinical prediction methods (Table 3). In the training, test, and external validation cohorts, the RF model consistently outperformed Logistic Regression and SVM across all metrics, including AUC, Accuracy, and F1-score. For instance, in the external validation set, the RF model achieved a higher AUC (0.883) compared to SVM (0.823) and Logistic Regression (0.795), demonstrating its robust predictive capability and relative advantage for IDH mutation prediction. The calibration curve and DCA showed that the predicted results were in good agreement with the actual outcomes, achieving the maximum clinical net benefit (Fig. 4A-F). Table 3Performance comparison of different machine learning algorithms across training, test, and external validation setsModelCohortAUCAccuracyPrecisionRecallF1-scoreLogistic RegressionTraining0.8500.7800.8000.7500.774Test0.8120.7500.7800.7200.748Validation0.7950.7350.7600.7000.728Support Vector Machine (SVM)Training0.8800.8100.8300.7800.804Test0.8560.7850.8100.7450.776Validation0.8230.7600.7850.7200.751 Random Forest Training0.9850.9500.9600.9300.945Test0.9220.8420.8750.7780.824Validation0.8830.8030.8000.9000.847 Table 3Performance comparison of different machine learning algorithms across training, test, and external validation setsModelCohortAUCAccuracyPrecisionRecallF1-scoreLogistic RegressionTraining0.8500.7800.8000.7500.774Test0.8120.7500.7800.7200.748Validation0.7950.7350.7600.7000.728Support Vector Machine (SVM)Training0.8800.8100.8300.7800.804Test0.8560.7850.8100.7450.776Validation0.8230.7600.7850.7200.751 Random Forest Training0.9850.9500.9600.9300.945Test0.9220.8420.8750.7780.824Validation0.8830.8030.8000.9000.847 Performance comparison of different machine learning algorithms across training, test, and external validation sets Fig. 4Assessment of predictive performance of the radiomics model for IDH mutation gliomas. ROC curves of the IDH mutation prediction model in the testing set (A) and validation set (D), as well as decision curve analysis in the testing set (B) and validation set (E), and calibration curves in the testing set (C) and validation set (F) Assessment of predictive performance of the radiomics model for IDH mutation gliomas. ROC curves of the IDH mutation prediction model in the testing set (A) and validation set (D), as well as decision curve analysis in the testing set (B) and validation set (E), and calibration curves in the testing set (C) and validation set (F) The radiomic features of IDH mutant gliomas exhibited good prognostic predictive performance This study demonstrated that the median survival time for patients with IDH mutant gliomas was 2.17 years, compared to 0.79 years for those with IDH wild-type gliomas, indicating that patients with IDH mutations had a longer survival time (Table 2). The results of the univariate Cox regression analysis indicated that all nine radiomic features were key factors in the prognosis of gliomas (P < 0.05). Further multivariate Cox regression analysis revealed that RF4 and RF6 were independent factors for glioma prognosis (HR = 5.352, P = 0.003; HR = 0.649, P = 0.036) (Table 4). This aligns with a growing body of evidence highlighting the clinical relevance of MRI-derived radiomic biomarkers in survival prediction. Previous studies have established that high-order texture features can serve as non-invasive surrogates for overall survival, often reflecting underlying intratumoral heterogeneity and immune cell infiltration. In the context of existing radiomic biomarkers, RF4 and RF6 are particularly valuable because they not only independently predict prognosis but also correlate significantly with key IDH mutation-associated genes. For instance, RF4 demonstrated a significant positive correlation with key IDH mutation-associated genes. Conversely, RF6 exhibited a significant negative correlation with MMP9, the significant negative correlation between RF6 and MMP9 suggests that this radiomic feature may non-invasively capture the transcriptomic alterations associated with tumor invasion and immune microenvironment remodeling. By mirroring these critical molecular events, RF4 and RF6 provide supplementary prognostic information that complements established clinical and genetic markers, reinforcing the translational potential of radiogenomic models in personalized glioma management. Table 4Univariate and multivariate Cox regression analysis identifying independent prognostic risk factors of radiomic featuresNameUnivariable CoxMultivariate CoxHR-(%CI)P-valueHR-(%CI)P-valueRF10.614 (0.389–0.968)0.0360.810 (0.435–1.509)0.506RF20.404 (0.235–0.694)0.0010.656 (0.346–1.247)0.199RF30.354 (0.222–0.565)0.0001.111 (0.393–3.138)0.842RF43.688 (2.231–6.097)0.0005.352 (1.780–16.094)0.003RF51.983 (1.460–2.694)0.0000.959 (0.484–1.898)0.904RF60.567 (0.345–0.934)0.0260.649 (0.433–0.973)0.036RF70.470 (0.333–0.665)0.0001.679 (0.670–4.209)0.269RF81.969 (1.413–2.746)0.0001.365 (0.563–3.309)0.491RF91.597 (1.134–2.250)0.0071.087 (0.730–1.620)0.681 Table 4Univariate and multivariate Cox regression analysis identifying independent prognostic risk factors of radiomic featuresNameUnivariable CoxMultivariate CoxHR-(%CI)P-valueHR-(%CI)P-valueRF10.614 (0.389–0.968)0.0360.810 (0.435–1.509)0.506RF20.404 (0.235–0.694)0.0010.656 (0.346–1.247)0.199RF30.354 (0.222–0.565)0.0001.111 (0.393–3.138)0.842RF43.688 (2.231–6.097)0.0005.352 (1.780–16.094)0.003RF51.983 (1.460–2.694)0.0000.959 (0.484–1.898)0.904RF60.567 (0.345–0.934)0.0260.649 (0.433–0.973)0.036RF70.470 (0.333–0.665)0.0001.679 (0.670–4.209)0.269RF81.969 (1.413–2.746)0.0001.365 (0.563–3.309)0.491RF91.597 (1.134–2.250)0.0071.087 (0.730–1.620)0.681 Univariate and multivariate Cox regression analysis identifying independent prognostic risk factors of radiomic features Certain radiomic features associated with IDH mutations were linked to MGMT demethylation, 1p/19qcodeletion, and recurrence To further investigate the relationship between the nine radiomic features and MGMT demethylation, 1p/19q codeletion, and tumor recurrence, we performed a Mann-Whitney U test. We found eight radiomic features (RF1-RF5 and RF7-RF9) were significantly differentially expressed across different MGMT methylation statuses (Fig. 5A-5I), suggesting a potential association with temozolomide treatment sensitivity. Additionally, RF4 was found to be lowly expressed in cases of 1p/19q codeletion (Fig. 5J), while RF3 was significantly downregulated and RF4 was significantly upregulated in recurrent gliomas (Figs. 5K and 5L). Fig. 5Radiomic features and clinical characteristics. Comparison of imaging features and MGMT methylation subtypes, 1p/19q codeletion subtypes and recurrences. (A-I) Comparison of imaging features and MGMT methylation subtypes. Comparison of imaging features and the relationship with 1p/19q codeletion subtypes (J) and recurrences (K, L). (M) Spearman correlation heatmap of radiomic features and clinical characteristics Radiomic features and clinical characteristics. Comparison of imaging features and MGMT methylation subtypes, 1p/19q codeletion subtypes and recurrences. (A-I) Comparison of imaging features and MGMT methylation subtypes. Comparison of imaging features and the relationship with 1p/19q codeletion subtypes (J) and recurrences (K, L). (M) Spearman correlation heatmap of radiomic features and clinical characteristics The correlation analysis of clinical characteristics and radiomic features indicated that older patients and those with IDH wild-type gliomas exhibited high expression levels of RF4, RF5, RF8, and RF9. Moreover, RF4, RF5, RF8, and RF9 were positively correlated with shorter survival times (Fig. 5M). Screening and validation of key genes associated with IDH mutations Based on the screening criteria, eight key genes associated with IDH mutations were identified: CD48, EN1, FGFBP2, HOXB2, HOXB7, HOXD9, IGF2BP3 and MMP9. These genes are upregulated in IDH wild-type gliomas and other subtypes of glioma (Fig. 6A, Supplementary Figure 2). We validated the expression of these genes in different IDH subtypes through Western blotting, and the experimental results were consistent with the bioinformatics analysis. Notably, CD48, FGFBP2, HOXB2, HOXD9, IGF2BP3, and MMP9 were upregulated in gliomas, with their expression levels in IDH wild-type gliomas being higher than those in IDH mutant types (Fig. 6B). Fig. 6The expression levels of key genes with different IDH mutation statuses in the CGGA dataset and glioma patients. (A) Comparison of expression levels of key genes with different IDH mutation statuses in the CGGA dataset. (B) Comparison of expression levels of key genes with different IDH mutation statuses in glioma patients. N = 3 The expression levels of key genes with different IDH mutation statuses in the CGGA dataset and glioma patients. (A) Comparison of expression levels of key genes with different IDH mutation statuses in the CGGA dataset. (B) Comparison of expression levels of key genes with different IDH mutation statuses in glioma patients. N = 3 The key genes associated with IDH mutations significantly affect the survival of glioma patients To investigate the correlation between key genes associated with IDH mutations and glioma prognosis, a nomogram survival model was constructed. The results indicated that all eight key genes were positively correlated with shorter overall survival (Fig. 7A and B). Among these, IGF2BP3, EN1, and MMP9 had higher weights in predicting poor prognosis. The ROC curve demonstrated that the model had good predictive ability for the three-year overall survival of glioma patients (AUC = 0.915, P < 0.05). Clinical decision curve and calibration curve analyses showed that the model had high clinical utility and could be used to predict the impact of key IDH mutation genes on prognosis (Fig. 7C and D). Fig. 7Key Genes as prognostic factors for survival in glioma patients. (A) Nomogram for predicting prognosis. (B) ROC curve to evaluate the efficiency of the nomogram. (C) Decision curve analysis assessing the clinical utility of the nomogram. (D) Calibration curve of the nomogram. Key Genes as prognostic factors for survival in glioma patients. (A) Nomogram for predicting prognosis. (B) ROC curve to evaluate the efficiency of the nomogram. (C) Decision curve analysis assessing the clinical utility of the nomogram. (D) Calibration curve of the nomogram. Key genes associated with IDH mutations are involved in the proliferation and immune infiltration of glioma GO enrichment analysis revealed that key genes associated with IDH mutations were significantly enriched in biological processes, including DNA-binding transcription repressor activity, multicellular organism growth, and embryonic skeletal system development (Fig. 8A). The KEGG enrichment analysis indicated that the relevant signaling pathways of these genes primarily involved the IL-17 signaling pathway, and endocrine resistance, suggesting that the key genes associated with IDH mutations were involved in the malignant proliferation and immune microenvironment response of glioma. Fig. 8Functional Enrichment, Immune Infiltration, and Clinical Distribution Heatmap Analysis of Key Genes with IDH Mutations. (A) GO and KEGG enrichment analysis of differentially expressed genes in glioma with IDH mutations. (B) Heatmap of clinical relevance for key genes with IDH mutations. (C) Correlation between expression levels of key genes with IDH mutations and immune infiltration levels in gliomas Functional Enrichment, Immune Infiltration, and Clinical Distribution Heatmap Analysis of Key Genes with IDH Mutations. (A) GO and KEGG enrichment analysis of differentially expressed genes in glioma with IDH mutations. (B) Heatmap of clinical relevance for key genes with IDH mutations. (C) Correlation between expression levels of key genes with IDH mutations and immune infiltration levels in gliomas Since immune infiltration was associated with the malignant proliferation and prognosis of glioma, the CIBERSORT algorithm was used to systematically assess the correlation between key genes associated with IDH mutations and the immune levels in glioma. The results showed that CD48, HOXB7, and HOXB2 were negatively correlated with naive B cells and plasma cells, while they were positively correlated with M1 macrophages, M2 macrophages, and neutrophils. HOXD9, EN1, IGF2BP2, and MMP9 exhibited a similar trend (Fig. 8C). Furthermore, clinical correlation analysis revealed that CD48, EN1, FGFBP2, HOXB2, HOXB7, HOXD9, IGF2BP3, and MMP9 were primarily expressed in IDH wild-type and high-grade gliomas (Fig. 8B). Additionally, the expression of these eight genes was also commonly observed in populations with MGMT demethylation and older patients. Visualization of the subgroup distribution and spatial localization of IDH mutation-associated keygenes in glioma In the single-cell sequencing data of glioma, key genes associated with IDH mutations were found to be highly expressed in specific cell populations. The single-cell sequencing data were categorized into six cell subpopulations through PCA analysis. Among these, CD48 exhibited the highest expression in microglia, while FGFBP2, EN1, HOXB2, HOXB7, and HOXD9 were most abundantly expressed in astrocytes. IGF2BP3 was widely expressed in astrocytes, mesenchymal glioblastoma cells, and neurons, whereas MMP9 was primarily found in the astrocyte and microglia subpopulations. Interestingly, these eight key genes were almost not expressed in endothelial cells and oligodendrocytes (Fig. 9A). In terms of spatial expression, the astrocyte, erythrocyte, and fibroblast cell subpopulations were widely distributed in glioma. Additionally, IGF2BP3 was concentrated in the regions of these cell subpopulations. MMP9 and HOXD9 were primarily localized in the astrocyte regions, with some distribution also observed in the erythrocyte and fibroblast areas. In contrast, FGFBP2, HOXB7, and HOXB2 exhibited minimal to no distribution (Fig. 9B-C). Fig. 9Analysis of key genes related to IDH mutations. (A). Analysis of Single-Cell Distribution Characteristics of Key Genes with IDH Mutations Based on Single-Cell Sequencing Data from the GEO Database. (B-C). Analysis of Spatial Expression Characteristics of Key Genes with IDH Mutations in Glioma Tissue Based on Public Spatial Transcriptomics Datasets Analysis of key genes related to IDH mutations. (A). Analysis of Single-Cell Distribution Characteristics of Key Genes with IDH Mutations Based on Single-Cell Sequencing Data from the GEO Database. (B-C). Analysis of Spatial Expression Characteristics of Key Genes with IDH Mutations in Glioma Tissue Based on Public Spatial Transcriptomics Datasets A notable correlation between radiomic features and key genes associated with IDH mutations Above all, we discovered that certain radiomic features were associated with the age of onset, prognosis, and treatment sensitivity of glioma, while key IDH mutation-associated genes were linked to glioma subtypes and poor prognosis. Therefore, we aimed to reveal the correlation between radiomic features and key IDH mutation-associated genes through correlation analysis. As shown in Fig. 10, RF3 and RF7 exhibited a significant negative correlation with the eight key IDH mutation-associated genes, while RF4 demonstrated a significant positive correlation with these genes. Additionally, RF2 and RF6 were significantly negatively correlated with MMP9. Notably, RF9 showed no significant correlation with any of the key IDH mutation-associated genes. Overall, the radiomic features we analyzed not only predicted IDH molecular subtypes but also reflected, to some extent, the imaging characteristics of certain key IDH mutation-associated genes on MRI. Fig. 10Association Between eight Gene Modules and nine Imaging Features Related to IDH Mutations Association Between eight Gene Modules and nine Imaging Features Related to IDH Mutations Discussion The key to limiting overall survival in glioma patients lay in the pathological grade of the lesions, morphological heterogeneity, and the abnormal expression of key genes [18, 19]. Through radiomics analysis, we found that nine radiomic features exhibited differences across various IDH molecular subtypes of gliomas and had a significant impact on prognosis, with RF4 and RF6 identified as independent predictors of poor prognosis. Transcriptomic analysis revealed that eight key genes associated with glioma development showed differential expression between IDH wild-type and mutant gliomas, all of which were linked to poor prognosis. Correlation analysis indicated that the remaining eight radiomic features, except for RF9, were significantly correlated with the key genes associated with IDH mutations. The 2021 WHO classification of central nervous system tumors incorporated molecular features into the histological diagnosis of gliomas, with IDH mutations serving as an important predictor of prognosis for glioma patients [20]. Compared to IDH mutant glioma patients, those with IDH wild-type tumors have a shorter survival time. This study analyzed 61 glioma samples from the TCGA database, revealing that the average survival time for IDH mutant glioma patients was 2.17 years, while for IDH wild-type patients, it was 0.79 years (P < 0.05). The reasons for the prognostic differences among the various IDH subtypes of gliomas remain not fully understood.The tricarboxylic acid cycle is a crucial pathway for cellular energy metabolism, responsible for the final metabolism of carbohydrates, lipids, and amino acids [21]. Isocitrate dehydrogenase (IDH) plays a key rate-limiting role in this cycle. Mutations in IDH1/2 not only result in the loss of normal catalytic function but also confer a novel ability to reduce NADPH-dependent α-ketoglutarate to the tumor metabolite D-2-hydroxyglutarate [22]. This alteration disrupts cellular metabolism and epigenetic regulation, thereby promoting tumorigenesis and progression. Some researchers suggest that IDH mutations may inhibit the expression of angiogenic factors in astrocytomas, while IDH wild-type gliomas show a high enrichment of hypoxia and angiogenesis-related genes [23]. Our analysis revealed that nine radiomic features and eight key genes associated with IDH mutations exhibited differential expression across various IDH molecular subtypes and were correlated with overall survival in patients. Notably, RF4 and RF6 were identified as independent prognostic factors for gliomas. MRI is the preferred tool for assessing intracranial tumors, providing detailed anatomical structures and morphological features of the tumors. However, it has limited effectiveness in distinguishing the molecular and pathological characteristics of gliomas. Radiomics is an analysis technique based on medical imaging that can rapidly, non-invasively, and accurately identify tumor types and molecular subtypes, as well as predict disease progression under different interventions [24]. Many researchers have successfully applied MRI-based radiomics methods to predict the IDH molecular subtype of gliomas. For instance, Yu et al. developed a predictive model for IDH1 status using genetic algorithms based on T2-FLAIR MRI, achieving an accuracy of 80%, sensitivity of 83%, and specificity of 74% [25]. Kim et al. demonstrated that machine learning models in radiomics hold promise for predicting the IDH mutation status in gliomas [26]. Our study yielded similar results, with RF4, RF5, RF8, and RF9 showing significantly lower expression in IDH mutant gliomas, while RF1, RF2, RF3, RF6, and RF7 exhibited significantly higher expression in IDH wild-type gliomas. In the test set, the IDH mutation model achieved an accuracy of 0.842 and precision of 0.875, while in the validation set, these values were 0.803 and 0.800, respectively. We found significant differences in the eight features RF1-RF5 and RF7-RF9 in the MGMT classification. Additionally, RF3 was significantly downregulated, while RF4 was significantly upregulated in recurrent gliomas. These findings suggest that our MRI-based radiomics model has the potential to differentiate IDH status, and some of these radiomic features are associated with treatment sensitivity and recurrence in gliomas. The rapid advancement of transcriptomics has provided significant value in disease diagnosis and differentiation, treatment and drug development, as well as prognostic assessment. In 2014, RNA sequencing was utilized to investigate the origins and transmission patterns of the Ebola virus during the West African outbreak [27]. Kang et al. performed bulk sequencing and transcriptomic analysis on 24 untreated gastric cancer patients and matched non-malignant gastric tissues, revealing differences between the gastric cancer tumor microenvironment and normal gastric stroma [28]. They quantified the abundance and cellular activity changes associated with malignant transformation for each stromal cell type. Activated fibroblasts and endothelial cells were particularly prominent in tumors, while direct ligand-receptor interactions between gastric cancer cells and the tumor microenvironment contributed to immune evasion. Comitani et al. categorized 13,313 tumor and non-tumor samples into 455 categories using a multiscale approach, constructing a tumor subtype map comprising 192 types derived from five tumor clusters [29]. Single-cell sequencing can reveal changes in gene expression during tumor progression. Integrating single-cell sequencing of adjacent normal tissues and adenomas at different stages in patients can comprehensively uncover genomic alterations, clonal structures, and metabolic dynamics during tumorigenesis, providing new insights for inhibiting tumor progression [30]. This study performed a differential analysis of gliomas compared to normal brain tissue, as well as between IDH wild-type and mutant gliomas, identifying eight genes were significantly differentially expressed in gliomas and across different IDH subtypes. Nomogram analysis indicated that all eight genes were associated with poor prognosis, with a 3-year prognostic ROC AUC value of 0.915. Decision curve analysis (DCA) demonstrated that the radiomics model provided a high clinical net benefit. Immune infiltration analysis showed that CD48, MMP9, HOXB7, and HOXB2 were significantly positively correlated with M2 macrophages and neutrophils, while being significantly negatively correlated with plasma cells and naïve B cells. Single-cell analysis revealed that FGFBP2, EN1, HOXB2, HOXB7, and HOXD9 were highly expressed in astrocytes, while IGF2BP3 was widely expressed in astrocytes, mesenchymal glioblastoma cells, and neurons, with a concentrated distribution in astrocyte and fibroblast subpopulations. Furthermore, these eight key genes are involved in biological processes such as chromosome segregation, mitotic nuclear division, and their regulation.The high expression of MMP9, IGF2BP3, and HOXD9 is associated with glioma invasion, migration, and malignant proliferation, and is related to immune infiltration. Literature has confirmed that MMP9, HOXB2, and HOXB7 are associated with IDH mutations, while CD48, HOXB2, HOXB7, HOXD9, EN1, and IGF2BP3 are linked to poor prognosis in gliomas. Furthermore, exploring the mechanistic roles of these key genes provides deeper insights into glioma pathogenesis. For instance, EN1, a homeobox transcription factor, has been implicated in promoting glioma cell proliferation and migration primarily through the activation of the Hedgehog signaling pathway [31]. Regarding FGFBP2, although current research in glioma is limited, its known function in modulating fibroblast growth factor (FGF) signaling suggests it could play a crucial role in tumor angiogenesis and microenvironment remodeling [32], which are critical for glioma progression. While our study successfully established the prognostic and diagnostic value of these genes using radiomics and validated their differential expression at the protein level via Western blotting, the specific in vitro and in vivo oncogenic mechanisms of FGFBP2, EN1, and other identified genes warrant further functional exploration. Genetic testing necessitates the acquisition of tissue samples, which carries certain risks. By integrating radiomics and transcriptomics for comprehensive analysis, predictive models can be constructed to predict gene expression levels of diseases through the analysis of radiomic features. In this study, RF3 and RF7 showed significant negative correlations with the expression of CD48, EN1, FGFBP2, HOXB2, HOXB7, HOXD9, IGF2BP3, and MMP9, while RF4 exhibited a significant positive correlation. Wang et al. combined radiomic features with prognostic genes related to hepatocellular carcinoma to construct a radiogenomic prognostic model with good predictive capability, outperforming the TNM staging model [33]. They found that a high radiogenomic score was associated with poorer overall survival and lower levels of immune infiltration. Similarly, Gao et al. created a model to predict the expression of immune-related genes in clear cell renal cell carcinoma by integrating radiomic features with tumor immune-related genes [34]. He et al. utilized a radiomic model to predict CTLA4 expression in clear cell renal cell carcinoma, revealing a significant increase in infiltrating M2 macrophages in the high-score group.These findings suggest that the combination of radiomics and transcriptomics can effectively predict disease progression and prognosis [35]. However, there is currently a lack of research on the pathological classification of gliomas in conjunction with radiomics and transcriptomics. Our results indicate that MRI radiomic features can, to some extent, predict IDH classification, prognosis, and the expression of key genes related to IDH mutations. Furthermore, our results showed that radiomic features could non-invasively reflect the molecular landscape of gliomas.In this study, we utilized MNI space registration to standardize multi-scanner data. While spatial interpolation can theoretically affect texture features, our ICC analysis confirmed that the features incorporated into our model maintained high stability (mean ICC = 0.85). This suggests that our radiomics signature captures intrinsic tumor heterogeneity rather than pre-processing artifacts. However, this study has some limitations. Firstly, while we validated the protein expression levels of the eight key genes (e.g., FGFBP2, EN1) using clinical glioma samples, this study primarily focused on macro-level radiomic feature extraction and transcriptomic profiling. We did not perform in vitro (e.g., cell proliferation or invasion assays) or in vivo (e.g., xenograft models) functional experiments to elucidate the exact mechanistic roles of these genes in glioma progression. Future studies comprising comprehensive molecular biology experiments are necessary to determine the causative mechanisms of these targets.Secondly, our radiomic feature extraction relied solely on T1CE sequences. Although T1CE provides clear visualization of tumor boundaries for ROI delineation across multi-center datasets, it may not capture the full extent of tumor heterogeneity compared to multi-parametric MRI. The exclusion of other available sequences, such as T2WI, FLAIR, and DWI, restricts our model’s ability to assess peritumoral edema and cellularity. Future prospective studies should employ multi-sequence radiomic models to comprehensively evaluate the tumor microenvironment and improve predictive accuracy.Thirdly, we acknowledge an imbalance in our external validation cohort between the IDH-wildtype (n = 40) and IDH-mutant (n = 26) groups. Given the study’s focus on IDH-mutant gliomas and the relatively modest expression differences observed in the eight key genes, this sample size disparity may limit the statistical power and generalizability of our current findings. Therefore, expanding the cohort with a larger, more balanced sample of IDH-mutant cases in future multicenter studies will be crucial to fully validate the robustness of our predictive models.Fourthly, the sample size of the TCIA cohort used for model development is relatively small (61 patients). Despite applying cross-validation, splitting this limited dataset at a 7:3 ratio to construct a nine-feature model poses a potential risk of overfitting.Furthermore, although our model showed robust performance in external validation, we did not explicitly control for tumor grade or perform subgroup analyses by WHO grade. Given the strong biological correlation between IDH mutations and lower histological grades, future studies utilizing larger multicenter cohorts are warranted to validate the model’s independent predictive power within specific grade categories and to further elucidate the complex interactions between radiomic features, tumor grade, and molecular subtypes. The clinical translation of our integrated radiomics-transcriptomics model holds significant promise for personalized neuro-oncology. Unlike traditional biopsy-based genotyping, which is the current diagnostic gold standard but inherently invasive and limited by sampling bias in heterogenous gliomas, our model provides a non-invasive, cost-effective alternative. By leveraging routine pre-operative MRI data and transcriptomic signatures, this ‘virtual biopsy’ allows for a comprehensive assessment of the entire tumor volume, rather than a localized tissue sample. This is particularly crucial for lesions located in eloquent areas where surgical biopsy poses high risks. Furthermore, while biopsy provides a static snapshot, our radiomics approach enables longitudinal monitoring of molecular evolution during therapy, offering a dynamic tool for real-time treatment adjustment. Although it does not currently replace histopathological confirmation, this model serves as a powerful adjunctive tool to refine surgical margins and stratify patients for targeted therapies prior to intervention. Conclusion This study identified nine radiomic features through MRI radiomics analysis that can effectively differentiate between various IDH subtypes of gliomas. These features are closely linked to key genes associated with IDH mutations and poor prognosis in gliomas. Notably, CD48, EN1, FGFBP2, HOXB2, HOXB7, HOXD9, IGF2BP3, and MMP9 are significantly overexpressed in wild-type IDH gliomas, suggesting a poor prognosis. In summary, the integration of radiomics and transcriptomics shows promise for developing more accurate predictive models for the differentiation, classification, and prognosis of gliomas, thereby providing robust support for personalized treatment. Supplementary Information Supplementary_Material for publication Supplementary_Material for publication