Radiomic Features of MRI Subcompartments Associate with Angiogenic and Inflammatory Transcriptomic Programs in Glioblastoma: An IvyGAP Exploratory Analysis

Simple Summary

Glioblastoma, the most aggressive brain cancer, consists of distinct biological regions that contribute to treatment resistance. Magnetic resonance imaging can differentiate tumor subregions based on their appearance, while gene expression profiling can uncover the molecular programs active in each area. Whether imaging-derived measurements accurately reflect these underlying molecular programs remains unknown. We combined two publicly available datasets for 28 glioblastoma patients, one providing imaging features from tumor subregions and another offering gene expression data from surgically isolated tissue zones. Of 24 molecular programs tested, inflammatory response was the only program supported by both analytical frameworks. Angiogenesis reached significance only in one framework and is reported as a tentative signal that requires independent validation. The remaining programs showed no detectable association; at this small sample size (28 patients vs. several hundred that would be needed for stable modeling), the absence of signal most likely reflects limited statistical power and the non-spatial nature of our zone-to-imaging mapping, rather than a true absence of biological association. Five of the 24 gene sets were derived from the same data used to test them and therefore do not constitute independent evidence. These findings delimit what can and cannot be inferred from imaging-based molecular profiling of brain tumors in small cohorts and motivate validation in larger, spatially co-registered datasets.

Abstract

Background: Glioblastoma exhibits profound intratumoral heterogeneity, with anatomically distinct tumor zones characterized by divergent molecular programs that drive therapy resistance. Whether magnetic resonance imaging (MRI)-derived radiomic features can capture these regional transcriptomic differences remains unknown. We aimed to determine whether subcompartment-level radiomic features associate with transcriptomic pathway enrichment scores derived from biologically approximate tumor zones. Methods: We matched 28 patients (mean age 58.5 years; 13/28 MGMT methylated) across the IvyGAP RNA-seq atlas and the IVYGAP-RADIOMICS datasets. Single-sample GSEA (ssGSEA) pathway scores were computed for 24 gene sets. Radiomic features (3920 per subcompartment) were reduced to 597. Nested leave-one-patient-out cross-validation (LOPO-CV) with Elastic Net served as the primary predictive analysis; linear mixed-effects models (LMM) provided exploratory associational analysis. Analyses used a biologically motivated but spatially non-co-registered zone-to-subcompartment mapping; all reported associations are zone-approximate. Results: Twenty-one of 24 pathways showed no predictive signal (R²_cv ≤ 0). Inflammatory Response (R²_cv = 0.185, 95% CI [0.071, 0.355], p = 0.008) was the only pathway supported by both the nested CV (FDR = 0.096) and the exploratory LMM (FDR = 0.024, ΔR² = 0.214 beyond subcompartment effects) analyses; the LMM association was robust to clinical covariate adjustment (likelihood ratio test p = 0.004). Angiogenesis (R²_cv = 0.209, 95% CI [0.028, 0.353], p = 0.006) reached nested CV significance (FDR = 0.096) but was not corroborated by the LMM (FDR = 0.445); it is therefore reported as a tentative single-framework signal requiring independent validation. T2-derived texture features were selected in 100% of folds for both pathways. Conclusions: Inflammatory Response is the only pathway supported by both analytical frameworks; Angiogenesis is a tentative nested-CV-only signal pending independent validation. The absence of signal for 21 of 24 pathways should not be interpreted as evidence of biological inaccessibility: at N = 28 (vs. N ≈ 240 required by Riley criteria), severe underpowering, attenuation from the non-spatial zone-to-subcompartment mapping, and methodological constraints each independently suffice to suppress real associations. Five of the 24 gene sets (the IvyGAP zone modules) are non-independent from the outcome data and cannot be interpreted as discovery. All reported associations are zone-approximate and may partly reflect macro-compartment (between-subcompartment) effects; validation in larger cohorts with spatially precise co-registration is essential.

1. Introduction

Glioblastoma (GBM) is the most common and aggressive primary malignant brain tumor in adults, with a median overall survival of approximately 15 months despite maximal safe resection, temozolomide chemotherapy, and radiotherapy [1,2]. A defining feature of GBM is profound intratumoral spatial heterogeneity, with molecularly and histologically distinct regions coexisting within the same tumor [3,4]. This heterogeneity drives therapy resistance by harboring treatment-refractory cellular subpopulations and creating diverse microenvironmental niches [5]. Understanding the molecular programs that govern these distinct tumor regions is therefore critical for developing spatially informed therapeutic strategies.

The Ivy Glioblastoma Atlas Project (IvyGAP) represents a landmark effort to characterize this spatial heterogeneity at the transcriptomic level [6]. In IvyGAP, laser microdissection (LMD) was performed on tumor sections from 41 patients to isolate RNA from five histologically defined anatomic zones: cellular tumor (CT), microvascular proliferation (CTmvp), pseudopalisading cells adjacent to necrosis (CTpan), infiltrating tumor (IT), and leading edge (LE). The resulting 270 RNA-sequencing samples provide a spatially resolved transcriptomic atlas of GBM. Separately, the IVYGAP-RADIOMICS companion dataset [7] provides 3920 International Biomarker Standardization Initiative (IBSI)-compliant radiomic features per MRI-defined subcompartment (enhancing tumor [ET], non-enhancing tumor [NET], and peritumoral edema [ED]) for 31 of these patients with available multiparametric MRI (T1, T1-gadolinium, T2, FLAIR). Despite both datasets being publicly available since 2020, no study has linked the IVYGAP-RADIOMICS feature set to the zone-level IvyGAP RNA-seq data to test whether radiomic features reflect the transcriptomic programs of specific anatomic zones.

Prior studies have explored imaging-transcriptomic associations in GBM [8,9,10,11,12], including radiomic prediction of immune enrichment scores [13] and radiomic–genomic survival models [14,15], but none have connected zone-level transcriptomic data with IBSI radiomic features. Park et al. [8] used ADC/CBV clustering in only five patients (r < 0.30); Le et al. [9] and Zhang et al. [12] used whole-tumor approaches; Hu et al. [10] achieved spatially matched biopsies in a different dataset; Beig et al. [11] linked IvyGAP imaging subcompartments to bulk TCGA gene expression and ssGSEA for survival prediction, but did not use zone-level RNA-seq or transcriptomic pathway scores as outcomes. No study has linked the IvyGAP atlas with the IVYGAP-RADIOMICS feature set.

In this study, we test whether radiomic features extracted from MRI-defined tumor subcompartments (ET, NET, ED) associate with transcriptomic pathway enrichment scores derived from the biologically approximate IvyGAP anatomic zones. We employ a biologically motivated but spatially approximate zone-to-subcompartment mapping (e.g., CT and CTmvp to ET; CTpan to NET; IT and LE to ED) and use both linear mixed-effects models (associational analysis) and nested cross-validated Elastic Net regression (predictive analysis) as complementary analytical frameworks. We explicitly frame this as a hypothesis-generating exploratory analysis, acknowledging that the absence of voxel-level spatial co-registration between LMD sites and MRI subcompartments is a fundamental limitation that attenuates all observed associations.

2. Materials and Methods

2.1. Datasets and Patient Matching

Two publicly available datasets were used. The IvyGAP RNA-seq dataset was obtained from the Allen Institute for Brain Science portal and comprises 270 LMD samples from 41 patients across five anatomic zones (CT, CTmvp, CTpan, IT, LE) [6]. All IvyGAP patients are IDH-wildtype GBM (pre-WHO 2016 cohort). Gene expression values were provided as fragments per kilobase of transcript per million mapped reads (FPKM). The IVYGAP-RADIOMICS dataset was obtained from The Cancer Imaging Archive (TCIA) and provides 3920 IBSI-compliant radiomic features per subcompartment (ET, NET, ED) across four MRI sequences for 31 patients [7]. All MRI data originate from a single institution, eliminating the need for ComBat harmonization. Radiomic features were pre-extracted by Pati et al. [7] from BraTS-style tumor segmentations.

Patient identifiers were matched across the two datasets, yielding 28 patients with both transcriptomic and radiomic data (mean age 58.5 years, SD 7.8; median Karnofsky Performance Status [KPS] 90; MGMT promoter methylated 13/28 [46%]; 27 primary tumors, 1 recurrent). Sex distribution is not reported in the public IvyGAP metadata. Zone availability varied across patients (CT available for all 28; other zones for 7–15), resulting in an unbalanced design after subcompartment aggregation (Section 3.1). Under WHO 2021 criteria, all patients classify as glioblastoma, IDH-wildtype (IDH status confirmed via the IvyGAP molecular annotations). Detailed prior treatment history (chemotherapy, radiation, dexamethasone) was not available from the IvyGAP metadata. Because CTmvp samples were available for only 9 of 28 patients, the ET transcriptomic score represents CT alone for 19 patients and mean (CT, CTmvp) for 9. This compositional inconsistency is particularly relevant for the Angiogenesis pathway, as CTmvp is the Angiogenesis-enriched zone; sensitivity analysis S1a (Section 2.9) addresses this by restricting ET to CT only. The overall study design is summarized in Figure 1.

2.2. Zone-to-Subcompartment Mapping

Because no spatial fiducials or validated registration pipeline links IvyGAP LMD sampling sites (identified post hoc on hematoxylin and eosin histology by neuropathologists [6]) to preoperative MRI voxels, we employed a biologically motivated but spatially approximate mapping between IvyGAP zones and MRI subcompartments (

Table 1. Zone-to-subcompartment mapping with biological rationale and empirical support.

IvyGAP Zone	MRI Subcompartment	Biological Rationale	Park et al. [8] Correlation
CT + CTmvp 1	Enhancing Tumor (ET)	Viable proliferating core and active Angiogenesis are the principal sources of gadolinium enhancement	r = 0.238 (CT-ET), r = 0.195 (CTmvp-ET)
CTpan	Non-Enhancing Tumor (NET)	Pseudopalisading necrosis regions are predominantly located within the non-enhancing tumor core	r = 0.241
IT + LE	Peritumoral Edema (ED)	Infiltrating tumor and leading edge extend into the FLAIR-hyperintense peritumoral zone	r = 0.294 (mean IT/LE-ED)

¹ CTmvp (microvascular proliferation) samples are transcriptomically heterogeneous and may share features with CTpan. Sensitivity analysis S1a tests the robustness of all results when CTmvp is excluded from the ET subcompartment.

).

The biological rationale for each mapping is as follows. CT and CTmvp were mapped to ET because the cellular tumor core is the principal site of blood–brain barrier disruption and active Angiogenesis, producing gadolinium enhancement on T1-weighted imaging; CTmvp (microvascular proliferation) directly drives the leaky neovasculature responsible for contrast enhancement. CTpan was mapped to NET because pseudopalisading necrosis surrounds the devitalized necrotic core, corresponding to the non-enhancing central region on MRI where tissue is no longer viable enough to support contrast uptake. IT and LE were mapped to ED because infiltrating tumor cells and the leading edge of invasion extend beyond the enhancing margin into the peritumoral region, which appears hyperintense on T2/FLAIR due to vasogenic edema from tumor-induced blood–brain barrier disruption.

Park et al. [8] reported weak but directionally consistent correlations between histologic zone proportions and MRI subcompartment volumes (mean r = 0.242), implying substantial signal attenuation (r² = 0.059). All reported associations must be interpreted as zone-approximate, not spatially precise, and the degree of attenuation cannot be precisely estimated.

2.3. Transcriptomic Target Definition

Twenty-four gene sets were used to compute single-sample Gene Set Enrichment Analysis (ssGSEA) pathway enrichment scores for each LMD sample [16,17]: (1) fifteen GBM-relevant Hallmark gene sets from the Molecular Signatures Database (MSigDB): Hypoxia, Angiogenesis, Epithelial–Mesenchymal Transition (EMT), Inflammatory Response, TNF-alpha/NF-kB Signaling, IL-6/JAK/STAT3 Signaling, Interferon Gamma Response, P53 Pathway, MYC Targets V1, E2F Targets, G2M Checkpoint, mTORC1 Signaling, Glycolysis, Oxidative Phosphorylation, and Complement; (2) four Neftel et al. [5] cellular state signatures: mesenchymal-like (MES; MES1 and MES2 collapsed), astrocyte-like (AC), oligodendrocyte progenitor-like (OPC), and neural progenitor-like (NPC; NPC1 and NPC2 collapsed, as sub-states share core transcription factor programs and the sample size limits power to distinguish sub-state nuances); and (3) five IvyGAP zone-specific gene modules defined as the top 200 differentially expressed genes per zone (Wilcoxon test, FDR < 0.05, |log₂ fold change| > 1).

ssGSEA was computed on log₂(FPKM + 1)-transformed expression values using the GSVA package [17] with Gaussian kernel cumulative density function estimation. Zone-level enrichment scores were aggregated to the subcompartment level by computing the mean across zones mapped to each subcompartment (Table 1).

Five of the 24 gene sets used in this analysis—the IvyGAP CT, CTmvp, CTpan, IT, and LE zone modules—were derived directly from the same IvyGAP transcriptomic atlas that provides our outcome data. This introduces a structural non-independence between gene-set definition and outcome evaluation: associations involving these five modules cannot be interpreted as independent biological discovery and are reported as internal-consistency checks only. The 15 Hallmark and four Neftel signatures—wholly external to IvyGAP—therefore constitute the interpretively primary gene sets for all downstream inference. This circularity is revisited as a dedicated subsection in the main Discussion (Section 4.5). Pairwise Jaccard similarity indices were computed for all 24 gene sets to quantify redundancy (Section 3.9).

2.4. Feature Reduction Pipeline

The 3920 radiomic features per subcompartment were reduced through a two-stage unsupervised filtering pipeline: (1) near-zero-variance filtering using the nearZeroVar function from the caret package [18], removing features with near-zero variance ratios (3920 to 3860 features); and (2) pairwise Spearman correlation filtering, removing one feature from each pair with |r| > 0.90 (3860 to 597 features). These two steps are unsupervised (outcome-agnostic) and do not introduce data leakage.

For the associational analysis (Section 2.5), a third supervised step was applied: per-pathway univariate Spearman correlation with each pathway enrichment score within each subcompartment separately, retaining features with minimum FDR < 0.10 across subcompartments (Benjamini–Hochberg correction [19] applied within each subcompartment) and selecting the top five per pathway. This step was performed on the full dataset, which is standard practice for exploratory associational analyses but introduces potential optimistic bias that must be acknowledged. After this pipeline, 12 of 24 pathways had at least one radiomic feature and were carried forward to mixed-effects modeling. For the predictive analysis (Section 2.6), supervised feature selection was instead performed inside each cross-validation fold to prevent data leakage. The complete feature reduction pipeline is illustrated as a flowchart in Figure S10, with dimensionality at each stage shown in Figure S9.

2.5. Exploratory Associational Analysis: Linear Mixed-Effects Models

The LMM analysis is exploratory and employs full-dataset feature pre-selection (Section 2.4), which may yield optimistically biased results. The nested CV (Section 2.6) serves as the primary evidentiary framework because feature selection occurs inside each fold, eliminating data leakage.

For each of the 12 pathways with at least one radiomic feature, a linear mixed-effects model (LMM) was fitted using the lme4 package [20] with lmerTest [21] for denominator degrees of freedom (Satterthwaite approximation):

Full model: z_pathway ~ z_rad_1 + … + z_rad_k + subcompartment + (1 | patient_id)

Null model: z_pathway ~ subcompartment + (1 | patient_id)

where z_pathway denotes the globally z-scored pathway enrichment score, z_rad_1 through z_rad_k denote the within-subcompartment z-scored radiomic features (k ≤ 5), subcompartment is a fixed factor (ET, NET, ED), and patient_id is a random intercept to account for repeated measures within patients.

The omnibus radiomic contribution was assessed via likelihood ratio test (LRT) comparing full and null models. Marginal R² and conditional R² were computed using the Nakagawa–Schielzeth method [22,23]. FDR correction was applied across all 24 pathways [19], assigning p = 1.0 to 12 pathways with zero features. This conservative approach inflates the FDR denominator, making it harder to achieve significant results, but does not correspond to a formal BH-FDR guarantee, since the assigned values are not true p-values under the null. For FDR-significant pathways, coefficient p-values were Holm-corrected [24]. Two FDR thresholds are used, reflecting the different statistical contexts. For the LMM, where only 12 pathways with surviving features were tested, conventional FDR < 0.05 was applied. For nested CV, the conservative 24-pathway denominator (assigning p = 1.0 to all pathways with R²_cv ≤ 0) yields stringent FDR values; FDR < 0.10 was therefore considered the interpretive threshold for this exploratory analysis. Feature pre-selection may inflate LMM significance; this is independently evaluated through nested CV (Section 2.6). For transparency, we also report the standard Benjamini–Hochberg FDR correction restricted to the pathways actually tested (12 for LMM; 3 with R²_cv > 0 for nested CV) as a comparison, in addition to the conservative 24-pathway correction.

2.6. Primary Predictive Analysis: Nested Cross-Validated Elastic Net

Elastic Net regression [25] was selected because it combines L1 and L2 penalties, promoting grouped selection of correlated features while maintaining sparsity [26], addressing the instability of LASSO [27] in settings with highly correlated radiomic features. The conservative lambda.1se rule was used to guard against overfitting [28].

To provide an unbiased assessment free from data leakage, we implemented nested leave-one-patient-out cross-validation (LOPO-CV) for all 24 pathways. Radiomic features and pathway scores were averaged across subcompartments within each patient, yielding N = 28 independent observations. Patient-level averaging sacrifices subcompartment-level resolution; the LMM (Section 2.5) directly models this variation and is complementary.

For each LOPO fold (28 folds per pathway): (1) one patient was held out for testing; (2) univariate Spearman correlations were computed between each of the 597 candidate features and the pathway score using only the 27 training patients; (3) features with FDR < 0.10 were retained, and the top five by raw p-value were selected; and (4) Elastic Net regression [25] was fitted on the selected features with alpha grid search (0.1 to 1.0, step 0.1) and lambda selected at the conservative one-standard-error rule (lambda.1se), using inner 5-fold CV within the training set.

Performance was evaluated via R²_cv, MAE, and Spearman correlation. Uncertainty was quantified via nonparametric bootstrap confidence intervals; the nested permutation p-values (Section 2.7) are the primary inferential tool. Implementation details of the bootstrap procedure and its interpretive limits are provided in Supplementary Methods S1.

Feature stability was assessed by recording which features were selected in each LOPO fold. Features appearing in more than 50% of folds were designated as “stability-selected” [29]. Hyperparameter selection details and stability of alpha/lambda across folds are reported in Supplementary Methods S1 and Supplementary Table S11.

2.7. Permutation Testing

Two permutation frameworks were employed. For the LMM associational analysis, patient-level permutation (N = 1000) shuffled the mapping between patients’ pathway scores and radiomic features, preserving the within-patient correlation structure. The LRT chi-squared statistic served as the primary test statistic.

For the nested CV predictive analysis, a fully nested permutation test (N = 1000) was performed for pathways with R²_cv > 0. In each permutation, pathway scores were shuffled across patients, and the entire nested CV pipeline, including feature selection inside each fold, was re-executed. This provides a truly unbiased permutation p-value that accounts for the adaptive feature selection process. Permutation p-values were computed as n_extreme/N_perm.

2.8. Legacy Analysis

A legacy analysis with pre-screened features (selected on the full dataset before cross-validation) is reported in Supplementary Table S5 for methodological comparison with the bias-free nested CV.

2.9. Sensitivity Analyses

Ten pre-specified sensitivity analyses tested robustness to alternative zone mappings (S1a, S1b), patient exclusion (S2), aggregation method (S3), FDR subset correction (S4a–c), random effects structure (S6), standardization approach (S7), denominator degrees of freedom (S8), and clinical covariate adjustment for Inflammatory Response (age, MGMT status, Verhaak molecular subtype; S9). Full details are in Supplementary Table S3.

2.10. Sample Size Justification

The minimum sample size for p = 5 predictors, R² = 0.20, and shrinkage S ≥ 0.90 was N_min = 240 (pmsampsize [30], Riley et al. [31], Criterion 4). With N = 28, the minimum detectable R² at S ≥ 0.90 was 0.641.

2.11. Software and Reproducibility

All analyses were conducted in R version 4.5.0 using the following key packages: lme4 v1.1-37 [20], lmerTest v3.1-3 [21], performance v0.13.0 [23], glmnet v4.1-8 [25], pmsampsize v1.1.3 [30], GSVA [17], caret v7.0-1 [18], and pbkrtest v0.5.5 (for Kenward-Roger correction). A random seed of 42 was used for all stochastic procedures. A CLEAR (CheckList for EvaluAtion of Radiomics research) [32] compliance table is provided as Supplementary Table S8.

3. Results

Throughout, R²_cv denotes cross-validated R², R²_m marginal R² (fixed effects), R²_c conditional R² (fixed + random), and ΔR²_m the radiomic increment beyond subcompartment effects.

3.1. Data Availability and Feature Reduction

Of 41 IvyGAP patients and 31 IVYGAP-RADIOMICS patients, 28 were present in both datasets and constituted the analysis cohort. Of these, 27 (96%) underwent primary surgery and one (4%) had recurrent tumor (mean age 58.5 years, SD 7.8; median KPS 90; MGMT methylated 13/28). These 28 patients contributed a total of 50 observations across three MRI subcompartments (ET: n = 28; NET: n = 15; ED: n = 7), with only 6 of 28 patients having data for all three subcompartments. The imbalance reflects differential availability of zone-level RNA-seq data across patients (Figure S8).

The unsupervised radiomic feature reduction pipeline progressively reduced dimensionality: 3920 initial features were filtered to 3860 after near-zero-variance removal, and to 597 after Spearman correlation pruning (|r| > 0.90) (Figure S9).

3.2. Nested Cross-Validation: Predictive Performance

Of the 24 pathways evaluated, three showed positive predictive signal (R²_cv > 0) in the nested LOPO-CV analysis with internal feature selection (

Table 2. Nested cross-validation results for pathways with R²_cv > 0.

Pathway	R2cv	95% CI e	MAE	Spearman Rho	Stable Features (>50% Folds)	Nested Perm p	FDR (24) 1
Angiogenesis	0.209	[0.028, 0.353]	0.702	0.581	5	0.006	0.096
Inflammatory Response c	0.185	[0.071, 0.355]	0.674	0.524	5	0.008	0.096
IvyGAP CTpan module d	0.133	[−0.079, 0.350]	0.740	0.348	4	0.013	0.104

¹ Two FDR correction approaches are reported for transparency. The conservative approach applies BH-FDR across all 24 pathways, assigning p = 1.0 to 21 pathways with R²_cv ≤ 0 (FDR = 0.096 for Angiogenesis and Inflammatory Response; Table S10). The standard approach applies BH-FDR restricted to the three pathways with R²_cv > 0 (FDR = 0.012 for both Angiogenesis and Inflammatory Response). We lead with the conservative correction to avoid selection bias, but the standard correction is provided as a transparency comparison. ^c Designated as primary based on convergent LMM evidence (Section 3.4). Both analyses use the same outcome data and are not independent; this represents consistency across different statistical models on the same dataset, not independent replication. ^d CI crosses zero; gene module scored on the same expression data from which it was derived (see Limitation 7); composition baseline test significant (p = 0.001). Interpret with caution. ^e Bootstrap CIs (B = 1000) condition on fitted predictions and capture metric sampling variability only; see Section 2.6 for interpretation.

, Figure 2).

After BH-FDR correction across all 24 pathways, Angiogenesis and Inflammatory Response both reached FDR < 0.10 (FDR = 0.096), while the CTpan module did not (FDR = 0.104). Under a standard BH-FDR correction restricted to the three pathways with R²_cv > 0, both Angiogenesis and Inflammatory Response reached FDR = 0.012; CTpan reached FDR = 0.013 but should still be interpreted cautiously due to its circularity and composition effect. Based on convergent evidence from the complementary LMM analysis (Section 3.4), which identified Inflammatory Response as the sole FDR-significant pathway (LMM FDR = 0.024) while Angiogenesis did not reach LMM significance (FDR = 0.445), Inflammatory Response was designated as the primary pathway. The Angiogenesis finding is retained and reported as a tentative nested-CV-only signal: because it is supported by only one of the two analytical frameworks and feature selection is known to be unstable at N = 28, it requires independent validation before any biological conclusion can be drawn.

The CTpan module had a confidence interval spanning zero, is circular by construction (Section 2.3, Limitation 7), and showed a significant composition effect (Spearman rho = 0.579, p = 0.001); it should not be considered robust. Composition baseline tests for Angiogenesis (rho = −0.105, p = 0.594) and Inflammatory Response (rho = −0.003, p = 0.988) were non-significant.

The remaining 21 of 24 pathways showed no predictive signal (R²_cv ≤ 0; Table S1). These included all four Neftel cellular state signatures (MES, AC, OPC, NPC), four of five IvyGAP zone-specific modules (CT, CTmvp, IT, LE), and 13 Hallmark pathways (Hypoxia, EMT, TNF-alpha/NF-kB Signaling, IL-6/JAK/STAT3 Signaling, Interferon Gamma Response, P53 Pathway, MYC Targets V1, E2F Targets, G2M Checkpoint, mTORC1 Signaling, Glycolysis, Oxidative Phosphorylation, and Complement). Of these, 12 pathways had zero features passing the univariate filter in any fold (median features per fold = 0). This pattern is consistent with an absence of detectable patient-level association under the present design but does not demonstrate biological inaccessibility: at N = 28 (vs. N ≈ 240 required), feature-selection instability and attenuation from the non-spatial zone-to-subcompartment mapping (Park et al. [8] r² ≈ 0.059) are each sufficient to suppress real associations of moderate effect size. The dominant negative finding and its alternative explanations are discussed in Section 4.

3.3. Feature Stability and Identification

All features contributing to the Inflammatory Response prediction were T2-derived (Table S12, Figure 3), consistent with the known sensitivity of T2-weighted imaging to inflammatory edema and tissue water content.

Two features were selected in all 28 folds (100% stability): T2 GLCM AutoCorrelation (rad_1707) and T2 GLSZM Large Zone Low Grey Level Emphasis (Bins-128, Radius-3; rad_1950). A third GLCM feature (Energy, rad_1930) was selected in 96% of folds (Table S12, Figure 3). Note that rad_1950 (Bins-128, Radius-3) and the LMM feature rad_1732 (Bins-128, Radius-1) are distinct radius configurations of the same GLSZM texture class, explaining their independent selection across the two analytical frameworks.

In imaging terms, GLCM AutoCorrelation quantifies the consistency of local T2 signal patterns—high values indicate homogeneous tissue, while low values reflect heterogeneous signal mixtures. GLSZM Large Zone Low Grey Level Emphasis captures the extent of large contiguous regions with low T2 signal intensity; its negative association with inflammatory enrichment (beta = −0.471 in the LMM; Section 3.4) suggests that greater inflammatory activity corresponds to more heterogeneous tissue architecture with fewer large homogeneous zones. GLCM Energy measures the uniformity of the T2 signal distribution. Together, these features quantify the heterogeneity of T2 signal within tumor subcompartments, consistent with the patchy edema, macrophage infiltration, and vascular permeability changes characteristic of the inflammatory microenvironment.

For the Angiogenesis pathway, stability selection identified five features in more than 50% of LOPO folds (Table S9), all T2-derived: three GLSZM texture features (100% stability) and two first-order/GLSZM features (96% stability).

A complete feature lookup table mapping all radiomic feature indices to their full IBSI names is provided in Supplementary Table S6.

3.4. Associational Analysis: Mixed-Effects Models

Of the 12 pathways tested in the LMM analysis (after supervised univariate pre-screening), only Inflammatory Response reached significance after FDR correction across all 24 pathways: FDR = 0.024, marginal R² = 0.384, of which ΔR² = 0.214 is attributable to radiomic features beyond the subcompartment effect (null model R²_m = 0.170), conditional R² = 0.687, LRT χ ² = 20.53 (df = 5, p = 0.001) (

Table 3. Linear mixed-effects model results for all 12 tested pathways, ordered by FDR. R²_m (null) = subcompartment + random intercept only; ΔR²_m = radiomic increment.

Pathway	Category	k	R2m (Null)	R2m (Full)	ΔR2m	R2c	LRT p	FDR (24)
Inflammatory Response	Hallmark	5	0.170	0.384	0.214	0.687	0.001	0.024 *
Angiogenesis	Hallmark	1	0.425	0.459	0.034	0.582	0.053	0.445
Hypoxia	Hallmark	5	0.837	0.847	0.010	0.918	0.085	0.445
P53 Pathway	Hallmark	5	0.470	0.533	0.063	0.716	0.101	0.445
Glycolysis	Hallmark	5	0.789	0.804	0.015	0.849	0.112	0.445
mTORC1 Signaling	Hallmark	3	0.671	0.694	0.024	0.709	0.133	0.445
Neftel MES	Neftel	2	0.546	0.562	0.016	0.699	0.133	0.445
Complement	Hallmark	5	0.059	0.208	0.149	0.462	0.148	0.445
EMT	Hallmark	5	0.339	0.409	0.070	0.589	0.176	0.469
TNFA/NF-kB	Hallmark	5	0.580	0.610	0.030	0.703	0.352	0.845
IvyGAP CTpan Module	IvyGAP	5	0.906	0.902	−0.004	0.922	0.773	1.000
Oxidative Phosphorylation	Hallmark	3	0.352	0.344	−0.008	0.557	0.800	1.000

* FDR < 0.05 across all 24 pathways (BH correction). Because the Nakagawa–Schielzeth R² is not strictly additive in mixed models, ΔR²_m values are approximate; negative values (CTpan, OxPhos) reflect variance repartitioning.

, Figure S1). Note that features were pre-selected on the full dataset; marginal R² values are likely optimistically biased (see Section 2.4). The intraclass correlation for Inflammatory Response was ICC = 0.492, indicating moderate within-patient correlation.

Pathways with high R²_m (full) but minimal ΔR²_m, such as Hypoxia (ΔR²_m = 0.010) and Glycolysis (ΔR²_m = 0.015), showed negligible radiomic increments. Inflammatory Response had the largest ΔR²_m (0.214).

Residual diagnostics (Figure S3) showed approximate normality (Shapiro–Wilk p = 0.384). One observation exceeded Cook’s D = 1.0 (D = 1.097); sensitivity analyses (S1a, S2) confirm robust results (Table S3).

Angiogenesis had only one feature surviving the univariate pre-screen and did not reach LMM FDR significance (FDR = 0.445), but the nested CV analysis identified a more informative feature set, underscoring the importance of the nested CV approach.

The Inflammatory Response model included five radiomic features (Table S13, Figure S2). The strongest individual contributor was T2 GLSZM Large Zone Low Grey Level Emphasis (Bins-128, Radius-1; rad_1732; beta = −0.471, 95% CI [−0.758, −0.184], Holm-adjusted p = 0.010), indicating a negative association between this texture feature and inflammatory pathway enrichment (Table S13). One additional feature, T2 CoLIAGe skewness of difference variance (ws = 5; rad_1080), showed nominal significance (beta = 0.308, 95% CI [0.007, 0.610], uncorrected p = 0.045) but did not survive Holm correction (adjusted p = 0.181).

The subcompartment fixed effect was also significant (F(2, 17.5) = 9.48, p = 0.002), indicating that inflammatory pathway enrichment scores differed across ET, NET, and ED subcompartments independently of radiomic features.

3.5. Permutation Testing

Patient-level permutation testing for the LMM (N = 1000 permutations) assessed whether the Inflammatory Response LRT χ² statistic exceeded chance expectation. The observed χ² of 20.53 yielded a permutation p-value of 0.055 for the χ² statistic and p = 0.050 for marginal R² (Figure S4). These values indicate that the Inflammatory Response association exceeds approximately 95% of the null distribution but does not reach conventional significance at α = 0.05. This permutation test used pre-selected features (i.e., the same features in every permutation), which does not fully account for the adaptive feature selection process.

The nested permutation test (Section 2.7), which re-selects features inside each fold of each permutation, provides a more rigorous assessment. All three pathways with R²_cv > 0 reached significance at the uncorrected level: Angiogenesis (p = 0.006), Inflammatory Response (p = 0.008), and IvyGAP CTpan module (p = 0.013) (Table 2). After BH-FDR correction across all 24 pathways (Table S10), Angiogenesis and Inflammatory Response reached FDR < 0.10 (FDR = 0.096 each), while CTpan did not (FDR = 0.104). The decision to test only pathways with R²_cv > 0 in the nested permutation is data-dependent; this is accounted for by reporting FDR-corrected p-values across all 24 pathways.

3.6. Clinical Covariate Adjustment

To assess whether the radiomic–transcriptomic association for Inflammatory Response was confounded by clinical variables, we progressively added covariates to the LMM (Table S7).

The radiomic–transcriptomic association for Inflammatory Response remained significant across all covariate models (LRT p < 0.005 in all cases). Adding age and MGMT methylation had minimal impact on the marginal R² (Models B-C), while adding Verhaak molecular subtype (Model D) increased the marginal R² to 0.479, suggesting that molecular subtype explains additional variance in inflammatory pathway activity beyond radiomic features alone. Critically, the radiomic contribution remained significant even in the fully adjusted model (LRT p = 0.004), indicating that the association is not driven by confounding from age, MGMT status, or molecular subtype. Model D results should be interpreted with caution, given the high parameter-to-observation ratio (12 parameters from 50 observations across 28 clusters).

3.7. Sensitivity Analyses

The Inflammatory Response association was robust across the majority of sensitivity analyses (Table S3, Figure S5). The LRT p-value remained below 0.01 under alternative zone mappings (S1a: p = 0.003; S1b: p = 0.003), exclusion of patients with single subcompartments (S2: p = 0.007), median aggregation (S3: p = 0.001), Hallmark-only FDR re-correction (S4a: FDR = 0.010), and Kenward–Roger denominator degrees of freedom (S8: p = 0.006).

Robustness under S1a (ET = CT only, excluding CTmvp) addresses the concern that CTmvp samples could confound the ET signal (S1a: p = 0.003, R²_m = 0.365).

Two analyses qualified the primary result. Global standardization (S7) weakened the association to p = 0.076, indicating that the radiomic signal captures within-subcompartment variation in tissue properties rather than between-subcompartment differences, which are already captured by the subcompartment fixed effect. The random slopes model (S6) could not be fitted due to insufficient sample size.

3.8. Legacy Pre-Screened Analysis

The legacy pre-screened Elastic Net yielded R²_cv = −0.104 for Inflammatory Response (Table S5), compared with R²_cv = 0.185 from nested CV, confirming that pre-screening introduced overfitting.

3.9. Gene Set Overlap

Jaccard similarity analysis confirmed that Angiogenesis and Inflammatory Response represent independent signals (maximum J < 0.10). Only the IvyGAP IT and LE modules showed high overlap (J = 0.653).

4. Discussion

This exploratory analysis tested whether MRI-derived radiomic features from tumor subcompartments associate with regional transcriptomic programs in GBM, leveraging the IvyGAP and IVYGAP-RADIOMICS public datasets. Using a nested cross-validation framework, not employed in prior IvyGAP radiomic studies [8,9,10,11,12], we identified Inflammatory Response (R²_cv = 0.185) as the only transcriptomic program supported by both the nested CV (FDR = 0.096) and the complementary LMM (FDR = 0.024) analyses. Angiogenesis (R²_cv = 0.209) reached nested-CV significance (FDR = 0.096) but was not corroborated by the LMM (FDR = 0.445) and is therefore reported as a tentative signal requiring independent validation. Twenty-one of 24 pathways showed no signal. In the complementary exploratory associational analysis, Inflammatory Response was the sole pathway reaching FDR significance (FDR = 0.024 across all 24 pathways, ΔR² = 0.214 beyond subcompartment effects), driven primarily by T2-derived texture features. Given the sample size of 28 patients, the zone-approximate mapping, and the absence of external validation, these associations should be regarded as preliminary hypotheses requiring prospective confirmation before any clinical translation. Furthermore, in the absence of voxel-level spatial co-registration between LMD sites and MRI voxels, the observed associations may partly reflect macro-compartment effects (systematic differences between ET, NET, and ED subcompartments) rather than true within-zone radiogenomic coupling; the R² decomposition (Table 3) helps quantify the subcompartment contribution, but it cannot fully separate macro-compartment from true zone-level signal.

4.1. Biological Interpretation

The identification of Angiogenesis and Inflammatory Response as the two pathways with predictive radiomic signals is biologically plausible. Both processes directly modulate MRI signals through mechanisms that alter tissue contrast.

Angiogenesis drives gadolinium enhancement through leaky neovasculature. However, the dominant features were T2-derived texture features. This likely reflects a methodological constraint: within the ET subcompartment (defined by enhancement), T1-gadolinium features have a restricted dynamic range because the region is already selected for high enhancement, whereas T2 features retain full dynamic range and capture internal heterogeneity of the enhancing core. Biologically, T2 texture within the enhancing tumor reflects downstream consequences of Angiogenesis (irregular vascular architecture, patchy edema, and microhemorrhage) rather than the vascular process directly. This interpretation is consistent with Dextraze et al. [14], who identified associations between Angiogenesis-related pathways and MRI-defined imaging habitats in 85 GBM patients. However, we emphasize that the Angiogenesis finding in our cohort is supported by only one of the two analytical frameworks (nested CV R²_cv = 0.209; LMM FDR = 0.445); at N = 28, cross-validated R² estimates and feature selection are both unstable, and asymmetric convergence between frameworks limits the robustness of any single-model result. The Angiogenesis signal should therefore be treated as hypothesis-generating rather than established, and any biological interpretation deferred to independent validation in a spatially co-registered cohort of adequate size.

Inflammatory Response involves tumor-associated macrophages, which constitute 30–50% of the GBM mass [33,34], modulating vascular permeability and blood–brain barrier integrity, processes that directly affect T2/FLAIR signal. The dominant feature (T2 GLSZM Large Zone Low Grey Level Emphasis, beta = −0.471) showed a negative association, suggesting that inflammatory activity corresponds to more heterogeneous tissue architecture.

4.2. Why These Pathways Survive Spatial Mismatch

The selective survival of Inflammatory Response is consistent with the spatial scale of this process. Tumor-associated macrophages constitute 30–50% of GBM mass [33,34], permeating the entire subcompartment and creating a spatially homogeneous alteration of tissue properties. Because inflammation modulates the full subcompartment volume, any LMD sample captures the inflammatory signal, and radiomic features integrate the same signal across the volume. The effective signal attenuation from spatial mismatch may therefore be lower than the 94% predicted by Park et al. [8] correlations for spatially localized processes.

Angiogenesis operates differently: neovascularization creates focal structures (glomeruloid microvascular proliferation bodies), but its downstream tissue-level consequences, such as edema, vascular permeability, and microhemorrhage, are diffuse and captured by T2 texture across the enhancing core. The mechanism is structural-consequential rather than diffuse per se.

4.3. Scope and Negative Results

The dominant finding is negative: 21 of 24 pathways showed no predictive signal (R²_cv ≤ 0). Before discussing patterns within this null, we state explicitly that the absence of signal should not be interpreted as evidence that these pathways are biologically inaccessible to MRI. Three alternative explanations—any of which is sufficient on its own to suppress real associations of moderate effect size—must be considered first. First, severe statistical underpowering: Riley criteria [31] for the intended effect size require N ≈ 240, whereas the present analysis uses N = 28; at this sample size feature selection and cross-validated R² are both unstable. Second, attenuation introduced by the non-spatial zone-to-subcompartment mapping: Park et al.’s [8] empirical correlations (mean r = 0.242, r² ≈ 0.059) imply roughly 94% loss of signal variance, so only very large true effects can survive. Third, methodological constraints: The restricted expressive capacity of hand-crafted IBSI features, the absence of voxel-level co-registration, and the structural non-independence of the five IvyGAP-derived zone modules (Section 4.5) all bias null findings toward spurious absence. A true biological null is possible but cannot be distinguished from these alternative explanations with the present data. The patterns discussed below are therefore presented as informative regularities, not as evidence of biological boundaries.

Four of five IvyGAP zone-specific gene modules showed zero features passing the univariate filter. Because these modules are structurally non-independent from the outcome data (Section 4.5), the absence of signal cannot be interpreted as a substantive biological finding about zone-module detectability and is most safely read as consistent with attenuation from the zone-to-subcompartment mapping; at N = 28, a genuine biological signal and a mapping-attenuation artifact cannot be empirically distinguished from the present data.

All four Neftel cellular state signatures (MES, AC, OPC, NPC) showed no signal in this cohort. These signatures represent cell-intrinsic transcription factor programs (e.g., CEBP/D for MES, OLIG1/2 for OPC) that are not expected to directly alter tissue-level MRI contrast at scales resolvable by conventional radiomics in a sample of this size, unlike Angiogenesis or inflammation, which physically modulate vascular permeability and tissue water content. Furthermore, single-cell data show that all four states coexist within the same histological zone, and bulk RNA-seq from that zone averages across these states, which may reduce spatial specificity. The spatial scale gap is also substantial: Neftel states vary at the 10–50 micron single-cell level, two to three orders of magnitude below the millimeter-scale radiomic features. The selective failure of cell-intrinsic signatures in our data is compatible with the interpretation that radiomics preferentially captures tissue-level biology in this cohort, although we emphasize that the absence of signal should not be interpreted as a definitive absence of biological association. The lack of signal for these pathways may equally reflect the limited statistical power (N = 28), the attenuation introduced by the zone-approximate mapping, or the restricted expressive capacity of hand-crafted IBSI features, rather than a true absence of cell-level radiogenomic coupling. Distinguishing these explanations would require spatially co-registered data, substantially larger cohorts, and deep-learning-derived features.

Several Hallmark pathways with a biological rationale for MRI detectability, including Hypoxia, EMT, and Glycolysis, also showed no radiomic association. For Hypoxia, the high null-model R²_m (0.837) indicates that subcompartment membership alone captures the Hypoxia gradient, leaving only ΔR²_m = 0.010 for radiomic features; once the categorical subcompartment label is known, within-subcompartment radiomic variation adds virtually no information about Hypoxia status.

Three mechanistic factors likely explain why most pathways showed no signal. First, spatial scale mismatch: cell-intrinsic programs (Neftel states) vary at the 10–50 micrometer single-cell level, three to four orders of magnitude below the centimeter-scale volumes over which radiomic features are computed. Second, insufficient tissue-level contrast modulation: transcriptomic programs that do not alter tissue water content, vascularity, or cellularity—the physical properties governing MRI signal—are invisible to radiomics regardless of how strongly they are expressed. Third, the zone-to-subcompartment mapping attenuation (Park et al. [8] mean r = 0.242, r² = 0.059) implies that approximately 94% of signal variance is lost in the mapping; only associations with very large true effect sizes could survive this attenuation at N = 28. Together, these factors explain why only tissue-scale microenvironmental processes (inflammation, Angiogenesis) that directly and diffusely modulate MRI contrast produced detectable signal, while the majority of pathways did not.

4.4. Methodological Considerations

Data leakage through feature pre-selection is a critical concern in radiomic studies [35]. Our legacy Elastic Net with pre-screened features produced R²_cv = −0.104 for Inflammatory Response, while the bias-free nested CV yielded R²_cv = 0.185, demonstrating that pre-selection introduced optimistic bias that paradoxically worsened predictions. We addressed this by implementing nested CV as the primary analysis, with feature selection inside each LOPO fold. For the LMM, features were pre-selected on the full dataset; the independent confirmation in nested CV provides convergent evidence that the Inflammatory Response association is not an artifact of data leakage. The discrepancy between LMM permutation p = 0.055 and nested CV permutation p = 0.008 for Inflammatory Response reflects different null hypotheses: the LMM permutation uses pre-selected features in every permutation, while the nested CV permutation re-executes the full pipeline, including feature selection under the null, generating a tighter null distribution. Importantly, the results are partly model-dependent. Angiogenesis reached FDR significance only in the nested CV analysis (FDR = 0.096 conservative; FDR = 0.012 standard BH on tested pathways) but not in the LMM (FDR = 0.445), while Inflammatory Response was significant in both frameworks (LMM FDR = 0.024). This asymmetric convergence—one pathway supported by both models, the other by only one—limits the robustness of the Angiogenesis finding and underscores that the results are not fully model-agnostic. Readers should interpret Angiogenesis as a nested-CV-only signal requiring independent replication before any biological conclusion can be drawn.

The designation of Inflammatory Response as the primary pathway was based on consistent results across the nested CV (FDR = 0.096) and LMM (FDR = 0.024) analyses. Because both analyses use the same 28 patients and the same ssGSEA outcome scores, the convergence of nested CV and LMM results represents consistency across different statistical frameworks applied to the same observations, not independent replication. Independent validation would require a separate patient cohort with spatially co-registered imaging and transcriptomics, which does not currently exist. The two approaches—one predictive with internal feature selection, one associational with pre-screened features—converge on the same pathway using different model structures and units of analysis (patient-level vs. observation-level), but this convergence increases internal consistency without substituting for external replication. The dependence of results on the standardization approach (S7: p = 0.076 under global standardization) suggests that the association is partly driven by within-subcompartment relative feature values rather than absolute magnitudes.

Five of the 24 gene sets we evaluated—the IvyGAP CT, CTmvp, CTpan, IT, and LE zone modules—were derived from the same IvyGAP transcriptomic atlas that provides our outcome data. The gene-set definitions and the ssGSEA scores on which we test radiomic association are therefore not independent sources of information, and any association involving these five modules is a consistency check on the zone definition, not a biological discovery. This circularity affects interpretation in three concrete ways. First, the IvyGAP CTpan module showed positive nested-CV R² (0.133) and a significant composition-baseline effect (Spearman rho = 0.579, p = 0.001) but did not survive the conservative 24-pathway FDR correction; we treat it as uninterpretable rather than as a borderline finding. Second, the zero-signal result for the four other zone modules (Section 4.3) is most safely read as consistent with mapping attenuation, since a substantive test of zone-module detectability would require externally defined gene sets. Third, the 15 Hallmark and 4 Neftel signatures, all wholly external to IvyGAP, remain the only gene sets from which independent radiogenomic inference can be drawn in the present data. All substantive biological claims in this manuscript are therefore restricted to these 19 external signatures, and the five IvyGAP modules are retained only for completeness of reporting.

4.5. Limitations

This study has several important limitations.

Firstly, the zone-to-subcompartment mapping is biologically approximate rather than spatially precise, with weak correlations (mean r = 0.242), as reported by Park et al. [8]. All results should be interpreted within this context. Although a classical measurement error attenuation model suggests true effect sizes could be larger, its applicability to categorical zone-to-subcompartment mapping is uncertain and cannot be verified with the current data. The study is severely underpowered for the scope of modeling. With N = 28 matched patients against a Riley criterion of N = 240 for p = 5, R² = 0.20, and shrinkage S ≥ 0.90, predictive performance estimates from LOPO-CV are intrinsically unstable: feature selection inside each fold is noisy, bootstrap confidence intervals are wide, and point estimates of R²_cv may vary substantially under small perturbations of the dataset. All predictive results must therefore be regarded as exploratory rather than as validated performance estimates, the study is hypothesis-generating rather than definitive, and the possibility of false-positive associations cannot be excluded. Additionally, there is no external validation cohort because IvyGAP is the only dataset that combines zone-level RNA-seq with matched MRI, preventing independent replication. The MRI data were obtained from a single institution, which may limit generalizability across scanners, though this also reduces batch effects. BraTS-style segmentations are radiological rather than biological boundaries. In the associational analysis, feature pre-selection on the full dataset may inflate significance, though nested cross-validation results help mitigate this concern. The R² decomposition now separates radiomic increments from subcompartment effects, offering a clearer interpretation. Since IvyGAP zone modules are scored on the same expression data used for their derivation, circularity affects certain modules like CTpan, which showed positive R²_cv but should be interpreted cautiously. The LMD samples represent microscopic tissue volumes—roughly 500–2000 cells—covering less than 0.01% of the subcompartment, across centimeters, which introduces a cross-scale gap between the transcriptomic measurements and radiomic features that integrate signals over full subcompartment volumes. FPKM normalization was used instead of TPM or raw counts, with rank-based ssGSEA partially mitigating this. Permutation p-values for the LMM were around 0.05, warranting cautious interpretation, while the nested CV permutation p-value of 0.008 offers a more rigorous bound; the true significance likely lies between these values (see Section 4). Treatment confounding cannot be fully assessed. Bevacizumab (anti-VEGF) alters enhancement patterns and suppresses angiogenic transcriptomic programs; however, its use is unlikely in this cohort, as 27 of 28 patients had primary tumors with no prior anti-angiogenic therapy. Dexamethasone, routinely administered perioperatively in GBM, suppresses NF-kB-mediated inflammatory gene expression while simultaneously reducing peritumoral edema and T2/FLAIR signal; this dual suppression would attenuate both the transcriptomic and radiomic components of the inflammatory signal, biasing the observed association toward the null rather than inflating it. Temozolomide and radiation effects on the tumor microenvironment cannot be excluded but are less likely to create spurious radiomic–transcriptomic correlations in primary tumors imaged before adjuvant treatment. Dexamethasone use at the time of MRI is unknown from the IvyGAP metadata. Overall, the direction of potential treatment confounding is toward attenuation (conservative bias) rather than inflation of observed associations. Finally, WHO 2021 molecular markers and the top feature caps were not available or were sensitivity-tested.

4.6. Future Directions

If confirmed with spatially precise co-registration, several clinical application scenarios emerge. First, radiomic identification of inflammatory-enriched subregions could inform pre-operative molecular stratification, enabling selection of patients for immunotherapy trials targeting the tumor-associated macrophage compartment. Second, T2 texture features that track inflammatory response could serve as non-invasive longitudinal biomarkers for treatment response assessment, avoiding the need for repeat biopsy. Third, radiomic profiling of Angiogenesis-enriched regions could guide surgical planning by identifying subregions likely to harbor residual anti-angiogenic disease. Fourth, if radiomic–transcriptomic mapping is validated with spatial co-registration, dose painting strategies could target molecularly aggressive subregions during radiotherapy planning.

Beyond conventional MRI sequences used in this study (T1, T1-gadolinium, T2, FLAIR), advanced acquisition techniques offer complementary molecular information. Quantitative susceptibility mapping (QSM) captures iron deposition and calcification patterns associated with tumor grade and molecular subtype; Rui et al. [36] demonstrated that QSM-derived features combined with deep learning achieved high accuracy (AUC = 0.91) for glioma grading and molecular subtyping. Perfusion MRI (dynamic susceptibility contrast, dynamic contrast-enhanced) provides hemodynamic information, and diffusion-weighted imaging (apparent diffusion coefficient, diffusion kurtosis) captures microstructural properties not reflected in conventional radiomics. Future studies should evaluate whether combining conventional radiomic features with these advanced sequences improves the prediction of transcriptomic programs beyond the T2-dominated signal identified here.

More broadly, future studies should prioritize spatially co-registered datasets extending the Hu et al. [10] approach, multi-institutional harmonized radiomic extraction, and deep learning-derived features that may capture nonlinear relationships beyond hand-crafted IBSI features.

5. Conclusions

In this exploratory analysis of 28 patients with matched MRI radiomic features and zone-level RNA-seq data from the IvyGAP atlas, Inflammatory Response (R²_cv = 0.185, 95% CI [0.071, 0.355]) was the only pathway supported by both the nested-CV predictive analysis (FDR = 0.096) and the exploratory LMM (FDR = 0.024, ΔR² = 0.214 beyond subcompartment effects), driven by T2-derived texture features. Angiogenesis (R²_cv = 0.209, 95% CI [0.028, 0.353]) reached significance only in the nested CV (FDR = 0.096) and was not corroborated by the LMM (FDR = 0.445); it is reported as a tentative signal requiring independent validation before any biological claim. The absence of signal for 21 of 24 pathways should not be read as evidence of biological inaccessibility: at N = 28 (vs. N ≈ 240 required), severe underpowering, attenuation from the non-spatial zone-to-subcompartment mapping (r² ≈ 0.059), and methodological constraints each independently suffice to suppress real associations. Five of the 24 gene sets (the IvyGAP zone modules) are structurally non-independent from the outcome data and are retained only as internal-consistency checks; all substantive biological claims are restricted to the 19 external (Hallmark and Neftel) signatures. All reported associations are zone-approximate; validation in larger cohorts with spatially precise co-registration is essential before any clinical translation.