Predicting glioma survival and extracellular matrix remodeling through MRI radiogenomics

Summary

Extracellular matrix (ECM) remodeling is essential for glioma invasion, yet lacks non-invasive assessment methods. This study employs radiogenomics to enable non-invasive survival prediction and ECM remodeling assessment in glioma. Utilizing a multi-dataset data (n = 891), an 11-feature radiomics signature is developed stratifying patients into low- and high-Rad-score groups (area under the receiver operator characteristic curve [AUC] = 0.886, 95% confidence interval [CI]: 0.807–0.964 in the training set from two local centers; AUC = 0.828, 95% CI: 0.796–0.893 in the validation set from five public datasets). Radiogenomic analysis (n = 572) reveals differentially expressed genes significantly associated with Rad-scores, particularly enriched in pathways associated with ECM remodeling, and identifies seven related hub genes (MMP2, MMP9, CXCL8, TIMP1, IL-6, COL1A2, and CCL2). These findings are validated using an external radiogenomic dataset and orthotopic (both syngeneic and xenograft) mouse models, where silencing MMP2 reduced Rad-scores and tumor infiltration. This study highlights the potential of MRI-based radiomics signatures in assessing ECM remodeling for survival prediction and improved glioma clinical management.

Graphical abstract

Highlights

• •An MRI radiomics signature non-invasively assesses glioma ECM remodeling
• •The radiomics signature robustly stratifies glioma patients into prognostic groups
• •Radiogenomics links the signature to ECM pathways and 7 hub genes (e.g., MMP2)
• •Silencing MMP2 reduces Rad score and tumor infiltration in in vivo mouse models

Bie et al. develop an MRI-based radiomics signature that non-invasively assesses extracellular matrix remodeling in glioma, predicts patient survival, and identifies MMP2 as a key therapeutic target. This approach offers a promising tool for improved clinical management.

Introduction

The extracellular matrix (ECM) serves as the structural scaffold of the brain, differing significantly from most organs.¹ The ECM plays a fundamental role in regulating glioma cells and surrounding stromal cells, thereby facilitating essential cellular functions throughout the process of glioma invasion.² The ECM integrates both structural elements—including fibronectin, hyaluronic acid, and collagens—and enzymatic modulators such as lysyl oxidase, as well as extracellular and membrane-bound matrix metalloproteinases (MMPs).³

ECM remodeling is a pivotal mechanism driving glioma migration and invasion, which includes dynamic alterations in the expression of enzymatic modulators and depositing and degrading of ECM elements.^1,2 Standard glioma treatments can prolong survival but do not prevent the tumor’s invasive growth pattern, which is closely related to tumor recurrence. Targeting ECM remodeling can enhance immune cell trafficking into the tumors, offering a promising strategy to reverse immune cell responses.^4,5 New approaches to control specific integrins, target mechano-sensitive receptors, and develop highly specific MMP inhibitors are currently being pursued in the ECM remodeling-targeting therapies.^2,3 MMPs were profoundly studied and considered a promising target in glioma invasion, showing broad and overlapping substrate specificities to degrade ECM substrates. In addition, increased MMP activity is common in glioma patients resistant to temozolomide.^6,7 Therefore, surveillance of MMPs can significantly enhance clinical decision-making, therapeutic response, and disease progression by providing non-invasive real-time insights into tumor microenvironment.⁸

Radiomics, now an established non-invasive modality, captures information on shape, intensity, and texture at the raw data level from radiographic medical images.^9,10 Through combined advances in MRI and high-throughput artificial intelligence-based analysis, radiomics can reflect the original state of the tumor and its microenvironment, allowing for the exploration of a tumor’s biological foundation.¹¹ Radiogenomics, which examines the relationship between imaging phenotypes and genomics, offers the potential for longitudinal application in active surveillance, monitoring, and treatment optimization.¹² Radiogenomics can reveal the microscopic nature of macroscopic phenotypes and help interpret the biological functions underlying imaging phenotypes, thereby facilitating the clinical translation of radiogenomics.¹³

This study aimed to improve the biological interpretation of individual radiomic phenotypes using a radiogenomic approach for ECM remodeling assessment and survival prediction. These results were validated through an independent radiogenomic dataset and an orthotopic mouse model.

Results

Clinical characteristics and radiomics signature construction

Overview of the study design was illustrated (Figure 1). For the radiomics analysis, 891 glioma patients were included: 389 in the radiomics discovery set (59.74 ± 13.01 years; 211 males, 178 females) and 502 in the radiomics validation set (50.03 ± 14.64 years; 287 males, 215 females). For the radiogenomics analysis, 572 glioma patients were included: 241 in the radiogenomics discovery set (52.76 ± 15.26 years; 136 males, 115 females) and 331 in the radiogenomics validation set (51.58 ± 14.51 years; 173 males, 158 females). Patient characteristics are shown (Table 1; Figure 2). The distributions of sex assigned at birth and tumor grade were similar between the discovery and validation sets. However, significant differences were observed in age, isocitrate dehydrogenase 1 (IDH1) status, and overall survival (OS). These baseline discrepancies could be explained by variations in data sources and a higher proportion of IDH-wild-type glioblastoma (GBM) patients in the validation set. To account for these potential confounders, all significant clinical variables were included as covariates in subsequent multivariate Cox regression analysis to ensure an independent assessment of the radiomic signature’s prognostic value.

Figure 1.

Graphical illustration of study design

TCIA, The Cancer Imaging Archive; TCIA-LGG, The Cancer Genome Atlas Low Grade Glioma; UCSF-PDGM, The University of California San Francisco Preoperative Diffuse Glioma MRI; TCGA-GBM, The Cancer Genome Atlas Glioblastoma Multiforme Collection; CPTAC-GBM, The Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multiforme Collection; IvyGAP, Ivy Glioblastoma Atlas Project; ROI, region of interest; THPA, The Human Protein Atlas.

Table 1.

Characteristics of radiomics and radiogenomics analysis sets

Characteristics	Radiomics analysis sets (n = 891)		Radiogenomics analysis set (n = 572)		p value
	Radiomics discovery set (n = 389)	Radiomics validation set (n = 502)	Radiogenomics discovery set (n = 241)	Radiogenomics validation set (n = 331)
Age, years	59.51 ± 13.33	52.67 ± 14.84	60.16 ± 12.70	52.89 ± 14.92	<0.001
Sex assigned at birth, n	–	–	–	–	0.461
Male	211	287	144	196	–
Female	178	215	97	135	–
Grade	–	–	–	–	0.291
Low-grade glioma	206	289	126	192	–
Glioblastoma	183	213	115	139	–
IDH1 status, n	–	–	–	–	0.025
Mutation	181	250	119	166	–
Wild type	187	206	109	153	–
Not available	21	46	13	12	–
Mean OS, months	36.84	24.18	39.10	22.31	<0.001
Mean PFS, months	25.18	18.31	27.45	15.94	<0.001

The radiogenomics discovery and validation sets were subsets of the radiomics sets by excluding cases without RNA sequencing data for further analysis.

OS, overall survival; PFS, progression-free survival.

Figure 2.

Performance of the radiomics prediction model in prognosis prediction

(A and D) The differences of survival and clinical characteristics (sex assigned at birth, age, tumor grade, and IDH status) in glioma patients in low- and high-Rad-score groups were shown. Patients in the radiomics discovery and validation sets are arranged in descending order of Rad-score values.

(B, C, E, and F) Kaplan-Meier curves showed the OS and PFS of patients in low- and high-Rad-score groups. Patients in the low-Rad-score group had better prognosis in both radiomics discovery and validation sets.

(G and J) The 1-, 2-, and 3-year OS and PFS of glioma patients were predicted by the Rad-score nomogram.

(H, I, K, and L) (H and K) The calibration plots showed the comparison between predicted and actual OS for 1-, 2-, and 3-year OS probabilities in the discovery set and (I and L) in the validation set.

For each patient, a total of 4,200 radiomics features were extracted (1,050 radiomics features per sequence from pre-contrast T1-weighted imaging [T1WI], post-contrast T1WI, T2-fluid-attenuated inversion recovery [FLAIR], and T2-weighted imaging [T2WI], respectively). The radiomics features were ranked after Z score normalization using three feature selection methods, as detailed in Data S1. We then evaluated and compared the prediction models built with different classifiers. Among these, the linear discriminative analysis (LDA) classifier, utilizing 11 selected radiomics features, demonstrated superior performance for survival prediction. It achieved areas under the receiver operator characteristic curve (AUCs) of 0.886 (95% confidence interval [CI]: 0.807–0.964) in the training set, 0.828 (95% CI: 0.796–0.893) in the validation set, 0.825 (95% CI: 0.756–0.882) in the test set, and 0.904 (95% CI: 0.871–0.937) in the 5-fold cross-validation (Table S1). Based on this result, an 11-feature radiomics signature was constructed, with each feature weighted by the coefficients derived from the LDA algorithm. The Rad-score for each patient was subsequently calculated using the following equation:

Rad-score = −0.618282262∗T2FLAIR_original_firstorder_Skewness-0.190926393∗T2FLAIR_wavelet-LLH_glcm_Correlation+0.796660069∗T2FLAIR_wavelet-LLH_glszm_SizeZoneNonUniformityNormalized-0.119149136∗T2FLAIR_wavelet-HLH_glrlm_LongRunEmphasis-0.863982266∗T2FLAIR_wavelet_HHL_glcm_Imc2+0.346714047∗T2_wavelet-LHL_firstorder_Mean + 0.291671189∗T2_wavelet-LHH_gldm_DependenceVariance-0.346085343∗T1ce_wavelet-LHL_glcm_InverseVariance-0.120428035∗T1ce_wavelet-LHL_glszm_ZonePercentage+0.701250879∗T1ce_wavelet-HLL_glcm_Imc2+0.44995381∗T1ce_wavelet-HLL_glcm_MCC.

A set of second-order texture features was extracted to quantitatively assess intratumoral heterogeneity from MRI. Specifically, features derived from the Gray Level Size Zone Matrix, Gray Level Co-occurrence Matrix, Gray Level Run Length Matrix, and Gray Level Dependence Matrix were employed to characterize the spatial distribution and relationships of pixel intensities. These metrics are hypothesized to reflect underlying pathological tissue characteristics, such as disorganized ECM and the presence of necrotic cores.

Key features comprising our radiomics signature include SizeZoneNonUniformityNormalized, reflecting heterogeneity in the size of homogeneous regions; LongRunEmphasis, indicating the prevalence of coarse textures; IMC2 and MCC, measuring the complexity and information content of texture patterns; and Skewness, Correlation, Mean, DependenceVariance, InverseVariance, and ZonePercentage, which collectively describe the statistical distribution, dependency, and coarseness of the image texture. These features were weighted to capture distinct spatial aspects of the tumor microenvironment, commonly altered by the presence of necrosis, edema, and hemorrhage.

The integrated radiomics signature demonstrated a significant association with patterns of image homogeneity and multi-scale texture complexity. This association suggests that the signature captures phenotypic manifestations of underlying biological processes in glioma, thereby bridging the gap between abstract features and tumor biology and enhancing its interpretability.

Performance of the radiomics prediction model in prognosis prediction

The relationship between the Rad-score of radiomics prediction model and the clinicopathological features of patients is shown in the heatmaps (Figures 2A and 2D). The median Rad-score of 0.187 in the radiomics discovery set was used as the cutoff value to stratify patients into low- and high-Rad-score groups. The prognostic performance of the radiomics signature in OS prediction was assessed using Kaplan-Meier curves across both discovery and validation sets (log rank test, hazard ratio [HR] = 2.668, p < 0.001; log rank test, HR = 2.545, p < 0.001, Figures 2B and 2E). Given that progression-free survival (PFS) is widely recognized as a surrogate endpoint for evaluating treatment responses in glioma patients receiving standard or experimental therapies, the predictive capability of constructed radiomics signature for PFS was accordingly evaluated (log rank test, HR = 1.746, p < 0.001; log rank test, HR = 1.673, p = 0.003, Figures 2C and 2F). Patients in the high-Rad-score group had significantly shorter survival and in both sets, indicating that a high Rad-score predicts a worse prognosis for glioma patients. Rad-score was identified as an independent prognostic factor for OS in the radiomics discovery and validation sets using univariate and multivariate Cox regression analysis (Table 2). Both the Rad-score nomogram for 1-, 2-, and 3-year survivals and calibration curves demonstrated satisfactory predictive performance (Figures 2G–2L). The Rad-score nomogram achieved a concordance index (C-index) of 0.890 (95% CI: 0.877–0.903) for OS prediction in the discovery set and 0.803 (95% CI: 0.786–0.819) in the validation set. To broaden the clinical applicability of the radiomic signature, we further evaluated its performance in predicting PFS in glioma patients, and it demonstrated robust predictive performance, with a C-index of 0.896 (95% CI: 0.885–0.907) in the discovery set and 0.780 (95% CI: 0.763–0.798) in the validation set.

Table 2.

Cox regression analysis of prognostic factors in radiomics discovery and validation sets

Characteristics	Radiomics discovery set				Radiomics validation set
	Univariate analysis		Multivariate analysis		Univariate analysis		Multivariate analysis
	HR (95% CI)	p value	HR (95% CI)	p value	HR (95% CI)	p value	HR (95% CI)	p value
Age	1.002 (0.993–1.007)	0.985	–	–	1.035 (1.015–1.056)	<0.001	1.034 (1.013–1.054)	0.001
Sex assigned at birth
Male	Ref.	–	Ref.	–	Ref.	–	–	–
Female	0.513 (0.413–0.638)	<0.001	1.335 (0.321–5.555)	0.691	1.274 (0.806–2.014)	0.299	–	–
Grade
Low-grade glioma	Ref.	–	Ref.	–	Ref.	–	Ref.	–
Glioblastoma	0.507 (0.407–0.630)	<0.001	0.287 (0.070–1.173)	0.082	0.429 (0.156–1.176)	0.1	0.429 (0.156–1.182)	0.102
IDH status
Mutation	Ref.	–	Ref.	–	Ref.	–	–	–
Wild type	0.610 (0.491–0.757)	<0.001	1.338 (0.941–1.902)	0.105	0.872 (0.540–1.408)	0.575	–	–
Not available	0.579 (0.392–0.857)	0.006	1.442 (0.830–2.507)	0.194	0.446 (0.108–1.841)	0.264	–	–
Rad-score	1.253 (1.128–1.391)	<0.001	1.280 (1.149–1.425)	<0.001	3.554 (2.573–4.910)	<0.001	3.526 (2.559–4.860)	<0.001

We observed a significant difference in the proportions of low- and high-Rad-score patients between low-grade glioma (LGG) and GBM in the discovery (Figures 3A and 3B, both p < 0.001). Rad-score was identified as an independent prognostic factor for OS in the radiomics discovery and validation sets using univariate and multivariate Cox regression analyses (Table S2, all p < 0.05).

Figure 3.

Grade-based analysis of the radiomics signature in glioma

(A and B) The proportions of low-grade glioma (LGG) and glioblastoma (GBM) differed significantly between the low- and high-risk groups in the discovery and validation sets, respectively. Significant differences in overall survival (OS) and progression-free survival (PFS) were observed between the low- and high-risk groups among both LGG (C–F) and GBM (G–J) patients across the discovery and validation sets.

Significant differences in OS and PFS were observed between the low- and high-Rad-score groups among LGG patients in the discovery (Figures 3C and 3E, log rank test, HR = 1.935, p < 0.001; log rank test, HR = 1.807, p < 0.001) and validation sets (Figures 3D and 3F, log rank test, HR = 1.406, p = 0.039; log rank test, HR = 1.521, p = 0.024). The constructed radiomics signature demonstrated satisfactory performance in identifying higher-risk LGG patients in the discovery (C-index of 0.770 [95% CI: 0.754–0.786] for OS and 0.763 [95% CI: 0.746–0.781] for PFS) and validation sets (C-index of 0.720 [95% CI: 0.706–0.774] for OS and 0.712 [95% CI: 0.699–0.725] for PFS).

Then, we analyzed the OS and PFS of GBM in the low- and high-Rad-score groups. Significant differences in OS and PFS were observed between the low- and high-Rad-score groups among GBM patients in the discovery (Figures 3G and 3I, log rank test, HR = 1.524, p = 0.011; log rank test, HR = 1.699, p = 0.003) and validation sets (Figures 3H and 3J, log rank test, HR = 1.708, p < 0.004; log rank test, HR = 1.937, p = 0.010). The constructed radiomics signature demonstrated satisfactory performance in identifying higher-risk GBM patients in the discovery set (C-index of 0.752 [95% CI: 0.732–0.771] for OS and 0.751 [95% CI: 0.732–0.770] for PFS) and validation set (C-index of 0.720 [95% CI: 0.706–0.745] for OS and 0.714 [95% CI: 0.693–0.734] for PFS).

Rad-score was identified as a specific radiomics indicator for ECM remodeling-associated pathways and ECM remodeling-related hub genes

To annotate the relationship between the radiomics signature and the biological functions in the discovery sets, the differentially expressed gene (DEG) analysis identified that genes significantly correlated with Rad-score (Supplement spreadsheet 2). The volcano plot and hierarchical clustering of DEGs between low- and high-Rad-score groups indicated that the landscape of genomics differed between low- and high-Rad-score groups (Figure 4C). In the Gene Ontology (GO)-Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis, significant clustering of ECM remodeling-associated pathways was observed, indicating that these pathways may be closely related to Rad-score, including collagen-containing ECM, positive cell adhesion regulation, and cell-cell adhesion regulation. GSEA similarly emphasized pathways associated with the ECM and metalloproteinases, including MMP activation, the fra pathway, proinflammatory and profibrotic mediators, the syndecan 1 pathway, collagen degradation, collagen fibrils and other multimeric structures, interleukin-4 and interleukin-13 signaling, matrisome-associated processes, and matrisome. The GO-KEGG and GSEA results demonstrated that radiogenomics-related biological processes were closely associated with ECM remodeling (Figures 4D and 4E).

Figure 4.

Functional analysis of radiogenomics landscape

(A) After feature selection, the 11-feature radiomics signature exhibited the best diagnostic performance. The areas under the receiver operator characteristic curve (AUCs) curve were 0.886 in the training set, 0.845 in the test set, 0.828 in the validation set, and 0.904 in the 5-fold cross-validated (CV) training set.

(B) The weighted coefficients of selected radiomics features.

(C) The volcano plot and hierarchical clustering of differentially expressed genes (DEGs) between low- and high-Rad-score groups indicated that the landscape of genomics differed between low- and high-Rad-score groups.

(D and E) The biological functional annotations for DEGs were performed by Gene Ontology (GO) function and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis, as well as gene set enrichment analysis (GSEA). The top 10 radiogenomics-related biological processes were significantly correlated with the formation of ECM and activation of MMPs in GO-KEGG and GSEA.

(F and G) Clustering and merging dendrogram of genes based on topological overlap performed by weighted gene co-expression network analysis (WGCNA) to identify the top 3 highest associated modules with Rad-score. Each row corresponds to a module eigengene with the corresponding correlation coefficient and p value from the left to the right side.

(H and I) The most significant sub-network module and the top 10 genes with the highest maximum clique centrality score were identified by the “MCODE” and “CytoHubba” plug-in in Cytoscape.

(J–M) (J and K) Hub genes had high connectivity. (L and M) Hub genes (MMP2, MMP9, CXCL8, TIMP1, IL-6, COL1A2, and CCL2) showed significantly up-regulated mRNA expression in the high-Rad-score group compared with the low-Rad-score group in both discovery and validation sets (∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001).

Then, the most significant sub-network module and top 10 genes were identified by “MCODE” and “CytoHubba” plug-in (Figures 4H and 4I), 38 co-expression modules were identified from the gene dendrogram (Figure S2), and after calculating the dissimilarity of module eigengenes, the turquoise module (correlation coefficient = 0.78; 2,261 genes), antiquewhite4 module (correlation coefficient = 0.63; 2,344 genes), and black module (correlation coefficient = −0.58, 2,326 genes) showed the strongest correlations with the Rad-score (Figures 4F and 4G). Finally, by intersecting the results from the two different screening methods, seven ECM remodeling-related hub genes (MMP2, MMP9, CXCL8, TIMP1, IL-6, COL1A2, and CCL2) were identified with high connectivity and significantly upregulated expression in the high-Rad-score group (Figures 4J–4M). The interaction network among these hub genes revealed strong enrichment using STRING (https://cn.string-db.org/, Table S3), which indicated that these proteins interact more frequently than a random set of proteins of similar size and degree distribution drawn from the genome and at least partially biologically connected as groups in glioma. The Cox regression analysis of the seven hub genes revealed a positive statistically significant correlation with OS in both the radiogenomics discovery and validation sets (Table S4).

Subsequently, based on the Rad-score, we identified DEGs in LGG and GBM. Using the same methodology applied to the pan-glioma cohort, we derived candidate hub genes for each grade (Supplement Spreadsheet 3). Among the seven previously defined ECM-associated hub genes, we found that three (CCL2, MMP2, and interleukin-6 [IL-6]) commonly constitute in both LGG and GBM. Additionally, two genes (TIMP1 and MMP9) were identified as GBM-specific hub genes. This grade-based stratification reveals that these hub genes associated with radiomics characteristics, are highly context-dependent and represent unique biological processes relied upon by tumors at different malignancy stages. The LGG-specific hub genes represent maintenance mechanisms that support the relatively slow-growing state of LGGs. Conversely, the GBM-specific hub genes are likely critical for conferring the highly malignant features of GBM, highlighting a fundamental molecular distinction from LGG.

Validation of the relationship between radiomics signature and seven ECM remodeling-related hub genes by histology

Compared to normal brain tissues, the expression of ECM remodeling-related seven proteins MMP2, MMP9, TIMP1, COL1A2, and CCL2 increased to varying extents (with medium to low antibody staining) in glioma, with expression rising as the tumor grade increased. However, protein expression of CXCL8 and IL-6 was rarely detected in The Human Protein Atlas (THPA). MMP2, TIMP1, and CCL2 were mainly localized in the cytoplasm or cell membrane, while MMP9 and COL1A2 were predominantly detected in the nucleus (Figure 5A).

Figure 5.

The protein expression patterns of radiogenomics-based hub genes

(A) The protein expression patterns of 7 hub genes between normal brain tissues and glioma were determined by immunohistochemistry (IHC) staining from The Human Protein Atlas (THPA) database. Compared with normal brain tissues, the protein expression levels of hub genes increased in glioma as the grading of the tumor increased, except for CXCL8 and IL-6 without obvious detection.

(B) IHC staining of acquired glioma specimens in proliferative core exhibited higher protein expression levels of MMP2, MMP9, TIMP1, COL1A2, and CCL2 in the high-Rad-score group. Compared with the hub gene labeling index (LI) in the low-Rad-score group, LIs of MMP2, TIMP1, and COL1A2 were significantly higher in the high-Rad-score group (∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001).

(C) The expression levels and localization of MMP2, TIMP1, and COL1A2 were shown by the multiplex immunohistochemistry (mIHC) staining. The biological markers represented by different colors are labeled in the figure. The expression of the three genes increased in the high-Rad-score group, and MMP2 exhibited the most significant difference.

In the radiogenomics set, immunohistochemistry (IHC) staining in the proliferative core of glioma specimens revealed that the protein expression patterns of MMP2, MMP9, TIMP1, COL1A2, and CCL2 in the high-Rad-score group were consistent with those observed in THPA and were upregulated compared to the low-Rad-score group.

Using the same IHC methodology applied in the proliferative core region, we performed IHC analysis for the seven core genes in the infiltrative margin area. The results are shown in Figure S3. We observed a significant decrease in the labeling indices of all seven core genes in the infiltrative margin compared to the proliferative core. When comparing the low-Rad-score and high-Rad-score groups, MMP2, TIMP1, and COL1A2 demonstrated statistically significant differences in the infiltrative margin.

These results demonstrate that Rad-score can predict MMP2 expression levels in glioma specimens. Compared with the hub gene LI of the low-Rad-score group, LI of the high-Rad-score group was significantly higher for MMP2, TIMP1, and COL1A2 (Figure 5B; Table S5).

Next, multiplex immunohistochemistry (mIHC) staining was used to detect the expression levels and localization of MMP2, TIMP1, and COL1A2. As expected, patients with a high Rad-score showed a higher density of MMP2, TIMP1, and COL1A2. Among these, MMP2 exhibited the most pronounced difference (Figures 5C and S4). These analyses provide compelling statistical evidence that the ECM, as characterized by mIHC, exhibits quantitative differences between the groups, thereby substantially reinforcing our conclusions.

Although the expression of three hub genes was elevated, MMP2 has been the most promising and extensively studied in glioma treatments, including direct MMP2-targeted therapies and nanoparticle drug delivery systems that leverage MMP2’s differential expression in the tumor microenvironment to minimize the impact of chemotherapeutic drugs on healthy tissues.^14,15,16,17 Therefore, non-invasive evaluation of MMP2 expression in the tumor microenvironment is pivotal to improving clinical practice.

Validation of Rad-score as a specific radiomics indicator for ECM remodeling hub gene MMP2 expression by an orthotopic xenograft mouse model

After confirming the predictive efficacy of Rad-score for MMP2 expression, we established an orthotopic glioma xenograft model in nude mice and combined it with dynamic MRI to further investigate the impact of MMP2 expression on the constructed radiomics signature (Figure 6A). We aimed to preliminarily verify species-specific differences in gene expression predictions based on the radiomics signature and thoroughly validate the correlations between MMP2 and the radiomics signature from bidirectional perspectives.

Figure 6.

Radiomic features correlate with MMP2 expression in an orthotopic glioma xenograft model

(A) Graphical illustration of the orthotopic xenograft model construction and exploration.

(B) Representative MR imaging sequences at different time points across experimental groups in a mouse xenograft model. The inset in the upper right corner illustrates the method for delineating the tumor region of interest (ROI).

(C) Radiomic features were extracted from MR imaging sequences at all time points across experimental groups, followed by Rad-score calculation and statistical analysis using line graphs.

(D) Body weight (g) changes in mice from different experimental groups monitored over the study timeline.

(E) The survival curve of tumor-bearing mice.

(F) Representative images of Ki-67 immunohistochemistry (IHC) staining showed different patterns of tumor invasion in four groups. The tumors in the high-Rad-score group (NC and Scramble group) infiltrated into normal brain tissue with a more obvious cross-growth pattern and an obscure tumor boundary than those in the low-Rad-score group (SB-3CT and shMMP2 group).

Data significance is denoted as follows: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

Tumor regions segmented from nude mouse MRI images were subjected to radiomics analysis (Figure 6B). SB-3CT is a potent competitive inhibitor of metalloproteinase MMP2 capable of penetrating the blood-brain barrier. Following establishment of the intracranial orthotopic transplantation model, mice in the SB-3CT group received daily intraperitoneal injections of 200 μL SB-3CT solution at 2.5 mg/mL concentration (approximately 25 mg/kg/day). Additionally, as detailed in the STAR Methods section, we employed a lentiviral MMP2 knockdown approach to modulate MMP2 gene expression in tumor cells.

Results demonstrated no significant differences in Rad-score among the groups at day 0; however, significantly lower Rad-scores were observed in the SB-3CT group compared to the NC group, along with reduced Rad-scores in the shMMP2 group versus the Scrambled group at both days 7 and 14 (Figure 6C). Evaluation of body weight and survival time revealed that both MMP inhibition and MMP2 suppression effectively delayed weight loss and prolonged survival in tumor-bearing mice (Figures 6D and 6E), consistent with previous studies on MMP2 functions.^18,19

The tumors in the low-Rad-score groups (SB-3CT and shMMP2) had relatively smaller volumes than those in the high-Rad-score groups (NC and Scramble). Besides, tumors in the low-Rad-score groups (SB-3CT and shMMP2) also exhibited clearer boundaries observed by Ki-67 IHC staining than those in the high-Rad-score groups (NC and Scramble), which might be associated with alterations in the tumor’s invasive capacity due to ECM remodeling (Figure 6F). A comparable phenomenon was evident, with some tumors in the NC and Scrambled groups showing slightly enhanced regions surrounding the primary lesion in the MR scans. Tumors in the high-Rad-score group showed more pronounced infiltration into normal brain tissue, characterized by cross-growth patterns and more obscure boundaries. However, there was no significant difference in Ki-67 LI among the four groups (Figure S5; Table S6), suggesting that the Rad-score reflected the tumor’s invasive capability rather than its proliferative activity.

Radiomic features link to MMP2 expression and invasive phenotype in an orthotopic syngeneic glioma model in mice

Although nude mice lack an adaptive immune system, ECM remodeling and MMPs are known to interact extensively with immune cells. This is particularly critical in glioma, where the immune system and tumor microenvironment play indispensable roles in tumor progression and invasion. Therefore, to enable a more physiologically relevant investigation, we established an immunocompetent orthotopic glioma model using murine GL261 cells in C57BL/6j mice and conducted a series of comparable experiments (Figure 7A).

Figure 7.

Radiomic features link to MMP2 expression and invasive phenotype in an orthotopic syngeneic glioma model in mice

(A) Graphical illustration of the orthotopic syngeneic model construction and exploration.

(B) Representative MR imaging sequences at different time points across experimental groups in a mouse syngeneic model. The inset in the upper right corner illustrates the method for delineating the tumor region of interest (ROI).

(C) Radiomic features were extracted from MR imaging sequences at all time points across experimental groups, followed by Rad-score calculation and statistical analysis using line graphs.

(D) Body weight (g) changes in mice from different experimental groups monitored over the study timeline.

(E) The survival curve of tumor-bearing mice.

(F) Representative images illustrating invasive phenotypes across experimental groups, including tissue MMP2 immunofluorescence, fibronectin immunohistochemistry, and N-cadherin immunohistochemistry.

Data significance is denoted as follows: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

Consistent with the findings in xenograft models, dynamic MRI revealed that the Rad-score in the SB-3CT group was significantly lower than in the NC group at both days 7 and 14, while the shMMP2 group also showed a reduced Rad-score compared with the Scrambled group (Figures 7B and 7C). Evaluation of body weight and survival time in nude mice demonstrated that both MMP inhibition and MMP2 knockdown effectively delayed weight loss and prolonged survival in tumor-bearing mice (Figures 7D and 7E). Furthermore, immunofluorescence staining of tissue samples confirmed significantly reduced MMP2 expression in the SB-3CT and shMMP2 groups compared with their respective controls. The ECM marker fibronectin showed moderate upregulation in both treatment groups, while the invasion marker N-cadherin expression was decreased (Figure 7F). These results indicate that the Rad-score correlates with MMP2 expression in glioma and is associated with tumor invasive phenotypes.

In summary, we demonstrated that the ECM remodeling hub gene (MMP2) expression impacted the constructed radiomics signature Rad-score, which served as an imaging biomarker for clinical diagnosis and prognosis prediction. Moreover, an interesting phenomenon was observed: Rad-score correlated with prognosis as a specific radiomics feature, not by implying tumor proliferation, but by indicating the tumor invasion.

Discussion

To personalize the diagnosis and treatment of glioma, significant efforts have focused on uncovering the molecular signatures of glioma subtypes by developing and validating robust clinical and laboratory pipelines. Radiomics has shown promise as a potential marker for glioma prognosis, and recent advances in radiogenomics have deepened the understanding of the relationships between imaging characteristics and gene expression patterns.⁹ This has solidified the role of radiomics in clinical decision-making.

In this study, we established a radiogenomic signature based on the survival of glioma patients. Through a series of cluster analysis and molecular biology experiments, we validated that this signature is closely associated with the remodeling of the tumor ECM and exhibits strong correlations with the expression of hub genes involved in ECM remodeling, especially MMP2.

The ECM of solid tumors is highly complex and undergoes dynamic changes during cancer progression, and ECM remodeling functioned pivotally in establishing the required environment for cancer cells.² MMP2 has been the most promising and extensively studied target in ECM remodeling and glioma management.^14,15,16,17 Various MRI manifestations of glioma can non-invasively reflect the intensity of MMP2 expression.²⁰ Therefore, we aimed to validate the sufficient and necessary correlations between MMP2 and the radiomics signature from integrated perspectives. We not only explored the predictive performance of Rad-score for MMP2 expression but also utilized an orthotopic xenograft mouse model of glioma to identify the sequential impact of MMP2 expression on the constructed radiomics signature, rather than conducting the simple correlation analysis.

Relevant studies are summarized in Table S7. The radiomics signature for survival was developed by using multimodal MRI data acquired from multiple public datasets and local centers under a standardized protocol.^18,21 Compared with previous studies,^19,21,22 the radiomics signature developed in this study benefited from integrated multi-dataset recruitment and an expanded sample size, demonstrating improved predictive performance (AUC = 0.886 [95% CI: 0.807–0.964] in the training set; AUC = 0.828 [95% CI: 0.796–0.893] in the validation set). Consistent with prior investigations,^11,19,21,23 both univariate and multivariate Cox regression analysis revealed that the Rad-score remained an independent prognostic factor for OS in glioma patients after adjustment for other clinical variables. Given that PFS is a good surrogate endpoint correlated with OS in assessing treatment responses for glioma patients undergoing standard or experimental therapies,²³ we further assessed the performance of the radiomics signature in predicting PFS to enhance prognostic evaluation in glioma. Prognostic stratification using Kaplan-Meier curves revealed clinically significant efficacy in both the discovery set and the validation set (both p < 0.05). The Rad-score was identified as an independent prognostic factor and subsequently incorporated into the construction of a nomogram with calibration curves, thereby enhancing the clinical applicability of the radiomic signature. The radiomics signature demonstrated robust predictive performance. For OS, it achieved a C-index of 0.890 (95% CI: 0.877–0.903) in the discovery set and 0.803 (95% CI: 0.786–0.819) in the validation set. For PFS, it achieved C-index values of 0.896 (95% CI: 0.885–0.907) and 0.780 (95% CI: 0.763–0.798) in the discovery and validation sets, respectively. Furthermore, it maintained acceptable predictive efficacy in grade-based subgroup analysis.

Radiomics-based DEGs were predominant in the high-Rad-score group, and pathway and enrichment analysis revealed significant roles in ECM remodeling. DEGs and functional analysis revealed biological pathways significantly associated with ECM remodeling in glioma. Seven hub genes were identified by integrating the protein-protein interaction (PPI) network and weighted gene co-expression network analysis results at the co-expressed protein level. These hub genes showed high connectivity and a significant correlation with the Rad-score. The protein expression patterns of the hub genes were examined using data from the THPA database and IHC staining of glioma specimens acquired from local centers. Through screening, MMP2 was determined as the target for further investigation by establishing its corresponding orthotopic xenograft glioma model in nude mice. Imaging data from experimental mice were acquired using ultra-high-field 9.4-T MRI with high spatial resolution. The data were analyzed using the constructed radiomics signature to calculate the Rad-score for experimental groups treated with MMP2 or inhibition of MMP2 expression. When evaluating the MRI images, we observed that experimental groups with low and high Rad-scores exhibited different imaging characteristics and that the Rad-score reflected tumor infiltration into the normal brain tissue rather than proliferation.

By reliably characterizing glioma using the radiomics signature, an in-depth genomic exploration of the pathophysiological foundation within glioma can be achieved non-invasively.⁹ We conducted an integrated analysis to systematically explore multi-omics data, gaining valuable insights into the glioma ECM microenvironment. Upregulated expression of DEGs was predominant in the high-Rad-score group, and pathway and enrichment analysis revealed significant roles in ECM remodeling. The ECM is an intricate and complicated network of macromolecules that provides biochemical and structural support to neighboring molecules and cells, regulating various physiological activities.⁸ ECM remodeling, driven by MMPs, has been linked to several pathological events, including carcinogenesis,^8,24 but has received less attention in radiogenomics. The ECM comprises heterogeneous components in formation, degradation, and immune response. Therefore, we hypothesized that tumors with high biological heterogeneity would display heterogeneous imaging characteristics, captured by the radiomics signature.

ECM remodeling hub genes share common functions. MMP2 and MMP9 are MMPs. Studies suggest that MMP2 is a key downstream component of tumor invasion. Functionally, MMP2 is involved in neurogenesis, migration, dendritic and axonal outgrowth, regeneration, and angiogenesis in the brain.^25,26 MMP9 plays an essential role in local proteolysis of the ECM, leukocyte migration, and osteoclastic resorption.²⁷ TIMP1, a metalloproteinase inhibitor, functions by forming complexes with target MMPs and as a growth factor, regulates cell differentiation, migration, cell death, and integrin signaling.²⁸ COL1A2, a collagen alpha-2(I) chain belonging to the fibrillar collagen family, is aberrantly expressed in various cancers and may contribute to tumor cell differentiation.²⁹ CXCL-8, also known as interleukin-8, is a chemotactic factor that attracts neutrophils, basophils, and T cells and is involved in neutrophil activation.³⁰ IL-6 is a potent inducer of the acute phase response and an essential part of the final differentiation of lymphocytes and monocytes.³¹ CCL-2, a C-C motif chemokine, functions as a ligand for the C-C chemokine receptor CCR2 by binding and activating it to induce a strong chemotactic response, contributing to the initiation and progression of tumorigenesis.³² In summary, these seven hub genes work together in the metabolism and degradation of the ECM and regulating MMP activity in glioma.

To elucidate the biological functions of radiomics, biological validation is expected to become standard practice, reinforcing its role in clinical decision-making and translation.¹³ To the best of our knowledge, few studies have validated radiomics or radiogenomics results using cellular or molecular confirmation or animal models. This study utilized an orthotopic model of glioma to strengthen the correlation between MMP2 expression and the constructed radiomics signature from bidirectional perspectives. Our study first screened hub genes that were closely associated with MRI radiomics features, aiming to identify a set of hub genes that are prevalently expressed across gliomas and are most relevant to malignant radiomics features. Subsequently, the grade-based analysis further explored the distinct characteristics underlying different tumor grades. The overall exploration across all gliomas provides an essential foundation for initial and general investigation, while the grade-specific research is tailored for precision medicine. Together, these two approaches form a comprehensive research framework for a systematic analysis of radiogenomic features.

Normalization and resampling were applied for radiomics exploration in the constructed mouse model, and Rad-score was compared between groups. However, the evaluation of diagnostic efficacy or a cutoff for the Rad-score was omitted due to the limited number of experimental animals. Nevertheless, the relative Rad-score was compared to explain the differences in imaging phenotypes across experimental groups. Furthermore, we used histology to uncover the pathological basis of imaging heterogeneity captured by the constructed radiomics signature.

Our infiltrative margin tissue was defined based on FLAIR-hyperintense regions (not contrast enhancing) on MRI and intraoperative appearance resembling normal cortical gray-white matter rather than the gray-red fish-flesh texture typical of tumor core. IHC confirmed markedly lower cellular density in the margin, with estimated tumor cell fraction below 30%.³³ The seven genes of interest are known to be highly expressed in the tumor microenvironment but low in normal brain tissue. Their reduced overall expression at the margin is, therefore, expected. However, the relatively low expression of MMP2 in this region may play a crucial role in glioma invasion—and this does not diminish their clinical relevance. Our current study is only an initial exploration of glioma infiltrative margin, and we aim to investigate this region more deeply in future work.

The inability to reliably detect CXCL8 and IL-6 via IHC indeed weakens the functional validation of all seven hub genes. First, the hub genes screened through radiogenomic correlation analysis were based on RNA-level data and high RNA expression or significant RNA-level changes do not necessarily translate directly to functional changes at the protein level, which is precisely why we conducted subsequent protein-level validation using IHC. Second, for secreted cytokines such as CXCL8 and IL-6, their protein expression may be transient—rapid release into the tumor microenvironment could result in concentrations below the detection threshold of conventional IHC. To further investigate the underlying reasons, we performed western blot analysis, which revealed detectable but low expression of both CXCL8 and IL-6 (Figure S6). Quantitative analysis showed no significant differences between the Rad-high and Rad-low groups. These findings suggest complexities in translating RNA-based screening results to protein-level functional validation. We acknowledge that the IHC and western blot results for CXCL8 and IL-6 were inconclusive. Future studies should incorporate more sensitive protein detection techniques and dynamic analyses of the tumor microenvironment to improve the validation framework.

In our animal experiments, we selected the SB-3CT inhibitor to target MMP2 expression. While this compound demonstrates activity against both MMP2 and MMP9, its inhibitory potency differs substantially between these two enzymes. The inhibition constant (Ki) values show 43-fold greater affinity for MMP2 (13.9 nM) than for MMP9 (600 nM).³⁴ Accordingly, we implemented a carefully optimized in vivo dosing regimen (25 mg/kg/day) that represents a well-established protocol for specific MMP2 inhibition in numerous previous studies.³⁵

The biological roles of ECM remodeling in the development, progression, and treatment of glioma have been extensively studied. Notably, ECM remodeling and MMPs have been identified as key drivers in glioma. The non-invasive evaluation of ECM remodeling via radiomics could lead to a substantial reduction in clinical costs, as it utilizes existing imaging protocols without additional invasive procedures.

This approach also enhances accuracy in the pre-operative prediction of tumor aggressiveness, provides quantitative and objective metrics for personalized response evaluation, and enables dynamic follow-up to monitor for early signs of recurrence or progression. In summary, radiogenomics offers another approach by linking imaging phenotypes to genetic alterations, further expanding the non-invasive tools available for clinical decision-making.

In this study, we used a radiogenomic approach for survival prediction in glioma to explore the links between imaging phenotypes and genomics. The correlations between ECM remodeling and radiomics signature were thoroughly validated, which we hope will provide promising clinical support for preoperative management, prognosis prediction, precision treatment, and longitudinal monitoring of glioma.

Limitations of the study

Several limitations of this study should be considered. First, while our study verifies the predictive value of the radiomics signature, robustly confirming its clinical applicability will require prospective validation in a larger and independent external cohort. Second, we made experimental efforts to identify the relationships between ECM remodeling hub genes and biological patterns, confirming tumor heterogeneity through tissue staining; however, there is still room for improvement in mapping radiomics features to the ground-truth characteristics of tumors in vivo. Third, it should be noted that the mechanisms underlying the invasive phenotype including ECM remodeling and the role of the tumor immune microenvironment represent highly complex processes. The investigation into the invasive phenotype and ECM remodeling remains preliminary, both in human specimens and murine models. We recognize that delving deeper into these mechanisms represents a highly promising and transformative area for future research.

Finally, precise image-tissue co-registration remains challenging due to irreversible deformations during surgical resection and tissue processing. While implementing image-guided stereotactic biopsy in glioma patients presents a potential solution, this approach entails considerable technical and ethical complexities. Given current technological limitations and time constraints, we have designated this as a key priority for future research.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Shicheng Sun (sunshicheng@sdfmu.edu.cn).

Materials availability

This study did not generate new unique material or reagents.

Data and code availability

De-identified human RNA-seq data have been deposited at the Genome Sequence Archive for Human (GSA-Human) as HRA017144. They are available upon request if access is granted. To request access, contact the GSA-Human Data Access Committee via the official portal at https://ngdc.cncb.ac.cn/gsa-human/request/HRA017144. This paper does not report original code. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This research was supported by the National Natural Science Foundation of China (grants 82303083 and 82203472) and the Natural Science Foundation of Shandong Province (grants ZR2023QH432, ZR2022QH152, and ZR2021MH387).

Author contributions

Conceptualization, Y.B. and S.S.; methodology, Y.B., X.C., Y.C., C.Z., L.C., W.H., M.S., and S.S.; investigation, Y.B., X.C., Y.C., and S.S.; visualization, Y.B., X.C., Y.C., C.Z., L.C., W.H., and S.S.; supervision, G.P., H.Z., B.Z., X.W., and S.S.; writing – original draft, review, and editing, Y.B., X.C., Y.C., X.W., and S.S.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
rabbit polyclonal anti-MMP2	Proteintech	RRID: AB_2250823
rabbit polyclonal anti-MMP2	Abcam	RRID: AB_10561597
rabbit polyclonal anti-MMP9	Proteintech	RRID: AB_10897178
rabbit polyclonal anti-CXCL8	Proteintech	RRID: AB_2861340
rabbit polyclonal anti-TIMP1	Proteintech	RRID: AB_2878292
rabbit polyclonal anti-IL-6	Proteintech	RRID: AB_11142677
rabbit polyclonal anti-COL1A2	Proteintech	RRID: AB_2082037
rabbit polyclonal anti-CCL2	Proteintech	RRID: AB_2918100
rabbit polyclonal anti-ki-67	Proteintech	RRID: AB_2756525
rabbit polyclonal anti-fibronectin	Abcam	RRID: AB_2262874
rabbit polyclonal anti-N-cad	Abcam	RRID: AB_2077524
Biological samples
Human glioma tissue	Shandong Provincial Hospital, the Second hospital of Shandong University
Chemicals, peptides, and recombinant proteins
SB-3CT	MCE	HY-12354
Human MMP2-targeting short hairpin RNA (shRNA) plasmids	Santa Cruz Biotechnology	sc-29398-SH
Mouse MMP2-targeting short hairpin RNA (shRNA) plasmids	Santa Cruz Biotechnology	sc-37264-SH
Deposited data
RNA-seq data	Genome Sequence Archive for Human in National Genomics Data Center (Nucleic Acids Res 2025)36,37	https://ngdc.cncb.ac.cn/gsa-human/browse/HRA017144
Experimental models: Cell lines
U-87MG	Cell Bank of Type Culture Collection of the Chinese Academy of Sciences
GL261	Cell Bank of Type Culture Collection of the Chinese Academy of Sciences
Experimental models: Organisms/strains
BALB/c nude mice	Charles River
C57BL/6j	Charles River
Software and algorithms
3D Slicer version 5.0.3	BWH and 3D Slicer contributors	https://www.slicer.org/
FeAture Explorer version 0.5.5	Song, Yang et al.38	https://github.com/salan668/FAE
Python version 3.7.6	Python Software Foundation	https://www.python.org
fastp software	Chen, Shifu et al.39	https://github.com/OpenGene/fastp
HISAT2	Kim, Daehwan et al.40	https://ccb.jhu.edu/software/hisat2
StringTie	Pertea, Mihaela et al.41	https://ccb.jhu.edu/software/stringtie
gffcompare	Pertea, Geo, and Mihaela Pertea42	https://github.com/gpertea/gffcompare/
R software version 4.2.1	R Foundation	https://www.r-project.org/
STRING	Global Biodata Coalition and ELIXIR	http://string-db.org
Cytoscape version 3.10.2	Shannon, Paul et al.43	https://cytoscape.org/
Fiji build of ImageJ version 1.54f	Schneider et al.44	https://imagej.net/software/fiji/
MedCalc version 22.014	MedCalc Software Ltd.	https://www.medcalc.org/
R software version 4.2.1	R Foundation	https://www.r-project.org/
QuPath version 0.5.1	Bankhead, Peter et al.45	https://qupath.github.io
Other
Database: The Cancer Imaging Archive	NCI Cancer Imaging Program	http://cancerimagingarchive.net/
Database: The Cancer Genome Atlas	National Institutes of Health	https://portal.gdc.cancer.gov/
Pyradiomics documentation	Pyradiomics community	https://pyradiomics.readthedocs.io/en/latest/
Database: The Human Protein Atlas version 24.0	Global Core Biodata Resource	https://www.proteinatlas.org/

Experimental model and study participant details

Study design and human participants

This retrospective multi-dataset study was performed in steps as follows (Figure 1): First, we developed an MRI radiomics signature for predicting overall survival (OS) in glioma. Second, we linked the prognostic radiomics signature with paired RNA-seq data to identify driving pathways and hub genes. Third, we validated the radiogenomic linkage through an integrated experimental approach by employing an orthotopic xenograft model of glioma in nude mice and in vivo imaging by 9.4T small-animal MRI.

The protocol of this study had been approved by the institutional review board of Shandong University (NSFC: NO.2022-8890), and the requirement for written informed consent was obtained. Patients were recruited from local centers (Shandong Provincial Hospital; the Second Hospital of Shandong University) and public databases (The Cancer Imaging Archive, TCIA; The Cancer Genome Atlas, TCGA). Inclusion criteria for patient were as follows: (a) pathology diagnosis of glioma, (b) available clinical data (age, sex assigned at birth, tumor grade, and IDH1 status) and follow-up information (overall survival and progression free survival), (c) preoperative MRI data including pre- and post-contrast T1-weighted (T1WI), T2-Fluid Attenuation Inversion Recovery (T2-FLAIR), and T2WI sequences. Exclusion criteria for patients were as follows: (a) history of preoperative treatment for glioma, (b) incomplete clinical data or lost to follow-up, (c) severe MRI image artifacts or incomplete imaging data. The influence of sex assigned at birth on the results was analyzed and is described in the Results section. Patients were allocated into subsets as follows (Table 1; Figure 2).

• (1)The radiomics analysis sets. The radiomics discovery and validation sets were to develop and validate a prognostic radiomics signature for glioma using preoperative MRI data. The radiomics discovery set retrospectively enrolled 496 patients with histologically confirmed glioma and preoperative MRI image data, from Shandong Provincial Hospital and the Second Hospital of Shandong University, between January 2022 and June 2024. After the application of exclusion criteria (n = 107 excluded), 389 cases were included and retained in the radiomics discovery set for subsequent analysis. The radiomics validation set included 5 collections of TCIA: The Cancer Genome Atlas Low Grade Glioma (TCIA-LGG: https://doi.org/10.7937/K9/TCIA.2016.L4LTD3TK), The University of California San Francisco preoperative Diffuse Glioma MRI (UCSF-PDGM: https://doi.org/10.7937/tcia.bdgf-8v37), The Cancer Genome Atlas Glioblastoma Multiforme Collection (TCGA-GBM: https://doi.org/10.7937/K9/TCIA.2016.RNYFUYE9), The Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multiforme Collection (CPTAC-GBM: https://doi.org/10.7937/K9/TCIA.2018.3RJE41Q1), and Ivy Glioblastoma Atlas Project (IvyGAP: https://doi.org/10.7937/K9/TCIA.2016.XLwaN6nL). After the application of exclusion criteria (n = 113 excluded), 502 cases with preoperative MRI image data were included and retained in the radiomics validation set for subsequent analysis.
• (2)The radiogenomics analysis sets. The radiogenomics discovery and validation sets were subsets of the radiomics sets by excluding cases without RNA-seq data for further identifying driving pathways and hub genes underlying prognostic radiomics phenotypes. Local glioma specimens were acquired from Shandong Provincial Hospital, with meticulous resection during surgery under preoperative planning, intra-operative neuro-navigation guidance, and storage in liquid nitrogen. After the pathological diagnosis of tumor specimens was confirmed by two neuropathologists, qualified samples were sent for RNA sequencing and consequent processing. A total of 241 glioma patients with available MRI data, corresponding RNA-seq data, and follow-up information were retained in the radiogenomics discovery set. As TCIA is linked to partial samples in TCGA, RNA-seq data from TCIA-GBM, CPTAC-GBM, and TCIA-LGG were downloaded from the Genomic Data Commons Data Portal of TCGA (https://portal.gdc.cancer.gov/). A total of 331 glioma patients with complete MRI data and corresponding RNA-seq data were included in the radiogenomics validation set after matching patient identifiers between TCIA and TCG.

Mouse studies

All animal experimental procedures were conducted strictly following the Guide for the Care and Use of Laboratory Animals and approved by the Animal Care and Use Committee of Shandong University (NSFC: NO.2022-8891). The sample size was determined based on previous well-established studies in glioma models with comparable effect sizes, without prior power analysis. Male mice were chosen for this study, as human glioblastoma shows notable sex differences with higher incidence and poorer outcomes in males, and the majority of preclinical studies in this field have similarly employed male animals. The use of female mice was not pursued, as investigating sex-specific factors such as estrogen was beyond the scope of this study.

Lentivirus: MMP2-targeting short hairpin RNA (shRNA) plasmids for human (sc-29398-SH) and mouse (sc-37264-SH) were purchased from Santa Cruz Biotechnology. Along with a nonspecific scrambled control shRNA (sequence: TTCTCCGAACGTGTCACGT), all constructs were packaged into lentiviral particles using the pGMLV-hU6-MCS-CMV-Blasticidin-WPRE vector.

Xenograft glioma model: Four-week-old male BALB/c nude mice were fed under specific pathogen-free conditions in the animal facility. Forty-eight nude mice were acclimatized for 7 days and then used to construct the orthotopic xenograft model of glioma with human glioma U-87MG cells. The U-87MG cell lines were authenticated by STR profiling and tested negative for mycoplasma contamination. Mice were divided into four groups randomly: negative control (NC), SB-3CT (MMP-2 inhibitor, HY-12354, MCE, USA), Scramble (non-targeting shRNA), and shMMP2 (MMP-2 knockdown). Each mouse was injected with 1×10⁵ U-87MG cells in the right hemisphere. Treatment regimen, medication duration, and dosage specifications for each experimental group are summarized in Figure 6A.

Syngeneic glioma model: Four-week-old male C57BL/6j mice were fed under specific pathogen-free conditions in the animal facility. Four-week-old male C57BL/6j mice were acclimated for 7 days and then used to construct the Syngeneic glioma model with human glioma GL261 cells. The GL261 cell lines were authenticated by STR profiling and tested negative for mycoplasma contamination. Each mouse was injected with 1×10⁵ GL261 cells in the right hemisphere. The modeling procedure and subsequent treatments followed the same protocol as the xenograft model described above and are detailed in Figure 7A.

Daily body weight measurements and survival status of nude mice were monitored throughout the experimental period, with mortality events recorded within 24 h of occurrence. After model establishment, MR imaging was performed at days 0, 7, and 14.

Method details

Radiomics features extraction and the radiomics signature construction

MRI images from multiple datasets were preprocessed with the ‘ANTsR’ package in R, which included N4 bias field correction (shrink factor = 2), resampling (isotropic voxel size = 1 mm × 1 mm×1mm), rigid transformation registration, and Z score intensity normalization. Then, the tumor region of interest (ROI) was segmented, based on post-contrast T1WI and T2-FLAIR MRI images of the glioma parenchyma, avoiding the necrotic area, cystic area, brain sulci, and ventricles, and the ROI was copied to other sequences on the same anatomical position after co-registration (Figure S1). ROIs were manually delineated by two radiologists with 8 years of experience in neuroradiology using 3D Slicer (version 5.0.3, https://www.slicer.org/). For each patient, 1050 radiomics features were extracted per MRI sequence (pre-contrast T1WI, post-contrast T1WI, T2-FLAIR, and T2WI) using the ‘SlicerRadiomics’ module (version aa418a5) implemented in 3D Slicer, divided into four types per sequence: first-order features (n = 18), shape and size features (n = 27), textural features (n = 75), filter-derived features (logsigma features, n = 186; wavelet features, n = 744). Therefore, a total of 4200 radiomics features extracted for each patient. The official website provides detailed information on radiomics features and formulas (https://pyradiomics.readthedocs.io/en/latest/). ComBat correction was applied to the extracted radiomics features to correct the batch effects with ‘SVA’ package in R.⁴⁶ To assess feature stability, two radiologists independently performed tumor segmentation on 50 randomly selected cases with a two-week interval. Inter- and intra-observer reproducibility were quantified using intraclass correlation coefficients (ICCs), with features demonstrating ICC >0.75 retained for subsequent analysis.

To ensure the stability of the glioma prediction model, workflows including data normalization, dimensionality reduction with feature selection, and classifier implementation were executed through the FeAture Explorer (FAE) pipeline (version 0.5.5, code available at https://github.com/salan668/FAE) based on Python (version 3.7.6, https://www.python.org).³⁸ To establish an interpretable model for early risk assessment, we defined the prediction target as a binary outcome of one-year OS. This endpoint was selected based on its clinical relevance as an approximate median survival benchmark in glioma, particularly for glioblastoma, thereby enabling straightforward patient stratification at diagnosis.⁴⁷ The radiomics signature was subsequently constructed through the following phases: First, the radiomics discovery set was divided into the training and test sets with a ratio of 7: 3, and normalized by Z score Normalization subtracting the mean value and dividing the standard deviation for each feature. Second, feature pairs with Pearson correlation coefficient >0.9 were considered collinear and removed. Analysis of Variance (ANOVA), Kruskal-Wallis (K-W), and Recursive Feature Elimination (RFE) were feature selectors (feature number range: 1–30) used for sequential feature selection. Third, machine learning algorithms were systematically evaluated to identify the optimal predictive model: Support Vector Machine (SVM), Linear Discriminative Analysis (LDA), Logistical Regression (LR), Random Forest (RF), Decision Tree (DT) and Logistical Regression-Least Absolute Shrinkage and Selection Operator (LR-LASSO). Model performance was iteratively assessed through 5-fold cross-validation during feature selection and hyperparameter tuning. The areas under the curve (AUC) of the receiver operator characteristic (ROC) curve of pipelines (n = 540) were all compared, and the pipeline with the highest Youden index was used for radiomics signature construction which was built by selected features and evaluated by Rad-score. Rad-score for each patient was weighted according to their coefficients through a linear combination of selected features (∑feature value × coefficient of feature), whereby a higher Rad-score represented a shorter survival, and vice versa.

RNA extraction, library construction, and processing of RNA-seq data

Total RNA was isolated and purified using TRIzol reagent (Invitrogen, Carlsbad, CA, USA) following the manufacturer’s procedure. The RNA amount and purity of samples were quantified using NanoDrop ND-1000 (NanoDrop, Wilmington, DE, USA). The RNA integrity was assessed by Bioanalyzer 2100 (Agilent, CA, USA) with RIN number >7.0, and confirmed by electrophoresis with denaturing agarose gel. Poly (A) RNA is purified from 1μg total RNA using Dynabeads Oligo (dT)25–61005 (Thermo Fisher, CA, USA) using two rounds of purification. Then the poly(A) RNA was fragmented into small pieces using Magnesium RNA Fragmentation Module (NEB, cat. e6150, USA) under 94°C 5-7min. Then the cleaved RNA fragments were reverse-transcribed to create the cDNA by SuperScript II Reverse Transcriptase (Invitrogen, cat. 1896649, USA), which were next used to synthesize U-labeled second-stranded DNAs with E. coli DNA polymerase I (NEB, cat.m0209, USA), RNase H (NEB, cat.m0297, USA) and dUTP Solution (Thermo Fisher, cat. R0133, USA). An A-base is then added to the blunt ends of each strand, preparing them for ligation to the indexed adapters. Each adapter contains a T-base overhang for ligating the adapter to the A-tailed fragmented DNA. Single- or dual-index adapters are ligated to the fragments, and size selection was performed with AMPureXP beads. After the heat-labile UDG enzyme (NEB, cat.m0280, USA) treatment of the U-labeled second-stranded DNAs, the ligated products are amplified with PCR by the following conditions: initial denaturation at 95°C for 3 min; 8 cycles of denaturation at 98°C for 15 s, annealing at 60°C for 15 s, and extension at 72°C for 30 s; and then final extension at 72°C for 5 min. The average insert size for the final cDNA library was 300 ± 50 bp. At last, we performed the 2 × 150bp paired-end sequencing (PE150) on an illumina Novaseq 6000 (LC-Bio Technology CO., Ltd., Hangzhou, China) following the vendor’s recommended protocol.

The fastp software (https://github.com/OpenGene/fastp)³⁹ were used to remove the reads that contained adaptor contamination, low quality bases and undetermined bases with default parameters. Then sequence quality was also verified using fastp software. We used HISAT2 (https://ccb.jhu.edu/software/hisat2)⁴⁰ to map reads to the reference genome of Homo sapiens GRCh38. The mapped reads of each sample were assembled using StringTie (https://ccb.jhu.edu/software/stringtie)⁴¹ with default parameters. Then, all transcriptomes from all samples were merged to reconstruct a comprehensive transcriptome using gffcompare (https://github.com/gpertea/gffcompare/).⁴² After the final transcriptome was generated, StringTie was used to estimate the expression levels for mRNAs by calculating FPKM (FPKM = [total_exon_fragments/mapped_reads(millions) × exon_length(kB)]).

Identification of driving pathways and hub genes related to radiomics phenotypes

RNA-seq DEGs analysis was conducted in groups based on the stratification of the Rad-score. The significant DEGs were selected with the p < 0.05 and |logFC| ≥ 1. The GO-KEGG analysis was performed to functionally annotate DEGs by the “clusterprofiler” package in R software (version 4.2.1, https://www.r-project.org/) with p < 0.05 and adjusted p < 0.1 considered. Gene Set Enrichment Analysis (GSEA) was to explore rich genes in biological processes. The annotated gene set c2.cp.v7.2.symbols.gmt [Curated] was selected as the reference gene set, while entries with false discovery rate (FDR) < 0.25 and p < 0.05 were considered. The normalized enrichment statistic (NES) was used to quantify the effect size of the gene set enrichment that was comparable between gene sets.

The Search Tool for the Retrieval of Interacting Genes (STRING, http://string-db.org) was used to construct the protein-protein interaction (PPI) network of co-regulated DEGs and analyzed the functional interactions between proteins with the setting as high confidence = 0.7. Cytoscape software (version 3.10.2, https://cytoscape.org/)⁴³ was used to visualize the PPI network. The ‘MCODE’ plugin in Cytoscape was used to screen significant sub-network modules from the PPI network, with the cutoff settings as Degree cutoff = 2, Node score cutoff = 0.2, K-core = 2, and Max.Depth = 100, and genes with high degrees of connectivity and specificity of function in the significant sub-network module were selected as candidate hub genes. Additionally, according to the maximum clique centrality (MCC), the ‘CytoHubba’ plugin in Cytoscape calculated the score of each node. The top 10 genes with the highest MCC score were also selected as candidate hub genes.

Weighted Gene Co-expression Network Analysis (WGCNA) was employed to construct a gene co-expression network to cluster highly interconnected genes into significant gene modules by “WGCNA” package in R.⁴⁸ The main steps were as follows: (a) Pick the soft-threshold power (β) for emphasizing strong correlations between genes and penalize weak correlations. (b) Calculate the adjacency and topological overlap matrix (TOM) to indicate the concurrence in shared adjacent genes. (c) Calculate the dissimilarity of module eigengenes and merge some similar modules with the cutoff settings as |module membership| > 0.8 and |gene significance| > 0.1. The relation between modules and Rad-score was used to identify candidate functional modules in the gene co-expression network.

Then, radiogenomics-related hub genes were obtained by intersecting the results of candidate hub genes and functional modules acquired both in the PPI network and WGCNA of DEGs.

Immunohistochemistry and multiplex immunohistochemistry of glioma specimens

The Human Protein Atlas (THPA, version 24.0, https://www.proteinatlas.org/) was used to identify general protein expression patterns of hub genes in glioma and normal brain tissues by IHC staining. Acquired glioma specimens were sent for IHC staining to explore different protein expression patterns of hub genes in low and high Rad-score groups. The multiplex immunohistochemistry (mIHC) staining visually demonstrated the expression of various genes in the tumor microenvironment.

Processes of glioma specimens were as follows: 10 glioma tissues of each group were embedded in paraffin, followed by treatment with xylene and ethanol; after antigen retrieval by microwave treatment, the samples were incubated with 3% hydrogen peroxide to remove endogenous peroxidase; subsequently, the samples were blocked using 10% goat serum and incubated with primary and corresponding secondary antibodies. The primary antibodies for IHC were: rabbit polyclonal anti-MMP2 (10373-2-AP, 1:400, Proteintech, China), rabbit polyclonal anti-MMP2 (ab92536, abacam), rabbit polyclonal anti-MMP9 (10375-2-AP, 1:500, Proteintech, China), rabbit polyclonal anti-CXCL8 (27095-1-AP, 1:1000, Proteintech, China), rabbit polyclonal anti-TIMP1 (16644-1-AP, 1:500, Proteintech, China), rabbit polyclonal anti-IL-6 (16644-1-AP, 1:500, Proteintech, China), rabbit polyclonal anti-COL1A2 (14695-1-AP, 1:1000, Proteintech, China), rabbit polyclonal anti-CCL2 (26161-1-AP, 1:200, Proteintech, China), rabbit polyclonal anti-ki-67 (27309-1-AP, 1:2000, Proteintech, China), rabbit polyclonal anti-fibronectin (ab2413, abacam), rabbit polyclonal anti-N-cad, (ab271856, abacam). All slides were scanned at 10X and 20X magnifications.

Cell counting of IHC images was performed in ‘Trainable WEKA segmentation’, a machine learning plugin in the Fiji build of ImageJ (version 1.54f, https://imagej.net/software/fiji/).⁴⁴ The hub gene labeling index (LI) was assessed as a percentage of positive staining of neoplastic nuclei or cytoplasmic out of at least 500 neoplastic cells in 5 non-overlapping fields.

The multiplex immunohistochemistry (mIHC) staining method identified the extracellular matrix hub genes in the TME, multiplex immunohistochemistry was obtained using PANO 7-plex IHC kit (cat0004100100, Panovue, Beijing, China). 10 glioma tissues of each group were processed using the same method as above. Primary antibodies were sequentially incubated, followed by horseradish peroxidase-conjugated secondary antibody incubation and tyramide signal amplification (TSA). The slides were treated with a microwave after TSA. Nuclei were stained with DAPI after all the human antigens had been labeled. Positive cells were measured in areas with the highest proliferation selected by scanning the entire IHC staining slides at low magnification.

mIHC images analysis, including data extraction, channel discrimination, and precise cellular quantification, was conducted using QuPath software. Briefly, five non-overlapping fields of view were randomly selected from each tumor sample. Optimal thresholds were established for positive cell detection to calculate the positive cell density for each marker. Density = total number of specific marker positive cells/analysis area (cells/mm²).

Scanning and analysis of orthotopic intracranial mouse model in vivo 9.4T small-animal MRI

After fourteen-day inoculation, tumor growth in the orthotopic xenograft model of glioma in nude mice was examined by a 9.4T small-animal MRI scanner (BioSpec 94/20, Bruker, Germany), including the pre-and post-contrast T1WI, T2-FLAIR, and T2WI sequences, with an 86 mm volume coil for transmission and a 2 cm × 2 cm phase array surface coil to receive the signal (ParaVision version 6.0.1). Anesthesia was maintained at a flow rate of 2L/min through inhalation of a mixture of 1.5% isoflurane and O₂. The body temperature of nude mice was maintained using a heating pad, and the respiration rate was monitored using a pressure sensor. Scanning parameters were according to standard protocols. The main scanning parameters were as follows: prone position and coronal image acquisition, (a) pre- and post-contrast T1WI: repetition time = 800 ms; echo time = 7.5 ms; field of view = 20 × 20 mm; slice thickness = 0.7 mm; flip angle = 180; spatial resolution = 0.08 × 0.08 × 0.7 mm3; scan resolution = 256 × 256. (b) T2-FLAIR: TR = 4140 ms; TE = 29.9 ms; inversion time = 29.9 ms; field of view = 20 × 20 mm; slice thickness = 0.7 mm; flip angle = 180; spatial resolution = 0.08 × 0.08 × 0.7 mm3; scan resolution = 256 × 256. (c) T2WI: repetition time = 2500 ms; echo time = 33 ms; field of view = 20 × 20 mm; slice thickness = 0.7 mm; flip angle = 180; spatial resolution = 0.08 × 0.08 × 0.7 mm3; scan resolution = 256 × 256. Magnevist^Tm Gd-DTPA was injected intraperitoneally at 0.5 mmol/kg. Post-contrast imaging started at 20 min after injection. The acquired MRI data were processed to extract radiomics features using the same workflow applied to the patient data. The corresponding Rad-score was then calculated for each group and compared between the NC and SB-3CT groups, as well as between the Scramble and shMMP2 groups.

After scanning, tumor-bearing nude mice were sacrificed, and brain tissues were fixed in 4% formaldehyde and embedded in paraffin for Ki-67 IHC analysis. Ki-67 LI was calculated in each group and compared between the NC and the SB-3CT groups, and between the Scramble and the shMMP2 groups.

Quantification and statistical analysis

The continuous data were expressed as the mean ± standard deviation and categorical variables were expressed as numbers or percentages. The analysis of variance (ANOVA), Kruskal-Wallis test, or Chi-square test was used for clinical characteristics. Predictive performances of constructed models were evaluated by the ROC curve using “pROC” and “ggplot2” package in R. The evaluation of patients’ survival and prognostic factors was explored by the Kaplan-Meier analysis and Cox regression analysis using “survival” package in R. Nomogram analysis and calibrate curves for 1-, 2-, and 3-year survival prediction was constructed by “rms” package in R. R, Python, and MedCalc (version 22.014, https://www.medcalc.org/) were used for statistical testing, and two-tailed p < 0.05 was considered statistically significant.

Published: May 12, 2026