Intercellular adhesion molecule 1 links glioblastoma biology to a scalable prognostic model integrating clinical, imaging, and molecular features

Abstract

Background

Glioblastoma (GB) exhibits marked tumor microenvironmental heterogeneity, contributing to both therapy resistance and poor survival outcomes. Intercellular adhesion molecule 1 (ICAM1) is an inflammation-associated adhesion molecule implicated in immune-stromal interactions, but its clinical and biological significance in GB remains incompletely defined.

Methods

We performed ICAM1 immunohistochemistry, bulk RNA sequencing, DNA methylation analysis, single-nucleus RNA sequencing, and preoperative MRI-derived morphologic feature analysis across multi-institutional cohorts of treatment-naïve IDH-wildtype GB.

Results

High ICAM1 expression at both transcript and protein levels is associated with shorter overall survival, particularly within the mesenchymal transcriptional subtype. ICAM1 expression is inversely associated with promoter methylation. Single-nucleus RNA-sequencing analysis shows that ICAM1-high tumors are enriched for mesenchymal, hypoxia, stress-associated, and immunosuppressive myeloid programs, suggesting that ICAM1 reflects a multicompartment inflammatory tumor ecosystem. ICAM1 expression, tumor surface irregularity, and patient age are independently prognostic and minimally correlated, capturing complementary clinical, imaging, and molecular features of GB biology. When combined, our Clinical-Imaging-Molecular (CIM) framework refines outcome prediction across independent GB cohorts using routinely available clinical data, conventional imaging, and standard immunohistochemistry.

Conclusions

ICAM1 characterizes an immune-remodeled, stress-adapted GB ecosystem associated with poor clinical outcomes. Integrating patient age, surface irregularity, and ICAM1 expression yields a scalable and clinically accessible prognostic framework for GB, supporting tailored outcome prediction, particularly in low- and middle-resource settings without routine access to advanced molecular profiling.

Key Points.

• ICAM1 expression marks a remodeled immune and tumor cell GB ecosystem

• ICAM1 promoter methylation, RNA, and protein are biomarkers of survival

• CIM model enables GB outcome prediction without advanced molecular profiling

Importance of the Study

We identify intercellular adhesion molecule 1 (ICAM1) as a marker of an immune-remodeled, stress-adapted tumor ecosystem associated with poor survival in glioblastoma (GB). We demonstrate that ICAM1 promoter methylation, along with RNA and protein levels, serves as biomarkers that can predict overall survival. We further delineate altered cellular composition and malignant cell states associated with ICAM1 expression, including enrichment for mesenchymal, hypoxia-, and stress-associated tumor states, alongside immunosuppressive myeloid programs, in ICAM1-high tumors. By integrating routinely accessible clinical data, conventional imaging, and standard immunohistochemistry, we developed a Clinical-Imaging-Molecular model for outcome prediction. These variables are independently prognostic and minimally correlated with one another, capturing complementary aspects of tumor biology. Our work establishes a robust, clinically accessible prognostic tool for GB that does not require advanced molecular sequencing and is broadly applicable across low-resource settings.

Glioblastoma (GB) is the most aggressive primary malignant brain tumor in adults, characterized by rapid progression, resistance to therapy, and near-universal recurrence.¹ Despite maximal safe surgical resection followed by chemoradiotherapy, median overall survival (OS) remains limited to 16-20 months.² A defining feature of GB is its highly heterogeneous tumor microenvironment (TME), which contributes to immune evasion and tumor progression.³ Bulk transcriptomic profiling classifies GB into proneural (PN), mesenchymal (MES), and classical (CL) subtypes, each associated with distinct molecular alterations, microenvironmental states, and therapeutic responses.⁴ Increasing evidence suggests that cell adhesion molecules (CAMs) play a critical role in mediating immune-stromal interactions within the TME and promoting tumor progression.^5,6

Intercellular adhesion molecule 1 (ICAM1) is a cell-surface glycoprotein that mediates cell adhesion and transendo­thelial migration of leukocytes during inflammatory responses,^6-8 dysregulated across multiple malignancies.⁹ Its prognostic significance is context-dependent across tumor types, with ICAM1 expression associated with favorable outcomes in cancers such as colorectal and non-Hodgkin’s lymphoma,^7,10,11 and unfavorable outcomes in lung, gastric, and breast cancers.^7,9,12,13 In GB, we and others have shown that ICAM1 expression is induced in tumor and endothelial cells by pro-inflammatory cytokines, including tumor necrosis factor-α (TNF-α), as well as by hypoxic signaling.^14-16 This induction is sustained through NF-κB activation,¹⁷ which promotes angiogenesis, leukocyte recruitment, and activation of immune checkpoint pathways.^18-20 Together, these observations suggest that ICAM1 contributes to a tumor-associated inflammatory state that facilitates interactions between malignant cells and myeloid populations, while promoting resistance to anti-angiogenic therapies.^7,16,21,22 The association, however, between ICAM1 expression, TME features, and clinical outcomes in GB remains incompletely understood.

In this study, we show that ICAM1 expression is associated with reduced OS in treatment-naïve GB and reflects an inflammatory tumor ecosystem. Through integrated multi-omic and single-cell analyses, we link ICAM1 to stress-associated and immunosuppressive microenvironmental programs. We further demonstrate that combining ICAM1 expression with standardized preoperative imaging features and routinely available clinical data yields a scalable prognostic model that refines outcome prediction across independent cohorts.

Methods

Patient Cohorts

We assembled a cohort of 149 treatment-naïve IDH-wildtype (IDH-wt) GB patients, defined according to WHO 2021 criteria,²³ from 7 medical institutions spanning across Spain (Ciudad Real, Hospital 12 de Octubre, Hospital Regional Universitario de Málaga, and Hospital Marqués de Valdecilla), Canada (Princess Margaret Cancer Center, and Toronto Western Hospital), and the United States of America (USA; Penn State Cancer Institute). Eligible cases included those with available tumor tissue, clinical data, and preoperative MRI scans meeting standardized resolution criteria: slice thickness ≤ 2.0 mm, no interslice gap, and in-plane pixel spacing ≤ 1.2 mm. Multifocal lesions were excluded from this study. The cohort was divided a priori into a training set of n = 129 patients used for model development, coefficient derivation, threshold selection, and model optimization, and an independent validation set of n = 20 patients that were withheld from all model development procedures and used exclusively for external testing of model performance. Both sets included patients from the participating institutions, representing geographically and institutionally distinct patient populations with separate clinical management, pathology workflows, and imaging acquisition procedures. A study overview diagram summarizing participating institutions, cohort allocation, and datasets used for each molecular, imaging, histopathologic, and computational analysis is provided in Table 1.

Table 1.

Patient cohort summary, and study overview and analysis plan

Patient cohort characteristics
Institution (n = 149)		Clinical variables (n = 149)	n (%)
Training set (n = 129)	n (%)	Age		Extent of resection
Ciudad Real General Hospital (Ciudad Real, Spain)	5 (3.9%)	≤65	77 (51.7%)	Biopsy	20 (13.4%)
12 de Octubre Hospital (Madrid, Spain)	3 (2.3%)	>65	72 (48.3%)	Sub-total	33 (22.2%)
Hospital Regional Universitario de Málaga (Málaga, Spain)	16 (12.4%)	Sex		Gross total	76 (51.0%)
Hospital Marqués de Valdecilla (Santander, Spain)	25 (19.4%)	Male	56 (37.6%)	Not available	20 (13.4%)
Princess Margaret Cancer Center (Toronto, Canada)	14 (10.8%)	Female	93 (62.4%)
Toronto Western Hospital (Toronto, Canada)	15 (11.6%)	KPS		MGMT status
Penn State Cancer Institute (Pennsylvania, USA)	51 (39.6%)	≤70	13 (8.7%)	Methylated (positive)	42 (28.2%)
Independent validation set (n = 20)		>70	99 (66.5%)	Non-methylated (negative)	39 (26.2%)
12 de Octubre Hospital (Madrid, Spain)	3 (15%)	Not available	37 (24.8%)	Not available	68 (45.6%)
Penn State Cancer Institute (Pennsylvania, USA)	11 (55%)
Toronto Western Hospital (Toronto, Canada)	6 (30%)

Study overview and analysis plan
Analysis	Cohort/set	Sample size (n)	Data required	Primary purpose
ICAM1 IHC survival analysis	Training and independent validation set	149	FFPE tissue RNA-seq survival analysis	Define ICAM1 prognostic cutoff
Epigenetic regulation	TCGA GB cohort	48	DNA methylation RNA-seq survival analysis	Link ICAM1 promoter methylation with expression and survival
Tumor microenvironment profiling	TCGA GB cohort	142	Bulk RNA-seq	Define ICAM1-associated immune landscape
Single-cell characterization	Subset of training set	11	Fresh/frozen nuclei snRNA-seq	Characterize ICAM1-associated cellular states and immune programs
Radiomic biomarker analysis	Training set	129	Preop MRI survival analysis	Identify complementary imaging biomarkers
CIM model development	Training set	129	Age ICAM1 IHC MRI	Integrate clinical-imaging-molecular (CIM) data to develop a unified model
CIM model validation	Independent validation set	20	Age ICAM1 IHC MRI	Validate the CIM model

Clinical variables collected for each patient included age, biological sex, MGMT methylation status (where available), KPS (where available), and extent of resection (EoR), determined by volumetric analysis of contrast-enhanced (CE) T1-weighted MRI acquired within 48 h of surgery. MGMT status and KPS were not uniformly available across the cohort due to variability in clinical testing and reporting practices across participating institutions at the time of sample collection. Therefore, MGMT status and KPS were included only in analyses restricted to patients with available data and were not ultimately included in the CIM model. OS was defined as the time interval from the preoperative MRI to the date of death or the last known follow-up for censored cases (Table 1). Institutional Review Board approvals were obtained at all participating centers in accordance with local regulations and ethical guidelines.

TCGA Data Analysis

We downloaded publicly available GB RNA-sequencing (RNA-seq) data from The Cancer Genome Atlas (TCGA),²⁴ accessed via the TCGA Data Portal (https://www.cancer.gov/about-nci/organization/ccg/research/structural-genomics/tcga).²⁴ Deidentified data were accessed under TCGA open-use provisions. A cohort of 142 primary, treatment-naïve IDH-wt GB samples with RNA-seq data was curated (Table 1). These datasets were analyzed to assess ICAM1 transcription and promoter methylation and to examine their association with transcriptional subtypes and OS outcomes. For integrative molecular correlation analyses, we generated a refined dataset in which ICAM1 gene expression (log₂-normalized counts) and DNA methylation β-values were matched by patient identifier, retaining only complete cases with paired data.

For the TCGA RNA-seq cohort, a multivariable Cox proportional hazards regression was additionally performed across the entire IDH-wt GB cohort to assess the independent association between ICAM1 transcript expression and overall survival. ICAM1 expression was modeled as a continuous variable, standardized per 1 standard deviation increase, and adjusted for age at diagnosis, biological sex, KPS, and MGMT methylation status.

Analysis of Cell-Type Proportions Using Bulk RNA-Seq Data

Using CIBERSORTx (https://cibersortx.stanford.edu/), TME composition was estimated from bulk RNA-seq data to infer immune and stromal populations associated with ICAM1 expression and across comparator groups. Bulk GB RNA-seq data from TCGA was analyzed using the LM22 signature matrix with 1,000 permutations and quantile normalization disabled. The resulting cell fractions were used for downstream correlation and OS analyses. In parallel, overall stromal and immune content was estimated using the ESTIMATE algorithm (v1.0.13; https://bioinformatics.mdanderson.org/estimate/), which calculates stromal and immune scores from gene expression signatures to infer tumor purity and abundance of nonmalignant cells. Spearman’s correlation was used to evaluate the relationship between ICAM1 expression and ESTIMATE scores, as described.^25,26

Differential Gene Expression and Pathway Enrichment Analysis

Differential gene expression (DGE) analysis was performed on stratified GB subgroups using normalized TCGA RNA-seq counts, as described.^27-29 Genes with |log₂fold change| > 1 and an adjusted P-value <.05 were considered statistically significantly differentially expressed. We used FindMarkers() with Wilcoxon rank-sum testing (log₂ fold change ≥0.25; minimum expression fraction ≥10%) within each cell type group and among ICAM1-high and ICAM1-low tumors. Highly regulated genes (|log₂ FC| ≥ 1, adjusted P ≤ .05) were subjected to pathway enrichment analysis across Gene Ontology (GO) biological processes (https://www.geneontology/org/), KEGG (https://www.genome.jp/kegg/), and Reactome (https://reactome.org/) databases using clusterProfiler and ReactomePA frameworks.

DNA Methylation Analysis

We downloaded publicly available bulk GB DNA methylation data from TCGA, generated using the Illumina Infinium HumanMethylation450k (450K) BeadChip (Illumina, San Diego, CA, USA). Raw data were processed with the Bioconductor minfi package³⁰ using ssNoob normalization after quality control and probe failure removal. Methylation data were filtered to remove sex-related probes on the X and Y chromosomes, cross-reactive probes,³¹ and probes overlapping known single-nucleotide polymorphisms (SNPs) using the dropLociWithSnps function in minfi. The ICAM1 gene on chromosome 19 was annotated by mapping DNA methylation probes using the Illumina HumanMethylation450kannoilmn12.hg19 database. Promoter and gene body regions were defined according to UCSC_RefGene_Group annotations.

Correlation of ICAM1 Transcription and Promoter Methylation

Cytosine-phosphate-guanine (CpG) probes mapping to the ICAM1 locus on chromosome 19 were annotated using the IlluminaHumanMethylation450kanno.ilmn12.hg19 database. Matched RNA-seq and 450K methylation data from 48 GB samples were used to compute Spearman’s correlation between CpG site β-values in the promoter or gene body regions and ICAM1 mRNA expression counts. CpG sites showing statistically significant correlations (P < .05) were retained for survival analysis. For each patient, a composite methylation score was calculated as the mean β-value across the 4 CpG sites most strongly correlated with ICAM1 expression. Patients were stratified into high- and low-methylation groups based on the median of these mean scores; these were subsequently used in Kaplan-Meier survival analyses, as described.^32,33

Single-Nucleus RNA-Seq Data Analysis

Fresh frozen tumor specimens (n = 11; Table 1) were mechanically dissociated using a Dounce homogenizer (sizes A and B, Sigma Aldrich) in ice-cold lysis buffer (0.32 M sucrose, 5 mM CaCl₂, 3 mM MgAc₂, 0.1 mM EDTA, 20 mM Tris-HCl, 40 U/mL RNase inhibitor, 0.1% Triton X-100 in DEPC-treated water) and centrifuged at 800 × g for 10 min at 4°C. Pelleted nuclei were resuspended in wash buffer (1× PBS, 1% BSA, 0.2 U/µL RNase inhibitor), filtered through 40 µm strainers (Flowmi, Sigma Aldrich), and sorted by FACS (BD Influx BRV, Becton Dickinson) with DAPI labeling to exclude debris and aggregates. Sorted nuclei were washed, counted, and loaded onto a 10× Chromium Controller using the Single Cell 3′ Library and Gel Bead Kit (v3; 10× Genomics), targeting approximately 6,000 nuclei per sample. Libraries were sequenced on an Illumina NovaSeq according to the manufacturer’s protocol, and at the Princess Margaret Genomic Center (Toronto, Ontario), achieving >50,000 reads per nucleus.

Sequencing data were processed with CellRanger (v7.0.0; 10× Genomics) for demultiplexing, alignment to the GRCh38 human reference genome, and UMI quantification. Downstream analysis was conducted in Seurat (v5.3.1). Cells with fewer than 200 detected genes, UMI counts outside 3 median absolute deviations from the median, gene counts exceeding 3 median absolute deviations above the median, or >20% mitochondrial transcripts were excluded. Quality control filtering removed 9,077 cells, leaving 120,651 high-quality cells for analysis. Transcriptome data from all samples were merged, normalized, and variance-stabilized using SCTransform, followed by principal component analysis on highly variable features. Batch correction and integration were performed using Harmony³⁴ (v1.2.3). Cells were clustered using shared nearest graph construction (FindNeighbors, FindClusters), and visualized with Uniform Manifold Approximation and Projection (UMAP; RunUMAP). Doublets were identified and removed using DoubletFinder (v2.0.4).

Malignancy Calling and Cell Type Assignment

Following preprocessing, samples were analyzed with inferCNV (Trinity CTAT Project; https://github.com/broadinstitute/inferCNV) using a normal human brain reference dataset from 10× Genomics (https://www.10xgenomics.com/). Inferred CNV scores were aggregated across samples, mean-centered at zero, and scaled by 0.10. Average CNV scores were then calculated for each chromosome arm in each cell.

Malignant cells were identified using 2 CNV-based metrics: (i) TotalCNV, defined as the absolute sum of all CNV scores per cell, and (ii) Chr7 + 10, defined as the absolute sum of positive CNV scores on chromosome 7 and negative scores on chromosome 10. Kernel density estimation was used to model score distributions. The Chr7 + 10 threshold was defined at the local minimum between bimodal peaks, and the TotalCNV cutoff at the seventh inflection point of its density curve, guided by distributions observed in the normal brain reference. Cells exceeding either threshold were classified as malignant; remaining cells were considered nonmalignant.

Nonmalignant cells were annotated using Azimuth (v0.5.0) with the built-in “humancortexref” reference. Cluster-level proportions were evaluated to identify dominant cell identities; discordantly labeled cells, including nonmalignant cells within malignant-dominant clusters or vice versa, were excluded. To harmonize annotations with established GB cell states, we applied FindTransferAnchors() using the GBMap scRNA-seq data as a reference.³⁵ Nonmalignant cells were subsetted prior to anchor identification and label transferred using shared integration features. Cells with a maximum prediction score ≥ 0.7 were retained as confidently annotated, while those below this threshold were classified as Uncertain_nonmalignant and excluded from downstream analysis.

Identification of Malignant and Immune Cell States

Malignant cells were scored for Nomura et al³⁶ cell-state signatures using AddModuleScore(). Following the classification framework described by Neftel et al,⁴ each cell was assigned to a discrete cellular state based on its highest-scoring transcriptional program. This same method was employed on CD4/CD8 T cells and tumor-associated macrophage-microglia (TAM-MGs) and TAM blood-derived monocytes (TAM-BDMs) for the scoring and classification of T cell and myeloid metaprograms (MPs), respectively, identified by Miller et al.³⁷

Immunohistochemical Staining and Digital Quantification

Formalin-fixed paraffin-embedded (FFPE) sections (5 µm thick) were rehydrated, subjected to heat-mediated antigen retrieval in sodium citrate dihydrate buffer, and blocked with 3% hydrogen peroxide in methanol, followed by 5% bovine serum albumin in PBS with 0.1% Triton X-100. Slides were incubated overnight at 4°C with anti-ICAM1 primary antibody (Abcam, ab53013; 1:100), followed by a secondary rabbit antibody (Agilent, K4003) for 1 h at room temperature. DAKO polymer-HRP system and DAB peroxidase kit were used for detection, and sections were counterstained with hematoxylin. Whole-slide images were digitally scanned and analyzed using HALO Image Analysis Platform (v3.0311, Indica Labs). Scoring was performed independently prior to outcome linkage. All reviewers (OS, SK, AG) were blinded to clinical outcomes, survival status, and ICAM1 classification during scoring. Membrane and cytoplasmic ICAM1 positivity were quantified as the proportion of positive cells relative to total detected cells using the Multiplex IHC (v2.0.1) algorithm (threshold = 0.185). The same analysis pipeline was applied to all annotated tissue sections uniformly across the training and independent validation sets (n = 149).

Imaging Data Analysis

Pretreatment CE T1-weighted MRI scans meeting standardized imaging criteria were processed by 4 image analysis experts (JPB, DMG, SG, BOT). MRI-derived morphologic features and extent of resection were independently assessed by 2 trained reviewers (SG, BOT), blinded to patient outcomes and ICAM1 expression status. All measurements were performed according to a standardized imaging protocol, and discrepancies were resolved by consensus adjudication. All imaging-derived measurements were performed independently prior to outcome linkage. Tumors were segmented using open-source platforms, including 3D Slicer (https://www.slicer.org/), ITK-SNAP (v2.2.0; www.itksnap.org), and FreeCAD (https://www.freecad.org/), which enabled surface-based analyses not measurable by ITK-SNAP alone. After DICOM import, CE regions were segmented using a threshold-based approach and manually refined slice by slice to ensure anatomical accuracy, consistent with prior literature.^38-41 Necrotic regions were defined as hypointense areas within CE tumor.

Morphological Measurements

Three-dimensional (3D) quantitative morphologic features were computed from segmented tumors, including:

Tumor Volume (V)

Estimated from total volume of CE (V_CE) and necrotic (V_I) components

$$(V\mathrm{ }=\mathrm{ }\mathit{VCE}\mathrm{ }+\mathrm{ }VI)$$

Contrast-Enhancing Rim Width (δs)

Estimated from CE and V_I as: $\delta \mathrm{s}\mathrm{ }=\mathrm{ }0.62[\sqrt[3]{(V_{CE}}+V_1)-\sqrt[3]{\left(V_1\right)]}$, approximating average thickness of enhancing rim under a spherical assumption.

Surface Area

Extracted from voxel-based segmentations in 3D Slicer or, for ITK-SNAP segmentations, computed in FreeCAD after reconstruction of a 3D surface mesh. Concordance between methods was confirmed by cross-comparing surface measurements from the same tumors.

SIR

Measured as deviation from sphericity and derived from surface regularity, comparing tumor volume to that of a sphere with equivalent surface area. Percentage values range from 0 (perfect sphere) to 100 (highly irregular surface), with higher scores reflecting greater morphological complexity.

$$\mathit{SIR}=\mathrm{ }\left(1-6\sqrt{\pi \mathrm{ }}\frac{\mathit{Total}\mathrm{ }\mathit{Volume}}{\sqrt{{\left(\mathit{Total}\mathrm{ }\mathit{surface}\right)}^3}}\right).\mathrm{ }100$$

Maximum 3D Diameter (D_max 3D)

Estimated as the greatest Euclidean distance between any 2 points on the tumor surface, providing a robust, orientation-independent estimate of tumor size.

Derivation of CIM Model Coefficients

The CIM coefficients were derived using multivariable Cox proportional hazards regression, implemented in Python (version 3.13.5) using the scikit-survival package (CoxPHSurvivalAnalysis class, version 0.23.1). The outcome was time to death, modeled as a right-censored survival endpoint (death status as the event indicator, survival time as the time-to-event variable). Regression coefficients (log hazard ratios) for each covariate were extracted directly from the fitted model and used as the weighting coefficients in the CIM scoring formula. Model discrimination was assessed using Harrell’s concordance index (C-index), computed on the training data.

To identify the optimal CIM score threshold for OS stratification, a sliding-window approach was applied across the observed range of CIM values. At each candidate threshold, patients were dichotomized into high- and low-CIM groups, and OS was compared using Kaplan-Meier analysis with the log-rank test, generating a P-value for that threshold. In parallel, the magnitude of separation between the resulting KM curves was quantified as the difference in median OS (in months) between the 2 groups. The optimal threshold (CIM = 217) was selected as the value that jointly minimized the log-rank P-value (P < .05) and maximized the survival separation between groups, rather than selecting strictly for the lowest P-value in isolation; this criterion favored a threshold with both robust statistical significance and clinically meaningful effect size, avoiding cutoffs driven by a single, isolated minimum in the P-value curve.

Statistical Analysis

Statistical analyses were performed using SPSS (v22.0). OS was estimated using Kaplan-Meier analysis, and significance was assessed using the log-rank test for non-crossing curves. A 2-tailed P-value <0.05 was considered statistically significant. Predictive performance was quantified using Harrell’s concordance index (c-index), applying the same threshold optimization strategy as described before. One-year survival prediction was evaluated using receiver operating characteristic (ROC) analysis, with area under the curve (AUC) summarizing model discrimination. Associations between variables were assessed using Spearman correlation (ρ > 0.7 considered strong). Normality was evaluated using the Kolmogorov-Smirnov test.

Results

ICAM1 Is a Prognostic Marker in GB

We first assessed ICAM1 transcription in GB using publicly available TCGA24 data. ICAM1 expression was significantly elevated in tumors relative to normal brain tissue (NBT; P < .001; Figure 1A) and markedly higher in the MES transcriptional subtype compared with both the CL and PN subtypes⁴² (P < 1.6 × 10⁻⁷; Figure 1B). Additionally, OS was statistically significantly lower in ICAM1-high relative to ICAM1-low MES tumors (median OS difference = 5.7 months; log-rank P = .021; Figure 1E).

Figure 1.

ICAM1 expression and promoter methylation in GB and their association with OS. (A) ICAM1 mRNA expression is statistically significantly elevated in GB tumors compared to NBT (P < .01). (B) ICAM1 mRNA expression varies significantly across GB transcriptional subtypes, with the highest expression in MES tumors compared to CL and PN subtypes. (ANOVA P < 1.6 × 10⁷). (C) High- and low-ICAM1 levels were determined through IHC staining (top); corresponding digital quantification (bottom) demonstrates heterogeneous expression across GB tumor samples. (D) ICAM1 promoter methylation levels differ by GB subtype, with statistically significantly lower methylation in MES tumors and higher methylation in both CL and PN subtypes. (E-G) Kaplan-Meier curves derived from TCGA data with 95% confidence intervals shaded. (E) MES GB patients with high ICAM1 mRNA expression (median difference = 5.7 months; P = .021; n = 24 for both high and low cohorts), (F) ICAM1 promoter methylation levels (median difference = 6.8 months; P = .0015, n = 25 for both high and low cohorts), (G) Patients with ICAM1-high (% positive cells > 27%) protein expression levels exhibit statistically significantly decreased OS (median difference = 4.7 months, P = .0084; ICAM1-high, n = 88; ICAM1-low, n = 61). *P < .01, **P < .001, ***<.0001.

To further evaluate the prognostic relevance of ICAM1 transcription across the entire TCGA GB IDH-wt cohort, we performed a multivariable Cox proportional hazards regression analysis. ICAM1 expression was modeled as a continuous variable (per 1 standard deviation increase) and adjusted for age at diagnosis, biological sex, KPS, and MGMT methylation status. Higher ICAM1 expression remained independently associated with worse overall survival after adjustment for these established clinical prognostic factors (HR, 1.48; 95% CI, 1.06-2.07; P = .023; Table 2), supporting its prognostic relevance beyond univariate Kaplan-Meier analysis.

Table 2.

Cox regression analysis of overall survival

Unmatched TCGA GBM cohort
	Variable	HR (95% CI)	P-value
Multivariable
	MGMT methylated vs unmethylated	0.75 (0.39-1.43)	.380
	KPS, per point	0.99 (0.96-1.01)	.310
	Male vs female	0.84 (0.45-1.58)	.588
	Age, per years	1.07 (1.04-1.11)	<.001
	ICAM1 expression, per 1 SD increase	1.48 (1.06-2.07)	.023

ICAM1 was scaled; HR represents hazard ratio per 1 standard deviation increase. Multivariable model adjusted for all listed variables.

We next examined ICAM1 protein expression by IHC across 3 different cohorts in the training set (n = 129; Table 1). Staining revealed membrane- and cytoplasmic ICAM1 expression throughout the tissue (Figure 1C; Supplementary Figure S1D), which was quantified using digital image-based scoring and validated by centralized pathology review (AG, SK). This assay represented a tissue-level composite signal and did not distinguish ICAM1 expression across malignant, endothelial, immune, or stromal compartments. Using a sliding-window approach, an ICAM1-positive (ICAM1⁺) threshold of 27% was identified as prognostic across our 3 patient cohorts (see Methods; Supplementary Figure S1A-C), with high expression (>27%) associated with statistically significantly reduced OS (median difference = 4.7 months; Log-rank P = .0084; Figure 1G). Our results established ICAM1 protein expression as a robust prognostic marker of GB.

ICAM1 Promoter Methylation Inversely Correlates With Transcription and Is Associated With Clinical Outcome

Using matched bulk GB transcriptomic and DNA methylation data (n = 48) from TCGA,²⁴ we assessed the relationship between ICAM1 mRNA expression and locus-specific DNA methylation. Analysis of 15 CpG loci spanning the ICAM1 promoter and gene body identified 6 promoter-associated loci (cg04754916, cg12151124, cg22874046, cg26549174, cg05106269, cg13720472) with significant inverse correlations with ICAM1 mRNA expression (Pearson r = −0.28 to −0.56; all P < 0.05; Supplementary Table S1). Among these, 4 CpG sites (cg04754916, cg12151124, cg22874046, cg26549174) exhibited the strongest negative correlations (r < −0.3), suggesting that ICAM1 promoter methylation likely contributes to transcriptional regulation of ICAM1. Stratification by transcriptional subtype further revealed that MES tumors displayed significantly lower DNA methylome scores in the ICAM1 promoter than CL and PN subtypes (P < .01; Figure 1D), consistent with their elevated ICAM1 expression. To assess the clinical relevance of these findings, we generated a composite methylome score based on the mean β-values of the 4 most informative CpG sites. Patients were then stratified into high- and low-methylation groups using the median score. High ICAM1 promoter methylation was associated with significantly improved OS, supporting a functional link between ICAM1 methylation, gene expression, and clinical outcome.

ICAM1 Expression Reflects an Inflammatory TME

To characterize the TME profile associated with ICAM1 expression, we analyzed bulk RNA-seq data from TCGA treatment-naïve GBs (n = 142). Tumors were stratified into ICAM1-high and ICAM1-low groups based on median ICAM1 transcript level, and differential expression analysis was performed to identify genes significantly up- or downregulated in ICAM1-high tumors. Among the top upregulated genes were mediators of myeloid recruitment and inflammation, including CCL7, CCL13, and ORM1 (log₂FC > 1, adjusted P < .05; Figure 2A). GO term analysis confirmed enrichment of immune-related processes, including leukocyte migration, chemotaxis, response to lipopolysaccharide, and positive regulation of inflammatory responses (Figure 2B), indicating that ICAM1-high tumors exhibit a proinflammatory, immune-enriched microenvironment. In contrast, ICAM1-high tumors showed downregulation of genes associated with neuronal differentiation and oxidative stress-related pathways,^42,43 including LHX5, NOX3, and C1orf105.

Figure 2.

ICAM1 expression is associated with stromal remodeling and leukocyte enrichment in GB. (A) Volcano plot showing differentially expressed genes between TCGA GB samples with high and low ICAM1 expression. Genes statistically significantly upregulated in ICAM1-high tumors (log₂FC > 1, adjusted P < .05) are highlighted in red. (B) GO term analysis of upregulated genes in ICAM1-high tumors. Dot size indicates the number of genes associated with each pathway, and color represents adjusted P-value significance from Fisher’s exact test. (C) Heatmap depicting immune cell type proportions estimated using CIBERSORTx in ICAM1-high and ICAM1-low tumors. Enrichment scores are color-scaled from low (lighter) to high (darker).*P < .01, **P < .001, ***<0.0001.

To further characterize immune composition, we deconvoluted bulk RNA-seq data from TCGA (n = 142). ICAM1-high tumors exhibited statistically significantly higher proportions of plasma cells, monocytes, macrophages, regulatory T cells (Tregs), and neutrophils compared with ICAM1-low tumors (P < .05; Figure 2C). Consistent with this, ICAM1 expression was strongly correlated with stromal abundance as determined by the ESTIMATE26 algorithm (Spearman ρ  =  0.57; P < .0001), supporting enrichment of both immune and stromal compartments. These results indicate that ICAM1 expression marks a reactive, inflammatory, and myeloid-enriched TME.

Altered Cellular Composition and Malignant Cell States Associated With ICAM1 Expression

To further characterize ICAM1 expression at the single-cell level, we analyzed snRNA-seq data from a subset of fresh frozen treatment-naïve GB samples within the training set (n = 11; Table 1), stratified into 6 ICAM1-high and 5 ICAM1-low tumors using the IHC-defined ICAM1-positivity cutoff of 27% (Figure 3A). We identified 51,469 malignant cells based on inferred copy-number alterations and an accompanying nonmalignant population, using reference mapping to the GBMap scRNA-seq dataset (Figure 3B). ICAM1-high tumors harbored significantly greater proportions of endothelial cells, TAM-MGs, TAM-BDMs, and oligodendrocytes, and lower proportions of malignant cells, neurons, and CD4/CD8 T cells (Figure 3C; Supplementary Table S2). Recent landmark^4,36 single-cell transcriptomic studies have delineated transcriptionally distinct malignant cellular states across GBs, including mesenchymal-like (MES-like), neural progenitor-like (NPC-like), neuronal-like (NEU-like), astrocyte-like (AC-like), oligodendrocyte progenitor-like (OPC-like), and glial progenitor cell-like (GPC-like) states.^4,36 Applying these programs to ICAM1-high and -low tumors, we found that ICAM1-high tumors were enriched for MES-like, NPC-like, NEU-like, hypoxia, and stress-associated programs. By contrast, ICAM1-low tumors were enriched for AC-like, OPC-like, and GPC-like states, consistent with more lineage-associated and less invasive phenotypes (Figure 3F). Our findings reflect an association of ICAM1 with aggressive, invasive, and stress-adapted malignant states.^21,43,44

Figure 3.

SnRNA-seq analysis reveals altered cellular composition across ICAM1-high and -low tumors. (A) UMAP of all GB cells colored by patient ID. (B) UMAP of all GB cells colored by cell type. (C) Pie chart depicting the proportion of cells in ICAM1-high or -low tumors. (D) UMAP of all GB cells colored by ICAM1 tumor type, where red is ICAM1-high and blue is ICAM1-low. (E) UMAP of all GB cells colored by ICAM1 expression. (F) Stacked barplot depicting the proportion of cells within a particular cellular state from either ICAM1-high or ICAM1-low tumors. (G) Heatmap depicting the proportions of TAM-BDM or TAM-MG cells within a particular myeloid cell state from either ICAM1-high or ICAM1-low tumors, respectively. (H) Heatmap depicting the proportion of CD4/CD8 T cells within T cell states from either ICAM1-high or ICAM1-low tumors.

Enrichment of Immunosuppressive Myeloid Programs in ICAM1-High Tumors

Recent landmark studies have described novel gene expression programs that capture both cell identity and functional states in GB immune cell populations.³⁷ We sought to determine how these programs were organized within the myeloid compartment and differed between ICAM1-high and low tumors at the single-cell level. TAM-BDMs in ICAM1-high tumors were enriched for scavenger-associated, immunosuppressive programs, whereas TAM-microglia were enriched for complement- and inflammation-related programs (Figure 3G; Supplementary Table S3). CD4 (helper) and CD8 (cytotoxic) T cells in ICAM1-high tumors showed statistically significantly higher proportions of effector programs (adjusted P = .01) and reduced enrichment of the naïve memory signature (adjusted P = .007; Figure 3H; Supplementary Table S4). Consistent with the chronic immune activation phenotype described by Miller et al,³⁷ our findings position high ICAM1 expression as a marker of an immune-remodeled TME enriched for immunosuppressive myeloid programs and shifts in T cell program composition.

To further identify coordinated biological processes associated with ICAM1 expression across malignant and myeloid cell types, we assessed enrichment of curated signaling and functional pathways using transcriptomic data, as described.⁴⁵ Malignant cells in ICAM1-high tumors were enriched for translational and biosynthetic processes, including peptide chain elongation, translation termination, and selenocysteine synthesis. By contrast, malignant cells in ICAM1-low tumors showed prominent upregulation of pathways related to synapse formation and synaptic transmission (P-value cutoff <.05; Supplementary Figure S2A-E). Similar pathways were observed in oligodendrocytes and CD4/CD8 T cells in ICAM1-high tumors, whereas these cells displayed upregulation of pathways linked to RHO-GTPase signaling, L1CAM interactions, and Netrin-1 signaling in ICAM1-low tumors. Within the myeloid compartment, TAM-MG in ICAM1-high tumors showed upregulation of IL-10 and IFN-γ signaling pathways, along with arachidonic acid metabolism, indicating a functionally polarized state characterized by simultaneous inflammatory signaling and immunoregulatory activity (Supplementary Figure S2A-E). These patterns suggest a shift of both malignant and nonmalignant cells toward a more metabolically active and inflammatory state in ICAM1-high tumors, and a concomitant loss of specialized or homeostatic functions.

Radiomic and Clinical Features Complement ICAM1 in a Prognostic Model

To assess whether the molecular and cellular features of GB associated with ICAM1 are reflected by tumor morphology as a noninvasive readout, we analyzed a comprehensive set of MRI-derived morphologic and morphometric features. These features captured complementary aspects of tumor size, geometry, and boundary complexity, including tumor volume, contrast-enhancing rim width, surface area, SIR, and maximum 3D diameter, serving as noninvasive prognostic markers. Manual annotation of tumor contours by 2 independent operators (SG, BOT) demonstrated inter-operator reproducibility for volume and surface area measurements (Pearson’s r = 0.99; Supplementary Figure S3A and B). Several size-related features, including tumor volume and surface area, were statistically significantly associated with OS, consistent with prior literature^46-49 (Supplementary Figure S3C and D), though their high collinearity limited prognostic utility. Contrast-enhancing rim width similarly reflected an estimate of tumor architecture but did not outperform size-based metrics. Consistent with prior work by our group and others,^46-49 tumor surface irregularity (SIR) emerged as an independent prognostic marker of OS (P < .001; c-index = 0.640)), capturing 3D tumor complexity and infiltrative growth patterns not fully reflected by bulk size measurements alone. Likewise, among routinely available clinical variables, age (P < .001; c-index = 0.617) was statistically significantly associated with OS in our cohort (Supplementary Table S5).

Motivated by these findings, we leveraged the aforementioned individually prognostic markers, including patient age, preoperative SIR, and ICAM1 expression, to develop a unified Clinical-Imaging-Molecular (CIM) model (CIM = 1.87 × age + 1.44 × % ICAM1⁺ + 2.09 × SIR). Each ­parameter was independently prognostic yet minimally correlated with the others (Figure 4A-D), highlighting the complementary information captured by integrating clinical, imaging, and molecular features. In our cohort, these variables were the strongest univariate predictors of OS (all P < .001; c-index 0.617-0.640), outperforming individual well-established clinical factors, including MGMT status, KPS, and EoR (Supplementary Table S5), although relative prognostic performance may differ across larger independent cohorts. The CIM score robustly stratified OS across the combined cohort, with an optimal cutoff of 217 (HR = 9.72; log-rank P < .001; c-index = 0.655, Supplementary Table S6). Receiver operating curve (ROC) analysis further demonstrated strong discrimination for 1-year survival (AUC = 0.779; sensitivity = 0.688; specificity = 0.807).

Figure 4.

Development, validation, and sample case visualization support the clinical utility of the CIM model. (A-C) Kaplan-Meier OS analyses demonstrate reduced OS associated with (A) high age ≥ 65 years (median difference = 9.5 months; P = .0002; age high n = 65, age low n = 64), (B) high SIR level > 22 (median difference = 6.8 months, P = .00067, SIR high n = 64, SIR low n = 65), and (C) high CIM score > 217 (median difference = 10.5 months, P < .001, CIM high n = 42, CIM low n = 87). (D) Spearman correlation matrix of model input features (age, ICAM1, SIR, area, volume) reveals minimal collinearity among age, ICAM1, and SIR—supporting their independent prognostic value and verified by (E) ROC analysis confirms strong 1-year survival prediction using the CIM model (AUC = 0.833, sensitivity = 0.917, specificity = 0.786). (F) The sliding threshold analysis using P-value that identifies CIM = 217 as the optimal prognostic cutoff. (G) Four representative patient cases from the validation cohort illustrate the range of CIM scores and the contributing variables: age, ICAM1 protein expression, and SIR derived from pre-operative MRI. Case examples highlight how non-redundant inputs contribute to individualized risk predictions. (H) Corresponding MRI scan overlays depicting segmented tumor regions (in red).

To assess whether postoperative variables further refined outcome prediction, we incorporated EoR to generate a 4-variable CIME model (CIME = 2.44 × age + 0.54 × %ICAM1⁺ + 1.99 × SIR—54.75 × EoR). Inclusion of EoR improved model performance, yielding a 14-month difference in median OS (HR = 3.95; P < .001; c-index = 0.686; AUC = 0.792; Supplementary Figure S4A-C). However, EoR is a postoperative variable influenced by surgical judgment, operative technique, and variability and availability in postoperative imaging across institutions.⁵⁰ We therefore retained the CIM model as the more readily implementable and resource-efficient framework, preserving broad clinical utility.

Cross-Cohort Validation and Independent Testing of the CIM Model

We next evaluated the robustness of the CIM model across geographically distinct institutional subcohorts within the training set (n = 129). Cross-cohort analyses demonstrated consistent outcome prediction, with median OS differences of 15 months in the Spain cohort (P < .001), 4.2 months in the USA cohort (P = .029), and 8.9 months in the Canada cohort (not statistically significant, likely reflecting smaller sample size; Supplementary Figure S4D-F). To further assess external performance, we applied the predefined CIM model, including fixed coefficients and the prespecified cutoff (217), to an independent validation set (n = 20) that was not used during model development or threshold optimization. The model stratified patients into high- and low-risk groups, yielding a 10.5-month difference in median OS (log-rank P < .001; Figure 4C). ROC analysis demonstrated strong discrimination for 1-year survival (AUC = 0.833; sensitivity = 0.917; specificity = 0.786; Figure 4E), consistent with performance observed in our discovery cohort (n = 129). These findings demonstrated that integrating clinical, imaging, and molecular features within the CIM framework provides a robust and scalable approach for prognostic stratification across independent GB test cohorts.

Case-Based Illustration of CIM Model Performance

To illustrate the application of the CIM model at the individual patient level, we present 4 representative cases randomly selected from the training set (n = 129) spanning a range of clinical, imaging, and molecular features (Figure 4G and H). In the example, 2 patients of similar age exhibit divergent outcomes. Patient 3 (63 years) had lower ICAM1 expression (3.4%), lower tumor SIR (19.5%), and a CIM score of 163.6; they survived 581 days. In contrast, Patient 4 (64 years) showed higher ICAM1 expression (53.5%), higher SIR (25.1%), and a CIM score of 249.6; they survived only 493 days (Figure 4G). Consistently, patients with concordant high-risk features clustered within the high-CIM groups and exhibited poorer outcomes, whereas those with lower CIM scores had favorable survival. These representative cases illustrate how the individual components of the CIM model (age, tumor SIR, and ICAM1 expression) contribute complementary, non-overlapping prognostic information for individual patients. By integrating these features, the CIM framework captures tumor biology and intra-tumoral heterogeneity at a more personalized level, enabling more refined prognostication than routinely captured clinical variables alone.

Discussion

In this study, we demonstrate that ICAM1 expression is markedly elevated in GB relative to normal brain tissue, with preferential enrichment in MES tumors, a transcriptional subtype associated with invasiveness, inflammatory activation, and therapeutic resistance.^51-53 Across independent cohorts, ICAM1-high tumors exhibited a statistically significant reduction in OS compared with ICAM1-low tumors at both the transcriptomic and protein levels. At the bulk level, this association remained significant after adjustment for established clinical prognostic factors, supporting the clinical relevance of ICAM1 expression in GB. We further show that hypomethylation of promoter-associated CpG sites correlates with higher ICAM1 expression and poorer OS, positioning ICAM1 promoter methylation as a prognostic biomarker in GB.

Beyond tumor-intrinsic expression, ICAM1 emerges as a characteristic feature of the immune-mesenchymal subtype. High ICAM1 levels are associated with upregulation of myeloid-recruiting chemokines, inflammatory and chemotactic pathways, and increased stromal content, collectively fostering a proinflammatory, immunomodulatory environment. ICAM1 is a broadly inducible adhesion molecule canonically associated with endothelial activation, but is also expressed across multiple cell types, including monocytes and macrophages in inflammatory and tumor-associated contexts.^7, 16 Our single-cell analyses reveal widespread ICAM1 expression across TAMs, oligodendrocytes, T cells, and malignant cells in GB, positioning ICAM1 as a feature of a multicompartmental tumor ecosystem. Malignant cells in ICAM1-high tumors are enriched for MES-like, NPC-like, NEU-like, hypoxia-, and stress-associated states, programs previously linked to increased cellular plasticity and invasiveness. By contrast, ICAM1-low tumors are enriched for AC-, OPC-, and GPC-like states, which more closely resemble defined neural lineage programs and reflect a comparatively differentiated malignant phenotype.^4,36

At the immune level, TAM-BDMs in ICAM1-high tumors exhibit a greater proportion of scavenger-associated immunosuppressive programs than in ICAM1-low tumors. Likewise, TAM-MGs are enriched for complement and systemic inflammatory programs in ICAM1-high tumors. Among CD4/CD8 T cell populations, there is a greater proportion of effector T cell programs and reduced enrichment of naïve cells memory signatures. Subsequent functional characterization of these immune populations using flow cytometry would further validate and extend these findings. Pathway-level analyses further demonstrate coordinated enrichment of translational, biosynthetic, and inflammatory processes across malignant, oligodendrocyte, and T cell populations in ICAM1-high tumors, whereas ICAM1-low tumors retain housekeeping neurodevelopmental signaling pathways. Collectively, these findings support a model in which ICAM1-high tumors mark an immune-suppressed, stress-adapted TME.

Prior studies have implicated ICAM1 in infiltrative tumor growth and structural heterogeneity, which are biological features that may manifest at the macroscopic level on imaging.^46-49 Given the availability of established clinical, imaging, and molecular prognostic markers in GB, we leveraged imaging data from our cohort to identify tumor SIR as a robust independent predictor of outcome. Unlike bulk measures such as volume or surface area, SIR captures 3D structural complexity and infiltrative growth patterns.^54,55 We thereby present a novel CIM model to enable clinically accessible outcome prediction without reliance on resource-intensive sequencing approaches.⁵⁶ Importantly, our objective was not to identify the optimal immunohistochemical biomarker among all possible candidates. Rather, ICAM1 was selected a priori based on its biological relevance to GB inflammation and mesenchymal biology,^14-16 together with its independent prognostic significance across transcriptomic, epigenetic, and protein-level analyses. By translating complex tumor biology into actionable prognostic insight, the CIM framework supports risk stratification and treatment planning, including consideration of preemptive treatment escalation or experimental therapies.

Our study should be interpreted in the context of several limitations. The retrospective design may introduce selection bias, and ICAM1 IHC quantification may be affected by batch effects or antibody variability, although centralized staining and standardized digital pathology workflows mitigate these concerns. Our analysis is also limited to 1 section per tumor, preventing capture of the full extent of intratumoral heterogeneity. In addition, our snRNA-seq analyses are observational and identify cellular states associated with ICAM1 expression rather than establishing causal relationships between ICAM1 and the observed malignant or immune programs. Functional studies will be required to determine whether ICAM1 directly regulates these cellular phenotypes or reflects broader inflammatory processes within the GB microenvironment. SIR estimation can be influenced by image quality and segmentation fidelity, which may vary across clinical settings. Future studies should prioritize prospective validation, standardization of ICAM1 scoring, and automated extraction of imaging features. Limited sample size may also misrepresent the prognostic weight of well-established clinical parameters, including MGMT status, KPS, and sex. Given ICAM1’s role in immune-mesenchymal signaling,¹⁸ its potential to predict treatment response, particularly to immunomodulatory therapies, warrants further investigation. All that said, our work establishes a robust, interpretable, resource-efficient, and scalable framework for refined prognostication in GB, capable of supporting outcome prediction and informing clinical decision-making, particularly across low-resource settings.

Supplementary Material

Supplementary material is available online at Neuro-Oncology Advances (https://academic.oup.com/noa).

Author Contributions

The affiliations reflect current institutional appointments of several authors and do not indicate sample acquisition, data generation, or patient care at Mayo Clinic. All patient samples and datasets were generated prior to author appointments at Mayo Clinic and were collected and analyzed at the University Health Network, Canada, and collaborating institutions in Spain and the United States (Penn State Cancer Institute). Substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data for the work (S.G., B.O.-T., M.M.B., O.S.S., K.D., S.K., Y.M., J.S., I.M., A.F.G., J.P.-B., A.R.-G., K.D.A., A.M., V.M.P.-G., G.Z., S.M.); Drafting the work or reviewing it critically for important intellectual content (S.G., B.O.-T., E.R.I., D.M.-G., Y.M., A.M., V.M.P.-G., G.Z., S.M.); Final approval of the version to be published (S.G., B.O.-T., M.M.B., O.S.S., K.D., E.R.I., D.M.-G., S.K., Y.M., J.S., I.M., A.F.G., J.P.-B., A.R.-G., A.H.-L., B.A., M.P.F.-R., P.S.G., K.D.A., C.S., K.J.D., I.F.P., C.V., A.M., V.M.P.-G., G.Z., S.M.); Agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved (G.Z., S.M.).

Conflict of Interest Statement

None declared.

Funding

S.M. and G.Z. received support from the Terry Fox Research Institute (TFRI-1090-P2).

Data Availability

Data will be made available upon request.