Proteogenomic analysis of pediatric and AYA high-grade glioma reveals age-dependent biology, female-male differences, and kinase targets

Summary

High-grade gliomas (HGGs) in children and adolescents and young adults (AYA) exhibit distinct biology across the neurodevelopmental spectrum. To dissect tumor-intrinsic molecular characteristics independent of developmental variation, we perform comprehensive proteogenomic analyses of tumors from 112 HGG patients aged 0–40 years. Our multi-omics analysis identifies two AYA subgroups—adolescents (aged 15–26 years) and young adults (aged 26–40 years)—with distinct molecular profiles and survival outcomes. Tumor-normal comparisons and survival modeling highlight roles of oxidative phosphorylation and neuronal system biology in glioma progression. Causal network analysis and cell line studies provide a rationale for personalized therapies targeting candidate kinases, such as CDK8. Survival modeling, clustering, and immune-landscape analyses identify proteins, post-translational modifications, and immune signatures linked to outcomes and reveal clinically relevant differences between male and female patients.

Graphical abstract

Highlights

• •Deep multi-omics characterization of 112 pediatric, adolescent, and young adult HGGs
• •Distinct genetic, transcriptomic, and proteomic features across age and sex groups
• •Druggable targets revealed by causal network analysis of kinase activities
• •Proteins and immune signatures linked to poor outcomes

Tignor et al. perform a comprehensive proteogenomic analysis of pediatric and AYA high-grade gliomas and identify age- and sex-dependent biology, candidate kinase targets, and proteins and immune signatures associated with poor survival outcomes.

Introduction

High-grade gliomas (HGGs) are highly aggressive primary brain tumors, with a 5-year survival rate below 10%. HGGs in children differ markedly from adult forms and are now classified separately in the 2021 WHO CNS5 guidelines,¹ based on features such as histone mutation status in children and IDH1/2 mutations in adults. Adolescents and young adults (AYA) with HGG represent a rare, understudied group whose tumors do not align with pediatric or adult disease,^2,3 reflecting transitional neurodevelopmental biology.^3,4,5

While genomic and anatomic classifications of pediatric and AYA HGG have revealed diverse cancer types with different outcomes,⁶ including infantile hemispheric glioma (IHG), diffuse hemispheric glioma (DHG), and diffuse midline glioma (DMG), few targeted therapies have been developed beyond those targeting IDH1-mutant tumors. Furthermore, the relationship between tumor biology and neurodevelopmental factors across age groups and in male and female patients remains poorly understood.

Differences between male and female patients, which influence both brain development and cancer biology, represent another underexplored dimension. A growing body of literature reveals male-/female-biased variation in HGG features, including cell-cycle regulation, tumor suppression, metabolism, and anti-tumor immunity,^7,8 underscoring the need for age-informed therapeutic approaches that account for differences between male and female patients.

Here, we analyzed pediatric and AYA HGG within a developmental framework to identify shared as well as age-specific or male- and female-associated biology, informing precision medicine and complementing established molecular classifications. We performed proteogenomic profiling in a cohort of 112 patients, with comparable numbers of males and females aged 0–40 years, including a discovery set of 87 and an independent validation set of 25. We integrated somatic mutations, DNA methylation, copy number alterations, gene expression, protein abundance, and post-translational modifications (PTMs). This multi-omics perspective enables us to discern biological similarities and differences across high-grade glial tumors without being constrained by legacy diagnostic boundaries, a strategy supported by prior work showing shared biology across histologies.⁹ We also combined these analyses with data from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) adult glioblastoma (GBM) cohort¹⁰ to obtain a comprehensive picture of tumor characteristics throughout the lifespan.

Results

Study patient cohorts and proteogenomic data overview

We assembled a discovery cohort of 87 patients (93 tumor samples) and an independent validation cohort of 25 patients, aged 0–40 years, with grade 3–4 HGG (Figure 1A; STAR Methods). For some patients, multiple tumors were retained to capture intra-patient heterogeneity. Proteogenomic profiling of the discovery cohort (34 AYA and 59 pediatric tumor samples) yielded comprehensive data for somatic mutations, copy number variation (CNV), DNA methylation, gene fusions, RNA expression, global proteomics, phosphorylation, and glycosylation (Figure S1A; Table S1). The validation cohort data included RNA sequencing (RNA-seq) data (n = 17) and global proteomics data (n = 25), after quality control (STAR Methods). To model tumor biology across the lifespan, we integrated the discovery cohort data with data from 99 adult GBM tumors (ages 24–88 years, PDC: PDC000204 and PDC000205)¹⁰ to form a combined discovery cohort (cDiscovery). To support prognostic analyses, we further compiled a reference cohort of 5,289 HGG patients (with mutation data and clinical annotations) from multiple large-scale studies^4,11,12 with survival data available for 1,615 individuals (STAR Methods; Table S1).

Figure 1.

Age-dependent DNA alterations and tumor molecular profiles

(A) Demographic and clinical annotations of the discovery cohort tumors (N = 93).

(B) Distribution of cancer groups stratified by PED and AYA. Cancer groups were aligned to WHO 2021 histologic classifications (STAR Methods). Five tumors labeled as “flag” (gray) were reclassified as other high-grade tumors. These were retained in the analyses as they were initially classified as HGGs by pathologists.

(C) Oncoprint of the frequently mutated genes (>5%) in the discovery cohort tumors.

(D) Mutation frequency profiles across age groups for selected genes in the reference cohort.

(E) Survival association of somatic mutations in different age groups in the reference cohort (STAR Methods; multivariable Cox models).

(F) Multi-omics consensus clustering of tumor temporal molecular profiles (T–TMPs; ages 0–80 years) in the cDiscovery cohort.

(G) Kaplan-Meier overall survival curves across age groups in the reference cohort. p values were calculated using the log-rank test.

Methylation- and genomically informed cancer subtype groups

Within the discovery cohort, we defined molecular subtypes⁶ by integrating histopathologic, epigenetic, and genomic features (STAR Methods; Figure 1B). Five tumors were reclassified as other high-grade tumors based on the WHO 2021 criteria (Table S1) but were retained, as they were historically diagnosed as HGGs.¹³

Subtype distributions differed between the PED and AYA groups. As expected, IHG (median age 2.8 months) occurred exclusively in PED cases, whereas DHG (age range of 15–19 years) was restricted to AYA patients.¹³ DMG, typically arising in individuals <20 years of age, was more prevalent in the PED group.

Age-dependent mutational and CNV landscape of HGG

The most frequently mutated genes in the discovery cohort were TP53, H3-3A, NF1, and ATRX (Figures 1C and S1B). Analysis of the larger reference cohort revealed higher frequencies of TP53, ATRX, and IDH1 mutations in AYA and the enrichment of H3-3A in PED (Figure 1D; STAR Methods), consistent with the known age-dependent patterns.^3,13,14 The distributions of the mutation types H3-3A and TP53, however, did not differ significantly by age (Figure S1C; Table S1).

These mutations also showed age-specific survival associations. In the reference cohort, IDH1 and ATRX mutations were associated with improved survival in AYA patients^3,14 (Figure 1E; STAR Methods), whereas TP53 mutations negatively impacted survival in AYA but not in PED patients.

CNV profiles were heterogeneous across HGG tumors in the discovery cohort (Figure S1D), aligning with prior findings. The most frequent alterations included gains in chromosome 1q and losses in chromosome arm 10q¹⁰ (STAR Methods; Table S1). GISTIC2.0 analysis¹⁵ showed more frequent deletions at 9p21.3, harboring CDKN2A, in AYA versus PED tumors¹⁶ (Figures S1E–S1H; STAR Methods).

Transcriptomic and proteomic profiles defining adolescent versus young adult age groups

We analyzed RNA and protein data from the cDiscovery cohort (ages 0–88 years), using tumor-temporal molecular profiles (T–TMPs), defined by smoothed local regression of the molecular abundance across age groups (STAR Methods).

The T–TMPs of RNAs, proteins, and phosphosites revealed pronounced age-dependent structure spanning PED, AYA, and adult tumors (Figure 1F). Within the AYA group, we identified two distinct molecular states separated by a transition around age 26, using consensus clustering of multi-omics T–TMPs (STAR Methods), defining adolescents (ADO, 15–26 years) and young adults (YA, 26–40 years). Among patients aged >40 years, tumors formed two additional clusters: 40–62 years (adult, ADU) and the >62 years (Senior, SEN).

Survival analysis in the reference cohort revealed significantly different outcomes across the PED, ADO, YA, ADU, and SEN groups, with the YA group showing the most favorable outcomes, and PED and SEN groups the poorest (Figure 1G). These associations remained after adjustment for histological subtype (e.g., DMG and GBM) and survival-associated mutations (H3-3A, IDH1, TP53, ATM, and ATRX) in Cox models (STAR Methods) and were recapitulated in a GBM subset lacking H3-3A mutations (Figure S1I).

Given our focus on brain development, subsequent analysis excluded the SEN group to minimize confounding from aging.

T–TMP-derived protein groups exhibiting different patterns among male and female patients

To reveal age-varying biological processes, we performed clustering and pathway enrichment analysis using T–TMPs.

In each male/female group, protein T–TMPs were categorized into four groups (STAR Methods) based on the temporal abundance patterns: monotonic decrease (T1), monotonic increase (T2), and valley (T3) or peak (T4) around the ADO/YA transition (Figures 2A and S2A). Pathway enrichment of these groups highlighted age-varying functions (Figure 2B). For example, “oxidative phosphorylation” (OXPHOS) and “neuronal system” pathways were enriched in T2 in both male and female groups, reflecting their increasing activity with age.

Figure 2.

TMPs reveal male- and female-associated patterns and tissue-specific differences between tumor and normal tissues

(A) Heatmaps shown separately for males and females, comparing normal–TMP (N–TMP) and tumor–TMP (T–TMP) profiles. Four T–TMP protein groups in males and females are indicated by the left-hand columns. Labeled proteins were selected from those above the 90th percentile for tumor-normal (T–N) difference and further filtered for strong male-female (M–F) differences in age-related tumor profiles.

(B) Pathways significantly enriched (Fisher’s exact test, FDR < 0.1) in each T–TMP protein group in males (left) and females (right). Thin bars indicate the pathways enriched across multiple groups, often in a female- or male-dependent manner.

(C) T–TMPs and N–TMPs for the selected glycoproteins, with shaded areas indicating 90% confidence intervals.

(D) Boxplots of protein abundances across age groups in normal and tumor samples for the selected glycoproteins. Significance: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001 (Wilcoxon rank-sum test).

(E) Heatmaps of the pathways enriched for proteins with large T–N differences (left) and/or male-female (M–F) differences (right), assessed using 545 normal and 9,026 tumor proteins. Dots indicate an FDR < 0.1 (Wilcoxon rank-sum test).

(F and G) TMPs of proteins showing large T–N differences in males and females within the “oxidative phosphorylation” (F) and “neuronal system” (G) pathways, including COX6C, PRKAR2B, and PRKACB (T–N > 90th percentile).

Overall, 27% of proteins showed discordant trajectories in male and female groups and were assigned to different T1–T4 groups between males and females (Figure S2A). Pathway enrichment of discordant groups—for example, proteins classified as T1 in males and T3 in females (denoted as MF-T13 in Figure S2A)—revealed functions differing between males and females (Figure S2B). X-linked proteins were enriched in MF-T41, with expression peaking in mid-development in males but declining with age in females. Glycosylated proteins, which influence the tumor microenvironment, epithelial-to-mesenchymal transition (EMT), and immune response,¹⁷ were enriched in T–TMP groups, with discordant patterns between males and females, including high-mannose glycopeptides in MF-T14 and fucosylated glycopeptides in MF-T24 (Figure S2B).

Tumor-normal differences revealed by TMPs

We compared T–TMPs with normal–TMPs (N–TMPs) derived from ∼500 proteins in normal brain tissues (ages 0–50 years)¹⁸ to distinguish tumor-specific processes from developmental processes. To enable cross-study comparison despite batch effects in mass spectrometry data, analyses were based on age-dependent trajectories rather than absolute abundance (STAR Methods).

In Figure 2A, we contrast T–TMPs and N–TMPs for proteins with marked tumor-normal (T–N) differences. Proteins in T2 (increasing in tumors) often showed flat or decreasing N–TMP trends. Among these, the neurodevelopmental proteins L1CAM, MAPT, and CNTN1 showed pronounced T–N differences in male tumors (Figures 2C, 2D, and S2D). Each has been implicated in the EMT across cancers.^17,19,20

Enrichment analysis of proteins and RNAs exhibiting T–N differences identified pathways with marked protein-specific T–N divergence, including OXPHOS and “neuronal system,” suggesting post-transcriptional and/or post-translational regulation (Figures 2E and S2E; STAR Methods).

OXPHOS protein abundances increased from PED to YA in tumors but not in normal tissues (Figure 2F), suggesting an age-associated shift toward greater reliance on this metabolic program in tumors.^21,22 In PED tumors, proteins upregulated in females relative to males showed significant enrichment for OXPHOS (Figure 2E), suggesting that female-biased OXPHOS regulation may emerge during tumorigenesis.

Similarly, “neuronal system” protein T–TMPs increased from PED to YA, particularly in males, but showed distinct temporal trends in the normal brain (Figures 2E and 2G). This pathway was upregulated in male ADO/YA tumors relative to those of females (Figure S2F), with cell type-specific analysis (STAR Methods) implicating tumor cells as a primary source of this signal (Figure S2G). PRKAR2B and PRKACB—regulatory and catalytic subunits of PKA—showed striking T–N differences in males, along with male-biased expression in tumors (Figures 2G and S2F), supporting differential reliance on cAMP/PKA signaling in male and female patients with glioma.^23,24

Age-related functional impact of frequent somatic mutations

Using multi-omics cis-regulation analysis with iProFun²⁵ (STAR Methods), we identified significant mutation-associated effects on RNA and/or protein levels for TP53, ATRX, and NF1 (Table S3). ATRX and NF1 mutations were linked to reduced RNA and protein levels (Figures 3A and S3A), whereas TP53 mutations increased protein abundance without affecting RNA levels (Figure 3A), consistent with mutation-induced protein stabilization.^10,26,27 These patterns were recapitulated in the HGG-NOS subgroup (Figure S3B), indicating that they were not attributable to heterogeneity across cancer groups.

Figure 3.

Age-related regulation: cis- and trans-effects of DNA alterations on functional molecules and pathways, as well as kinase-TF interactions

Significance levels are based on t tests, unless otherwise noted (∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001).

(A) Boxplots of RNA and protein abundances of TP53 (top) and ATRX (bottom) stratified by mutation status.

(B) Heatmap of trans-regulatory effects of TP53 and ATRX mutations on pathways across age groups (methylation, RNA, and protein). Dots indicate an FDR < 0.1 (linear model).

(C) Boxplots of RNA and protein abundances for three E2F target genes differentially affected by ATRX mutation in PED vs. ADO, measured by TMT (column 2) and validated by MRM/SRM (column 3).

(D) Heatmap of trans-regulatory effects of H3-3A and NF1 mutations on pathways across age groups (methylation, protein, and RNA). Dots indicate an FDR < 0.1 (linear model).

(E) Heatmap of trans-regulatory effects of IDH1 mutations on pathways (methylation, RNA, and protein). Dots indicate an FDR < 0.1 (linear model).

(F) Boxplots of protein abundances of GOLGA2 and GUCY1B1, two AYA-specific CNV-RNA/protein cascade genes, stratified by their CNV status across age groups.

(G) Kinase-TF modules from the top kinase-TF pairs for ADO and YA tumors. Arrowheads indicate upregulation; blunt-ended arrows indicate downregulation (STAR Methods). Each module is labeled according to the most relevant pathway identified by Enrichr. Genes in pathways are highlighted in bold.

Pathway enrichment of genes/proteins trans-regulated by TP53, ATRX, H3-3A, and NF1 mutations (Figure S3C) revealed biological processes impacted by these alterations (Figures 3B and 3D). While TP53-associated trans-regulatory effects were relatively consistent across age groups, the impact of ATRX, H3-3A, and NF1 mutations varied between the PED and ADO/YA tumors (Figures 3B, 3D, and S3D). Notably, ATRX mutations were associated with downregulation of the E2F target pathway in PED tumors but upregulation in ADO tumors (Figure S3D). We validated the age-dependent effects using targeted mutiple/selected reaction monitoring (MRM/SRM), confirming differential regulation of key E2F pathway proteins, including CDKN1B, ILF3, and MCM4 (Figure 3C; STAR Methods).

IDH mutations have been reported to induce genome-wide DNA hypermethylation in gliomas.^28,29 Given the limited number of IDH1-mutant samples in the cDiscovery cohort, age groups were combined to assess cis-/trans-regulatory effects. IDH1 mutation did not show significant cis-effects on IDH1 RNA or protein levels in HGG tumors but were associated with widespread trans-regulatory effects across methylation, RNA, and proteomic layers (Figure S3C). Proteomic analysis indicated that IDH1 mutation was associated with the downregulation of glycosylation pathways and upregulation of metabolic pathways, including OXPHOS (Figure 3E).

In a genome-wide screen for associations between gene-level DNA methylation and IDH1 wild-type (WT) protein abundance, no significant associations were detected after multiple-testing correction, suggesting a limited role of IDH1 WT protein in mediating epigenetic reprogramming.

Age-related cis-regulation of CNVs

iProFun²⁵ analysis also identified genes whose transcript and protein abundance were influenced by their own CNVs, termed CNV-RNA/protein cis-cascade events (Figure S3E; Table S3; STAR Methods). We observed age-specific cis-cascade events. For example, GUCY1B1 (4q32.1) and GOLGA2 (9q34.11) showed a CNV-RNA/protein cis-cascade effect in AYA but not in the PED or ADU groups (Figures 3F and S3F). GUCY1B1, a nitric oxide sensor, has been reported as a prognostic marker for multiple cancer types.³⁰ We further observed that the downregulation of GUCY1B1 protein was associated with poorer survival outcomes in a cross-population survival analysis using the reference cohort (Table S4; STAR Methods).

Age-related kinase and transcription factor regulations

We analyzed kinase activity in each HGG, using KEA3,³¹ which integrates kinase-substrate interactions and kinase-gene co-expression across twenty databases (STAR Methods), and derived the transcription factor (TF) activity scores from RNA-seq, using ChEA3,³² to identify shared or age-specific signaling modules (Figures 3G and S3G; STAR Methods).

Several kinase-TF modules were uniquely enriched in YA tumors compared with ADO tumors (Figure 3G; STAR Methods). In YA tumors, a DNA repair-related TF module (Table S3; STAR Methods) was upregulated by a cell cycle-related kinase module (CDK8, CDK9, and CLK2) and downregulated by a calmodulin-binding kinase module. CDK8, previously shown to promote glioblastoma stemness via cell-cycle response genes,³³ may play a distinct role in YA tumors. In ADO tumors, the cell cycle-related kinase module inhibited a neuronal TF module (Table S3). ZSCAN18, a member of this TF module, was downregulated in glioblastoma cell lines; its overexpression suppressed tumor progression. These findings suggest potentially distinct roles of CDK8, CDK9, and CLK2 in ADO and YA HGGs.

In PED vs. ADO tumors, multiple kinase modules showed age-specific enrichment (Figure S3G). For example, a DNA repair TF module (Table S3) was suppressed by a CDK7/9-containing kinase module in PED tumors. Moreover, a senescence-related TF module (Figure S3G; Table S3) was activated by different kinase modules in ADO and PED tumors.

Male-/female- and age-specific proteomics-based prognostic signatures

Robust prognostic signatures for PED/AYA HGG could significantly improve clinical decision-making. We leveraged proteomics to identify candidate prognostic markers associated with patient overall survival (OS) outcomes.

Validating a prognostic score based on WT IDH1 and IDH2 proteins

Our previous study⁹ established a prognostic score for H3-3A^WT HGG patients (ages 0–40 years) based on the protein abundances of IDH1^WT and IDH2^WT. To assess potential male-female differences, we applied separate multivariable Cox regression in males and females in the cDiscovery cohort (STAR Methods). The association in males is consistent with our earlier findings⁹ (p = 0.026, Figure 4A, left), but no association was observed in females (Figure S4A). We confirmed this male-specific association using MRM/SRM measurements of IDH1^WT and IDH2^WT in a subset of the discovery cohort (Figure 4A; STAR Methods).

Figure 4.

Male-/female-specific prognostic proteins and pathways

(A) OS in male H3 WT patients stratified by high IDH1/2 score vs. low IDH1/2 score from TMT (left) and SRM (right). p values are from multivariable Cox models.

(B) Heatmap of OS-associated glycopeptides; the top-associated glycopeptide per protein is shown. Dots indicate an FDR < 0.10 (multivariable Cox models).

(C) Heatmap displaying the overlap of genes significantly associated with OS (FDR < 0.10, multivariable Cox models) across age groups in males and females from CPSA, for protein and RNA.

(D) Concordant OS associations from CPSA (protein). Left: pathway enrichment; right: proteins significant within each pathway. Black squares indicate concordant associations across the reference (FDR < 0.10) and cDiscovery cohorts (p < 0.01) in multivariable Cox models; black dots indicate significance in the reference cohort only (FDR < 0.10).

(E and F) Coefficients with 90% confidence intervals (CIs) from penalized multivariable Cox models (Ridge; multi-marker + covariates) for selected protein markers in males and females. Markers significant in single-marker multivariable Cox models (FDR < 0.10) are labeled ∗.

(G and H) Hazard ratios with 90% CIs from multivariable Cox models in the cDiscovery (G) and validation (H) cohorts. OS risk scores are derived from (E) and (F).

(I and J) Model-adjusted OS by high PHKG2 levels vs. low PHKG2 levels (median split) in the cDiscovery (I) and validation (J) cohorts. p values are from multivariable Cox models of PHKG2 and clinical variables.

Glycosylation associated with OS outcomes

We identified male-/female- and age-dependent prognostic roles of glycopeptides. Separate Cox regression analyses in males and females revealed 218 OS-associated glycopeptides in males and 4 in females (false discovery rate [FDR] < 0.1; Table S4; STAR Methods), and most remained significant after adjusting for their global protein abundance (Figure 4B). In male ADO patients, more glycopeptides were linked to poorer OS than improved outcomes (p < 0.001), whereas the opposite trend was observed among male PED patients (Figure 4B). Enriched pathways included “nervous system development” and “ECM organization” (STAR Methods), with the latter containing CD44 and ITGAV (Figure 4B), whose protein and glycosylation are implicated in tumor cell adhesion and migration.³⁴

Cross-population survival analysis

To investigate how age-related neurodevelopment interacts with tumor progression, we introduce cross-population survival analysis (CPSA), which uses age-dependent molecular trajectories learned in one cohort to derive age-matched, expected molecular values in another cohort for survival analysis (Figure S4B; STAR Methods). The CPSA framework leverages information across populations to boost the statistical power for identifying robust prognostic features. Applying CPSA to the cDiscovery and reference cohorts, we identified a large set of age- and male-/female-specific prognostic markers, with pronounced RNA-protein discordance that underscores post-transcriptional and/or post-translational regulation and highlights the added value of proteomics (Table S4; Figure 4C).

Protein-based OS risk score for HGG

OS-associated pathways across age strata in males and females

We next performed pathway enrichment analysis separately in the reference and cDiscovery cohorts to identify pathways with prognostic relevance in both (STAR Methods). We focused on pathways with consistent OS-associations (FDR < 0.1) across both cohorts in at least one age stratum in males or females and identified 138 concordant protein pathways (Table S4; Figure 4D). Notably, 77% of these pathways showed associations only in males. Several of these pathways also exhibited significant male-specific T–N differences in prior analysis, including OXPHOS, the “neuronal system,” and “monatomic ion transmembrane transporter activity” (Figure 2E).

In contrast, 20% of OS-associated pathways were exclusive to females, such as “ECM organization” and “coagulation” (Figure 4D). Many TMP groups differing between males and females (defined in Figure S2A) were enriched with OS-associated proteins, including the TMP groups MF-T13, T23, T14, and T24. These findings suggest that the temporal dynamics of TMPs, shaped by neurodevelopment processes in the brain, capture biologically meaningful programs that differentially influence survival in males and females.

Among the OS-associated protein pathways shown in Figure 4D, only a limited subset, including the “neuronal system” pathway, showed a significant association with OS in RNA-based CPSA (Figure S4C). Our finding that the “neuronal system” pathway was prognostic aligns with a recent GBM study,³⁵ which linked a neuronal transcriptional profile to longer survival in recurrent glioblastoma.

Construction and validation of protein-based OS risk scores

We constructed male-/female-specific OS risk scores (OS-RSs), using prognostic proteins identified by CPSA, focusing on patients aged 0–26 years where survival associations were the strongest and most consistent, and validated them in an independent validation cohort (Figures 4E–4H; STAR Methods). Based on regularized Cox regressions with adjustment for key clinical and molecular covariates, we defined male- and female-specific OS-RSs as weighted averages of protein abundances, where the weights were the coefficients from multivariable Cox models (STAR Methods). OS-RSs were significantly associated with survival in both male- and female-stratified analyses as well as in the combined analysis and remained robust across sensitivity analyses (Table S4).

Among the male OS-RS panel, MORC2 and PHKG2 were leading proteins associated with the poorer and better survival, respectively (Figures 4E, 4I, and S4I). MORC2 and PHKG2 are from the “chromatin binding” and “energy derivation by oxidation of organic compounds” pathways, respectively. Both pathways showed strong associations with OS in male patients (Figures S4G and S4H). The findings in the validation cohort further supported these trends: MORC2 was associated with poorer survival (p = 0.095) (Figure S4J), and PHKG2 was associated with better survival (p = 0.119) (Figure 4J). Consistently, MORC2 has established oncogenic roles in chromatin remodeling and DNA repair,³⁶ whereas PHKG2, a component of phosphorylase kinase, regulates glycogen metabolism and may inhibit tumor progression through ferroptosis.³⁷

Causal kinase network construction

To pinpoint potential treatment targets for PED/AYA HGG tumors, we performed causal network analysis to identify kinases with activity causally linked to patient OS outcomes.

We focused on 186 kinases whose activities were linked to patient OS in the reference cohort by using CPSA (STAR Methods), with a subset shown in Figure S5A. Employing DAGBagM, a causal network construction method,³⁸ we identified candidate causal kinases associated with patient OS based on cDiscovery data (STAR Methods). The network model includes nodes for a binary OS outcome (“better” vs. “worse”), kinase activity scores, key prognostic mutations (ATRX, TP53, H3-3A, and IDH1; Figure 1E), and clinical covariates. Performing sensitivity analyses using various OS-time cut-offs, with and without male/female stratification, we obtained similar results (Table S5). Aggregating network results, we identified 26 kinases consistently inferred as causal parent nodes of OS status or their first- or second-degree neighbors (Figure 5A). From the inferred network, we identified multiple modules (Figures 5B and S5B). Module-I includes seven kinases—CDK8, ATM, ATR, LCK, NTRK2, PLK2, and GUCY2D—with relevant drugs available for all except GUCY2D. While ATR and ATM are well-established therapeutic targets,³⁹ supporting the validity of our approach, CDK8 and LCK represent promising but less well-characterized candidates in the context of HGG.

Figure 5.

Kinases with causal association to survival

(A) Subnetwork topology showing inferred causal relationships among 26 kinases associated with OS in the cDiscovery cohort (ages 0–40 years). “Available drugs” indicates kinases with associated drugs in the Children’s Cancer Model Atlas.

(B) Network topology for module I. Node and edge annotations are as in (A).

(C) Tumor cell-specific correlations among KEA3 kinase activity scores for the seven module I genes. Dots indicate significant correlations at FDR < 0.10 in iProMix results.

(D) Pathways upregulated or downregulated by six druggable module I kinases in cancer cells based on iProMix results. Asterisks indicate validation by CRISPR-Cas9 L1000 data; triangles indicate kinases in the pathway.

(E) Volcano plot of tumor cell-specific associations between oxidative phosphorylation proteins and CDK8 activity (iProMix). The enrichment p value indicates significant depletion (Wilcoxon rank-sum test). Downregulation of NDUFA4, NDUFB5, COX7A2, and ACAA2 is validated by CRISPR-Cas9 L1000.

(F) Heatmap of locCSN-inferred Z scores for kinase-kinase interactions in module I across tumors. The samples are ordered by the decreasing proportion of significant interactions (p < 0.05, locCSN).

(G) Schematic of the precision treatment strategy and validation experiments. Tumors with elevated kinase activity are hypothesized to be more sensitive to targeted inhibition, supported by greater proliferation suppression in the corresponding cell lines.

(H) Concordance (negative Spearman’s correlation) between kinase activity and proliferation under drug inhibition (red) and CRISPR knockout (blue) in tumor-derived cell lines for ATM, CDK8, and LCK .

(I) Scatterplots showing negative correlations between tumor kinase activity and drug response Z scores for ATM, CDK8, and LCK (Spearman’s r and p values are shown).

Notably, in male- and female-stratified analyses, modules like module-I and module-II were identified in male tumors (Figures S5C and S5D), but not in female tumors, suggesting a potential male-specific causal association between these kinases and prognosis.

Tumor cell-specific kinase regulation

To determine whether the kinase causal network reflects cancer cell regulation rather than the tumor microenvironment, we performed cancer cell-specific co-expression tests for selected kinase pairs, using iProMix,⁴⁰ a statistical method for inferring cell type-specific associations from bulk data (Table S5; STAR Methods). As illustrated in Figure 5A, 53.3% of the edges in the causal kinase network appeared active in cancer cells. Notably, all seven edges in module-I were determined to be cancer relevant (Figure 5C). However, in module-II, only the JAK2-FLT1 interaction met this criterion (Figure S5E).

To further characterize the pathways perturbed by module-I kinases, we applied iProMix to screen cell type-specific associations between the kinase activity scores and global protein abundance (Figures 5D and 5E). ATM, ATR, CDK8, and PLK2 were found to influence numerous pathways, including those involving cell cycle, DNA repair, metabolism, neuronal programs, and transcription. While the roles of ATM and ATR in DNA repair and transcription pathways are well documented,^41,42 the role of CDK8 in many pathways shown in Figure 5D has not been reported.

Using L1000 CRISPR perturbation data from central nervous system cell lines, we confirmed that CDK8 significantly impacts three pathways, with OXPHOS exhibiting the strongest association (p = 1.47e−55). Remarkably, OXPHOS downregulation by CDK8 was more prominent in cancer cells (Figures 5E and S5F; STAR Methods). Moreover, association analysis stratified by the IDH1 mutation status confirmed that CDK8 kinase downregulates the OXPHOS pathway regardless of the IDH1 mutation status or IDH1 WT protein abundance (Figure S5H, STAR Methods). Our earlier analysis (Figure 4D) linked the OXPHOS pathway to favorable OS outcomes. These findings suggest that CDK8 downregulates OXPHOS in cancer cells, aligning with its association with poor survival in both the reference (Figure S5A) and validation cohorts (Figure S5G).

Cell line-based investigation of precision treatment strategies targeting causal kinases

We hypothesized that the inhibitory drugs targeting CDK8/ATM/ATR/PLK2/LCK may benefit patients with elevated module-I activity. To characterize module-I activity per tumor, we used locCSN⁴³ to estimate the subject-specific interaction strengths between each pair of kinases that showed causal associations in module-I (STAR Methods). The heatmap in Figure 5F shows Z scored interaction strengths (Z scores) for the seven module-I edges across tumors. In 33% of tumors, at least 5 out of the 7 interactions were significant, indicating a subset of tumors with coordinated activation of this regulatory program.

We hypothesized that tumors with high module-I activity may be more sensitive to CDK8/ATM/ATR-targeted therapy. If a kinase promotes tumor growth, then inhibiting it should reduce proliferation in those cell lines with high kinase activity (Figure 5G). To test this, we selected tumor-derived cell lines with varying module-I activity (Figure 5F) and performed CRISPR knockout and/or drug inhibition experiments⁴⁴ targeting CDK8, ATM, LCK, ATR, NTRK2, and PLK2 (STAR Methods). Proliferation scores from CRISPR knockout and drug inhibition of CDK8, ATM, and LCK across various cell lines showed a negative correlation with their kinase activity levels in the corresponding tumors, supporting the hypothesis (Figures 5H, 5I, and S5J). We also performed parallel perturbation analysis for a subset of markers in module-II. FLT1, RAF1, and MAP2K1—each associated with worse OS—showed a similar concordance between the cell line proliferation scores and corresponding kinase activity scores in tumor samples (Figures S5I and S5J). Together, these findings suggest that tumors with elevated module-specific kinase activity, particularly those specific to module-I, define a distinct subset of PED/AYA HGG patients that may be preferentially sensitive to targeted inhibition.

Protein pathway activities reveal distinct subtypes of HGGs

To characterize molecular heterogeneity in PED/AYA HGGs, we conducted a protein-based subtyping analysis. We performed male-and female-stratified consensus clustering using single-sample gene set enrichment analysis (ssGSEA) scores of 8,167 biological pathways derived from global protein abundance of 190 tumors (87 females and 103 males) in the cDiscovery cohort (Figure 6A; STAR Methods). A similar set of three protein clusters, labeled M1/M2/M3 and F1/F2/F3, was identified in males and females, respectively. For simplicity, we refer to the combined male and female clusters as C1, C2, and C3 (e.g., C1 is the union of M1 and F1). In a sensitivity analysis without male/female stratification, we identified three clusters like those found in the stratified analysis (Figure S6A). Given the observed heterogeneity between male and female tumors, we focused on the male-/female-stratified subtypes to interrogate male-/female-specific signatures in more detail.

Figure 6.

Protein pathway clusters revealing male-female differences associated with disease outcomes

Significance is indicated as ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

(A) Protein pathway activity across male- and female-stratified protein clusters in the cDiscovery cohort (ages 0–62 years), shown by ssGSEA scores for the selected pathways. Right tracks indicate the direction and significance of cluster associations in tumors from patients aged 0–40 years (linear model). Causal kinases linked to cluster-associated pathways are annotated.

(B) Protein pathway ssGSEA scores for the same pathways in the validation cohort. Protein clusters were derived by consensus clustering as in cDiscovery.

(C) Correlation of the average pathway scores between the cDiscovery and validation clusters (Pearson’s correlation).

(D) Distribution of H3-3A mutations across protein clusters. H3 WT tumors are enriched in C3 (one-tailed Fisher’s exact test).

(E) Activity of the kinases ATM, CDK8, LCK, and NTRK2 across protein clusters in cDiscovery (ages 0–40 years) (Wilcoxon rank-sum test).

(F) CD276 (B7-H3) abundance stratified by male and female groups and protein clusters (cDiscovery, ages 0–40 years). Bottom: cluster vs. others; top: male-female differences within clusters (Wilcoxon rank-sum test).

(G) OS comparison of F2, M2, and other clusters in cDiscovery (ages 0–40 years). p value is from the male-by-C2 interaction term in a multivariable Cox model.

(H) Differential protein abundance between F2 and M2 for interferon γ response genes (cDiscovery; linear model).

(I) Boxplots of PLA2G4A protein (left) and phosphosite (right) abundances across male and female groups and protein clusters. Bottom: cluster vs. others; top: male-female differences within clusters (Wilcoxon rank-sum test).

(J) Correlation between LCK kinase activity and PLA2G4A expression in males and females in the cDiscovery cohort (ages 0–40 years; Pearson r and p values are shown).

(K) Pathway enrichment comparing F2 vs. M2 based on glycopeptide abundance (discovery cohort; linear model).

(L) Differential glycopeptide abundance between F2 and M2 for epithelial-mesenchymal transition genes. Glycoforms: F + S, fucosylation + sialylation; HM, high mannose; F, fucosylation; S, sialylation. The top association is an ITGAV glycopeptide with N-linked sialylated glycan N3H5F1S1G0 (discovery cohort; linear model).

(M) ITGAV glycopeptide (N3H5F1S1G0) abundance across male and female groups and protein clusters. Bottom: cluster vs. others; top: male-female differences within clusters (discovery cohort, Wilcoxon rank-sum test).

(N) Correlation between PD-L1 and ITGAV glycopeptide (N3H5F1S1G0) abundances in female (red), male (blue), and combined (black) tumors (discovery cohort, Pearson’s r and p values are shown).

As shown in Figure 6A, C1, C2, and C3 exhibit distinct pathway activities. C1 is characterized by upregulated rRNA metabolism and double-stranded break repair pathways, C2 by immune response pathways, and C3 by neuronal and energy metabolism functions. Although all clusters include samples from every age group, PED samples were enriched in C2 and ADU samples in C3 (Figure S6B). H3-3A mutation, rare in adults, was consistently depleted in C3 (Figure 6D).

The three clusters were recapitulated in our validation cohort via unsupervised clustering (Figures 6B and 6C; STAR Methods), supporting the robustness of these protein pathway-based subtypes.

Comparison with literature subtypes and RNA subtypes

Comparing our HGG protein clusters with the adult GBM subtypes from Wang et al.,¹⁰ we observed partial concordance: C2 was enriched for the “mesenchymal” subtype, while C1 and C3 were scattered across multiple subtypes (Figure S6B). In comparison with subtypes from Garofano et al.,⁴⁵ C1 and C3 largely aligned with the “proliferative/progenitor” and the “neuronal” subtypes, respectively, whereas C2 samples represented a mixture of multiple subtypes (Figure S6B; STAR Methods). Similarly, RNA-based clustering showed limited alignment with these classifications of adult GBM (Table S6). Furthermore, 29% of tumors were assigned to different clusters based on RNA vs. protein data (Figure 6A), consistent with previous studies,^9,46 indicating that proteomic data capture post-transcriptional patterns not evident in transcriptomes.

Moreover, protein clusters showed only partial overlaps with methylation- and genomically informed cancer groups (Figure S6B), with C1 and C2 spanning multiple entities. When evaluating heterogeneity across cancer types versus protein clusters using proteomic ssGSEA pathway scores, we obtained substantially lower silhouette scores for cancer-type groupings than for protein-cluster groupings (Figure S6C), indicating that cancer types do not fully capture the underlying molecular heterogeneity. Notably, while histological classification and mutation profiles define the baseline tumor identity, dominant proteomic programs and signaling states are broadly shared across tumor entities, with entity-specific differences primarily reflected in their distribution across clusters and the underlying mutation landscapes.

Causal kinases and CAR-T markers associated with protein subtypes

We focused on patients aged 0–40 years and identified kinase activities (KEA3 scores) linked to each cluster (FDR < 0.1): ATM and CDK8 in C1, LCK in C2, and NTRK2 in C3 (Figure 6E). Elevated activities of ATM in C1, LCK in C2, and NTRK2 in C3 are supported by their global protein abundance in the cDiscovery and/or validation cohorts (Figure S6D). However, CDK8 abundance does not correlate with its kinase activity, highlighting the importance of phosphoproteomics for kinase activity characterization. Notably, these kinase-cluster associations were consistent in males and females and across cancer groups (Table S6).

We further investigated the association of CAR-T markers relevant to glioma with protein subtypes (STAR Methods). CD276 (also known as B7-H3) showed elevated protein abundance and RNA expression in the subtypes C1 and C2 compared with C3 in both the cDiscovery and validation cohorts (Figures 6F, S6E, and S6F; Table S6). B7-H3, a member of the B7 protein family, is an immune checkpoint molecule and a key player in cancer progression, with emerging targeted therapies including CAR-T therapy.⁴⁷

OS outcome differences across protein subtypes

Significant OS differences were observed across the six male-/female-stratified protein clusters (p = 0.002, Figure S6G; STAR Methods). Notably, the M2 cluster had the poorest OS outcomes, while F2 was among those with the best. The significant male-female difference in OS outcomes within the C2 cluster was also supported by Cox models with a male-female cluster interaction term (STAR Methods) for both 0–62 (p = 0.007, Table S6) and 0–40 (p = 0.004, Table S6; Figure 6G) age group cohorts. No significant male-female differences in OS were detected for other clusters (Table S6).

The observed OS gap between M2 and F2 was explained neither by clinical covariates, mutations, or DNA-level alterations nor by the differences in cancer-group composition based on multivariable adjustment (Table S6). Proteomic analysis (STAR Methods) detected upregulated pathways (Figures S6H, S6K, and S6L; Table S6), including interferon γ and α response pathways, in F2 tumors relative to M2 tumors (Figure 6H). These protein pathway differences were confirmed in the validation cohort (Figures S6H and S6I) and the observed cDiscovery RNA data (Table S6).

At the marker level, PLA2G4A, in the interferon γ pathway, was significantly upregulated in M2 relative to F2 (Figures 6H and 6I; FDR < 0.10). A similar pattern was observed for the PLA2G4A_S377 phosphosite (Figure 6I) and its RNA expression (Figure S6J). PLA2G4A (cPLA2) triggers eicosanoids—lipid-based hormones that regulate inflammation and physiology. Its gene expression has been linked to poorer prognosis in gliomas.⁴⁸ A female-specific correlation between PLA2G4A_S377 and LCK kinase activity (Figure 6J), which was upregulated in C2 relative to C1/3 (Figure 6E), suggests differential regulation of these components in males and females.

By contrast, collagen formation and EMT pathways were upregulated in M2 relative to F2 in the cDiscovery and validation cohorts (Figure S6H). This was corroborated by glycoproteomic data (Figure 6K; STAR Methods), showing the enrichment of upregulated glycopeptides in M2 in these pathways. In the EMT pathway, a sialylated glycopeptide of ITGAV (alpha V) (N3H5F1S1G0) exhibited the strongest upregulation in M2 relative to F2 (Figures 6L and 6M). ITGAV was also found to be significantly associated with M2/F2 in gene-level analysis of glycopeptides based on observed abundances (FDR < 0.1; Figure 6M) and protein-adjusted glycopeptide levels (p = 0.02; Figure S6M; Table S6). However, no difference between M2 and F2 was found for the unsialylated counterpart (N3H5F1S0G0) (Figure S6M).

ITGAV is well documented in cancers and has been proposed as a therapeutic target.^49,50 Using pan-cancer glycoproteomic data,²⁶ we detected elevated abundance of this sialylated glycopeptide of ITGAV in tumors compared with adjacent normal tissues across several solid tumors among adult patients (Figure S6N; STAR Methods), indicating a role in cancer biology. Previous studies in low-grade gliomas have suggested a role of ITGAV in regulating immune checkpoint molecules such as CD274 (PD-L1).⁵¹ Our data revealed a significant negative correlation between the PD-L1 abundance and the sialylated glycopeptide abundance of ITGAV based on both observed (Figure 6N) and protein abundance-adjusted data. This association was absent in RNA, global protein, or unsialylated glycopeptide data (Figure S6O). These results suggest that the sialylated ITGAV glycopeptide may help shape immune environments across protein clusters and contribute to the survival disparity between M2 and F2.

Immune signaling in PED/AYA HGGs

Deconvolution of the integrated proteomic and RNA data from the cDiscovery cohort using BayesDeBulk⁵² revealed heterogeneous cell type composition across tumors (Figure 7A; STAR Methods). Several cell types exhibited age-dependent infiltration patterns: SEPP1 Mo tumor-associated macrophages (TAMs) and T cells were less enriched in PED tumors than in YA tumors, whereas pericytes were more enriched in PED tumors (Figures 7B and S7A).

Figure 7.

Immune landscape characterization

Significance levels based on the Wilcoxon rank-sum test are indicated as follows: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001, unless otherwise noted.

(A) Heatmap of BayesDeBulk-estimated cell type fractions (Z scored) in cDiscovery tumors. Bottom: RNA/protein expression of cell type markers and ssGSEA scores for immune pathways.

(B) Boxplots of T cell and SEPP1 Mo TAM fractions by age class.

(C) Boxplots of T cell and SEPP1 Mo TAM fractions by protein clusters.

(D) Cox survival analysis (ages 0–40 years) comparing low SEPP1 Mo TAM/high T cell vs. high SEPP1 Mo TAM/low T cell infiltration in the cDiscovery tumors (STAR Methods). Thresholds are defined by the median cell fractions. ANOVA p value for the combined effect is shown.

(E) Survival analysis in the validation cohort, performed as in (D).

(F and G) Boxplots of T cell and SEPP1 Mo TAM fractions by NF1 (F) or H3-3A (G) mutation status in cDiscovery.

(H) Heatmap of associations between T cell and SEPP1 Mo TAM fractions and immune checkpoint gene/protein abundance across age strata in males and females (<40 and >40 years). Color indicates signed –log₁₀(FDR) for significant associations only (FDR < 0.1; linear model).

Correlating the cell type fractions with protein clusters, we found that myeloid cells—including hypoxic monocyte-like tumor-associated macrophage (Hypoxic Mo TAM), interferon-responsive monocyte-like tumor-associated macrophage (IFN Mo TAM), and SEPP1-expressing monocyte-like tumor-associated macrophage (SEPP1 Mo TAM)—were more abundant in F2/M2 (Figures 7C and S7B), whereas neurons and astrocytes were enriched in F3/M3 (Figure S7B). This pattern aligned with the pathway activity scores (ssGSEA from global proteomics): immune pathways (e.g., “complement” and “interferon γ”) were elevated in F2/M2, while cell-cycle and metabolic pathways were upregulated in less immunogenic F1/M1 and F3/M3 clusters (Figure 7A).

We then investigated the associations between OS and cell type fractions, adjusting for age and H3-3A mutation (STAR Methods). Among females aged 0–40 years, the high SEPP1 Mo TAM fraction and the low T cell fraction were associated with worse prognosis (p = 0.02; Figure 7D); this pattern was not detected in male patients (Figure S7C). These trends were replicated in the validation cohort (p = 0.01, Figure 7E; STAR Methods). As reported by Antunes et al.,⁵³ SEPP1 Mo TAMs were enriched in recurrent tumors—consistent with their higher infiltration in more aggressive tumors here.

We next examined the associations between somatic mutations and cell type fractions. SEPP1 Mo TAMs were elevated in NF1-mutant tumors, and T cell infiltration was reduced in H3-3A-mutant tumors (Figures 7F and 7G). In addition, IFNG Mo TAMs and microglia-derived tumor associated macrophages (Mg TAMs) were enriched in BRAF-mutant tumors, while oligodendrocytes and astrocytes were more enriched in IDH1-mutant tumors (Figure S7D).

To investigate immune regulation, we correlated SEPP1 Mo TAM and T cell fractions with the immune checkpoint and regulatory markers using RNA and proteomics data (Figure 7H). In adults (aged > 40 years), T cell fractions were significantly associated (FDR < 0.1) with the RNA expression levels of PD-L1, PD-1, and ANXA1 in both males and females. Among PED/AYA patients, T cells were significantly associated with PD-1 and ANXA1 expression levels only in males (FDR < 0.1), and SEPP1 Mo TAMs showed a significant negative association with PD-1 expression levels exclusively in male PED/AYA patients, with no association in either group among adult patients.

Discussion

This study provides an integrated multi-omics analysis of HGGs across development, establishing a framework in which age and developmental context shape tumor biology. We identified a molecular breakpoint at age 26 that distinguishes AYA HGG by molecular profiles and survival and defined the age-linked regulatory programs through an integrated analysis of cis-/trans-regulatory effects, kinases, and TFs.

Male- and female-biased features emerged across the metabolic and neuronal pathways, with OXPHOS proteins showing greater T–N deviation in males and IDH1/2 protein scores associating with improved survival only in males. These differences in OXPHOS regulation^54,55,56 may have implications for metabolic targeting.

CPSA, leveraging male- and female-specific developmental trajectories, identified male- and female-specific protein markers and OS-RSs that were validated in an independent cohort and remained robust across models. Glycopeptides were associated with OS, particularly in males, consistent with female- and male-biased tumor biology and suggesting a role of glycosylation in immune evasion in male tumors, consistent with our immune landscape analysis.

Causal network analysis nominated key kinases—including CDK8, ATM, and LCK—as potential treatment targets in PED/AYA HGGs. CDK8 was identified as a causal regulator of the OXPHOS pathway, and CRISPR knockout and drug screening in cell lines supported an association between kinase activity and drug sensitivity. Together with prior work showing that phosphoproteomics-inferred kinase activity predicts sensitivity to targeted inhibitors,^57,58 these findings support the use of these kinase activity profiles to inform personalized therapeutic targeting.

Protein pathway-based clustering identified subtypes associated with survival and distinct kinase dependencies. We identified kinases and CAR-T cell target epitopes upregulated in specific clusters. Subtype-specific differences in ATM/CDK8 and LCK activation, along with enrichment of the CAR-T target CD276 (B7-H3), suggest actionable therapeutic targets that may inform patient stratification. A previously unrecognized male cluster (M2) exhibited the poorest survival and was characterized by distinct pathway and glycopeptide features.

Our integrated analysis also revealed immune cell types associated with survival. T cell infiltration was linked with favorable outcomes in female PED/AYA HGG patients but not in male counterparts and correlated with PD-1 expression only in males, a pattern absent in adult tumors (>40 years). These findings suggest that PD-1-mediated suppression may limit T cell effectiveness in male PED/AYA tumors, positioning PD-1 inhibition as a promising strategy in this group. Consistent with these observations, male-female differences have emerged as an important determinant of immune response in glioblastoma,^59,60 with evidence that male patients may exhibit greater sensitivity to the immune checkpoint blockade.^60,61 As PD-1 inhibition therapies are under active investigation in clinical trials involving PED/AYA HGG patients, our findings support the development of male- and female-specific treatment strategies. Additional mechanisms of immune suppression, including tumor glycosylation and integrin signaling, may contribute to specific subgroups (e.g., M2) and represent potential therapeutic targets.

Limitations of the study

While the findings of this study support age-informed precision oncology tailored to male and female patients, limitations remain. Stratified analyses reduced power, particularly in underrepresented subgroups such as YA patients with limited glycoproteomic data, and the small validation cohort limited male-/female-specific validation. Sample size also precluded systematic replication of pan-HGG findings within individual cancer groups, especially rare histological entities, where findings should be interpreted cautiously and validated in larger, subtype-focused cohorts.

The mechanistic basis of the observed male-female differences remains to be defined, and in vivo validation is required to establish the therapeutic relevance of targets such as CDK8. Despite these limitations, our study advances understanding of the developmental patterns in male and female HGGs and identifies molecular and immunologic features with translational potential.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Pei Wang (pei.wang@mssm.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The raw proteomic data from this study are available at PDC: PDC000497, PDC000498, and PDC000499. All pre-processed proteogenomic datasets are provided as supplemental data via PDC and can also be explored via a ProTrack Web-portal: http://hope.cptac-data-view.org/.
• •Raw and processed genomic and transcriptomic data are available via dbGaP: phs002517.v5.p3, as well as through the Gabriella Miller Kids First Data Resource Center (https://kidsfirstdrc.org/) and the Children’s Brain Tumor Network (https://cbtn.org).
• •Software code for major data processing, analysis, and figure generation is available at GitHub (https://github.com/WangLab-MSSM/Pediatric-AYA-high-grade-glioma).
• •The temporalCPSA R package, which provides a generalized framework for CPSA and related trajectory analyses, is publicly available on GitHub at https://github.com/lashbroz/temporalCPSA.
• •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Consortia

The members of CPTAC are Nicole L. Tignor, Shrabanti Chowdhury, Francesca Petralia, Marina A. Gritsenko, Weiping Ma, Giacomo B. Marino, Xiaoyu Song, Jeffrey R. Whiteaker, Dmitry Rykunov, Yingwei Hu, Noshad Hosseini, Felipe da Veiga Leprevost, Lijun Chen, Yi-Ting Wang, Karl K. Weitz, Rosalie K. Chu, Ronald J. Moore, Azra Krek, Eden Z. Deng, Chia-Feng Tsai, Tyler Sagendorf, Vladislav A. Petyuk, Tujin Shi, Thomas L. Fillmore, Rui Zhao, Matthew E. Monroe, Lei Zhao, Richard G Ivey, Uliana J. Voytovich, Boris Reva, Richard D. Smith, Mehdi Mesri, Ana I. Robles, Li Ding, Karin D. Rodland, Bing Zhang, Alexey I. Nesvizhskii, Marcin Cieslik, Joseph E. Ippolito, Avi Ma’ayan, Hui Zhang, Amanda Paulovich, Tao Liu, Pei Wang, Jasmin Bavarva, Xu Zhang, Eunkyung An, Tara Hiltke, Henry Rodriguez, Liang-Bo Wang, Song Cao, Bo Wen, Xinpei Yi, Zhiao Shi, William E. Bocik, Sandra S. Garcia-Buntley, Ratna R. Thangudu, Alexander Pilozzi, Karen A. Ketchum, Shuang Cai, Nathan J. Edwards, Steven A. Carr, Michael A. Gillette, Shankha Satpathy, Daniel W. Chan, D.R. Mani, Gad A. Getz, Saravana M. Dhanasekaran, Jeffrey W. Tyner, Brian J. Druker, Paul D. Piehowski, Akhilesh Pandey, Rafael Fonseca, Diwakar Davar, Ramaswamy Govindan, Peter McGarvey, Toan Le, Yin Lu, Kristen Nyce, Kenneth Aldape, and Antonio Colaprico.

The members of CBTN are Mateusz Koptyra, Yuankun Zhu, Jo Lynne Rokita, Komal S. Rathi, Marius V. Dannappel, Paul Daniel, Mariarita Santi, Demetri Dupal, Jena Lilly, Adam Kraya, Joseph M. Dybas, Bo Zhang, Chuwei Zhong, Miguel A. Brown, Saksham Phul, Eric Wafula, Alvin Farrel, Zhuangzhuang Geng, Ryan J. Corbett, Ammar S. Naqvi, Daniel P. Miller, Jennifer Mason, Tatiana S Patton, Stephanie McGrory, Shannon Robins, Allison Heath, Catherine Sullivan, Noel Coleman, Allison Morgan, Phillip B. Storm, Ron Firestein, Adam Resnick, Brian R. Rood, Alicia Lenzen, Angela Waanders, Javad Nazarian, Jason Fangusaro, Tobey J. MacDonald, Eric Raabe, Eric Jackson, Stacie Stapleton, Toba Niazi, Shelly Wang, Matija Snuderl, Jessica Clymer, Sarah Leary, Nicholas Vitanza, Michelle Monje, Laura M. Prolo, Sonia Partap, Donald Parsons, Austin J. Stuckert, Phillip B. Storm, Michael Fisher, Jane Minturn, Jessica Foster, Peter Madsen, Cassie Kline, Ian F. Pollack, Nalin Gupta, Mike Prados, Joanna Phillips, Sabine Mueller, Jeff Greenfield, Nadia Dahmane, Samuel Cheshier, Angela Viaene, Thinh Q. Nguyen, Zeinab Helili, Bailey K. Farrow, Nathan Young, Nick Van Kuren, and Elizabeth Appert.

The members of PC4C are Jennifer Mason, Allison Morgan, Adam Resnick, Phillip B. Storm, Jasmin Hussain, Elizabeth Appert, Timothy Prior, Eileen Maloney, Atom Sarkar, Alan Turtz, Michael Weaver, Scott Shepard, Kevin Judy, Donald O'Rourke, Nduka Amankulor, Iyad Alnahhas, Pascal Jabbour, Valeria Lopez-Gil, Noel Coleman, and Zev A. Binder.

Acknowledgments

The Clinical Proteomic Tumor Analysis Consortium (CPTAC) is supported by the National Cancer Institute (NCI), NIH (award numbers: U24CA270823, U24CA271079, U24CA271012, U24CA271114, U24CA271037, U24CA271076, U24CA271075, U01CA271402, U01CA271410, U01CA271407, U01CA271412, U24CA210993, U24CA210985, U01CA214114, U24CA210955, U24CA210972, U01CA214116, U24CA210954, U24CA210967, U24CA210986, U24CA210979, and U01CA214125). This project was also supported by NCI, other NIH awards/contracts (HHSN261201500003I, HHSN26100064, U24CA224260, U24CA264250, U01CA271407, R01CA235575, R33CA263705, P30ES017885, T32GM136542, 1F30CA265288, F30CA271622, and R50CA211499), and the Aven Foundation (A.G.P.). Portions of the proteomics work were performed at the Environmental Molecular Sciences Laboratory (grid.436923.9), a DOE user facility at PNNL, operated by Battelle under contract DE-AC05-76RL01830. We acknowledge the American Association for Cancer Research (AACR) and the Project GENIE consortium for data sharing. Computational resources were provided by Scientific Computing and Data at the Icahn School of Medicine at Mount Sinai, supported by CTSA grant UL1TR004419 from the National Center for Advancing Translational Sciences. Additional support was provided by NIH infrastructure awards S10OD026880 and S10OD030463. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the US Department of Health and Human Services, nor do trade names, commercial products, or organizations imply endorsement by the US Government. We honor our co-author and long-time colleague Uliana J. Voytovich, who recently passed away while in Ukraine performing humanitarian work in the service of those affected by the war. We thank Aleha Syed for assistance with the GitHub repository.

Author contributions

Conceptualization, B.R.R., A.R., and P.W.; experiment and data generation, A.F., A.G.P., A. Morgan, A.R., A.S.N., B.R.R., B.Z., C.-F.T., C.Z., D.D., D.P.M., E.W., E.Z.D., G.B.M., H.Z., J.L., J.M., J.M.D., J.R.W., K.D.R., K.K.W., L.C., L.Z., M.A.B., M.A.G., M.K., M.S., M.V.D., N.A., P.D., R.F., R.G.I., R.J.C., R.J.M., R.K.C., S. McGrory, S.P., S.R., T.L., T.L.F., T.S., T.S.P., T.M.Y.-B., U.J.V., V.A.P., W.M., X.S., X.W., Y.T.W., Y.Z., and Z.G.; methodology, data analysis, and visualization, A. Kraya, A. Krek, A.R., C.-F.T., D.R., E.M.H., E.Z.D., F.D.V.L., F.P., G.B.M., J.L.J., J.L.R., K.S.R., M.C., M.K., N.H., N.L.T., P.D., P.W., S.C., S. Migliozzi, W.M., X.S., X.W., Y.H., and Y.Z.; software, N.A.,Y.-c.L., and N.L.T.; resources, A.I.R., A. Ma'ayan, A.I.N., A.P., A.R., B.R.R., B.Z., H.Z., K.D.R., M.C., M.M., P.W., R.F., and T.L.; biological interpretation and analysis, A.I., A.R., B.R.R., D.R., E.E.S., J.E.I., J.L.J., J.R., M.K., M.M., M.V.D., N.L.T., P.D., P.W., R.F., S. Migliozzi, and J.L.R.; writing – original draft, A.I., A. Krek, A. Ma'ayan, A.I.N., A.P., A.R., B.R.R., C.F.T., D.R., E.E.S., F.P., G.B.M., H.Z., J.E.I., J.L.R., J.M.D., J.R., K.S.R., M.A.G., M.C., M.K., M.V.D., N.H., N.L.T., P.D., P.W., R.F., S.C., S. Migliozzi, T.L., W.M., X.S., and X.W.; writing – review & editing, all authors

Declaration of interests

L.C.C. is a founder and board member of Agios Pharmaceuticals; founder and research-supported investigator with Petra Pharmaceuticals; inventor on a Weill Cornell Medicine patent (WO2019232403A1) related to PI3K-targeted combination therapies; co-founder and shareholder of Faeth Therapeutics; equity holder in and consultant to Cell Signaling Technologies, Volastra, Larkspur, and 1 Base Pharmaceuticals; and consultant to Loxo-Lilly. J.L.J. has received consulting fees from Scorpion Therapeutics and Volastra Therapeutics. T.M.Y.-B. is a co-founder of DeStroke.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT to improve English grammar. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit monoclonal anti-TP53	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-TP53-1; RRID:AB_2877664
Rabbit monoclonal anti-CDK1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-CDK1-1; RRID: AB_2877655
Rabbit monoclonal anti-TUBB	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-TUBB-1; RRID: AB_2877665
Rabbit monoclonal anti-RAF1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-RAF1-3; RRID: 2827856
Rabbit monoclonal anti-KRT7	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-KRT7-1; RRID: AB_2820264
Rabbit monoclonal anti-KRT5	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-KRT5-1; RRID: AB_2820262
Rabbit monoclonal anti-KRT5	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-KRT5-2; RRID: AB_2820263
Rabbit monoclonal anti-FBN1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FBN1-1; RRID: AB_2877658
Rabbit monoclonal anti-FBN1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FBN1-2; RRID_2877659
Rabbit monoclonal anti-PCNA	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-PCNA-2; RRID: AB_2877661
Rabbit monoclonal anti-ATM	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-ATM-3; RRID: AB_2877654
Rabbit monoclonal anti-ATM	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-ATM-1; RRID: AB_2877653
Rabbit monoclonal anti-FAAP100	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FAAP100-1; RRID: AB_2877656
Rabbit monoclonal anti-FAAP100	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FAAP100-2; RRID: AB_2877657
Rabbit monoclonal anti-RIF1	Fred Hutchinson Cancer Research Center	custom monoclonal
Mouse monoclonal anti-PTEN	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-PTEN-4; RRID: AB_2722096
Mouse monoclonal anti-PTEN	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-PTEN-7; RRID: AB_2722099
Mouse monoclonal anti-CCND1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-CCND1-5; RRID: AB_2722035
Mouse monoclonal anti-RPTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-RPTOR-2; RRID: AB_2722104
Mouse monoclonal anti-RPTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-RPTOR-1; RRID: AB_2722103
Mouse monoclonal anti-FOS	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FOS-2; RRID: AB_2722059
Mouse monoclonal anti-FOS	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FOS-3; RRID: AB_2722060
Mouse monoclonal anti-AKT2	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-AKT2-3; RRID: AB_2722017
Mouse monoclonal anti-CDH1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-CDH1-1; RRID: AB_2722037
Mouse monoclonal anti-CDH1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-CDH1-2; RRID: AB_2722038
Mouse monoclonal anti-BRAF	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-BRAF-1; RRID: AB_2722024
Mouse monoclonal anti-RAF1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-RAF1-1; RRID: AB_2877663
Mouse monoclonal anti-MTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MTOR-5; RRID: 2722089
Mouse monoclonal anti-MTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MTOR-4; RRID: AB_2722088
Mouse monoclonal anti-MTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MTOR-3; RRID: AB_2722087
Mouse monoclonal anti-MTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MTOR-7; RRID: AB_2722091
Mouse monoclonal anti-ERBB2	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-ERBB2-3; RRID: AB_2722053
Mouse monoclonal anti-ERBB2	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-ERBB2-2; RRID: AB_2722052
Mouse monoclonal anti-AKT1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-AKT1-1; RRID: AB_2722012
Mouse monoclonal anti-AKT1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-AKT1-3; RRID: AB_2722014
Mouse monoclonal anti-AKT3	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-AKT3-6; RRID: AB_2814779
Rabbit monoclonal anti-EGFR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-EGFR-8; RRID: AB_2827847
Rabbit monoclonal anti-EGFR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-EGFR-9; RRID: AB_2827848
Rabbit monoclonal anti-GSK3B	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-GSK3B-10; RRID: AB_2827862
Rabbit monoclonal anti-GSK3B	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-GSK3B-9; RRID: AB_2827863
Rabbit monoclonal anti-MAPK1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MAPK1-2; RRID: AB_2827853
Rabbit monoclonal anti-MAPK3	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MAPK3-2: RRID: AB_2722081
Rabbit monoclonal anti-MTOR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MTOR-9; RRID: AB_2820270
Rabbit monoclonal anti-MAPK1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MAPK1-4; RRID: AB_2827841
Rabbit monoclonal anti-MAPK3	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-MAPK3-4; RRID: AB_2827841
Rabbit monoclonal anti-RIF1	Fred Hutchinson Cancer Research Center	custom monoclonal
Rabbit monoclonal anti-ERBB2	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-ERBB2-1: RRID: AB_2617255
Rabbit monoclonal anti-JUN	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-JUN-4: RRID: AB_2877660
Rabbit monoclonal anti-RB1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-RB1-2; RRID: AB_2820275
Rabbit monoclonal anti-ANXA1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-ANXA1-4; RRID: AB_2617223
Rabbit monoclonal anti-CALR	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-CALR-1; RRID: AB_2617228
Rabbit monoclonal anti-GSTP1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-GSTP1-1; RRID: AB_2617266
Rabbit monoclonal anti-PCNA	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-PCNA-1; RRID: AB_2617307
Rabbit monoclonal anti-RRM2	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-RRM2-1; RRID: 2617333
Rabbit monoclonal anti-FSCN1	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-FSCN1-1; RRID: AB_1553796
Rabbit monoclonal anti-PRDX4	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-PRDX4-5; RRID: AB_2877662
Rabbit monoclonal anti-UBE2C	CPTAC Antibody Portal (antibodies.cancer.gov)	Cat# CPTC-UBE2C-2; RRID: AB_2617367
Biological samples
Fresh frozen tissue samples	See STAR Methods	N/A
Chemicals, peptides, and recombinant proteins
Papain	SciQuest
Ovomucoid	SciQuest	Catalog: 542000
DNase	Sigma	Catalog: 10104159001
RNase A	Qiagen	Catalog: 19101
TMT 11plex reagents	Thermo Fisher Scientific	Catalog: A34808
IMAC resin kit	Thermo Fisher Scientific	Catalog: A32992
BCA protein assay	Thermo Fisher Scientific	Catalog: 23221
Stable isotope-labeled synthetic peptide standards	New England Peptide	Custom synthesis, >95% pure by HPLC
Aprotinin	Sigma	Catalog: A6103
Leupeptin	Roche	Catalog: 11017101001
Phenylmethylsulfonyl fluoride	Sigma	Catalog: 93482
Phosphatase Inhibitor Cocktail 2	Sigma	Catalog: P5726
Phosphatase Inhibitor Cocktail 3	Sigma	Catalog: P0044
PUGNAc	Sigma	Catalog: A7229
Sodium butyrate	Sigma	Catalog: 303410
Ni-NTA Superflow Agarose Beads	Qiagen	Catalog: 30410
Lysyl Endopeptidase	Wako Chemicals	Catalog 129-02541
Sequencing grade modified trypsin	Promega	Catalog: V517
Urea	Sigma-Aldrich	Cat#U0631
Trizma base (Tris)	Sigma-Aldrich	Cat#T2694
iodoacetamide (IAM)	Sigma-Aldrich	Cat#A3221
EDTA	Sigma-Aldrich	#E7889
EGTA	Sigma-Aldrich	#E0396
Protease Inhibitor Cocktail	Sigma-Aldrich	#P8340
phosphate buffered saline	Thermo Fisher Scientific	#BP-399-20
tris(2-carboxyethyl)phosphine	Thermo Fisher Scientific	Cat#77720
10 mg Sep-Pak solid-phase extraction	Waters	Cat#1860001280
10 mg Sep-Pak solid-phase extraction	Waters	Cat#1860000309
Acetonitrile	Sigma-Aldrich	Cat#A955
Water	Sigma-Aldrich	Cat#W6
ammonium bicarbonate	Sigma-Aldrich	Cat#A6141
ammonium hydroxide solution	Sigma-Aldrich	Cat#320145
Formic acid	EMD Millipore	Cat#1.11670.1000
Trifluoroacetic Acid	Sigma-Aldrich	Cat#302031
Critical commercial assays
TruSeq RNA Sample Prep Kit	Illumina	N/A
KAPA Library Preparation Kit	Roche	Catalog: KK8201
DNA/RNA AllPrep Kit	Qiagen	Catalog: 80204
AllPrep DNA/RNA/miRNA Universal kit	Qiagen	Catalog: 80224
QIAsymphony DSP DNA Midi Kit	Qiagen	Catalog: 937255
KAPA Hyper prep kit	Roche	Catalog: KK8541
KAPA Stranded RNA-Seq with RiboErase kit	Roche	Catalog: KK8484
Micro BCA assay	Thermo Fisher Scientific	Cat#23235
Deposited data
Proteomics datasets	This paper	https://pdc.cancer.gov under accession numbers PDC000497, PDC000498, and PDC000499
Genomics datasets	This paper	dbGaP (accession number: phs002517. v5.p3).
PhosphoSitePlus	Hornbeck et al., 201562	https://www.phosphosite.org
TCGA – GBM	The Cancer Genomic Atlas Research Network, 2008	https://portal.gdc.cancer.gov/
Software and algorithms
MSFragger-3.1.1	Nesvizhskii Lab, Kong63	http://msfragger.nesvilab.org/
Philosopher-v3.3.12	Nesvizhskii Lab, da Veiga Leprevost.64	https://philosopher.nesvilab.org/
TMT-Integrator-v1.1.10	Nesvizhskii Lab, Djomehri et al.65	http://tmt-integrator.nesvilab.org/
PeptideProphet	Keller66	http://tools.proteomecenter.org/wiki/index.php?title=Main_Page
PTMProphet	Shteynberg et al.67	http://tools.proteomecenter.org/wiki/index.php?title=Main_Page
ProteinProphet	Nesvizhskii68	http://tools.proteomecenter.org/wiki/index.php?title=Main_Page
ComBat (v3.20.0)	Johnson et al.,69 2007	https://bioconductor.org/packages/release/bioc/html/sva.html
DreamAI	Pei Wang Lab	https://github.com/WangLab-MSSM/DreamAI
COSMO	Yoo et al.,70 2021	https://github.com/bzhanglab/COSMO
GISTIC2.0	Mermel et al., 201115	ftp://ftp.broadinstitute.org/pub/GISTIC2.0/GISTIC_2_0_23.tar.gz
iProFun	Song et al., 201925	https://github.com/WangLab-MSSM/iProFun
Locfdr	Efron et al., 201571	https://CRAN.R-project.org/package=locfdr
GSVA	Hänzelmann et al., 201372	https://bioconductor.org/packages/3.21/bioc/html/GSVA.html
ESTIMATE	Yoshihara et al., 201373	https://bioinformatics.mdanderson.org/public-software/estimate/
Joint Random Forest	Petralia et al., 201674	https://rdrr.io/cran/JRF/man/JRF.html
TSNet	Petralia et al., 201875	https://github.com/WangLab-MSSM/TSNet
xCell	Aran et al., 201776	http://xcell.ucsf.edu/
CPTAC Data Viewer	Pei Wang Lab	http://pbt.cptac-data-view.org/
ConsensusClusterPlus	Monti et al., 2003,77 Wilkerson and Hayes, 201078	http://bioconductor.org/packages/release/bioc/html/CancerSubtypes.html
Strelka2 v2.9.3	Kim et al., 201879	https://github.com/Illumina/strelka
CNVkit v. 2.9.3	Talevich et al., 201680	https://github.com/etal/cnvkit
STAR v2.10a	Dobin et al., 201381	https://github.com/alexdobin/STAR
GENCODE v27	GENCODE consortium	https://www.gencodegenes.org/human/release_27.html
RSEM v1.3.1	Li and Dewey, 201182	https://github.com/deweylab/RSEM
ENSEMBL		N/A
Sumer	Savage et al., 201983	https://github.com/bzhanglab/sumer
TCGAbiolinks	Colaprico et al., 201684	https://bioconductor.org/packages/release/bioc/html/TCGAbiolinks.html
MoonlightR	Colaprico et al., 202085	https://bioconductor.org/packages/release/bioc/html/MoonlightR.html
Music	Wang et al., 201986	https://github.com/xuranw/MuSiC
Ascore v1.0.6858	Beausoleil et al., 200687	https://github.com/PNNL-Comp-Mass-Spec/AScore
MASIC	Monroe et al., 200888	https://github.com/PNNL-Comp-Mass-Spec/MASIC
MS-GF + v9981	Kim and Pevzner, 201489	https://github.com/MSGFPlus/msgfplus
mzRefinery	Gibbons et al., 201590	https://omics.pnl.gov/software/mzrefinery
DAGBagM	Chowdhury et al., 202238	https://github.com/WangLab-MSSM/dagbagM
iProMix	Song et al., 202340	https://github.com/songxiaoyu/iProMix
BayesDeBulk	Petralia et al., 202352	http://bayesdebulk.wanglab.cloud/; https://github.com/WangLab-MSSM/BayesDeBulk
BioRender	BioRender.com	http://BioRender.com
temporalCPSA	This paper	https://github.com/lashbroz/temporalCPSA

Experimental model and study participant details

The tumor and blood samples of the study patient cohort were obtained from the Children’s Brain Tumor Network (CBTN) at the Children’s Hospital of Philadelphia (CHOP) and through the Hospital of the University of Pennsylvania - Children’s Hospital of Philadelphia (HUP-CHOP) Neurosurgery Tumor Tissue Bank (NTTB) collaborative at the Hospital of University of Pennsylvania (UPENN). CBTN samples were collected under the CHOP Institutional Review Board-approved “Children’s Brain Tumor Network (CBTN) Collection Protocol” (CHOP IRB 09–007316). The cases included in this study were diagnosed with high-grade glioma tumor with available flash-frozen tumor tissue. All samples were collected at the time of surgery, freshly frozen and shipped and stored in the Biorepository Resource Center (BioRC) at Children’s Hospital of Philadelphia.

Discovery cohort consisted of 96 cases, obtained from 88 HGG patients aged between 0 and 40 years (CHOP: 87 samples from 81 patients; UPENN: 9 samples from 7 patients). 50 mg tissue pieces were cut/chipped off using disposable scalpels on dry ice and delivered to the CPTAC Biospecimen Core Resource (BCR) for sample preparation.

Validation cohort consisted of 29 samples from 29 patients aged between 3 and 23 years, with 4 of the 29 patients overlapping with the Discovery cohort (to facilitate data harmonization). Tissue samples were freshly frozen and stored in BioRC at CHOP. 50–150mg chunks of tissue pieces were delivered to the CPTAC BCR center for further sample processing. The proteogenomic data of this cohort was generated as a part of the NCI funded Kids First Proteomic X01 sequencing program.

The age, male/female label, race, ethnicity, and tumor annotations of all patients in both the Discovery and Validation cohorts were provided in Table S1. Age and male/female labels were carefully considered in all the statistical analysis.

The study was initiated with the Discovery cohort, and the additional validation cohort was subsequently included as further funding became available. Thus, assignment to the Discovery and Validation cohorts was determined by the timing of sample inclusion and funding availability rather than by biological or clinical characteristics.

Method details

Sample processing for both the discovery and the validation cohort

The flash frozen human brain tumor tissue and peripheral blood samples used in the study were coordinated through Children’s Brain Tumor Network (CBTN) consortium at the Children’s Hospital of Philadelphia and through the Hospital of University of Pennsylvania (HUP)–Children's Hospital of Philadelphia (HUP-CHOP) Brain Tumor Tissue Bank joint initiative, part of the Philadelphia Coalition for Cure (PC4C) collaboration, following ethical and technical guidelines on the use of human samples for biomedical research purposes. All patient samples were de-identified before processing. DNA and RNA quantity and quality was assessed by PerkinElmer DropletQuant UV-VIS spectrophotometer (PerkinElmer) and an Agilent 4200 TapeStation (Agilent, USA) for RNA Integrity Number equivalent (RINe) and DNA Integrity Number equivalent (DINe) measurement.

Whole genome sequencing for the Discovery cohort

Library preparation and sequencing was performed by the NantHealth sequencing center. Briefly, DNA sequencing libraries were prepared for both tumor tissue and matched-germline (blood) DNA using the KAPA Hyper prep kit (cat# KK8541, Roche); Whole genome sequencing (WGS) was performed at an average coverage of 60X for tumor samples and 30X for matched-germline. The tumor panel sample was sequenced to 470X and the normal panel sample was sequenced to 308X.

RNA-seq experiment for the Discovery cohort

Tumor RNA-Seq libraries were prepared using KAPA Stranded RNA-Seq with RiboErase kit (cat# KK8484, Roche) or with TruSeq™ Stranded mRNA Sample Preparation Kit (Illumina). RNA samples were sequenced to an average of 200M reads. All sequencing was performed on the Illumina HiSeq platform (X/400) (Illumina) or with NovaSeq 6000 platform (Illumina) with 2 × 150bp read length.

Single cell RNA-seq experiment for a small subset of the Discovery cohort

Library preparation and sequencing of single nuclei were performed using Chromium Single Cell 3′ Reagent kit v.3.1 (https://www.10xgenomics.com/support/universal-three-prime-gene-expression/documentation/steps/library-prep/chromium-single-cell-3-reagent-kits-user-guide-v-3-1-chemistry-dual-index). Agencourt RNAClean XP beads (Beckman Coulter) were used for RNA purification. Oligo-dT primed reverse transcription (RT) was performed using Maxima H Minus reverse transcriptase (Life Technologies) and a template-switching oligonucleotide (TSO; Qiagen). PCR amplification (22 cycles for snRNA-seq) was performed using KAPA HiFi HotStart ReadyMix (KAPA Biosystems), followed by Agencourt AMPure XP bead (Beckman Coulter) purification. Libraries were generated using the Nextera XT Library Prep kit (Illumina). Libraries from 768 cells with unique barcodes were combined and sequenced using a NextSeq 500/550 High Output Kit v2.5 (Illumina).

Proteomics and phosphoproteomic experiment for the Discovery cohort

Protein extraction and Lys-C/trypsin tandem digestion

Approximately 50 mg of each of frozen brain tumor tissue was cryopulverized and lysed separately in 800 μL of lysis buffer (6 M urea, 25 mM Tris, pH 8.0, 1 mM EDTA, 1 mM EGTA, 1:100 v/v Sigma protease inhibitor, 1:100 v/v Sigma phosphatase inhibitor cocktail 2, and 1:100 v/v Sigma phosphatase inhibitor cocktail 3). Lysates were sonicated in a cup horn probe (filled with ice water) three times at 50% power for 30 s, with incubation on ice for 10 s between each sonication step. Lysates were then cleared by centrifugation at 20,000 × g for 10 min at 4°C. The supernatant was transferred to a cryovial, and the protein quantitation was measured by BCA assay (Thermo). After adjusting the protein concentration to approximately 1.5 μg/μL, proteins were reduced with 5 mM dithiothreitol for 1 h at 37°C and subsequently alkylated with 10 mM iodoacetamide for 45 min at 25°C in the dark. Samples were diluted to 2 M urea concentration with 25 mM Tris, pH 8.0 and digested with Lys-C (Wako) at 1:50 enzyme-to-substrate ratio. After 2 h of digestion at 25°C, an aliquot of sequencing grade modified trypsin (Promega, V5117) at 1:25 enzyme-to-substrate ratio was added to the samples and further incubated at 25°C for 14 h. The digested samples were then acidified with 100% formic acid (FA) to 1% of the final concentration of FA and centrifuged for 15 min at 1,500 × g at 4°C before transferring samples into new tubes leaving the resulting pellet behind. After 3-fold dilution with 0.1% FA, tryptic peptides were desalted on C18 SPE (Waters tC18 SepPak, WAT054925) and dried using Speed-Vac.

TMT-11 multiplex design

We profiled 96 tumor samples using 10 11-plex tandem mass tag (TMT) experiments. For quality control, we included 3 replicates for two of the tumor samples, which resulted in 100 tumor samples in total. Each TMT multiplex consisted of 10 tumor samples and one common pooled reference sample. This reference sample was identical to the one used in the TMT experiments of the CPTAC GBM studies,^10,91 allowing us to effectively bridge the two datasets and enhance the power in the downstream analysis. Additionally, all tumor samples were randomized across the 10 TMT-plexes to avoid confounding effects from multiplex batch variations. The global, phospho and glyco-proteomics profiling used the same TMT-labeled samples using an integrated workflow with sequential enrichment of the PTMs.^91,92

TMT-11 labeling of peptides

Desalted peptides from each sample were labeled with 11-plex TMT reagents (Thermo Fisher Scientific). Peptides (400 μg) from each of the samples were dissolved in 80 μL of 50 mM HEPES, pH 8.5 solution, and mixed with 400 μg of TMT reagent that was dissolved freshly in 20 μL of anhydrous acetonitrile according to the optimized TMT labeling protocol described previously.⁹³ Channel 126 was used for labeling the internal reference sample¹⁰ throughout the sample analysis. After 1 h incubation at RT, 60 μL 50 mM HEPES pH8.5, 20% acetonitrile (ACN) solution was added to dilute the samples, and 12 μL of 5% hydroxylamine was added and incubated for 15 min at RT to quench the labeling reaction. Peptides labeled by different TMT reagents were then mixed, dried using Speed-Vac, reconstituted with 3% ACN, 0.1% FA and desalted on tC18 SepPak SPE columns.

Peptide fractionation by basic reversed-phase liquid chromatography (bRPLC)

Approximately 3.5 mg of 11-plex TMT labeled sample was separated on a reversed phase Agilent Zorbax 300 Extend-C18 column (250 mm × 4.6 mm column containing 3.5-μm particles) using the Agilent 1200 HPLC System. Solvent A was 4.5 mM ammonium formate, pH 10, 2% acetonitrile and solvent B was 4.5 mM ammonium formate, pH 10, 90% ACN. The flow rate was 1 mL/min and the injection volume was 900 μL. The LC gradient started with a linear increase of solvent B to 16% in 6 min, then linearly increased to 40% B in 60 min, 4 min to 44% B, 5 min to 60% B and another 14 min of 60% solvent B. A total of 96 fractions were collected into a 96 well plate throughout the LC gradient. These fractions were concatenated into 24 fractions by combining 4 fractions that are 24 fractions apart (i.e., combining fractions #1, #25, #49, and #73; #2, #26, #50, and #74; and so on). For proteome analysis, 5% of each concatenated fraction was dried down and re-suspended in 2% ACN, 0.1% FA to a peptide concentration of 0.1 mg/mL for LC-MS/MS analysis. The rest of the fractions (95%) were further concatenated into 12 fractions (i.e., by combining fractions #1 and #13; #3 and #15; and so on), dried down, and subjected to immobilized metal affinity chromatography (IMAC) for phosphopeptide enrichment.

Phosphopeptide enrichment using IMAC

Fe³⁺-NTA-agarose beads were freshly prepared using the Ni-NTA Superflow agarose beads (QIAGEN, #30410) for phosphopeptide enrichment. For each of the 12 fractions, peptides were reconstituted in 500 μL IMAC binding/wash buffer (80% ACN, 0.1% trifluoroacetic acid (TFA)) and incubated with 20 μL of the 50% bead suspension for 30 min at RT. After incubation, the beads were sequentially washed with 50 μL of wash buffer (1X), 50 μL of 50% ACN, 0.1% TFA (1X), 50 μL of wash buffer (1X), and 50 μL of 1% FA (1X) on the stage tip packed with 2 discs of Empore C18 material (Empore Octadecyl C18, 47 mm; Supelco, 66883-U). Phosphopeptides were eluted from the beads on C18 using 70 μL of elution buffer (500 mM K₂HPO₄, pH 7.0). Sixty μLs of 50% ACN, 0.1% FA was used for elution of phosphopeptides from the C18 stage tips after two washes with 100 μL of 1% FA. Samples were dried using Speed-Vac and later reconstituted with 12 μL of 3% ACN, 0.1% FA for LC-MS/MS analysis.

LC-MS/MS analysis

Fractionated samples prepared for global proteome and phosphoproteome analysis were separated using a nanoACQUITY UPLC system (Waters) by reversed-phase HPLC. The analytical column was manufactured in-house using ReproSil-Pur 120 C18-AQ 1.9 μm stationary phase (Dr. Maisch GmbH) and slurry packed into a 25-cm length of 360 μm o.d. x 75 μm i.d. fused silica picofrit capillary tubing (New Objective). The analytical column was heated to 50°C using an AgileSLEEVE column heater (Analytical Sales and Services). The analytical column was equilibrated to 98% Mobile Phase A (MP A, 0.1% FA/3% ACN) and 2% Mobile Phase B (MP B, 0.1% FA/90% ACN) and maintained at a constant column flow of 200 nL/min. The sample was injected into a 5-μL loop placed in-line with the analytical column which initiated the gradient profile (min:%MP B): 0:2, 1:6, 85:30, 94:60, 95:90, 100:90, 101:50, 110:50. The column was allowed to equilibrate at start conditions for 30 min between analytical runs.

MS analysis was performed using an Orbitrap Fusion Lumos mass spectrometer (Thermo Fisher Scientific). The global proteome and phosphoproteome samples were analyzed under identical conditions. Electrospray voltage (1.8 kV) was applied at a carbon composite union (Valco Instruments) coupling a 360 μm o.d. x 20 μm i.d. fused silica extension from the LC gradient pump to the analytical column and the ion transfer tube was set at 250°C. Following a 25-min delay from the time of sample injection, Orbitrap precursor spectra (AGC 4 × 10⁵) were collected from 350 to 1800 m/z for 110 min at a resolution of 60K along with data dependent Orbitrap HCD MS/MS spectra (centroid) at a resolution of 50K (AGC 1 × 10⁵) and max ion time of 105 ms for a total duty cycle of 2 s. Masses selected for MS/MS were isolated (quadrupole) at a width of 0.7 m/z and fragmented using a collision energy of 30%. Peptide mode was selected for monoisotopic precursor scan and charge state screening was enabled to reject unassigned 1+, 7+, 8+, and >8+ ions with a dynamic exclusion time of 45 s to discriminate against previously analyzed ions between ±10 ppm.

Construction and utilization of the Comparative Reference samples

As a quality control measure, two different types of “Comparative Reference” (“CompRef”) patient-derived xenograft (PDX) samples were generated as previously described^94,95 and used to monitor the longitudinal performance of the proteomics workflow throughout the course of this study. Briefly, the PDX tumors from established basal and luminal breast cancer intrinsic subtypes were raised subcutaneously in 8-week-old NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ mice (Jackson Laboratories, Bar Harbor, ME) using procedures reviewed and approved by the Institutional Animal Care and Use Committee at Washington University in St. Louis. Xenografts were grown in multiple mice, pooled, and cryopulverized to provide enough uniform material for the duration of the study. Full proteome and phosphoproteome process replicates of each of the two types of CompRef samples were prepared and analyzed as standalone 11-plex TMT experiments alongside every five TMT-11 experiments of the study samples, using the same analysis protocol as the patient samples. These interstitially analyzed CompRef samples were evaluated for depth of proteome and phosphoproteome coverage and for consistency in quantitative comparison between the basal and luminal models.

Glycoproteomics experiment for the Discovery cohort

Enrichment of glycopeptides

All unbound peptides from phosphopeptide enrichment were desalted on reversed phase C18 SPE column (Waters). The glycopeptides were enriched with OASIS MAX solid-phase extraction (Waters). The MAX cartridge was conditioned with 3 × 1 mL ACN, then 3 × 1 mL of 100 mM triethylammonium acetate buffer, followed by 3 × 1 mL of water, and finally 3 × 1 mL of 95% ACN (1% TFA). The peptides were loaded twice. The cartridge was washed with 4 × 1 mL of 95% ACN (1% TFA) to remove non-glycosylated peptides. The glycopeptide fraction was eluted with 50% ACN (0.1% TFA), dried down, and reconstituted in 3% ACN, 0.1% FA prior to ESI-LC-MS/MS analysis.

ESI-LC-MS/MS for glycoproteome analysis

The TMT-labeled glycoproteome fractions were analyzed using an Orbitrap Fusion Lumos mass spectrometer (Thermo Scientific). Approximately 0.8 μg of glycopeptides were separated on an in-house packed 28 cm × 75 mm diameter C18 column (1.9 mm Reprosil-Pur C18-AQ beads (Dr. Maisch GmbH); Picofrit 10 mm opening (New Objective) lined up with an Easy nLC 1200 UHPLC system (Thermo Scientific). The column was heated to 50°C using a column heater (Phoenix-ST). The flow rate was set at 200 nL/min. Buffer A and B were 3% ACN (0.1% FA) and 90% ACN (0.1% FA), respectively. The peptides were separated with a 6%–30% B gradient in 84 min. Peptides were eluted from the column and nanosprayed directly into the mass spectrometer. The mass spectrometer was operated in a data-dependent mode. Parameters for glycoproteomic samples were set as follows: MS1 resolution - 60,000, mass range – 500 to 2000 m/z, RF Lens – 30%, AGC Target – 5.0e5, Max injection time – 50 ms, charge state include – 2–6, dynamic exclusion – 45 s. The cycle time was set to 2 s, and within this 2 s the most abundant ions per scan were selected for MS/MS in the orbitrap. MS2 resolution – 50,000, HCD activation energy – 35, isolation width (m/z) – 0.7, AGC Target – 1.0e5, Max injection time – 100 ms.

MRM experiment for the Discovery cohort

MRM assay development

Heavy cleavable and tryptic stable isotope-labeled (SIS) peptides were synthesized by Biosynth (Gardner, MA) with a fully atom-labeled 13C and 15N on the tryptic C-terminal arginine (R) or lysine (K) residue (if available; otherwise, the incorporation occurred at another amino acid residue (leucine (L), alanine (A), or phenylalanine (F)). Cleavable SIS peptides incorporated additional 3–6 amino acids native to the protein on both C- and N-terminal ends of the tryptic sequence. All SIS peptides were obtained at purity >95%, resuspended in 5% ACN with 1% FA and stored at −80°C. Optimum transitions were selected using synthetic peptides and collision energies taken from Skyline targeted proteomics software. The final method consisted of 121 peptides to 59 proteins.

Sample processing for MRM

500 μL of lysis buffer (25 mM Tris, 6 M Urea, 1 mM EDTA, 1 mM EGTA, 1% (v/v) protease inhibitor, 1% (v/v) phosphatase inhibitor cocktail 2, and 1% (v/v) phosphatase inhibitor cocktail 3) was added to the cryopulverized tissue and the sample was vortexed for 10–15 s and sonicated three times in a cup horn probe (filled with ice water) at 50% power for 30 s. The samples were stored in the vapor phase of liquid nitrogen until digestion. Samples were thawed and transferred to a deep-well plate for processing on an EpMotion 5057 (Eppendorf). 200 fmol of a mix of extended stable isotope-labeled peptide standards was added to each sample and then each sample was reduced in 30 mM TCEP for 30 min at 37°C with shaking, followed by alkylation with 50 mM IAM at room temperature. Samples were then diluted with 0.8 mL 200 mM TRIS before Lys-C endopeptidase was added at a 1:50 enzyme:substrate ratio by mass and incubated for 2 h at 37°C with mixing at 600 rpm (Thermomixer, EpMotion 5057). After 2 h, trypsin was added at a 1:50 enzyme:substrate ratio. Digestion was carried out overnight at 37°C with mixing at 600 rpm. After 16 h, the reaction was quenched with FA (final concentration 1% by volume). Following FA quench of protease digestion, samples were desalted using a positive pressure manifold using an Oasis HLB μElution 96-well plate (30 μm, 2mg, Waters, 186001828BA). Oasis plates were first conditioned with 50:50 LC-MS ACN:water with 0.1% FA and then equilibrated with 100% LC-MS water with 0.1% FA prior to loading samples. Digested samples were loaded into the Oasis wells and washed with 100% water/0.1% FA. Peptides were eluted from Oasis wells (<6 psi) using 50:50 ACN:water with 0.1% FA, lyophilized and stored at −80°C.

LC-MRM analysis

Samples were resuspended in 10 μL of 2% ACN/0.1% FA and injected onto the column of an Eksigent Ultra 425 nanoLC system equipped with a chipFLEX system. Sample injection volume was 1 μL. Peptide separation was performed by reversed-phase chromatography using a cHiPLC ReproSil C18 chip column (3 μm particles, 15 cm × 75 μm; temperature 45°C). A flow rate of 300 nL/min was used over a gradient of mobile phase A (0.1% formic acid) and mobile phase B (90% acetonitrile, 0.1% formic acid in water). The separation of peptides started at 2% B for 0–4 min and then ramped from 5 to 40% B over the course of 45 min. The gradient then switched to 90% for 5 min followed by a ramp back to starting conditions of 2% B for 5 min. The Eksigent nanoLC was coupled to QTRAP 6500 hybrid triple quadrupole/ion trap mass spectrometer (SCIEX) equipped with an OptiFlow source and SteadySpray nano probe (SCIEX). All targeted LC-MRM analyses were carried out in positive ion mode at a spray voltage of 3000 V. The source curtain gas was 20 and interface heating temperature set to 150°C. Transitions were scheduled with a retention time window of 300 s.

Data preprocessing

MRM data were analyzed by Skyline.⁹⁶ Peak integrations were reviewed manually, and transitions from analyte peptides were confirmed by the same retention times and relative transition areas of the light peptides and heavy stable isotope-labeled peptides. Transitions with detected interferences were not used in the data analysis. Integrated raw peak areas were exported from Skyline and total intensity was calculated using Peak Area + Background. Transitions were summed for each light/heavy pair and peak area ratios were obtained by dividing peak areas of light peptides by that of the corresponding heavy peptides. All measurements were filtered by the LLOQ.

SRM experiment for the Discovery cohort

Stable isotope-labeled peptides

Proteotypic peptides for the target proteins were selected for targeted proteomics analysis based on the TMT identification results and well-accepted criteria,⁹⁷ and the corresponding crude heavy stable isotope-labeled peptides were synthesized with ¹³C/¹⁵N on C-terminal lysine or arginine (New England Peptide, Gardner, MA). The heavy peptides were dissolved individually in 15% ACN and 0.1% FA at a concentration of 2 mM and used for creating a SIS peptide mixture with a final concentration of 5 μM for each peptide.

SRM assay development

The heavy peptides in the peptide mixture were evaluated for peptide response and fragmentation pattern using LC-SRM. For each peptide, transition settings were as follows: (1) precursor charges: 2, 3 and 4; (2) fragment ion charges: 1, 2 and 3; (3) ion type: y ions; and (4) m/z window: 250–1500. Transition lists were generated with optimal collision energy values by Skyline software110. LC-SRM was then used to evaluate all heavy peptides for stability of peptide retention time, reliable heavy peptides identification, transition interferences, and endogenous peptide detectability. In the end, 3 transitions per peptide were selected for the final assay configuration for targeted quantitation of a total of 108 peptides of 79 target proteins.

LC-SRM data acquisition

For direct analysis, the digested tissue samples were reconstituted in 2% ACN/0.1% FA and spiked with 50 fmol/μL heavy peptides for a final concentration of 0.25 μg/μL, and 2 μL of the resulting samples were analyzed by LC-SRM using a Waters nanoACQUITY UPLC system coupled to a Thermo Scientific TSQ Altis triple quadrupole mass spectrometer (Thermo Scientific). A 100 μm i.d. × 10 cm, BEH 1.7-μm C18 capillary column (Waters) was operated at a temperature of 44°C. The mobile phases were (A) 0.1% FA in water and (B) 0.1% FA in ACN. The peptide samples were separated at a flow rate of 400 nL/min using a 110-min gradient profile as follows (min:%B): 7:1, 9:6, 40:13, 70:22, 80:40, 85:95, 93:50, 94:95 and 95:1. Data were acquired in time-scheduled SRM mode (retention time window: 15 min). The parameters of the triple quadrupole instrument were set with 0.7 FWHM Q1 and Q3 resolution, and 1.2-s cycle time.

SRM quantification

SRM data were analyzed using the Skyline software.⁹⁶ The total peak area ratios of endogenous light peptides and their heavy isotope-labeled internal standards (i.e., L/H peak area ratios) were exported for quantitation. Peak detection and integration were carried out according to two criteria: (1) same retention time and (2) similar peak area ratios for the transitions. All data were manually inspected to ensure correct retention time, peak detection and accurate integration. All SRM results including the assay characterization data are organized as Skyline files on the Panorama server^98,99 and can be accessed via https://panoramaweb.org/CPTAC_GBM_AYA.url.

DIA proteomics experiment for the validation cohort

All tissue samples were cryo-pulverized and lysed according to previously published CPTAC protocols. ¹EvosepOne LC system (Evosep Biosystems) coupled with timsTOF-HT (Bruker) mass spectrometry was used for the data-independent acquisition mass spectrometry (DIA-MS). After tryptic digestion, 1 μg peptides were loaded on Evotip according to the EvosepOne protocol. A PepSep C18 column of 15 cm × 150 μm (1.5 μm, Bruker) was used for peptide separation of 30 samples per day at 50°C with 44 min gradient length, 0.5 μL/min flow rate, LC gradient from 3% to 35 (mobile phase A is 0.1% FA in HPLC water and mobile phase B is 0.1% FA in ACN).The global proteomic data was acquired using the DIA-PASEF mode on timsTOF-HT (Bruker) with settings as follows: MS1 scan range of 100–1700 m/z; MS2 scan range of 338–1338 m/z, mass width 25.0 Da without mass overlap, 1 mobility window, 1/K0 range of 0.70–1.45 V.s/cm², ramp time 85.0 ms.

RNAseq experiment for the validation cohort

Tumor RNA-Seq libraries were prepared using KAPA Stranded RNA-Seq with RiboErase kit (cat# KK8484, Roche) or with Illumina TruSeq™ Stranded mRNA Sample Preparation Kit Illumina). RNA samples were sequenced to an average of 200M reads. All sequencing was performed on the Illumina HiSeq platform (X/400) (Illumina) or with NovaSeq 6000 platform (Illumina) with 2 × 150bp read length.

Quantification and statistical analysis

Genomic data preprocessing

Somatic mutation calling

Since not all tumor samples had a paired normal available, all tumor WGS data were run through the following tumor only workflow (https://github.com/kids-first/kf-tumor-workflow/tree/v0.4.1-beta) which utilizes Mutect2 for SNV and InDel calling. Samples with both tumor and normal WGS were run through a four SNV/InDel variant caller workflow (https://github.com/kids-first/kf-somatic-workflow), followed by hotspot annotation and consensus calls. Mutect2 v4.1.1.0, Strelka2 v2.9.3, VarDict Java v1.7.0, and Lancet v1.0.7 were used and consensus calls were made if the variant was called by two of the four algorithms. Variants were annotated with VEP v105. Additional details on variant calling can be found in the OpenPedCan manuscript.¹⁰⁰

Additional filtering of somatic mutations

Somatic mutations were required to be classified as one of the following categories (“Variant_Classification") for a specific gene: Missense_Mutation, Splice_Region, Frame_Shift_Del, Splice_Site, Nonsense_Mutation, In_Frame_Del, In_Frame_Ins, Frame_Shift_Ins, Translation_Start_Site, Nonstop_Mutation. Additionally, the sequencing depth at each mutation locus (“t_depth") was required to be at least 30, and the mutated allele count at least 5 (“t_alt_count"). For samples that had no normal tissue paired, “HotSpotAllele" = 1 was required. If a gene had at least one mutation that passed these filters, it was considered a mutated gene.

Mutation table derivation for the cDiscovery cohort

As noted in the result section, we refer to the combined datasets from our Discovery cohort and that from the proteogenomic GBM study¹⁰ as the cDiscovery (combined-Discovery) cohort data.

After the derivation of the gene-level mutation table (0/1) for the Discovery cohort, we focused on the 30 genes with mutations present in at least 5 tumors for further analysis.

To derive the mutation data for the cDiscovery cohort, we obtained the mutation data of these 30 genes from Wang et al.¹⁰ and combined it with our Discovery cohort data. Additionally, we included mutation information for ‘PTEN’, a gene frequently mutated in older GBMs, in the combined mutation table.

DNA copy number calling and segmentation

Copy number calls were generated using Control-FREEC v11.6 (tumor/normal and tumor only) and CNVkit v0.9.3 (tumor/normal). Code can be found in the workflows above.

Copy number variation segmentation

CNV analysis (i.e., segmentation, purity and ploidy estimate and absolute copy number calls) was performed using CNVEX.^46,101 Briefly, CNVEX estimates coverage within fixed genomic intervals and calculates Log2-ratio values utilizing coverage from tumor and matched normal samples. Further, CNVEX uses variant calls to compute B-allele frequencies (BAFs) at germline variant positions. Log2-ratio values are then normalized for GC-bias using LOESS smoothing across targeted regions within the GC range of 0.3 and 0.7, using the span = 0.5. All the GC normalized log2-ratio values and BAFs are then jointly segmented using a custom algorithm based on Circular Binary Segmentation (CBS).¹⁰² The resulting segmentation track was iteratively ‘‘pruned’’ by merging segments that had similar Log2-ratios and BAFs, short lengths, enrichment in blacklisted regions, and a high variation in coverage among whole cohort germline samples. Given the relatively high variance of the sequencing data in this cohort, we noticed that some samples were over-segmented using the default settings. To adjust the segmentation algorithm to the variance of each sample individually, we utilized the method previously mentioned in Chowdhury et al.⁴⁶ and performed an adaptive procedure of pruning. In this setting, we defined 3 types of pruning. Default pruning: Perform pruning as described above for all samples. Strict pruning: Count the number of segments for each sample, take the samples with the top 20th percentile number of segments and perform pruning with stricter parameters (i.e., more pruning which leads to higher rate of adjacent segments merging). Relaxed pruning: Take the samples with the lowest 20th percentile number of segments and perform more loose pruning on them (i.e., less pruning which leads to lower rate of adjacent segments merging). The above procedure was repeated two times, effectively adjusting segmentation pruning parameters to the noise level of each sample. The quality of the resulting segmentations was manually evaluated. The resulting segmented copy-number profiles were then subject to joint inference of tumor purity and ploidy and absolute copy number states, implemented in CNVEX, which is most similar to the mathematical formalism of ABSOLUTE¹⁰³ and PureCN.¹⁰⁴ Because the copy-number inference problem can have multiple equally likely solutions, further biological insights are necessary to choose the most parsimonious result. The solutions were reviewed by two independent field experts and the most likely solution was selected.

Deriving gene-level CNV data

Copy number variation on genome segments was further aggregated to gene-level. For genes containing multiple segments, the weighted average of the variation was calculated as the gene-level summary. Weights of the segments in each gene were proportional to the reverse of the segment length.

CNV data from the GBM cohort¹⁰ were processed and aggregated to the gene-level using the same genome reference. To assemble the CNV data for the cDiscovery cohort, CNV data from the Discovery cohort and GBM cohort were merged by genes.

RNAseq preprocessing

Alignment and quantification of RNAseq data of the Discovery and Validation cohort

For both the Discovery and Validation cohort data, STAR⁸¹ v2.7.10a was used in two-pass mode to align paired-end RNA-seq reads against the GRCh38 human genome and annotation was performed using the ENSEMBL GENCODE 39 gene annotation reference (https://www.gencodegenes.org/human/release_39.html). RSEM⁸² v1.3.1 was then used to calculate FPKM and TPM values for both transcript- and gene-level expression. Additional pipeline details can be found at (https://github.com/kids-first/kf-rnaseq-workflow).

Normalization and harmonization of RNAseq data for the cDiscovery cohort

We performed normalization and batch effect correction to harmonize the RNAseq data from our Discovery cohort and the GBM cohort from Wang et al.¹⁰ Median TPM counts of each sample were aligned to the cohort median level by rescaling each sample accordingly. Log2(x+1) transform was then applied to the rescaled count data for both cohorts. To align baseline differences between the two cohorts we further applied batch effect correction using ComBat.⁶⁹ Age group index (0–15, 15–26, 26–40, and ≥40) was adjusted during the batch effect correction.

Preprocessing of RNAseq data of the validation cohort

FPKM data for 61,587 isoforms corresponding to 60,324 genes were obtained for 20 out of 29 samples in the Validation cohort. For two samples with technical replicates, average FPKM values were used. FPKM values were then Log2(x+1)-transformed, and isoforms with more than 50% missing data were filtered, leaving 25,908 isoforms. Gene-level data were derived by averaging across isoforms, resulting in RNA abundance data for 25,845 genes. Finally, global normalization via median-centering was performed.

Detection of gene-fusion based on RNAseq data

RNA fusions were called using STAR-Fusion v1.10.1 and Arriba v2.2.1 and quality-control filtered using annoFuse 0.92.0¹⁰⁵ using the following workflow: https://github.com/kids-first/kf-rnaseq-workflow.

Methylation preprocessing

Derivation of probe level beta-values

Methylation data were obtained from the OpenPedCan⁷⁵ v11 data release (v2.0.0 repository release). Briefly, raw Illumina 450K and EPIC 850K Infinium Human Methylation Bead Array intensities were processed using the minfi Bioconductor package¹⁰⁶ using the following module: https://github.com/d3b-center/OpenPedCan-analysis/tree/dev/analyses/methylation-preprocessing. Probes containing common SNPs in dbSNP (minor allele frequency >1%) were removed if present at CpG interrogation sites or single nucleotide extension sites.

Gene-level aggregation of DNAm

DNA methylation levels for both our Discovery cohort and the GBM cohort from Wang et al.¹⁰ were quantified using Illumina’s EPIC methylation array. Probe genomic location coordinates for the EPIC methylation array were obtained from IlluminaHumanMethylationEPICanno.ilm10b2.hg19. To convert genome coordinates from hg19 to hg38 genome assemblies, UCSC’s LiftOver tool was utilized https://genome.ucsc.edu/cgi-bin/hgLiftOver. Genic and CpG-related annotations were updated based on hg38 genomic coordinates using the CpG island tracks from AnnotationHub; the genic annotations come from the txdb.hsapiens.ucsc.hg38 and org.Hs.eg.db packages. Entrez IDs were mapped to study-wide ApprovedGeneSymbols. Methylation data from our Discovery and GBM cohorts were then aggregated separately to the gene level. Only probes mapping to a CpG island, and to either a 5′UTR or associated with a promoter were utilized. Probes with a missing rate >0.9 were filtered out prior to gene-level aggregation. Gene-level methylation was derived by taking the average beta value over all probes mapping to the same ApprovedGeneSymbol.

Harmonization of gene-level methylation for the cDiscovery cohort

Methylation data (M-values) of our Discovery cohort and the GBM cohort from Wang et al.¹⁰ were combined and corrected for batch effects (cohort differences). To prevent outlier samples with abnormally high methylation M-values in the GBM cohort from being skewed by the batch correction process, we first identified the Extremely Hypermethylated sample group using hierarchical clustering based on Euclidean distance. By setting the cluster number to 2, we identified 8 Extremely Hypermethylated samples (C3N-00662, C3N-01505, C3N-01368, C3L-01046, C3N-01814, C3N-01815, C3N-02188, C3N-02788).

Batch effect correction was performed on the combined data matrix, excluding these samples, using the ComBat method⁶⁹ with adjustment to the age groups (0–15, 15–40, and >40). After batch correction, the Extremely Hypermethylated samples were realigned to the corrected data matrix. Since all Extremely Hypermethylated samples belonged to the same age group (>40), we applied the same linear transformation used in batch correction within this age group to adjust their methylation values for the corresponding genes.

The final harmonized dataset was obtained by merging the batch-corrected data matrix for non-hypermethylated samples with the linearly transformed data matrix for the Extremely Hypermethylated samples.

Proteomics data pre-processing

Global- and phospho-proteomics TMT data pre-processing (cDiscovery cohort)

To ensure proper data integration between our Discovery cohort and the GBM cohort from Wang et al.,¹⁰we performed harmonized database search for both datasets simultaneously.

Database search and quantification

Proteomic data analysis was conducted using the Philosopher pipeline version 3.3.12,⁶⁴ as previously described.^107,108 The Discovery cohort data and the GBM cohort¹⁰ were processed independently, and their quantification matrices were combined at the conclusion of the analysis.

For each cohort, MS/MS spectra were searched with MSFragger version 3.1.1⁶³ using the CPTAC3 harmonized protein sequence database, which included decoy sequences for Validation (2019-09-26-td-RefSeq.20180629_Human_ucsc_hg38_cpdbnr_mito_264contams.fasta). The whole proteome data analysis employed a precursor ion and fragment mass tolerance of 20 ppm, while permitting C12/C13 isotope errors. Fixed modifications were set for cysteine carbamidomethylation, lysine TMT labeling, and peptide N-terminal TMT labeling. Variable modifications involved methionine oxidation, protein N-terminal acetylation, and serine TMT labeling. The analysis focused on tryptic peptides, allowing up to two missed cleavage sites. Additionally, phosphorylation of serine, threonine, and tyrosine residues was considered for phosphopeptide-enriched data, excluding serine TMT labeling.

The MSFragger results were processed with PeptideProphet⁶⁶ to calculate the probability of correct peptide-to-spectrum matches (PSM). For phosphopeptide-enriched sets, PTMProphet⁶⁷ was utilized to identify phosphorylation sites. Results from PeptideProphet (or PTMProphet) across TMT-plexes were assembled into protein groups using ProteinProphet,⁶⁸ creating a consolidated protein group file. This file and individual PSM lists from each TMT 11-plex were refined using Philosopher’s filter function. Peptides unique to a protein group or assigned as razor peptides to the group with maximal evidence were identified. Protein groups underwent filtering to maintain a 1% protein-level False Discovery Rate (FDR) using a target-decoy approach and best peptide strategy. PSM lists within each TMT 11-plex were similarly screened, retaining only those with a PeptideProphet probability of 0.9 or higher and mapping to proteins within the 1% protein-level FDR threshold. MS1 precursor ion intensity for each passing PSM were obtained through Philosopher’s label-free quantification function.

For TMT-labeled peptides, eleven TMT reporter ion intensities were extracted from MS/MS scans within a 0.002 Da window. Precursor ion purity scores were derived from the the balance of sequenced precursor to interfering ions in MS1 data, using a 0.7 Da isolation window. All significant PSM data, including protein/gene accessions and quantification info, were compiled into PSM.tsv files per TMT 11-plex.

TMT-Integrator⁶⁵ processed PSM.tsv files to generate reports at different levels (gene, peptide, protein for both global and phosphopeptide data, and additional site reports). Criteria for PSM inclusion involved (1) presence of a TMT label, (2) reference channel intensity, (3) precursor ion purity over 50%, (4) reported ion intensity not within the lowest percentiles, and (5) peptide phosphorylation in phosphopeptide data. Only the PSM with the highest TMT intensity from redundant records was retained. PSMs linked to contaminants were excluded, employing unique and razor peptides for quantification. Log2 transformation and normalization of reporter ion intensities relative to the reference channel (in both the Discovery and the Wang et al. cohort, the same GBM pooled sample was used as reference sample) were performed, producing log2-based ratios. PSMs were grouped by predefined levels, outlier removal was carried out using the interquartile range method, and remaining ratios were median-centered. Finally, normalized ratios were reconverted to abundances via the weighted sum of the MS1 intensities of the top three peptide ions by intensity.

In the resulting protein/phosphosite abundance table of the Discovery cohort, the number of proteins and phosphosites measured per sample for tumors ranged from 8296 to 9496 (median, 9134) and 8502 to 23298 (median, 19351), respectively. In total, we identified and quantified 11027 proteins and 53164 phosphosites. In the proteomics data from the GBM study by Wang et al.,¹⁰ the number of proteins and phosphosites measured per sample for tumors ranged from 9961 to 10873 (median, 10509) and 31424 to 39092 (median, 36899), respectively, and 12213 proteins and 88635 phosphosites were identified in total.

For phosphoproteomics data, we used both site-level and gene-level abundance tables. The gene-level phosphorylation data was calculated by taking the average abundance values across all sites mapping to the same ApprovedGeneSymbol for each sample.

Normalization, outlier removal, batch-correction and imputation

For both the Discovery cohort and the GBM study,¹⁰ we first applied sample median normalization and TMT-multiplex outlier detection/removal for each dataset (proteomics and phosphoproteomics) separately. Specifically, to detect the TMT-multiplex outliers, we applied an inter-TMT-plex t test on every protein or phosphosite for each cohort separately, as performed in our previous work.⁹ TMT-plexes with statistically significant measurement differences were removed from the datasets. No outlier TMT-plex was identified in the Discovery cohort, while 10 genes and 1 gene in GBM proteomic/phosphoproteomic data, respectively. The observations within the outlier TMT-plex of corresponding genes was removed (and marked as missing in the data matrix).

We then harmonized the proteomic and phosphoproteomic datasets between the two cohorts. Since the same internal reference sample was used in the TMT-experiments of both studies, we aligned the abundance levels of the two studies based on the measurements of these common reference samples. Specifically, for each data type (global or phosphoproteomics), we considered the common features detected in both studies and fitted linear transformations between the reference sample measurements in our Discovery cohort data and those in the GBM study. We then applied the inferred linear transformation to the tumor sample measurements.

Following the above adjustment, we applied batch effect correction to the combined proteomic and phosphoproteomic data to remove the technical variation across different TMT-multiplexes and studies. The ComBat method⁶⁹ was applied with adjustments for histology subtypes and age groups (0–15, 15–40, and >40).

To impute missing values in the datasets, we employed an ensemble imputation algorithm: DreamAI (https://github.com/WangLab-MSSM/DreamAI).¹⁰⁹ Imputation was done separately for each study after TMT-plex batch correction, focusing on markers (proteins or phosphosites) with less than 50% missing data.

The imputed proteomic and phosphoproteomic data were further annotated with HGNC-approved gene symbols (2022–06 version). In the global proteomic data, measurements for proteins mapped to the same gene symbol were aggregated using the average. In the end, 1533 phosphosites from ischemia-induced proteins¹¹⁰ were excluded from the phosphoproteomics data to avoid any artificial effect induced by variations in sample collection.

MRM/SRM data preprocessing for the Discovery cohort

SRM data were first log10-transformed. For both MRM and SRM data, outliers above and below 4 times the interquartile range (IQR) from the median for each sample were truncated to their respective limits. SRM/MRM data were aggregated to the gene-level by taking the mean value across sequences mapping to the same protein or protein variant.

Glycoproteomics data pre-processing for the Discovery cohort

Database search and quantification for glycoproteomic data

Glycoproteomic raw data files were converted to universal format mzML files using the msconvert tool from ProteoWizard and searched using the GPQuest search engine (version 3) with the following modifications: dynamic oxidation (+15.9949 Da) on Met, and static carbamidomethylation (+57.021464 Da) on Cys residues. GPQuest was applied to identify intact N-linked glycopeptides from MS/MS spectra using two approaches: searching spectra containing oxonium ions (‘oxo-spectra’) and identifying intact N-linked glycopeptides. The oxonium ions were used as the signature features of the glycopeptides from the MS/MS spectra, which result from fragmentation of glycans attached to intact glycopeptides in the mass spectrometer. In this study, the MS/MS spectra contained the oxonium ions (m/z 204.0966) in the top 10 abundant peaks (N-linked glycopeptide search). The intact N-linked glycopeptides were identified by using GPQuest to search against the GENCODE protein sequence database (V42) containing 58,597 proteins and a glycan database collected from the public human N-linked glycan database containing 695 glycan compositions.¹¹¹ Each tandem mass spectrum was first processed in a series of preprocessing procedures, including removing reporter ions, spectrum de-noising, oxonium ions evaluation, deisotoping and glycan type prediction.¹¹² The top 300 peaks in each preprocessed spectrum were matched to the fragment ion index generated from a peptide sequence database to identify all the candidate peptides. All the qualified (≥6 fragment ion matchings) candidate peptides were compared with the spectrum again to calculate the Morpheus scores¹¹³ by considering all the peptide fragments, glycopeptide fragments, and their isotope peaks. The peptide with the highest Morpheus score was then assigned to the spectrum. The mass gap between the assigned peptide and the precursor mass was searched in the glycan database to find the associated glycan. The best hits of all ‘oxo-spectra’ were filtered by precursor isotopes distribution fitting score and then ranked by the Morpheus score in descending order, in which those with glycopeptide-spectrum matches (GPSMs) FDR <1% and covering of >10% total intensity of each tandem spectrum were reserved as qualified identifications. The precursor mass tolerance was set as 10 ppm, and the fragment mass tolerance was 20 ppm.

Peptide-spectrum matches (PSMs) were quantified using the MS-PyCloud proteomics pipeline (https://bitbucket.org/mschnau1/ms-pycloud/src/main/). TMT correction factors were applied to correct the MS2 intensity of each PSM. Only fully-tryptic peptides with up to two missed cleavages were retained. Glycopeptide (peptide + glycan) false discovery rate (FDR) was restricted to ≤1% by applying a PSM-level FDR filter of ≤0.05%, requiring a minimum of two PSMs per peptide, and a minimum of one peptide per protein. Sample PSM log2 ratios were calculated for each PSM relative to the PSM measurement in the pooled reference from that sample’s TMT-plex. PSM log2 ratios were then rolled up to glycopeptide-level by taking the median of PSMs that map to the same glycopeptide. The glycopeptide log2 ratio matrix was then median normalized across all samples. The glycopeptide abundance matrix was derived from the log2 ratio matrix by adding the median log2 value of all TMT-plex pooled reference summed MS2 intensities to each sample log2 ratio for a given glycopeptide.

QC filtering, normalization, batch-correction and imputation

In the resulting data of 10 TMT-plexes, the 5th and 8th TMT-plexes showed particularly low identification of glycopeptides (2742 and 2975 respectively) and were removed from the downstream analysis. Among the remaining 8 TMT-plexes, the number of glycopeptides identified and quantified per sample ranged from 3801 to 9036, and in total 20931 glycopeptides were quantified for the entire cohort. We then filtered away low abundance observations (less than log2(50)) from the data matrix, to avoid noisy measurement of low-abundance glycopeptides.

After the data quality filtering, we applied TMT-outlier detection (same procedure applied on proteomics data) and TMT-plex/batch effect correction to remove the technical variation. No TMT outlier was detected for the data matrix. ComBat⁶⁹ was used to remove TMT-plex/batch effect with the batch index defined by TMT-plex index.

To impute missing values, we retained 3,440 glycopeptides with at least 25% observed values within each batch and at least 50% observed overall. Missing values were then imputed using DreamAI.¹⁰⁹

Additionally, 2 samples (7316–6283, non-HGG histology; 7316–2751, sample labeling error) were removed from glycoproteomics data prior to downstream analysis. Details of histology QC and sample labeling error checks are provided below.

In the final preprocessed glycopeptide data, one tumor had 2 replicates. The Pearson correlation coefficient for this replicate pair was 0.839, indicating good reproducibility.

Global proteomics data preprocessing for the validation cohort

A total of 446 samples from nine diagnostic subtypes were measured by DIA experiments. The number of proteins measured in each sample ranged from 2506 to 11884 (median:10862.5), with 14,231 proteins measured overall.

Raw protein abundances were first rescaled across samples to align each sample median with the global median in the dataset. A set of 366 blood related proteins were excluded from the rescaling factor calculation. Following median alignment, a log2(x+1) transformation was applied to the raw abundances values for each sample.

Samples with severe ischemia contamination were removed, including one Craniopharyngioma tumor (running order: 303), one Medulloblastoma tumor (running order: 253), and one Choroid Plexus Papilloma tumor (running order: 152).

Missing values were imputed separately within each diagnostic subtype using the DreamAI method.¹⁰⁹ Prior to imputation, proteins with >50% missing values within each subtype were excluded.

Male/female annotation, histology QC and omics-data sample labeling check for the discovery cohort

A total of 96 samples were collected in this Discovery cohort. Male and female annotations were determined using genomics data. Specifically, we utilized expression patterns of chromosome-linked marker genes, including expression of DDX3Y and RPS4Y1 from the Y chromosome and XIST from the X chromosome to perform the classification.

Histology QC were performed manually by clinicians in the team. Two samples (7316–4065, 7316–6283), corresponding to a single patient whose tumor did not meet the criteria for HGG classification, were excluded.

We then applied the COSMO⁷⁰ tool to systematically assess sample labeling by evaluating consistency across four omics data types: global proteomics, phosphoproteomics, RNA-seq, and CNV. Some samples showed inconsistencies across multi-omics profiles. As a result, we excluded (1) all omics data for sample 7316–2751, an autopsy sample from a patient whose surgical sample data were retained in the study, and (2) proteomics and phosphoproteomics data for samples 7316–3000 and 7316–1106.

Handling multiple tumor samples per patient in the discovery cohort

In the Discovery cohort, for six patients, multiple tumor samples were collected at different time points (one patient with three samples and five with two). Among these samples, about half of the primary-recurrent tumor pairs differed in protein subtype assignments and/or overall protein landscape, echoing the primary-recurrent heterogeneity we previously reported.⁹ To maximize the use of these replicates while minimizing potential confounding factors, we applied analysis-specific inclusion criteria as follows.

• •Age-dependence of genomic alterations (Figure 1): all available specimens per patient were retained to capture developmental trends.
• •Temporal expression profiling (Figure 2): only sequentially proximate samples were included, under the assumption that recurrence-associated variability is smaller than the overarching developmental signal.
• •Cis- and trans-regulation modeling (Figure 3) and immune-landscape characterization (Figure 7): each tumor event (initial vs. recurrent) was treated as a separate sample, reflecting distinct mutation, copy-number, and microenvironmental profiles.
• •Survival analyses (Figure 4) and the causal kinase analysis (Figure 5): only the first diagnosed specimen per patient was used, with kinases selected based on those survival results.
• •Protein-subtype analysis (Figure 6): surgical (non-postmortem) samples from five patients with multiple tumor events were included to assess clustering heterogeneity (noting subtype discordance in two cases), while survival comparisons remained restricted to each patient’s first diagnostic sample.

Gene symbol alignment across datasets

To facilitate integrative analysis across different data modalities, we harmonized gene symbols across all datasets to HUGO-approved gene symbols (updated 2022–06). Genomic data (RNAseq, CNV, mutation, methylation, gene fusion) were aligned by HGNC ID annotation in the GENCODE v39 database. Proteomics data (g;obal proteomics, phosphoproteomics, and glycoproteomics), as well as gene names in pathway databases, were aligned by gene symbols directly.

Comprehensive pathway database for enrichment analysis

Multiple databases, including HALLMARK (N = 50), KEGG (N = 618), REACTOME (N = 1691), BIOCARTA (N = 292), GOBP (N = 7643), and GOMF (N = 1767) (version v2023.2) were downloaded from GSEA (https://www.gsea-msigdb.org/gsea/msigdb/index.jsp). We also utilized a mitochondrial specific database (MitoCarta3.0, N = 149) obtained from Broad Institute (https://www.broadinstitute.org/mitocarta/mitocarta30-inventory-mammalian-mitochondrial-proteins-and-pathways). These databases were consolidated into a single study-pathway database provided along with the processed data files (see data and code availability section), which was utilized for our analyses.

In each pathway enrichment analysis, we focused on the subsets of pathways/gene-sets whose overlap with the interrogating geneset contained strictly more than 5 and fewer than 250 genes. Following this selection based on overlapping criteria, redundant gene sets were eliminated, retaining only the first pathway as listed in the study database. We applied FDR correction on a per-database basis using the Benjamini–Yekutieli method,¹¹⁴ as the pathway databases vary in size.

Molecular subtyping of HGG for the discovery cohort

AYA HGG tumors with methylation data were classified using the DKFZ v12.5 brain classifier⁶ and integrated into final molecular subtypes using the World Health Organization 2021¹³ classification criteria as per OpenPedCan.¹⁰⁰ Briefly, tumors with high-confidence methylation classification (score ≥ 0.8) for histone H3 K28 mutation or containing H3-3A K28M, H3C2 K28M, H3C3 K28M, H3C14 K28M, H3-3A K28I, H3C2 K28I, H3C3 K28I or H3C14 K28I mutations were classified as “DMG, H3 K28”. Tumors with high-confidence methylation classification (score ≥ 0.8) for histone H3 G35 mutation or H3-3A G35V or G35R mutation were classified as “HGG, H3 G35”. Tumors with IDH1 R132 mutation were classified as “HGG, IDH”. If tumors had a “pathology_free_text_diagnosis” containing “infant type hemispheric glioma” or had high-confidence methylation cns_methylation_subclass = = “IHG", they were classified as “IHG, followed by the defining fusion (e.g.,: NTRK, ROS1, ALK, MET). If tumors had a “pathology_free_text_diagnosis” suggestive of a pleomorphic xanthoastrocytoma diagnosis or had a high-confidence (score ≥ 0.8) methylation_subclass = = “PXA” and had a pathogenic BRAF mutation or other MAPK pathway gene alteration with homozygous deletion of CDKN2A and/or CDKN2B, they were classified as “PXA". Any tumors with WGS that did not meet any of these criteria were marked as “HGG, H3 wildtype” (i.e., there was no canonical variant the DNA or methylation classification subtype). Remaining samples with no DNA or methylation were categorized as “HGG, To be classified”. TP53 status appended to subtypes were based on the results from ‘tp53_nf1_score` module. Code for the subtyping module can be found in the project repository: https://github.com/d3b-center/hope-cohort-analysis/tree/master/analyses/molecular-subtyping-HGG.

Telomere content calculations

TelomereHunter¹¹⁵ was used to extract telomere reads using the default filtering criteria (non−consecutive repeats, mapq threshold = 8, repeat types = TTAGGG, TGAGGG,TCAGGG, TTGGGG). TelomereHunter provides telomere content as: the number of intratel_reads ∗ 1,000,000/total_reads_with_tel_gc. We calculated a tumor/normal telomere content ratio as:

$$\mathrm{t}\_\mathrm{n}\_\text{telomere}\_\text{content}=\text{tel}\_\text{content}\_\text{Tumor}/\text{tel}\_\text{content}\_\text{Normal}$$

We derived positive or negative alternative lengthening of telomeres (ALT) based on our previous study¹¹⁶ using t_n_telomere_content ≥ 1.068 as “ALT+" or t_n_telomere_content <1.068 as “ALT-". The TelomereHunter workflow can be found at: https://cavatica.sbgenomics.com/public/apps/cavatica/apps-publisher/telomerehunter, and ALT status computation at https://github.com/d3b-center/hope-cohort-analysis/tree/master/analyses/alt-analysis.

Compilation and analysis of the external reference cohort

Clinical and mutation data from patients with diagnoses including “Glioblastoma”, “Astrocytoma”, "Anaplastic astrocytoma”, “High grade glioma”, and “Diffuse midline glioma” were obtained from the Dana-Farber glioma study (glioma_dfci_2020 dataset, 1157 samples),¹¹ GENIE dataset version 12 (GENIE-12, 4248 samples),¹¹⁷ multicenter study of 1000 pediatric gliomas (1000_gliomas_2017 dataset, 853 samples),⁴ TCGA (471 samples),^104,118 and from German Cancer Research Center (DKFZ) study (pediatric_dkfz_2018 dataset, 120 samples).¹¹⁹ In compiling our external Reference cohort (Table S1), we required age at diagnosis and male or female annotation for all datasets. Survival data were also required for all datasets except GENIE-12 and pediatric_dkfz_2018. Samples with more than 200 mutations were excluded.

Mutation profile clusters in the reference cohort

Based on the reference cohort data, we derived gene mutation profiles, which are the mutation frequences as functions of age, using the following steps: We rounded age at diagnosis to the whole year. For each age year in the range of 0–90, we considered the samples within ±5 years sliding window. Then for each gene, we calculated the percentage of samples in the selected age window that exhibited a mutation (encompassing all mutation types) in the gene. Note that when calculating these percentages, we considered the total number of samples in the age window in which such mutations could be detected. This approach is necessary because mutations in the GENIE-12 and glioma_dfci_2020 datasets were identified using various targeted panels rather than comprehensive whole exome sequencing.

We then focused on 6215 genes with a mean mutation profile value greater than 10% across the whole age range and clustered them into two groups based on their mutation profiles using the ConsensusClusterPlus R package.⁷⁸ Consensus clustering was performed using Spearman correlation and the PAM algorithm, with over 25 iterations, sampling 90% of items per iteration. We selected K = 2 clusters based on the consensusMatrix values, as K = 2 gave the lowest sum of between-cluster consensus values. The plots in Figure 1D illustrate the genes with the highest mean profile values in each mutation group, alongside known driver genes from Figure 1C.

Survival analysis to assess age-dependent impacts of mutations

To evaluate the prognostic impact of high-frequency mutations on overall survival, 1,604 reference cohort samples with complete data were analyzed using a multivariable

Cox proportional hazards model. Covariates included age and binary indicators for glioblastoma, diffuse midline glioma, male patients, and mutation status (WT/Mutant) for 11 mutations with frequencies exceeding 5% in the [0,40) and (40,62] age groups. The model was applied to the PED, AYA, and ADU age groups. Bonferroni-adjusted p-values were calculated for each model, with covariates showing adjusted p-values <0.10 in the PED or AYA groups highlighted in Figure 1E. Additional details, including survival associations specific to the ADU age group, are provided in Table S1.

Chromosome arm level DNA copy number gain/loss derivation

We first assign the gain and loss status to each gene by applying the cutoff of +0.5 and −0.5 to the gene-level CNV log2 ratio scores. We then calculated the proportion of samples with gain or loss on each gene, then further averaged these proportions across genes within each chromosome arm. The top 5 chromosome arms with the highest average proportion of gain or loss were listed in Table S1.

Gistic plot analysis for the CNV data of the Discovery cohort

Recurrent copy number alteration analysis was performed using GSITIC2 with the following parameters: -rx 0 -genegistic 1 -twosize 1 -armpeel 1 -savegene 1 -maxseg 10000 -conf 0.99 -brlen 0.3. To avoid batch effects in CNV segmented data inherent to CNV segmented data due to platform differences, we focused on the Discovery cohort data.This analysis was performed separately for patients aged <15 and those aged 15–40 years.

For each deletion peak called by GISTIC, we counted the number of samples in each age group having a lower copy number compared to the median copy number of corresponding chromosome-arm (i.e., arm-median-CN - Peak-CN < 0). Fisher’s exact test was used to evaluate whether one group had a significantly higher rate of deletions than the other.

The same framework was used for gain peaks, where we instead calculated the number of samples with higher copy numbers compared to the median copy number of chromosome arm (i.e., arm-median-CN - Peak-CN > 0).

Temporal molecular profiles for tumor and normal tissues

Tumor - temporal molecular profile (T–TMP) derivation

We focused on patients aged 0 to 80 in the cDiscovery cohort and derived Age-dependent tumor molecular profiles using their protein, phosphosite, RNA, methylation, and kinase activity data. Specifically, for this analysis, we considered multi-omics datasets that comprised 189 protein samples (9,026 genes), 189 phosphosite samples (10,766 phosphosites), 184 RNA samples (13,079 genes), 176 methylation samples (15,399 genes).

Firstly, for each type of omics data and marker, we standardized the abundance measurements of the 0–62 age range to have a mean zero and a unit variance. We then stratified the data by male/female and fitted localized regression curves as a function of patient age using the R packages ‘loess’ and ‘gcv’. To determine the optimal span parameter for smoothing, we employed generalized cross-validation. We considered span parameters ranging from 0.5 to 3 for protein, phosphosite, and methylation data types, and from 1 to 3 for RNA data. Wahba’s generalized cross-validation score was calculated at intervals of 0.1 within the range span parameters, allowing us to select the value that yielded the minimum score for fitting the localized regression curve for each feature. The resulting regression curve of each marker was referred to as its Tumor - temporal molecular profile (T–TMP).

We also derived kinase T–TMP using the same strategy based on KEA3 Kinase activity scores for 519 kinases across 179 surgical tumor samples (excluding autopsy samples) in the cDiscovery cohort.

Normal-temporal molecular profile (N–TMP) derivation

To characterize differences of temporal molecular profiles between tumor and normal tissues, we utilized an external reference dataset comprising 69 global protein and 44 RNAseq profiles from normal tissues of the dorsolateral prefrontal cortex for subjects aged 0–50.¹⁸ We then derived the N–TMPs as follows: We first standardized the data within the 0–50 age range. Protein and RNA profiles were subsequently adjusted for covariates including PMI (postmortem interval), ethnicity, and pH by fitting regression models incorporating these covariates and then taking the residual, as suggested by the original publication.⁷⁸ In selecting the span, a parameter for loess fitting of both normal RNA and protein data, values between 1 and 3 were considered. Note that when comparing T–TMP and N–TMP, we re-derived the T–TMP using the same age span to ensure a robust comparison.

Interpolation of T–TMP and N–TMP

Localized regression curves were estimated separately in males and females and for tumor and normal datasets. To facilitate the comparative analysis among them, interpolation was employed to align age coordinates of various T–TMPs and N–TMPs. Specifically, we used the age points from the full cDiscovery cohort and evaluated the inferred localized regression curves (TMPs) at these age points. N–TMPs were also interpolated using the same age points to facilitate comparison between N–TMPs and T–TMPs.

In the Cross Population Survival Analysis (described below), we generated Cross-T–TMPs by evaluating the inferred TMPs (localized regression curves) from the cDiscovery cohort data at the age points of the Reference cohort.

Age-group identification among HGG patients based on T–TMP classification

We first displayed the age-aligned T–TMPs of RNA, protein and phosphoprotein abundance data from 174 surgical samples (excluding autopsy samples) with complete information across all data types in the cDiscovery cohort using a heatmap (Figure 1F). Note, for phosphoproteomics data, T–TMPs derived based on the gene-level phosphoprotein abundance were used in this analysis.

Significant heterogeneity was observed in the heatmap, indicating the presence of age groups with distinct molecular characteristics among HGG tumors within the AYA (15–40) and Adult (40–80) categories. This observation motivates us to conduct clustering analysis to identify additional omics-driven age groups within the currently defined clinical age categories.

For RNA and protein data, we selected the top 2,500 genes whose T–TMPs showed the greatest variability, measured by standard deviation, separately in males and females. For phosphoprotein data, we selected the top 1,810 features, amounting to 50% of the phospho-proteins with complete abundance information across the cDiscovery cohort. The age-aligned T–TMPs for the select RNA, protein, and phosphoproteins were then combined row-wise to create a matrix comprising 5,000 RNA T–TMPs (2500 from male and 2500 from female), 5,000 protein T–TMPs, and 3,620 phosphoprotein T–TMPs.

Consensus clustering was performed on the T–TMP matrix to identify age groups with distinct patterns. We first consider the conventionally defined AYA (Adolescent and Young Adult) age range of 15–40 years. Applying consensus clustering on the sub-T–TMP matrix of this age range, we detected two groups: one group consisting of all age time points from [15, 26], and the other for [26, 40]. Strikingly, even if we perform consensus clustering analysis for the entire 0–80 age range, the leading (most consistent) breakpoint emerged at age 26, regardless of the number of clusters chosen (ranging from 2 to 9). These results indicate the presence of two distinct age groups among AYA patients.

To further characterize patients aged over 40, we clustered the sub-matrix corresponding to the 40–80 age range. This analysis identified age 62 as a significant breakpoint across multiple cluster configurations. We labeled the 62+ group as SEN (Senior), noting that tumor tissues in this group were likely influenced by age-related changes in brain tissue. The other groups were designated as PED (Pediatric, 0–15), ADO (Adolescent, 15–26), YA (Young Adult, 26–40), and ADU (Adult, 40–62). These refined age groupings, evident in Figure 1F, form the basis for our subsequent analyses and provide a foundation for interpreting age-related effects across the dataset.

All clustering analyses were conducted using the ConsensusClusterPlus package in R, employing the k-means clustering algorithm with 200 iterations and a sampling percentage of 80% per iteration. Silhouette scores for all 174 tumors, based on this age classification, were calculated using the silhouette function from the cluster package in R (Figure 1F, Table S1).

Association of derived age groups with overall survival in HGG patients

To assess the association of age group with overall survival, we conducted a Kaplan-Meier survival analysis on 2,243 patients from the Reference cohort, using the survfit function from the survival package in R (Table S1). Log rank tests were employed to compare survival outcomes across different age groups (Figure 1G; Table S1). To test the robustness of the age-dependent survival pattern, we analyzed a Reference cohort subgroup of glioblastomas lacking H3-3A mutations. The age-survival association persisted in this restricted subset, indicating that the pattern is robust to histologic subtype heterogeneity (Figure S1I; Table S1).

Protein subgrouping based on normal and tumor temporal molecular profiles

Because we are interested in the interplay between brain development and tumor biology, we focused on age range 0–62 and excluded the SEN group, in which aging-related processed introduce additional biological complexity.

To identify groups of proteins with similar temporal expression profiles and to compare their temporal dynamics between males and females, T–TMPs for age range 0–62 were clustered separately in each male/female group. We included only proteins with T–TMPs having a standard deviation above 0.15 (Table S2) in both males and females, yielding 5,058 features per group. These T–TMPs were standardized and clustered using the ConsensusClusterPlus package in R, with the K-means clustering algorithm 500 iterations. We evaluated cluster solutions ranging from two to five groups and selected four clusters, which maximized concordance between the male and female clustering solutions as measured by the Rand Index. A similar approach was used to cluster N–TMPs with the cluster number fixed at four. These resulting clusters, referred to as Male and Female T–TMP or N–TMP groups, are illustrated in Figures 2A and S2A.

Pathway enrichment analysis of T–TMP groups

A one-tailed Fisher’s exact test was used for pathway enrichment analysis to identify biological functions enriched within T–TMP groups. For this analysis, we expanded our study-wide pathway database by adding gene sets that correspond to the proteins encoded on the X and Y chromosomes, as well as those linked to the various glycan types represented in our glycosylation data. Specifically, proteins were categorized based on their glycan structures as Sialylated (S), Fucosylated (F only), Fucosylated and Sialylated (F + S), High Mannose (HM), and Other.

Enrichment analysis was performed separately for male and female clusters (as shown in Figure 2B), as well as for groups defined by the combined male and female cluster assignments (Figures S2A and S2B). For each analysis, up to the top 100 significant pathways (adjusted p < 0.10) per T–TMP group were consolidated using the Sumer tool,⁸³ which groups pathways with significant overlaps and similar enrichment results into modules. Representative pathways from each module were then manually selected to be shown in Figures 2B and S2B.

T–N distance for identifying molecules with abnormal expression

The Euclidean norm was used to measure the distances between T–TMPs and N–TMPs for each protein separately in males and females. The analysis was restricted to ages 0–40, as the sample numbers in the ADU group in the normal data¹⁸ is small. T–N distances were evaluated for 545 proteins and 8,150 transcripts and were standardized using min-max scaling to the [0,1] interval. Proteins exceeding the 90th percentile of standardized distances were highlighted in Figure S2C.

M–F difference for identifying molecules with male or female-biased expression

For each protein or gene, their age aligned T–TMPs derived from the male and female groups were compared within each age segment (PED, ADO, YA), as well as across the entire age range of 0–40 years. Mean M–F T–TMP differences were obtained for 9026 proteins and 13,079 transcripts. These M–F differences were subsequently transformed into z-scores within each age segment. Proteins with M–F differences exceeding the 90th percentile Z score in at least one age group were classified as male- or female-biased, according to the up-regulation direction. The same procedure was also applied to the N–TMPs for 545 proteins and 8,658 transcripts.

Pathway enrichment of T–N distance and M–F difference

Pathway enrichment analysis was performed for T–N Distances and M–F Differences based on the protein and RNA data, respectively (Table S2). We utilized one-tailed Wilcoxon rank-sum tests to screen for pathways enriched with proteins exhibiting large T–N Distances. Two-tailed Wilcoxon rank-sum tests were applied on M–F Differences to detect pathways enriched with proteins showing large differences between males and females within each age stratum, separately for tumor and normal data. We considered canonical pathways, gene sets defined by T–TMP groups, gene sets comprising genes encoded on X and Y chromosomes, and specific glycan modification types identified in our glycosylation data. Specifically, pathways and gene sets containing 5 to 250 observed genes or proteins were considered for the T–N comparison. For the M–F comparison, we included pathways within the same size range and those identified as significant in the T–N comparison.

To select representative pathways for highlighting in Figures 2E, S2E, and S2G, we considered the more significant T–N comparison FDR across males and females for each pathway (or the most significant FDR across different age strata in M–F comparison). We then selected up to the top 100 most significant ones (all with an FDR <0.10) and consolidated them using the Sumer tool,⁸³ which groups pathways with overlapping genes and similar enrichment results into modules. Then, from each module, we manually selected the representative pathways that contained proteins with the most extreme T–N differences (as shown in Figure S2C) and/or were relevant to survival as determined in later analyses (see the CPSA section below). A similar analysis was performed for the RNA domain for comparison (Figure S2E).

Tumor cell specific differential analysis between male and female tumors

We implemented a pipeline for cell-type specific differential analysis to determine which proteins/RNA were differentially expressed in tumor cells between male and female tumors. Specifically, we used BayesDeBulk to estimate the protein/RNA levels in tumor cells and immune/stromal cells of “female" versus “male" tumors. Considering this stratification, the mean expression level of the j-th gene was modeled as the following function:

$$E\left(y_{i,j}\right)={\pi }_i{x}_{j,female}1\left(i\in female\right)+{\pi }_i{x}_{j,male}1\left(i\in male\right)+\left({1-\pi }_i\right)x_{j,S/I}$$

with π_i being the estimated tumor purity for sample i; 1(i∈female) being an indicator function equal to 1 if the i-th sample is female and 0 otherwise; 1(i∈male)being an indicator function equal to 1 if the i-th sample is male and 0 otherwise; and x_j,female, x_j,male, x_j,S/I being the mean expression of the j-th gene in tumor cells among “female” tumors, tumor cells among “male” tumors and immune stromal cells, respectively. Here, purity parameters were considered as fixed and estimated via TSNet. For each gene j, BayesDeBulk was utilized to estimate the parameters of this model, i.e., x_j,female, x_j,male, x_j,S/I. In BayesDeBulk, 10000 Markov Chain Monte Carlo iterations (MCMC) were considered. The parameter estimates were derived as the median across the MCMC iterations after discarding a burn-in of 1,000 iterations. For each MCMC iteration, molecular pathway scores were computed as the mean of genes belonging to that pathway. Confidence intervals (CI) of the molecular pathway scores across MCMC iterations were computed. A pathway was considered upregulated in tumor cells of female (male) if the CI for that group was statistically greater (non-overlapping) than the pathway score for tumor cells of male (female) as well as the pathway score of immune/stromal cells. For CI calculation, we considered different thresholds such as 95% and 99%. To compute the estimate of the pathway scores shown in Figure S2F, the median across MCMC iteration was computed.

Detecting cis-regulation of mutation/fusion/CNV/methylation through integrative proteogenomic analysis

In this study, we investigated the functional molecular quantitative traits (mRNA, protein, and phosphoprotein abundances) that are affected by DNA-level alterations (somatic mutations, copy number variations (CNVs), gene fusions, and DNA methylation). We used an integrative analysis tool called iProFun²⁵ to jointly model the multi-omics outcomes and enhance the power for detecting significant cis-regulatory associations by sharing information across different omics data types.

The cis-association analysis included 153 samples with age<62 from the cDiscovery cohort (the union of our Discovery cohort and the GBM cohort from Wang et al.¹⁰). The mRNA expression levels were available for 17,344 genes on 147 samples, and the global and phospho-proteomics were measured on 152 samples for 10,557 genes and 32,506 phosphosites, respectively. As for the predictors, somatic mutations were captured on 31 genes for 150 samples, gene fusions were collected on 355 genes for 147 samples, CNVs were obtained for 18,439 genes on 150 samples, and gene-level DNA methylation were derived for 15,386 genes on 139 samples.

To remove confounding impacts on the molecular quantitative trains by clinical status, we controlled for male/female, age, tumor purity, and DASC status as covariates in iProFun analysis. Additionally, to understand how the associations differ among different age groups (pediatric, adolescent, young adult, and adult), we included interactive effects between age groups and two DNA-level alterations (CNV and DNA methylation) in each regression.

To infer cis regulations for multi-omics data simultaneously, we applied further adjustment in the statistical inference in the iProFun analysis. Family-wise error rate (FWER) was calculated for the cis associations with mutation and gene fusion, due to the limited number of features. Significance was determined by the cutoff of FWER at 10%. For association with CNV and DNA methylation, we estimated empirical false discovery rate (eFDR) from 20 permutations. We then selected the gene with a significant eFDR (<10%) and a sufficient posterior probability (>75%) as a significant marker.

The significant cis association markers from mutation, gene fusion, CNV and DNA methylation are presented in Table S3. Finally, we called a gene-level alteration as a cis-cascade event when it simultaneously regulates RNA and protein levels.

Detecting trans-regulation of mutation/Fusion through integrative proteogenomic analysis

We employed multiple linear regression models to assess trans-regulatory effects of somatic mutations of TP53, H3-3A, NF1, ATRX, and IDH1. In each model, we treated the transcript levels, protein and phosphosite abundances, kinase activity score (KEA3), and gene-level DNA methylation of other genes as responses. We performed age group-stratified analysis and included male/female, tumor location, tumor grade, and tumor purity as covariates. Additionally, we incorporated the CNV of the response gene as a predictor for modeling transcript and protein levels, while the protein abundance of the response gene was used as a predictor for phosphosite abundances. After performing multiplicity correction (for each mutation separately), we declared significant trans regulation if the adjusted p < 0.1 and the absolute log2 fold change (mutant/wild type) exceeds 0.4.

We also performed pathway enrichment analysis for each age group and each somatic mutation (except for IDH1 mutation where we did age-combined analysis). Pathway databases we considered include KEGG, Hallmark, Reactome, Biocarta, GOBP, GOMF and MITO3. After multiplicity correction, several hundreds of pathways were found to be significantly perturbed by the mutation of the above-mentioned genes in different age groups (adjusted p < 0.1). We then applied Sumer tool⁸³ to further consolidate the pathways perturbed by each mutation into different modules of similar pathway functions. Broadly, the significant pathways perturbed by ATRX, TP53, NF1 and H3-3A mutations were grouped into functions like cell cycle, spliceosome, ribosome and translation, synapse and neuron, epigenetic modifications, metabolic and immune. On the contrary, the pathways perturbed by IDH1 mutation lacked cell cycle and spliceosome pathways but included several glyco-related pathways.

Derivation of kinase and transcription factor activity scores via KEA3 and chEA3

We derived kinase activity scores for 519 kinases by applying KEA3 to phosphoproteomics data. Specifically, we firstly selected phosphosites from non-ischemia protein,¹¹⁰ and removed the samples from the postmortem sample-collection. We derived the normalized tumor sample intensity z-scores for all phosphosites by standardizing the distribution across all tumor samples in the cDiscovery cohort. Next, we selected the 500 most upregulated phospho-proteins or 500 most downregulated phospho-proteins from the normalized phosphosite Z score data for each sample. We performed KEA3 rank analysis using Appyter on the selected protein lists³¹ (https://appyters.maayanlab.cloud/KEA3_Consensus_Kinases/).

Similarly, we derived transcription factor (TF) activity scores using the 500 most upregulated or 500 most downregulated genes in standardized RNA data of each tumor sample, via chEA3³² (https://appyters.maayanlab.cloud/#/ChEA3_Consensus_Transcription_Factors).

The rank outputs from KEA3 or chEA3 were further normalized using the same approach from¹²⁰ to adjust the baseline distribution of each kinase or TF.

Identification of kinase – TF regulations differing between age groups

Following the strategy used in our previous work,¹²⁰ we constructed Kinase-TF regulatory networks to characterize shared as well as differing Kinase-TF regulation patterns between age groups based on the ChEA3³² and KEA3³¹ scores derived above. Specifically, we focused on two sets of age group comparisons: PED (0–15) vs. ADO (15–26), and ADO vs. YA (26–40). For each comparison, we denote the two age groups as AG1 and AG2.

For a given kinase-TF pair, we assigned a concordant score to each patient based on the ranks of the corresponding KEA3 and ChEA3 scores in that patient: the smaller the rank was, the more significant the KEA3 and Chea3 enrichment results were. The concordant score is then calculated as:

$$H\left(76-\left[\text{KEA}3\text{Rank}\right]\right)/75+H\left(76-\left[\text{ChEA}3\text{Rank}\right]\right)/75\text{,}$$

where H(x) = xfor non-negative values, and 0 otherwise.

Considering the KEA3 (ChEA3) scores derived from both the upregulated and downregulated phosphosites (TF) respectively, we had four kinds of concordance scores for a pair of kinase and TF.

• 1.(Up-up) Up-phosphoprotein-enriched-KEA3 scores vs. up-gene-ChEA3 scores.
• 2.(Up-down) Up-phosphoprotein-enriched-KEA3 scores vs. down-gene-ChEA3 scores.
• 3.(Down-up) Down-phosphoprotein-enriched-KEA3 scores vs. up-gene-ChEA3 scores.
• 4.(Down-down) Down-phosphoprotein-enriched-KEA3 scores vs. down-gene-ChEA3 scores.

For each kind, the kinase-TF pairs with the top 0.5% concordance scores were extracted and retained. We then constructed a Kinase-TF-AG matrix for these selected kinase-TF pairs (rows for kinases and columns for TFs): pairs appearing in the top 0.5% only for AG1 tumors received a value of 2; pairs appearing in this top ranked group only for AG2 tumors received a value of −2; pairs occurring in top ranked lists from both AG1 and 2 received a value of 1; and all other pairs received a value of 0.

Clusters of the Kinase-TF-AG matrix were used to manually produce the summary bipartite graph. Specifically, we focused on the Up-up and Up-down regulation patterns. Kinases and TFs common in multiple clusters across the Kinase-TF-AG matrices were grouped (if more than half of the kinases/TFs are shared, then clusters are merged). These groups were further connected by arrows or blunt-ended arrows indicating inferred activation (up-up) or inhibition (up-down) between kinases and TF in AG1, AG2, or both groups. Each cluster, augmented with gene-gene coexpression with Geneshot,³¹ was submitted to Enrichr^31,121 and Rummagene¹²² and the most relevant enriched term was used to label the cluster. For clusters with less than five genes were labeled with domain knowledge and bolded genes in the clusters denote membership to the cluster label.

Survival analysis of omic data

Sample and covariates inclusion in survival analyses

In survival analyses discussed below, the sample set was restricted to treatment-naïve tumors, excluding any autopsy samples. Survival times exceeding 2,000 days were truncated at 2000, with the corresponding survival status designated as ‘Alive'. For patients with multiple clinical events/samples, only the first collected sample was used. An overview of the sample sizes for different survival analyses discussed below is included in Table S4.

In the Cox regression analyses, the covariates considered (wherever appropriate) included age-group (or age), male or female group, tumor grade, tumor location, secondary/recurrent tumor status, and the mutation status of IDH1, TP53, ATRX, and H3-3A. In the analyses involving specific sample strata, covariates with a frequency less than 3 were excluded.

Pathway enrichment analysis details for identifying survival associated pathways

Pathway enrichment analysis was performed to identify biological pathways associated with survival within each age stratum in males and females. In addition to the comprehensive databases of canonical pathways outlined in previous sections, we included TMP protein groups (Figure S2A), as well as glycoprotein gene groupings categorized by the (typically non-exclusive) glyco-peptide modification types observed in our glycoproteomic data: high mannose, fucosylated and sialylated (F + S), sialylated only (S), and fucosylated only (F), or Other.

Pathway enrichment analysis was conducted based on the Wilcoxon rank-sum test (two-tailed test), as well as the Fisher exact test (one-tailed test), if there were ≥10 individual markers with FDR <0.10. In the case of Fisher’s exact test, analysis was performed separately for positive and negative associations and the test result with the smallest FDR was considered. For each pathway, the association with the lowest FDR across tests was selected as representative.

For all enrichment tests, the Benjamini–Yekutieli (BY) FDR correction procedure was used to adjust for multiple comparisons. For the analysis of protein data, for which the number of pathways associated with survival was large, SUMER was used to identify pathway modules.⁸³ This consolidation step was performed for males and females separately using the minimum of the adjusted p-values across age-strata considering up to the 100 topmost significant pathways (FDR<0.10).

Validating a prognostic score based on wild-type IDH1 and IDH2 proteins

We evaluated the prognostic model of IDH1 and IDH2 wild-type protein abundances from our previous work, Petralia et al. 2020⁹:

$$\text{Survival}\sim \text{age}\text{group}+\text{tumor}\text{location}\left(\text{cortical}\text{vs}.\text{others}\right)+\text{tumor}\text{purity}+\text{IDH}1/2^{\text{WT}}\text{,}$$

where, IDH1/2^WT is calculated as IDH1/2^WT = w1∗IDH1^WT + w2∗IDH2^WT, with w1 = 0.34 and w2 = 0.66 as specified in Petralia et al., 2020. Note that in this analysis, we considered only the H3^WTIDH1/2^WT samples and fitted separate models for male (Figure 4A) and female (Figure S4A) using both the TMT protein and SRM protein data.

Male-, female-, and age-class-specific survival associations with glycopeptides

Multivariable Cox proportional hazard regression models were employed to assess the association between OS outcomes and abundances of 3,440 glycopeptides. Survival analyses were performed separately for 21 male patients and 22 female patients aged 0–26 years. In the analysis of male data, tumor grade, as well as mutation status of IDH1, TP53, ATRX, and H3-3A were all included as covariates. In the female data, only tumor grade, H3-3A mutation and TP53 mutation were included (Table S4) due to their frequency being greater than 3.

We considered the below Cox Proportional Hazard regression model

$$\text{Survival}\sim {\mathrm{\beta }}_1\left(\text{Glycopeptide}\text{Abundance}\right)+{\mathrm{\beta }}_2\left({\mathbf{I}}_{\text{ADO}}\times \text{Glycopeptide}\text{Abundance}\right)+{\mathrm{\beta }}_3\left(\text{Age}\right)+\mathrm{\Sigma }{\mathrm{\beta }}_{\mathrm{i}}\left(\text{Covariates}\right)\text{,}$$

where I_ADO represents the ADO age group indicator. The combined effects of glycopeptide abundance (β₁) and its interaction with age group (β₂) were assessed using the ANOVA test respect to the null models: Survival ∼ β′₃(Age) + Σβ′_i(Covariates).

Based on the above model, the log-HR of the glycopeptide in the ADO age group is β₁ + β₂, while is β₁ in the PED age group. Consequently, when calculating the signed –log10 of the ANOVA p-value, the sign was determined by β₁ + β₂ for ADO and by β₁ for PED.

For the pathway enrichment analysis, we labeled each protein using the minimal association p-value among all glycopeptides linked to the protein and performed the enrichment test as described in the previous section.

Beside using the observed abundances of glycopeptides, we performed the same analysis using protein-adjusted glycopeptide abundances. Specifically, for 3,289 glycopeptides, the corresponding global protein abundance measurements were available. We adjusted the glycopeptide abundances for protein expression using a linear regression model without additional covariates, and the residuals from these models were used to represent the protein-adjusted glycopeptide abundances.

We identified pathways significantly associated with OS based on either the observed glycopeptide data or the protein-adjusted glycopeptide data of the cDiscovery cohort, and further screens for Concordant Pathways through overlapping with the results of the CPSA protein analysis (see next section). We then selected proteins within these pathways that had an UniProt annotation indicating membrane association and displayed the most significant OS-associated glycopeptide per protein in Figure 4B.

Cross-population survival analysis (CPSA)

We leveraged the data from the Reference cohort to facilitate the identification of OS-associated proteins/gene/pathways via the Cross-population survival analysis (CPSA) (Figure S4B).

Survival Associations of Cross-T–TMPs of RNAs and proteins in the Reference cohort: Firstly, we derived Cross-T–TMPs for tumors in the Reference cohort (650 males and 517 females) based on both gene expression T–TMPs and protein T–TMPs from the cDiscovery cohort (please see the Section of T–TMP derivation for details). These Cross-T–TMPs served as proxies for molecular abundances, and we evaluated survival associations for Cross-T–TMPs with a standard deviation greater than 0.15 across the 0–62 age range. Details on sample filtering for these analyses are provided in Table S4.

For each group S (s = male or female), to investigate age group-specific associations between Cross-T–TMPs of a given gene/protein and OS in the Reference cohort, we implemented the below Cox proportional hazards model:

$${\text{Survival}}_{\mathrm{S}}\sim {\mathrm{\beta }}_1{\text{TMP}}_{\mathrm{S}}+{\mathrm{\beta }}_2\left({\mathbf{I}}_{\text{PED}}\mathrm{x}{\text{TMP}}_{\mathrm{S}}\right)+{\mathrm{\beta }}_3\left({\mathbf{I}}_{\text{ADO}}\mathrm{x}{\text{TMP}}_{\mathrm{S}}\right)+{\mathrm{\beta }}_4\left({\mathbf{I}}_{\text{YA}}\mathrm{x}{\text{TMP}}_{\mathrm{S}}\right)+{\mathrm{\beta }}_5{\text{age}}_{\mathrm{S}}+{\mathbf{\beta }}_6^{\mathrm{T}}{\text{Covariates}}_{\mathrm{S}}\text{,}$$ (Equation 1)

where TMP_S represents the inferred Cross-T–TMP levels of the gene/protein for patients in the S group; I_PED and I_ADO represents the PED and ADO age group indicator, respectively. The covariates included binary indicators for diffuse midline glioma (DMG) and glioblastoma (GBM) histological statuses, as well as survival-associated mutations of H3-3A, IDH1, TP53, ATM, and ATRX (Figure 1E).

ANOVA tests were then used to identify significant age-class specific effects of Cross-T–TMPs on OS. For example, for the PED age group, we compare the above full model with the below reduced model through an ANOVA test:

$${\text{Survival}}_{\mathrm{S}}\sim \mathrm{\beta }{’}_3\left({\mathbf{I}}_{\text{ADO}}\mathrm{x}{\text{MP}}_{\mathrm{S}}\right)+\mathrm{\beta }{’}_4\left({\mathbf{I}}_{\text{YA}}\mathrm{x}{\text{MP}}_{\mathrm{S}}\right)+\mathrm{\beta }{’}_5{\text{age}}_{\mathrm{S}}+\mathbf{\beta }{’}_6^{\mathrm{T}}{\text{Covariates}}_{\mathrm{S}.}$$

A similar strategy was used to evaluate the effect of Cross-T–TMPs on OS in the ADO and YA age groups. The logHR of Cross-T–TMPs in each age group are calculated as: β_PED = β₁ + β₂, β_ADO = β₁ + β₃, and β_YA = β₁ + β₄.

We used the R package ‘locfdr'¹²³ to evaluate and control False Discovery Rates (FDRs). This method is more robust to strong dependence structure among features (genes, proteins) compared to other FDR estimating techniques, such as the Benjamini–-Yekutieli (BY) and Benjamini–Hochberg (BH) procedures. Specifically, within each male or female group, Z scores were derived from the one-sided ANOVA p-values of all genes/proteins, and the density distribution f(z) was estimated using the parameter settings detailed in the GitHub repository. Features associated with survival were identified by applying a threshold of 0.10 to the left-tailed FDR estimate at the midpoint of each bin in the discretized f(z) distribution. Numbers of significant proteins/genes in each age stratum in males and females were illustrated in Figure 4C.

Pathway enrichment analysis was performed based on Cross-T–TMP association results using Fisher’s exact test and Wilcoxon rank-sum test (please see the Section for pathway enrichment test for additional details). Specifically, Fisher’s exact test was performed for each age stratum in males and females based on proteins/genes with FDR <0.10. The Wilcoxon rank-sum test was applied using the signed –log10 p-values.

Survival Associations of RNAs and proteins in the cDiscovery cohort: Survival analysis was conducted for 63 male and 52 female patients aged 0–62 in the cDiscovery cohort. For male samples, covariates included binary indicators for Grade (3/Other), tumor location (Cortical/Midline/Other), and mutation statuses of IDH1, TP53, ATRX, and H3-3A. For female samples, the same covariates were used, except for IDH1 due to low frequency (Table S4).

The Cox proportional hazards models and ANOVA tests described above were applied to the protein abundance and gene expression of each gene in either the male or female group. FDRs were calculated using the BY procedure. A small set of proteins/genes that achieved statistical significance (FDR <0.1) in at least one age stratum in males or females are summarized in Table S4. Specifically, VASP and HBG2, showed significant association with OS among female patients; while three proteins, HEXA, RARRES2, GALT, showed significant association with OS among male patients.

Concordant proteins/genes and Pathways: OS-associated pathways were identified through concordant association results from both the Reference cohort, which utilized Cross-T–TMPs, and the cDiscovery cohort, which used gene/protein abundances. Concordant Pathways were defined as those that achieved significance in both the Reference (FDR <0.10) and cDiscovery (FDR <0.10) survival analyses, exhibiting consistent directions of associations in at least one age stratum in males or females.

Concordant Proteins/genes were defined using more relaxed criteria, requiring significance in the Reference cohort (FDR <0.10) and marginal evidence in the cDiscovery cohort (p < 0.01), along with a consistent direction of association. This relaxed criterion was implemented to ensure a sufficient panel of protein markers for developing protein-based multivariable OS risk score in the next section.

A protein-based overall survival risk score (OS-RS)

As the above analyses indicated strong prognostic signals in the protein data for PED and ADO patients, we aimed to develop a multi-protein panel to predict OS risk for patients aged 0 to 26. For each male/female group, we considered two sets of candidates. First, we included the Concordant Proteins defined in the previous section. Second, we incorporated proteins that were significantly associated with survival in the cDiscovery cohort (FDR <0.10) but were not already represented among the concordant proteins, this including two proteins for females — VASP and HBG2. Note, all the three proteins that showed significance in males based on the cDiscovery cohort analysis (HEXA, RARRES2, GALT) were already included among concordant proteins. Additionally, for male patients, we included proteins of IDH1 and IDH2, which were identified to be associated with OS among H3-3A WT patients (Figures 4A and S4A). From the resulting set—195 proteins in males and 19 in females—we further identified the subset present in the Validation data, yielding 85 proteins for males and 10 for females, which were used in multivariable survival risk modeling.

For each male/female group, Ridge-regularized Cox regression models were fitted via the ‘glmnet’ package in R, using data from the cDiscovery cohort, restricted to patients aged 0–26, incorporating covariates of age group (ADO/Other), tumor grade (Grade 3/Other), tumor location (Midline/Other), secondary tumor status (Secondary or Recurrent versus Other), H3-3A mutation status, and protein abundances of the selected proteins. A 10-fold cross-validation procedure was used to determine the optimal tuning parameter in the ridge regression. We then derived the 90% confidence intervals (CI) for the regression coefficients using bootstrap procedure with 10,000 iterations. A total of 28 proteins for male patients and 4 proteins for female patients, whose 90% confidence intervals did not include zero, were selected to calculate the male- and female-specific OS-RSs.

Specifically, for males and females separately, OS-RS were defined as the weighted average of selected protein abundances:

$$\text{OS}-\text{RS}={\beta }_1\times {\text{Protein}}_1+{\beta }_2\times {\text{Protein}}_2+\dots +{\beta }_k\times {\text{Protein}}_{k,}$$

where β_i represents the average coefficient of the ith protein in the Ridge-regularized Cox regression models, and Protein_i represents the abundance of the ith protein in the sample.

To illustrate the association of OS-RS_M and OS-RS_F with OS in the cDiscovery cohort (ages 0–26), we evaluated the following Cox model:

$$\text{Survival}\sim \text{OS}-{\text{RS}}_{\mathrm{M}}+\text{OS}-{\text{RS}}_{\mathrm{F}}+{\mathbf{I}}_{\text{ADO}}+{\mathbf{I}}_{\text{Secondary}/\text{Recurrent}}+{\mathbf{I}}_{\text{Midline}}+{\mathbf{I}}_{\text{Grade}}+{\mathbf{I}}_{\mathrm{H}3-3\mathrm{A}}+{\mathbf{I}}_{\text{Male}}\text{,}$$

where I represents the indicator function for categorical variables. Specifically, I_H3–3A represents the H3-3A mutation status.

Sensitivity analyses were conducted in the cDiscovery cohort using the same model, adding either cancer group (categorized as Diffuse Midline Glioma, High-Grade Glioma–Not Otherwise Specified, or Other) or treatment status (second radiation treatment: yes or unavailable/no) as additional covariates. The survival impact of a second radiation treatment was initially evaluated in a prescreening analysis that examined all treatment-related variables—including first and second instances of both chemotherapy and radiation (coded as yes, no, or unavailable; see Table S4)—along with male/female, secondary or recurrent tumor status, H3-3A mutation status, grade, and midline tumor status. Among these variables, only receipt of a second radiation treatment was significantly associated with survival in the prescreening analysis. However, including this variable (coded as a binary yes or no/unavailable) in the Cox model with OS-RS scores did not meaningfully change the conclusions for either the male/female stratified analyses or combined analyses (Table S4).

We next performed an analysis restricted to tumors classified as HGG-NOS. Due to limited sample size, the model was fit in the combined cohort, using the male- and female-specific OS-RS terms. The model adjusted for male group, adolescent age group, and grade; variables with <5 non-zero entries (H3-3A mutation and midline status) were excluded. In this restricted, histologically heterogeneous subgroup, the OS-RS remained significantly associated with overall survival, demonstrating the robustness of the trajectory-based signal (Table S4).

Finally, we validated the prognostic values of OS-RS_M and OS-RS_F in the Validation cohort (ages 0–26), using a similar model, but excluding I_Grade, as this variable was not available in the Validation cohort:

$$\text{Survival}\sim \text{OS}-{\text{RS}}_{\mathrm{M}}+\text{OS}-{\text{RS}}_{\mathrm{F}}+{\mathbf{I}}_{\text{ADO}}+{\mathbf{I}}_{\text{Secondary}/\text{Recurrent}}+{\mathbf{I}}_{\text{Midline}}+{\mathbf{I}}_{\mathrm{H}3-3\mathrm{A}}+{\mathbf{I}}_{\text{Male}}$$

Additionally, we assessed the association of PHKG2 and MORC2 in the Validation cohort using the similar model: Survival ∼ Protein+ I_ADO + I_{Secondary/Recurrent} + I_Midline + I_H3–3A + I_Male.

Survival analysis result visualization

To visualize associations between survival and combined protein scores or individual proteins in males and females (Figures 4I, 4J, S4I, and S4J), scores were dichotomized at the median within the 0–26 age range and incorporated into survival models. Covariate-adjusted survival curves were generated using ‘ggsurvplot’ from the ‘survminer’ package, with other covariates set to their reference categories. For the combined analysis in Figure S4F, survival was evaluated under two scenarios: (1) high/low male score with a low female score and (2) high/low female score with a low male score, with other covariates set to their reference (Other) categories. Groupings were dichotomized at the median value. The reported p-values correspond to the significance of male and female scores modeled as continuous variables rather than as dichotomized ones.

Identifying survival associated kinases through CPSA

We applied the aforementioned CPSA to the KEA3 kinase activity scores to screen for kinases associated with OS outcomes (Figure S5A), leveraging the Reference cohort data. The selected kinases were then inputted to the causal network analysis (please see the next section) for further investigation.

To enhance the robustness of kinase selection, we implemented a stability selection procedure within the CPSA framework. Specifically, we analyzed male and female tumors separately. Of the 519 kinases, those exhibiting relatively flat Cross-T–TMPs (SD < 0.15) were excluded from consideration, resulting in the removal of 273 kinases in the male analysis and 232 kinases in the female analysis. Then, for each kinase, survival associations were evaluated for its Cross-T–TMP over 100 iterations, each using a random subset of 80% of samples in the Reference cohort. The stability of associations was assessed by the number of times each kinase was selected (p-value<0.01) with the same association direction. In total, 186 kinases were selected, including 36 kinases unique to females and 139 unique to males. These kinases were then used in the DAGBagM analysis.

Identification of kinases causally associated with OS via DagBagM

Focusing on the 186 OS-associated kinases from the analysis described above, we applied DAGBagM³⁸ to screen for kinases whose activities were causally associated with patient OS in the cDiscovery cohort.

We first defined the good and bad prognosis groups using the OS days: patients with OS > 1 year were assigned to the better prognosis group, while those with a death event within 6 months were assigned to the poor prognosis group. This categorization resulted in 90 patients in the better prognosis group, and 11 in the poor prognosis group.

We also included the mutation status of frequently mutated genes as well as those associated with OS, including ATRX, TP53, H3-3A and IDH1, in the node space as the parent nodes for the kinases. In addition, we included male/female, age groups (PED, ADO and YA), tumor locations (cortical, midline and others) and tumor grade (3 and others) as the root nodes. Specifically, we incorporated a hierarchy structure among these nodes: 7 tier 1 nodes representing male/female, age, tumor locations and grade; mutation status of 4 genes in tier 2; kinase activities of 186 kinases in tier 3; and finally, the prognosis status in tier 4. Since DAGBagM can take prior information in network learning, we used this feature of DAGBagM to allow the root nodes in tier 1 to impact any node in the following tiers and suppressed the edges from tiers 2,3 and 4 to tier 1. Again, we only allowed the nodes in tier 2 to impact any nodes in the tiers 3 and 4 but suppressed the edges backward. We didn’t put any prior on the edges between kinases and prognosis, so the kinases can be inferred as either parent or child nodes for prognosis. With 198 nodes and 101 samples, we were able to infer 314 directed edges in the network, of which 6 kinases were inferred to be direct causal parents of prognosis and ∼30 kinases were inferred as their immediate parents.

We next performed sensitivity analysis to obtain causal kinases with robust signals, by defining the prognosis groups (better and poor) using multiple cut-offs on the overall survival (OS) days. Specifically, the patients with OS > c1 months were defined to have better prognosis while those with a death event within c2 months were defined to have poor prognosis. We varied c1 and c2 and constructed causal networks with 198 nodes using both male/female-stratified analyses and a combined analysis (the cut-offs and sample sizes in different prognosis groups are given in Table S5). Finally, we obtained 26 kinases that were frequently inferred from most settings as direct causal nodes to prognosis, or their first/second degree neighbors (Figure 5A).

Estimating the interaction strength between kinases connected in the DAG using locCSN

The locCSN algorithm⁴³ estimates the connection strength between each pair of genes in each cell, resulting in a gene-gene network for each cell. This connection strength is based on a local independence test that is applied to each gene pair and each cell. In the application of locCSN to kinase datasets, kinase activity scores substituted for gene expression, and tumor samples were analogously treated as individual cells.

Investigation of OS-associated kinases through cell-type specific association analysis

We studied the pathways associated with the 26 causal prognosis-associated kinases in cancer and non-cancer cell components using our recently developed cell-type-specific proteomics association analysis tool iProMix.⁴⁰ Analysis was conducted in cDiscovery data (ages 0–40), to evaluate their associations using a total of 9028 proteins, adjusting for age groups, male/female, and common mutations as covariates. We first estimated the cell-type specific kinase-protein correlations, and then used these estimates, together with their likelihood ratio (LR) statistics, to perform pathway enrichment analysis. The analysis used Wilcoxon rank-sum test to compare the signed LR statistics within versus outside of a pathway. To identify kinase-associated pathways, we evaluated seven databases, including Hallmark, KEGG, Reactome, Biocarta, GO Biological Process, GO Molecular Function, and MITO3, and adjusted FDR in each database separately. As these kinases were very active with ubiquitous associations, we further performed an enhanced enrichment analysis imposing a considerable effect size requirement, where we tested the one-sided signed LR statistics within versus outside of a pathway after the statistics outside of the pathway had shifted 20% in rank toward the direction of statistics inside of the pathway. Additionally, we also investigated the associations between each of these 26 kinases and the other 518 kinases in cancer and non-cancer cell components using.⁴⁰ This analysis estimated the cell-type specific kinase-kinase correlations and employed permutation to identify significant associations at 10% empirical false discovery rate (FDR).

Next, we validated the causal effects of these kinases on significant pathways using the LINCS L1000 CRISPR Perturbations database. One cell line from the central nervous system (U251MG) was identified and used for validation. For pathways positively associated with a kinase, we tested whether knocking out this kinase in the cell line can down regulate a group of genes that were enriched in these pathways. Similarly, for pathways negatively associated with a kinase, we tested whether knocking out this kinase can up regulate a group of genes that were enriched in these pathways. A 2 by 2 table to contrast the perturbed gene sets and genes in each pathway was used to identify enriched pathways, and pathways with odds ratio (OR) > 1 and Fisher’s exact test p-value <0.05 in the 2 by 2 table were considered as validated.

Effect of IDH1 mutation and IDH WT proteins on CDK8 kinase-dependent regulation of OXPHOS

To assess whether the regulation between CDK8 kinase activity and OXPHOS was influenced by IDH1 mutation status, we conducted a stratified linear regression analysis by IDH1 mutation:

Abundance of OXPHOS protein ∼ CDK8 kinase activity + age-class + male + tumor location + tumor grade.

As shown in Figure S5H, a significant percentage of OXPHOS proteins were down-regulated by CDK8 kinase activity in both the IDH1 mutant and WT tumor groups. In addition, we also assessed the correlation between the single-sample GSEA score of the OXPHOS pathway and CDK8 kinase activity in IDH1 mutant and WT samples. We found strong negative correlations in both groups: r = −0.6, p = 0.03 for the IDH1 mutant group and r = −0.4, p < 0.0001 for the IDH1 WT. These results suggest that the observed negative regulation of CDK8 on OXPHOS was observed in both IDH1 mutant and WT groups.

To further evaluate whether IDH1 WT protein may influence the regulation between CDK8 kinase activity and the OXPHOS pathway, we performed the following linear regression analysis focusing on IDH1 WT samples:

Abundance of OXPHOS protein ∼ CDK8 kinase activity + IDH1 protein abundance + age-class + male/female status + tumor location + tumor grade

Based on this model, after correcting multiple hypothesis testing, 127 out of 181 OXPHOS proteins showed significant association with CDK8 kinase activity, while only 7 proteins showed significant association with IDH1 WT protein abundance. In addition, although IDH1 is a member of the OXPHOS pathway, IDH1 WT protein abundance itself did not show a strong association with CDK8 kinase activity (Figure S5H). Furthermore, CDK8 kinase activity was not associated with IDH1 mutation status. Collectively, these results suggest that CDK8 kinase activity negatively regulates OXPHOS via an IDH1-independent mechanism.

Cell line-based investigation of the precision treatment strategy using kinase inhibition

We leveraged the CRISPR data from 4 cell-lines and drug response data from the 5 cell lines corresponding to the respective tumor samples in the Discovery cohort for the 14 kinases with available drugs. Perturbation response was obtained from the Childhood Cancer Model Atlas (CCMA⁴⁴). CRISPR gene dependency scores (β-score) for each of the 14 kinases were derived by comparing gRNA abundance at day 21 to plasmid control (day 0). Drug effect (Z score) was evaluated following 3-day treatment with 1 μM of drug and response normalized to controls to derive a Z score. Out of the 14 kinases, 7 showed consistency with our hypothesis. Among these, CDK8, ATM, LCK are from module I and FLT1, RAF1, MAP2K1 are from module II. All six kinases were associated with poor survival in the Discovery cohort and showed negative correlations between drug response/CRISPR scores and kinase activity scores (Figures 5H, 5I, S5I, and S5J), validating our hypothesis. The only other candidate was BTK which was associated with better survival and showed a positive correlation. However, since there was no booster drug available for BTK, we focus on the other 6 kinases (associated with poor survival), all of which have inhibitor drugs available and hence could be deemed potential drug targets to pursue possible follow-up experiments.

Derivation of patient subtypes through proteomic clustering

Proteomic clusters were obtained by clustering ssGSEA scores derived for 190 samples across 8,167 pathways, each with more than 5 and less than 250 members among the 9026 proteins assayed via TMT. ssGSEA scores were generated using the gsva package from Bioconductor with default parameter settings. Clustering was performed with R’s ConsensusClusterPlus, using the k-means algorithm with Euclidean distance and default parameters. Both male/female-stratified and unstratified clustering were conducted. Multiple cluster numbers (k = 2–5) were evaluated, with three clusters selected in each case based on the highest mean silhouette score, calculated using the R silhouette package. To compare silhouette score distributions for groupings based on cancer type and proteomic clusters, as in Figure S6C, this same ssGSEA score matrix was used.

Clustering results from unstratified and male/female-stratified analyses were aligned by maximizing sample overlap across different clustering configurations (see Figure S6A). Specifically, M1 and F1 correspond to male and female groups aligned to the same unstratified cluster, and similarly for the remaining clusters. A similar procedure was applied to derive protein clusters for the Validation cohort of 29 PED/AYA HGG samples. Specifically, ssGSEA scores were derived for 8,031 pathways using global protein abundance measurements of 10,775 proteins. Due to the limited sample size in the Validation cohort, only unstratified clustering was performed.

Mapping of the GBM classification from Garofano et al. to the cDiscovery cohort

Classification of CPTAC- pediatric high-grade gliomas (PG-HGG) tumors in functional subtypes has been recently reported with the identification of subtype-specific protein signatures.¹²⁴ The 50 highest-scoring proteins from these signatures were used to calculate the enrichment of each functional GBM subtype (normalized enrichment score of subtype activity, NES) for each tumor using single-sample MWW-GST (ssMWW-GST). Each tumor was assigned to a subtype only if NES was higher than 0.3 and the enrichment significance (FDR) < 0.05, otherwise the tumor was classified as undefined. If a tumor has shown statistically significant enrichment of multiple subtypes, the subtype with the highest score has been assigned.

Identification of proteins, pathways, and kinases associated with proteomic subtypes

To characterize subtypes derived from male/female-stratified clustering, we considered the following linear model without an intercept:

$$\text{Abundance}\sim {\mathrm{\beta }}_1\left(\mathrm{M}1\right)+{\mathrm{\beta }}_2\left(\mathrm{M}2\right)+{\mathrm{\beta }}_3\left(\mathrm{M}3\right)+{\mathrm{\beta }}_4\left(\mathrm{F}1\right)+{\mathrm{\beta }}_5\left(\mathrm{F}2\right)+{\mathrm{\beta }}_6\left(\mathrm{F}3\right)$$

In this model, M1, M2, M3, F1, F2, and F3 represent the binary indicator variables for each subtype. P-value distributions for each beta-coefficient were not heavily skewed. Given this, the localized FDR adjustment procedure used in the cross-population survival analysis above, was deemed unnecessary, and the BY¹¹⁴ adjustment was applied instead. Using this model, we identified proteins and kinase activities (KEA3 scores) associated with each subtype. To evaluate age-related heterogeneity within subtypes, this analysis was performed separately for strata of patients aged 0–40 (Figure 6A; Table S6) and 0–62 (not shown).

Pathway association testing was performed using the Z score associated with each coefficient in the above model. Both enrichment and depletion (one-sided) tests were conducted, leading to twelve sets of pathway enrichment tests results for the 6 clusters. P-value adjustments using BY were applied to each set separately. The pathways included in Figure 6A were selected based on specific enrichment in the 0–40 age range after pathway consolidation using Sumer.⁸³ Additionally, a handful of pathways enriched in at least one subtype and previously identified as relevant—through T–N enrichment, association with prognosis, or significant male–female differences within clusters (see below)—were also included. Specifically, for each pathway, the minimum of the adjusted p-values across the 12 pathway enrichment tests performed in the 0–40 age range was used as the input weight when running Sumer, including up to 100 of the topmost enriched pathways. From each of the resulting Sumer modules, a representative pathway was selected based on the minimum adjusted p-value. Kinases belonging to each of these pathways whose kinase activities were also associated with subtypes in the 0–40 age range are shown.

For each strata-pathway-cluster, combined p-values were calculated using Fisher’s method to represent overall enrichment or depletion across males and females mapping to the same cluster. The smallest combined p-value determined the pathway association score, calculated as the absolute value of the log10 p-value, with a negative sign for depletion and a positive sign for enrichment.

Additionally, to further evaluate heterogeneity between male and female groups within each cluster, the protein abundance was used to fit the following model:

$$\text{Abundance}={\mathrm{\beta }}_1\left(\text{Male}\right)+{\mathrm{\beta }}_2\left(\text{age}\right)+\mathrm{\epsilon }$$

In this model, Male was included as a binary indicator variable, age is a continuous variable, and ε is the error term. Pathway analysis was then performed using the Z score derived from the coefficient β₁. The results from these various analyses are summarized in Table S6.

Association of proteomic subtype and overall survival

We firstly fitted a Cox regression model including all six male/female-stratified protein clusters, adjusting for H3-3A mutation, midline tumor location, tumor grade, secondary/recurrent tumor status, and age group across the 0–62 age range (Figure S6G). The survival analysis included 115 treatment–Naïve patients who were alive at the time of sample collection. Survival times exceeding 2,000 days were administratively censored at this threshold. Only the first sample collected from patients with multiple clinical events was used. Protein subtypes were significantly associated with overall survival (likelihood ratio test comparing models with and without subtype, p = 0.02), with M2 showing the worst survival.

To directly compare OS outcomes between M2 and F2 (Figure 6G), we employed the following Cox proportional hazards model including an interaction between cluster C₂ and male status:

$$\text{Survival}\text{time}\text{object}\sim C_2+\left(C_2\mathrm{x}\text{Male}\right)+\text{Age}+\text{Male}+\text{Covariates,}$$

where C₂ and Male are binary indicator variables, and the same covariates were included as in previous. For analysis of patients ages 0–40, the hazard ratio for the interaction term was 9.93 (95% CI: 2.06–47.87, p = 0.004), indicating that the male patients in C₂ cluster experienced a significant survival disadvantage.

We also fitted analogous Cox models with Cluster × Male interaction term for C₁ and C₃, but didn’t detect significant interaction effects (p > 0.05, Table S6).

Identification of glyco-peptides/pathways associated with proteomic subtypes

Glycosylation data was available for 77 patients, ranging from age 0 to 40. Due to the smaller sample sizes available for glycosylation, we used a single linear model without an intercept to identify markers showing differences between males and females within the C₂ cluster:

$$\text{Abundance}\sim {\mathrm{\beta }}_1{\mathrm{C}}_1+{\mathrm{\beta }}_2{\mathrm{C}}_2+{\mathrm{\beta }}_3{\mathrm{C}}_3+{\mathrm{\beta }}_4\left({\mathrm{C}}_2\times \text{Male}\right)+{\mathrm{\beta }}_5\left({\mathrm{C}}_1\times \text{Male}\right)+{\mathrm{\beta }}_6\left({\mathrm{C}}_3\times \text{Male}\right)+{\mathrm{\beta }}_7\left(\text{Age}\right)$$

In this model, C_k represents the binary indicator variables for each cluster; and C_k × Male, represent interaction terms between specific clusters and the Male indicator variable. This analysis was performed for both glycopeptide (77 samples with 3440 glycopeptides) and adjusted glycopeptide data (75 samples with both protein and glycopeptide data with 3289 glycopeptides).

Using the p-values associated with the β₄ term in the model above, we obtained a gene-level association result by combining p-values across glycopeptides mapping to the same gene. One-tailed p-values were computed separately for positive and negative associations and separately combined across glycopeptides mapping to the same gene using Fisher’s method. These p-values were adjusted across genes using BY.¹¹⁴ The association direction with the smaller adjusted p-value was selected for each gene (Table S6).

Pathway enrichment analysis was performed using the z-scores of the β4 term. For pathway enrichment analysis we selected the glycopeptide with the highest absolute Z score per gene. We then applied a two-sided Wilcoxon rank-sum test comparing pathway-level z-scores to background. We identified 8 pathways that were significant (FDR <0.10) after BY correction (Table S6). A curated selection of non-redundant pathways appears in Figure 6K. For comparison, we performed the same analysis using both protein and RNA-seq data for the Discovery and Validation cohorts, as summarized in Table S6.

Cross-cancer analysis of ITGAV-ANTTQPGIVEGGQVLK glycopeptide expression in tumor vs. normal adjacent tissue

We generated an expression matrix of intact N-linked glycopeptides using CPTAC phospho-enriched datasets from CCRCC, HNSCC, LSCC, LUAD, and PDAC (with PDAC data combined with its glyco-enriched cohort). For this, we employed GPQuest 2.1 and MS-PyCloud to search against a reference of 29,188 glycopeptides from the N-glycositeAtlas and 252 common N-linked glycans.^125,126 We then performed differential expression analysis comparing tumor samples to normal adjacent tissues in each cohort using the Wilcoxon rank-sum test, and specifically extracted the result for the glycopeptide ITGAV-ANTTQPGIVEGGQVLK.¹²⁷ These results are summarized in Figure S6N and Table S6.

Single cell RNA seq data preprocessing

Gene expression was derived as counts-per-million (CPM). We divided the CPM values by 10 as the size of single-cell libraries is estimated to be in the order of 100,000 transcripts. Low quality cells were filtered-out based on low number of detected genes (n = 500) or high expression of mitochondrially-encoded genes (a proxy for dying cells; cut-off = 5%). Gene expression was log2 transformed and genes with low expression were filtered out.

Cells were clustered using graph-based clustering (Louvain’s algorithm) on a 2-dimensional UMAP space. Malignant and non-malignant cells were distinguished by inferring the copy-number from gene expression using inferCNV tool.¹²⁸ Myeloid, T cells, OPC, oligodendrocytes, neurons, astrocytes, endothelial and pericytes cells were characterized based on the expression of known marker genes.

Malignant cells were assigned to pathway-based transcriptomic states based on the highest significant score using single sample MWW-GST (ssMWW-GST)^45,129 and the subtype-specific gene signatures reported in Migliozzi et al.¹²⁴

scRNA based cell-type specific marker derivation

Cell-type specific markers were defined as the first 50 highest scoring (MWW-score) of the ranked-list of genes comparing the expression profiles of the cells in each cell type/cluster with all other cells using the Wilcoxon rank-sum test. MWW-score for each gene has been defined in Garofano et al.⁴⁵

Estimation of stromal and immune scores and tumor purity from gene expression

ESTIMATE⁷³ was utilized to infer immune and stromal scores based on RNA-seq data; while tumor purity was estimated via TSNet⁷⁵ using immune and stromal signatures from.⁷³

Deconvolution analysis

To estimate the fraction of different cell types in the tissue microenvironment, we performed a multi-omic based deconvolution integrating proteomic and RNA-seq data via BayesDeBulk.⁵² To perform the deconvolution, BayesDeBulk requires a list of cell-type specific markers for each cell type. To derive cell-type-specific markers, we leveraged the single-cell RNA data collected from tumor samples of 10 patients in the Discovery cohort to distinguish the following cell types: astrocytes, T cells, myeloid cells, neurons, pericytes and endothelial cells. BayesDeBulk took as input the list of markers differentiating each pair of these cell types. For each pair of cell types (i,j), the top 20 markers upregulated in the i-th cell type compared to the j-th cell type were considered as a list of markers. Then, to better characterize the myeloid compartment, we obtained cell-type-specific markers from Antunes et al.⁵³ BayesDeBulk was implemented using 5,000 Markov Chain Monte Carlo iterations and 1,000 burn-in. To account for tumor cell percentage, the resulting cell-type fractions were normalized to sum to one minus the tumor cell percentage derived using TSNet.¹²¹ The above strategy was applied to the cDiscovery cohort and the Validation cohort data separately.

Association between mutation profiles and cell type fractions

We derived the association between cell type fraction and mutation profiles (mutant versus wild-type) via the Wilcoxon rank-sum test. For this analysis, the top 31 mutations were considered. P-values from Wilcoxon rank-sum test were adjusted by Benjamini–Hochberg adjustment and only association with adjusted p-values <0.10 were considered as significant.

Association between overall survival and cell type fraction

cDiscovery cohort

Association between cell type fractions and overall survival was derived via Cox Proportional-Hazard regressions in male and female population separately. For each model, overall survival was modeled as a function of SEPP1 Mo TAM and T cell fraction after adjusting for H3-3A mutation status and patient age. ANOVA p-value comparing the two models below is reported in Figure 7D:

model 1 covariates: H3-3A, age

model 2 covariates: SEPP1 Mo TAM, T cell, H3-3A, age

Validation cohort

Association analysis between SEPP1 Mo TAM and T cell fractions and overall survival were derived via Cox Proportional-Hazard regressions. Given the small sample size of the Validation cohort (n = 26), we included both female and male tumors, with indicators of male and H3-3A mutation status incorporated as covariates in the model. ANOVA p-value comparing the two models below is reported in Figure 7E:

model 1 covariates: H3-3A mutation and male

model 2 covariates: SEPP1 Mo TAM, T cell, H3-3A, and male

Association between checkpoint inhibitors and T cell/SEPP1 Mo TAM fractions

Association analysis between gene/protein level data and T cell/SEPP1 Mo TAM fractions was derived via linear regression model. Each marker was modeled as a function of SEPP1 Mo TAM and T cells fractions. P-values were adjusted via Benjamini–Hochberg adjustment and only associations significant at 10% FDR were considered significant.

Additional resources

Data portal web server

All processed proteogenomic datasets as well as clinical meta information can be queried, visualized and downloaded from an interactive ProTrack data portal (http://hope.cptac-data-view.org/).

The HOPE Cohort Analysis Repository (d3b-center) contains computational pipelines, data, and statistical methods used in this study, focusing on the molecular and clinical characterization of high-grade gliomas. This resource ensures reproducibility and enables further exploration of the dataset. The repository is available at: github.com/d3b-center/hope-cohort-analysis

Other proteogenomic integrative analysis code used in this study is available at: github.com/WangLab-MSSM/Pediatric-AYA-high-grade-glioma.

Published: September 17, 2026