Abstract

   Gliomas exhibit high heterogeneity and poor prognosis. Despite
   substantial progress has been made at the genomic and transcriptomic
   levels, comprehensive proteomic characterization and its implications
   remain largely unexplored. In this study, we perform proteomic
   profiling of gliomas using 343 formalin-fixed and paraffin-embedded
   tumor samples and 53 normal-appearing brain samples from 188 patients,
   integrating these data with genomic panel information and clinical
   outcomes. The proteomic analysis uncovers two distinct subgroups:
   Subgroup 1, the metabolic neural subgroup, enriched in metabolic
   enzymes and neurotransmitter receptor proteins, and Subgroup 2, the
   immune subgroup, marked by upregulation of immune and inflammatory
   proteins. These proteomic subgroups show significant differences in
   prognosis, tumorigenesis, microenvironment dysregulation, and potential
   therapeutics, highlighting the critical roles of metabolic and immune
   processes in glioma biology and patient outcomes. Through a detailed
   investigation of metabolic pathways guided by our proteomic findings,
   dihydropyrimidine dehydrogenase (DPYD) and thymidine phosphorylase
   (TYMP) emerge as potential prognostic biomarkers linked to the
   reprogramming of nucleotide metabolism. Functional validation in
   patient-derived glioma stem cells and animal models highlights
   nucleotide metabolism as a promising therapy target for gliomas. This
   integrated multi-omics analysis introduces a proteomic classification
   for gliomas and identifies DPYD and TYMP as key metabolic biomarkers,
   offering insights into glioma pathogenesis and potential treatment
   strategies.

   Subject terms: CNS cancer, Proteomics, Cancer genomics, Tumour
   heterogeneity, Tumour biomarkers
     __________________________________________________________________

   Comprehensive molecular characterisations could shed light on the high
   heterogeneity and poor prognosis of gliomas. Here, the authors perform
   proteomic profiling of 188 glioma patients, revealing immune and
   metabolic neuron-related subgroups as well as metabolic biomarkers
   linked to prognosis.

Introduction

   Gliomas, accounting for around 81% of all primary central nervous
   system (CNS) malignancies, constitute the most common type of intrinsic
   brain tumor with high heterogeneity and poor prognosis^[60]1. Recently,
   gliomas have been classified into distinct subgroups according to an
   integrated set of molecular and histological parameters, which was
   initially introduced in 2016 and subsequently updated in 2021 (WHO
   CNS5)^[61]2,[62]3.

   High-throughput research has been conducted with the aim of uncovering
   major oncogenic molecular events, identifying therapeutically
   actionable alterations, and delineating prognostic and predictive
   subgroups among glioma patients^[63]4–[64]6. Notably, these efforts
   have predominantly concentrated on the genomic, transcriptomic, and
   epigenetic dimensions^[65]7. At the genomic level, isocitrate
   dehydrogenase (IDH) mutation is a widely recognized biomarker^[66]8.
   The GLASS consortium has described other genetic alterations during
   glioma recurrence, such as CDKN2A deletion and hypermutations^[67]9. At
   the transcriptomic level, the Cancer Genome Atlas Program (TCGA)
   classified glioblastoma (GBM) into classical, mesenchymal, proneural,
   and neural subtypes^[68]10. The GLIOTRAIN discovery cohort also
   identified three unique patient transcriptome clusters, characterized
   by tumor microenvironment compositions. Although continuous efforts
   have been made in multi-omics research, clinical applications remain
   confined to a few biomarkers, such as IDH mutation, MGMT promoter
   methylation, and 1p/19q co-deletion^[69]11–[70]13. Consequently, the
   development of targeted therapies and refinement of glioma
   classification for patient stratification remain ambitious, and are
   long-term goals necessitating significant progress.

   Despite the substantial insights gained from DNA- and RNA-based
   classifications, it is anticipated that proteomic patterns, as
   discerned through mass spectrometry (MS), will serve as the primary
   conduits for elucidating biological functions in
   glioma^[71]7,[72]14,[73]15. In the meantime, emerging studies have
   demonstrated the limited correlation between protein abundance and DNA
   or RNA level (r = 0.23–0.45) in breast, colorectal, and ovarian
   cancer^[74]16–[75]18. Intriguingly, proteins have shown superior
   prognostic predictive capabilities compared to their upstream mRNA
   counterparts^[76]18. Hence, the proteome, which reflects the functional
   consequences of genomic modifications, holds the potential to offer
   insights into the real state of tumor cells. In intrahepatic
   cholangiocarcinoma, researchers have successfully identified four
   subgroups characterized by unique genetic alterations, microenvironment
   dysregulation, tumor microbiota composition, and potential therapeutics
   features, with the identification of SLC16A3 and HKDC1 as prognostic
   biomarkers and therapeutic targets^[77]19. The proteomic classification
   of GBM has demonstrated the capacity to better reflect immune and
   metabolic biological processes associated with survival as compared
   with transcriptomic profiling^[78]14. Oh et al. similarly identified
   two proteomic clusters based on distinct Warburg-like and oxidative
   phosphorylation-related protein profiles, paving the way for the
   identification of the mTORC1/2 dual inhibitor AZD2014 as a treatment
   modality for GBM^[79]7. Collectively, these results demonstrate the
   value of proteomes of glioma. However, comprehensive proteomic
   subtyping across all four WHO-graded gliomas is still lacking, which
   may enhance the predictive power of clinical outcomes.

   The critical role of nucleotide metabolism in driving cancer
   progression across various malignancies, characterized by rapid cell
   proliferation and genomic instability, has gained increasing
   recognition^[80]20–[81]22. Chemotherapeutic agents such as
   5-fluorouracil (5-FU) target nucleotide metabolism, underlining its
   significance in cancer treatment^[82]23,[83]24. Recent studies have
   highlighted how dysregulated nucleotide metabolism affects the tumor
   microenvironment and immune response, offering potential for enhancing
   immunotherapy^[84]25,[85]26 Nucleotides and their metabolites can
   activate the immune system through host receptors, including Toll-like
   receptors, NOD-like receptors, adenosine receptors, and purinergic
   receptors^[86]27–[87]29. Thus, understanding the interplay between
   nucleotide metabolism and the immune response is crucial for optimizing
   cancer therapies. In glioma research, further investigation is required
   to understand the diverse biological roles of hyperactive nucleotide
   metabolism, with the aim of developing effective combination
   therapies^[88]20,[89]30.

   In this work, we present a comprehensive analysis of data from 188
   glioma patients, integrating MS-based proteomic data, genomic panel
   data, and clinical information. We identify two distinct proteomic
   subgroups, the immune subgroup (S-Im) and the metabolic-neural subgroup
   (S-Mn), each exhibiting unique immune and metabolic-neural
   characteristics that correlate with patient survival. Additionally, we
   discover two metabolic biomarkers, dihydropyrimidine dehydrogenase
   (DPYD) and thymidine phosphorylase (TYMP), which have significant
   prognostic and therapeutic implications, highlighting the relevance of
   nucleotide metabolism in glioma. Our results demonstrate the potential
   of proteomic classification and targeted metabolic reprogramming as
   promising strategies for glioma treatment.

Results

Overview of glioma proteomics analyses

   Proteomic profiling was performed using 343 tumor samples (including
   154 biological replicates) and 53 normal-appearing brain samples from a
   cohort of 188 glioma patients, including 6 pilocytic astrocytoma, 51
   diffuse astrocytomas (27 WHO grade 2, 14 grade 3, 9 grade 4, 1 grade
   NA), 47 diffuse oligodendrogliomas (28 WHO grade 2, 19 grade 3), 82
   glioblastomas (WHO grade 4), and 2 not elsewhere classified^[90]3. Our
   analysis employed a well-established MS workflow, as depicted in
   Fig. [91]1. To ensure the purity of samples (tumor or normal-appearing
   tissue) for proteome analysis, two neuroscientists independently
   confirmed the punch sites, and specimens of size 0.5 × 0.5 mm were
   extracted for proteomic analysis, as shown in Supplementary
   Fig. [92]S1. Clinical information, including gender, age, tumor
   location, WHO 2021 histology, grade, and survival, is summarized in
   Supplementary Data [93]1. The median age of the patients from whom the
   tumors were excised was 49.03 ± 14.48 years (range: 13–84 years), with
   a male-to-female ratio of 1.76:1 (male: N = 120; female: N = 68).
   Predominantly, the tumors were located in the frontal (99 cases) and
   temporal (46 cases) lobes. We obtained molecular pathology reports for
   129 samples from the Department of Pathology at Huashan Hospital. These
   reports detail the status of IDH mutation, O6-alkylguanine DNA alkyl
   transferase (MGMT) promoter methylation, 1p/19q co-deletion, and
   telomerase reverse transcriptase promoter mutation. The remaining 59
   samples were tested by immunohistochemistry only, and some marker
   statuses were unavailable. As of the last follow-up in September 2019,
   58 patients (30.8%) had passed away, 127 (76.6%) were still alive, and
   3 (1.6%) did not attend the follow-up. The median follow-up for
   patients alive at the last follow-up was 26.7 months.

Fig. 1. Overview of glioma proteomics analyses.

   [94]Fig. 1
   [95]Open in a new tab

   A comprehensive proteomic analysis is performed on a cohort of 188
   glioma patients, classified by WHO grades, including 343 tumor samples
   and 53 normal-appearing brain samples. Tissue specimens (0.5 × 0.5 mm)
   are extracted from formalin-fixed paraffin-embedded (FFPE) samples for
   proteomic analysis. Mouse liver samples are used as quality controls,
   alongside pooled samples for LC–MS/MS analysis. The analysis is
   conducted in 27 randomly assigned batches. The study integrates glioma
   genetic data, proteomic profiles, and clinical information, quantifying
   8561 proteins for downstream analyses. Proteomic profiling reveals two
   distinct subgroups: S-Mn (metabolic-neural) and S-Im (immune), which
   show significant differences in prognosis within both the glioma cohort
   and an independent validation cohort. Metabolic pathways are explored,
   leading to the identification of potential biomarkers, which are
   validated using cellular and mouse models to confirm their accuracy and
   relevance. Cartoons are created with BioRender.com.

   The proteomic analysis of all glioma samples and normal-appearing brain
   tissues was conducted by randomly assigning them to 27 batches
   (Supplementary Data [96]2). As part of quality control measures for
   sample preparation and MS, we conducted correlation analyses involving
   mouse liver samples (Supplementary Fig. [97]S2A) and pooled samples
   (Supplementary Fig. [98]S2B), which demonstrated a strong and
   satisfactory correlation, affirming the reliability and robustness of
   our experimental procedures. The stability of the MS analysis was
   confirmed by the Pearson correlation between replicates’ protein
   abundances, which was over 0.9 (Supplementary Fig. [99]S2C). The t-SNE
   plots indicated the absence of any batch effect among three MS machines
   (Supplementary Fig. [100]S2D), and the variance among the 27 batches
   was also represented in Supplementary Fig. [101]S2E. After removing
   isoforms, 8561 proteins were quantified for further analyses in
   combination with clinical information.

Proteomic features vary among glioma clinical classifications

   Next, we conducted an analysis to identify proteins with significantly
   different expression levels between glioma and normal-appearing brain
   samples. Our investigation demonstrated 570 proteins with significant
   differential abundance (Student’s t-test, Benjamini–Hochberg (B–H)
   adjusted P value < 0.001, fold change > 2) (Fig. [102]2A). Utilizing
   Ingenuity Pathway Analysis (IPA), we uncovered enrichment in proteins
   associated with neurotransmission, synaptogenesis, neuroinflammation,
   tumor necrosis, and the PD-1 signaling pathway within the glioma
   samples, indicating the vital role of neuronal and immune-associated
   proteins (Fig. [103]2B). Recent research has reported that neuron
   interaction with glioma cells can directly stimulate the progression
   and recurrence, with activation of WNT/PCP pathway and BRAF
   kinase^[104]31. Recognizing the significance of altered cellular
   metabolism as a communication interface between glioma cells and immune
   cells within the microenvironment^[105]18,[106]19, our subsequent focus
   centered on proteins related to immune regulation and metabolic
   processes. Among these 570 differentially expressed proteins between
   glioma and normal-appearing brain tissues (Student’s t-test, B–H
   adjusted P value < 0.001, fold change > 2), we identified 28
   immune-related and 51 metabolism-related proteins (Fig. [107]2C).
   Notably, the metabolism-related protein GPX7 and SOAT1, which are
   significantly upregulated in tumors, have been reported to lead to poor
   prognosis of glioma and participate in the immune regulatory process.
   For example, SOAT1 expression has a trend of positive correlation with
   Treg infiltration^[108]32, and GPX7 may be involved in immune
   mechanisms of innate immunity and adaptive immunity, type I interferon
   production, and regulation of synaptic transmission in low-grade
   glioma^[109]33.

Fig. 2. Proteomic features vary among glioma clinical classifications.

   [110]Fig. 2
   [111]Open in a new tab

   A Heatmap illustrates 570 differentially expressed proteins between 343
   tumor samples and 53 normal-appearing brain samples (Two-sided unpaired
   Welch’s t-test, B–H adjusted p < 0.001, fold change > 2). B Pathways
   are significantly enriched by IPA using the 570 differentially
   expressed proteins between tumor and normal-appearing brain samples
   (right-tailed Fisher’s Exact Test, p < 0.01). C Volcano plot shows 570
   differentially expressed proteins between 343 tumor and 53
   normal-appearing brain samples (Two-sided unpaired Welch’s t-test, B–H
   adjusted p < 0.001, fold change > 2). Immune-related and
   metabolism-related proteins are highlighted in red and purple,
   respectively (28 immune-related, 51 metabolism-related). D Manhattan
   plot displays 1047 differentially expressed proteins across four grades
   (one-way ANOVA test, B–H adjusted p < 0.001). Proteins associated with
   immunity and metabolism are marked in red and purple, respectively. E
   Enriched immune and metabolic processes, including nucleotide
   synthesis, antigen presentation, interferon signaling, and TGF-β
   pathways, are presented. Cartoons are created with BioRender.com.
   Statistical details for the relevant proteins are provided in
   Supplementary Data [112]3 (Welch’s t-test, B–H adjusted p < 0.001, fold
   change > 2). Source data are provided as a [113]Source Data file.

   In addition to the obvious protein differences between normal-appearing
   brain and glioma tissue, we also noticed the variations in glioma
   protein expression across different grades and between LGG and GBM. The
   t-SNE plots visually depicted distinctions among these various
   categories (Supplementary Fig. [114]S3A–C). Regarding histological
   types, the heatmap provided a clear illustration of how the
   differentially expressed proteins within distinct histological types of
   gliomas can be categorized into two distinct clusters (Supplementary
   Fig. [115]S3D). Cluster 1 exhibits significant enrichment in pathways
   related to protein translation, such as eIF2 and eIF4, while Cluster 2
   is notably enriched in pathways associated with synaptic synthesis and
   signal transduction pathways (Supplementary Fig. [116]S3E, F).
   Additionally, there were obvious distinctions in metabolic pathways,
   such as amino acid, nucleotide, and lipid metabolism, between these two
   clusters (Supplementary Fig. [117]S3G, H). We also observed distinct
   metabolic pathways, such as amino acid, nucleotide, and lipid
   metabolism, between these two groups (Supplementary Fig. [118]S3G, H).

   To uncover critical proteins contributing to tumor progression across
   various glioma grades, we identified 1047 differentially expressed
   proteins (ANOVA test B–H adjusted P value < 0.001) among WHO grades 1–4
   glioma. Notably, these proteins associated with immune and metabolic
   processes displayed considerable abundance (Fig. [119]2D). Immune and
   metabolism-related proteins can be clustered into four clusters with
   different expression trends as the tumor grade increases (Supplementary
   Fig. [120]S3I, J). They were predominantly enriched in pathways related
   to nucleotide synthesis, antigen presentation, interferon pathway, and
   TGF-β pathways, indicating intricate and complex interactions between
   metabolic characteristics of glioma cells and the immune
   microenvironment (Fig. [121]2E and Supplementary Data [122]3). Research
   has suggested the cross-talk between tumor metabolism and immune
   responses through immunometabolic rewiring^[123]34–[124]36, indicating
   that exploring the relationship between immunity and metabolism in
   glioma is worth studying.

Proteomic metabolic-neural and immune subtypes of glioma could predict
clinical outcomes

   Within our glioma cohort, we employed subtype classification based on
   proteome data, which is known to better reflect the disease state and
   underlying biology compared to genome data^[125]14. We firstly filtered
   355 the most significantly differential expression proteins between
   tumor and normal-appearing brain samples (proteins with missing rates
   over 20% were removed, missing values were imputed by substituting them
   with 80% of the minimum value, B–H adjusted P value < 0.05, fold
   change > 3). Then through the unsupervised clustering using these
   proteins, the 309 tumor samples were optimally classified into two
   groups when the Silhouette coefficient is closer to 1 (Supplementary
   Fig. [126]S4A), designated as subgroup 1 (N = 133) and subgroup 2
   (N = 176). Out of 355 proteins, a total of 312 proteins showed
   significantly differential expression between subgroups 1 and 2
   (adjusted p value < 0.05, Supplementary Data [127]4). Interestingly,
   subgroup 1 demonstrated enrichment in both metabolic processes and
   neurotransmitter receptor activity, including proteins like glutaminase
   (GLS), glutamate receptors, GABA receptors, and calcium channels.
   Consequently, we have designated this group as the metabolic-neural
   subgroup, abbreviated as S-Mn. In contrast, subgroup 2 displayed a
   significant upregulation of immune and inflammatory proteins, including
   AKT2, CDK2, TGFB1, and IRF3, which has been named as the immune
   subgroup, designated as S-Im (Fig. [128]3A). These subtypes exhibited
   distinct molecular and clinical features. The proteins enriched in the
   S-Mn subgroup were primarily associated with synaptogenesis signaling
   pathway and signal transduction, among other pathways. While the S-Im
   subgroup exhibited an enrichment of proteins mainly involved in the
   PD-1 cancer immunotherapy pathway and death receptor pathways. In
   particular, patients in S-Mn demonstrated better clinical outcomes than
   those in S-Im (Fig. [129]3B). The principal component analysis of the
   diffused glioma proteome effectively stratified the two subtypes, with
   the first principal component distinguishing them (Fig. [130]3C).
   Single-factor logistic analysis indicated there were differences in IDH
   status, WHO grade, Ki-67 index between the two-subgroup proteomic
   groups (Supplementary Data [131]5), and further multivariate logistic
   analysis revealed that IDH status and WHO grade has little impact on
   two-subgroup proteomic classification (Supplementary Fig. [132]S5).

Fig. 3. Proteomic metabolic-neural and immune subtypes of glioma could
predict clinical outcomes.

   [133]Fig. 3
   [134]Open in a new tab

   A Heatmap shows two proteomic subtypes within our study cohort, based
   on 355 differentially expressed proteins between tumor and
   normal-appearing brain samples (proteins with missing rates over 20%
   were removed, missing values were imputed by substituting them with 80%
   of the minimum value, Two-sided unpaired Welch’s t-test, B–H adjusted P
   value < 0.05, fold change > 3). Subgroup 1 enriches both metabolic
   processes and neurotransmitter receptor activity proteins, designated
   as the metabolic-neural subgroup (S-Mn). Subgroup 2 exhibits
   significant upregulation of immune and inflammatory proteins,
   designated as the immune subgroup (S-Im) (two-sided unpaired Welch’s
   t-test, B–H adjusted p < 0.05, fold change > 3). Pathways are
   significantly enriched by IPA using the proteins in S-Mn (upper) and
   S-Im (lower) groups, with key proteins listed. B Kaplan–Meier survival
   curve with 95% confidence interval (CI) from our glioma cohort shows
   better clinical outcomes for S-Mn (N = 133) subtype compared to the
   S-Im subtype (N = 176) (log-rank test, p < 0.0001). C PCA analysis of
   the glioma proteome effectively stratifies the two subtypes. D Heatmap
   of TCGA gene expression data validates our two-subtype proteomic
   classification using the same 355 differentially expressed proteins as
   in (A). E Kaplan–Meier survival curve with 95% CI based on TCGA data
   shows better OS for the S-Mn subtype (n = 267) and poor OS for the S-Im
   subtype (N = 308) (log-rank test, p < 0.0001). F Heatmap of proteomic
   data from NC 2022 displays the top 2000 differentially expressed
   proteins, classified by our proteomic subtype. G Kaplan–Meier curves
   with 95% CI reveal a better prognosis for the S-Mn group (N = 36)
   compared to the S-Im group (N = 151) based on NC 2022 data (log-rank
   test, p < 0.0001). Source data are provided as a [135]Source Data file.

   To validate the accuracy of our proteomic classification, we utilized
   TCGA gene expression data (N = 575). Our analysis revealed that the
   subtypes could effectively differentiate between glioma patients
   (Fig. [136]3D), with patients from the S-Mn subtype showing a favorable
   prognosis, while those from the S-Im subtype displayed a poorer
   prognosis (Fig. [137]3E). Furthermore, we corroborated our proteomic
   classification using proteomics data from 187 glioma patients,
   published by Qie et al. in Nature Communications in 2022^[138]37. The
   heatmap displayed the top 2000 differentially expressed proteins
   (Fig. [139]3F), and Kaplan–Meier curves revealed significant prognostic
   differences between the S-Mn (N = 36) and S-Im (N = 151) subtypes,
   thereby further validating our findings (Fig. [140]3G). Considering the
   different prognoses of patients with different WHO grades
   (Supplementary Fig. [141]S6A) and IDH status, we also stratified the
   data and performed subgroup clustering of grade 1–3 (Supplementary
   Figs. [142]S4B and [143]S6B), grade 4 (Supplementary Figs. [144]S4C
   and [145]S6C), and IDH wild-type (Supplementary Figs. [146]S4D
   and [147]S6D) patients. All the optimal grouping of patients was
   achieved by dividing them into two categories, with those in S-Mn
   demonstrating better clinical survival regardless of their WHO grade
   and IDH status, indicating the stability and predictive power of our
   proteomics subtype clustering, suggesting that two-subgroups
   classification is applicable to all diffuse gliomas.

   Numerous studies have revealed the importance of the IDH status on the
   metabolic and epigenetic reprogramming of gliomas^[148]38,[149]39. The
   significance of IDH status was officially recognized in the WHO
   classification of CNS tumors in 2016, where it was established as a
   biomarker with prognostic outcomes. In our investigation, we conducted
   a comparative analysis between IDH mutant (N = 87) and IDH wild-type
   (N = 81) gliomas across varying WHO grades, resulting in the
   identification of 203 differentially expressed proteins (Supplementary
   Fig. [150]S7A). The volcano plot distinctly illustrated differential
   protein expression profiles between IDH mutant and wild-type gliomas
   (Supplementary Fig. [151]S7B). Among them, the mitochondrial enzymes
   glutamate dehydrogenase GLUD1 and GLUD 2 are significantly upregulated
   in IDH mutant tumors, and literature also shows that high glutamate
   neurotransmitter flux renders cells resistant to growth-inhibitory
   effects of IDH mutation by supplying α-KG to fuel the citric acid cycle
   and support lipid synthesis^[152]40. In addition, glioma progression
   markers, such as LGALS3, FABP5, MSN, and TAGLN2 were significantly
   down-regulated in the IDHmut group^[153]41–[154]44. Notably, our
   analysis revealed that metabolic processes and cytokine signaling
   pathways, such as lipid metabolism, fatty acid transport, DNA
   biosynthetic process, IL-12 signaling pathway, and IL-6 signaling
   pathway, were among the significantly affected pathways (Supplementary
   Fig. [155]S7C). Next, we investigated the differential proteins within
   IDH mutant tumors at different stages. Previous studies have
   illuminated the substantial prognostic disparities between grade 4 and
   grade 2/3 astrocytomas^[156]45,[157]46. Our findings indicated a
   significant decrease in CDK2 and IRF3 expression in grade 2/3
   astrocytoma compared to grade 4, which is consistent with the
   established consensus that the loss of CDKN2A/B, accompanied by
   increased CDK activity, is intimately associated with the malignant
   progression of gliomas (Supplementary Fig. [158]S7D). Finally, the cell
   cycle control of chromosomal replication pathway was identified as the
   primary differential pathway indicating the different proliferation
   characteristics between grade 2/3 and grade 4 astrocytomas
   (Supplementary Fig. [159]S7E).

Pyrimidine metabolism is associated with glioma patients’ survival

   Metabolic reprogramming stands as a hallmark of cancer and represents
   an early event in the gliomagenesis^[160]47,[161]48. For instance, GBM
   cells adapt to their distinctive microenvironment by upregulating
   enzymes associated with pyrimidine metabolism, facilitating rapid cell
   proliferation and DNA damage repair^[162]49,[163]50. To narrow down the
   proteins that were more associated with survival, we selected 89
   patients with long-survival (≥30 months, N = 48) and short-survival
   (≤12 months, N = 41). The overall survival distribution of different
   grade and IDH wild-type glioma was depicted in Supplementary
   Fig. [164]S8A. As illustrated in the flow chart in Supplementary
   Fig. [165]S8B, our initial screening process focused on identifying the
   differential expression proteins between tumor and normal-appearing
   brain samples associated with tumor metabolism in Fig. [166]3A.
   Subsequently, we conducted further screening to identify metabolic
   proteins that also showed a significant association with prognosis
   (log-rank p value < 0.05) in this cohort of 89 patients (14 metabolic
   proteins and 11 metabolic pathways). Meanwhile, we selected the
   differentially expressed proteins among four groups and were also
   associated with metabolism (47 metabolic proteins and 34 metabolic
   pathways). The intersection of these two screening stages resulted in
   the identification of 11 pathways, including 25 proteins. Following
   screened by log-rank p value < 0.05, 18 proteins were found to be both
   metabolically relevant and prognostically significant (Fig. [167]4A).
   All the expression levels of these 18 proteins were found to be closely
   associated with patients’ prognosis (Supplementary Fig. [168]S8C). We
   proceeded to establish a risk-scoring model based on the expression of
   these proteins, and Kaplan–Meier curves affirming that patients in the
   low-risk score group exhibited more favorable clinical outcomes
   (Fig. [169]4B). Upon further analysis of these metabolic pathways, we
   discovered their predominant involvement in pyrimidine metabolism, a
   process closely related to DNA and RNA synthesis (Fig. [170]4C). To
   validate the robustness of this model, we utilized TCGA data, yielding
   consistently the same prognosis conclusion as our own data
   (Fig. [171]4D). Thus, our analysis further confirmed the crucial
   involvement of nucleotide metabolism in controlling glioma cell
   behavior and patient survival and how the metabolic reprogramming
   especially the nucleotide metabolism contributes to tumor malignancy
   remained to be explored.

Fig. 4. Pyrimidine metabolism associate with glioma patients' survival.

   [172]Fig. 4
   [173]Open in a new tab

   A Heatmap shows 18 different expressed proteins between the
   long-survival (N = 48) and short-survival (N = 41) groups, filtered
   through the methods shown in Supplementary Fig. [174]S8B. B
   Kaplan–Meier survival curve with 95% CI (high-risk group, N = 41;
   low-risk group, N = 48) illustrates the risk-scoring model based on
   these 18 proteins (log-rank test, p < 0.0001). C Pathway enrichment
   analysis indicates that pyrimidine metabolism is predominantly
   involved. D Kaplan–Meier survival curve with 95% CI using TCGA data
   (high-risk group, N = 288; low-risk group N = 287) validates the
   risk-scoring model (log-rank test, p < 0.0001). Source data are
   provided as a [175]Source Data file.

DPYD and TYMP are required for GSCs proliferation and self-renewal

   As illustrated in the comprehensive nucleotide metabolism map depicted
   in Fig. [176]4C highlighting proteomic-identified differential
   nucleotide enzymes, we found the significant roles played by catabolic
   enzymes within this process. The homeostasis of nucleotide metabolism
   requires the balance between nucleotide biosynthesis and degradation
   and affecting the balance of nucleotide pools could cause genomic
   instability and cancer progression. Among them, DPYD and TYMP are
   enzymes that regulate the catabolism of pyrimidine nucleosides. To
   explore their potential functions and mechanisms in glioma, we focused
   on their roles in patient-derived glioma stem cells (GSCs), which are
   at the top of the glioblastoma cell hierarchy and have a self-renewing
   effect that contributes to treatment resistance and tumor
   recurrence^[177]51.

   Initially, we divided patients into high- and low-expression groups
   based on the levels of DPYD and TYMP expression derived from our
   proteomic profiling. To validate the accuracy of our proteomic
   profiling findings, we conducted immunohistochemical (IHC) analyses on
   seven randomly selected patient samples from each group and the protein
   levels obtained from IHC were consistent with our proteomic profiling
   (Fig. [178]5A and Supplementary Fig. [179]S9A). Subsequently, we
   employed two independent shRNAs to efficiently knock down the DPYD and
   TYMP mRNA levels in T4121 and Mes28 cells, as confirmed by real-time
   PCR (Fig. [180]5B and Supplementary Fig. [181]S9B).

Fig. 5. DPYD is required for GSC proliferation and self-renewal.

   [182]Fig. 5
   [183]Open in a new tab

   A IHC validation of dihydropyrimidine dehydrogenase (DPYD) expression
   based on proteomic profiling. The optical density of IHC staining is
   measured quantitatively, confirming the different expression between
   the two groups (high group, N = 21; low group, N = 21; Student’s
   t-test, p < 0.0001; scale bar = 50 μm). B DPYD mRNA levels and protein
   levels in T4121 and Mes28 cells after knockdown show an efficiency
   greater than 50% (mRNA, mean ± SEM, N = 4; Student’s t-test; in T4121,
   shCont vs shDPYD-1, p < 0.0001 and shCont vs shDPYD-2, p < 0.0001; in
   Mes28, shCont vs shDPYD-1, p < 0.0001 and shCont vs shDPYD-2,
   p < 0.0001). C Proliferation assay demonstrates that targeting DPYD
   impairs proliferation, as assessed by cell number on day 3 (mean ± SEM,
   N = 4, Student’s t-test; in T4121, shCont vs shDPYD-1, p = 0.0104 and
   shCont vs shDPYD-2, p = 0.0198; in Mes28, shCont vs shDPYD-1,
   p = 0.0406 and shCont vs shDPYD-2, p = 0.0599). GSC proliferation is
   significantly inhibited on day 5 (mean ± SEM, N = 4, Student’s t-test;
   in T4121, shCont vs shDPYD-1, p = 0.0001 and shCont vs shDPYD-2,
   p < 0.0001; in Mes28, shCont vs shDPYD-1, p < 0.0001 and shCont vs
   shDPYD-2, p = 0.0014). D In vitro limiting dilution assays (analysis
   was performed using software available at
   [184]http://bioinf.wehi.edu.au/software/elda with simple line
   regression; in T4121, shCont vs shDPYD-1, p < 0.0001 and shCont vs
   shDPYD-2, p = 0.0001; in Mes28, shCont vs shDPYD-1, p < 0.0001 and
   shCont vs shDPYD-2, p < 0.0001) and (E, G) tumorsphere formation reveal
   impaired self-renewal of DPYD knockdown GSCs. Sphere number (E,
   mean ± SEM, N = 8, Student’s t-test; in T4121, shCont vs shDPYD-1,
   p < 0.0001 and shCont vs shDPYD-2, p < 0.0001; in Mes28, shCont vs
   shDPYD-1, p < 0.0001 and shCont vs shDPYD-2, p < 0.0001) and size (G,
   scale bar = 20 μm) are significantly decreased in DPYD knockdown GSCs.
   H, F Immunofluorescence shows increased phospho-γH2A.X foci in the
   nucleus of GSCs after DPYD knockdown, indicating severe DNA damage (F,
   mean ± SEM, N ≥ 4, Student’s t-test in foci > 15 as positive; in T4121,
   shCont vs shDPYD-1, p < 0.0001 and shCont vs shDPYD-2, p < 0.0001; in
   Mes28, shCont vs shDPYD-1, p = 0.0009 and shCont vs shDPYD-2,
   p = 0.0028). The percentage of phospho-γH2A.X positive cell percentage
   is calculated, indicating elevated DNA damage after DPYD knockdown. I
   Western blot shows increased protein levels of phospho-γH2A.X and
   Cleaved Caspase-3 (CC3) proteins following DPYD knockdown. J Knockdown
   of DPYD in T4121 (each group, N = 8) and Mes28 (each group, N = 6)
   significantly increases mouse survival (log-rank test; in T4121, shCont
   vs shDPYD-1, p = 0.0171 and shCont vs shDPYD-2, p = 0.0041; in Mes28,
   shCont vs shDPYD-1, p = 0.0005 and shCont vs shDPYD-2, p = 0.0024). H&E
   staining of mouse brain shows that DPYD inhibition suppresses GSC
   growth (N = 3, scale bar = 1 mm). K An external validation of another
   cohort of glioma patients (N = 43) confirms differential expression by
   IHC, and the Kaplan–Meier survival curve shows poor prognosis in the
   high DYPD expression group (log-rank test, p < 0.0001; scale
   bar = 100 μm). Source data are provided as a [185]Source Data file.

   Knocking down DPYD and TYMP in GSCs markedly impaired their
   proliferation ability, as evidenced by reduced cell numbers on the 3rd
   and 5th days following knockdown (Fig. [186]5C and Supplementary
   Fig. [187]S9C). Moreover, the self-renewal capacity of GSCs was
   impaired, as assessed through in vitro limiting dilution assays
   (Fig. [188]5D and Supplementary Fig. [189]S9D) and neurosphere
   formation (Fig. [190]5E and Supplementary Fig. [191]S9E). Specifically,
   the number and size of spheres significantly decreased in DPYD and TYMP
   knockdown GSCs (Fig. [192]5G and Supplementary Fig. [193]S9F).

   Given the close connection of these two proteins to DNA metabolism, and
   the well-known DNA damage repair capacity of GSCs, which contributes to
   treatment resistance (chemo drug and radiation), we evaluated the level
   of phospho-γH2A.X in GSCs through western blot and immunofluorescence.
   The knockdown of DPYD resulted in severe DNA damage and apoptosis, as
   indicated by an increase in phospho-γH2A.X foci within the nucleus
   (Fig. [194]5F, I) and elevated protein levels of phospho-γH2A.X and
   cleaved caspase-3 (CC3) (Fig. [195]5I). Collectively, these findings
   support the critical function of DPYD and TYMP in GSC maintenance and
   DNA metabolism.

   To address the potential benefit of therapeutic targeting of DPYD and
   TYMP in vivo, mice bearing intracranial patient-derived GSCs tumors
   were transduced with a non-targeting, control shRNA (shCont) or shRNA
   targeting DPYD and TYMP. In particular, we aimed to investigate their
   roles in tumor initiation and growth. Knockdown of DPYD and TYMP
   significantly prolonged the survival of mice (Fig. [196]5J and
   Supplementary Fig. [197]S9G). Subsequently, the mice were sacrificed,
   and their brains were subjected to hematoxylin and eosin (H&E) staining
   to assess glioma presence, confirming the absence of glioma growth in
   GSCs lacking DPYD and TYMP (Fig. [198]5J and Supplementary
   Fig. [199]S9G). Additionally, immunohistochemical analysis of
   pathological sections from 43 glioblastoma (GBM) patients revealed that
   high-expression levels of DPYD and TYMP were associated with poor
   prognosis (Fig. [200]5K, Supplementary Fig. [201]S9H and Supplementary
   Data [202]6). Together, these findings have revealed the significance
   of DPYD and TYMP in the initiation and growth of glioma in vivo, there
   thereby reinforcing the importance of targeting nucleotide metabolism
   as a potential strategy for the treatment of glioma.

Discussion

   In this study, we conducted a large-scale proteomic analysis of
   gliomas, with a specific focus on uncovering the immune and metabolic
   heterogeneity within these tumors, while also exploring proteome-based
   subtyping. Our investigations led to the identification of two distinct
   proteomic subtypes within diffuse glioma, each characterized by unique
   immune and metabolic features. Importantly, these subtypes demonstrated
   a robust association with clinical outcomes. In addition, we validated
   the effectiveness of our proteomic subtyping approach using data
   sourced from the TCGA database. Our biological experiments also
   revealed the critical roles of DPYD and TYMP in promoting the
   proliferation and self-renewal of GSCs, making them potential targets
   for precision treatment strategies.

   For many years, adult malignant gliomas have been managed using uniform
   approaches, regardless of their molecular characteristics, often
   following the conventional Stupp regimen^[203]50. While significant
   progress has been made in understanding glioma development at the DNA
   and RNA levels, the establishment of targeted therapies tailored to
   specific molecular profile, such as IDH mutant or VEGF expression
   gliomas, has remained elusive. We propose that, given the primary role
   of proteins in executing physiological functions within organisms, a
   proteome-based approach may uncover signaling pathways that could pave
   the way for the development of innovative therapeutic strategies.
   Indeed, there is increasing evidence in the literature that supports
   the existence of distinct metabolic proteomic subtypes in glioblastoma
   based on aerobic/anaerobic energy metabolism^[204]7,[205]52.

   We have made several significant discoveries in our study. Firstly, we
   identified key proteins that distinguish between normal-appearing brain
   tissue and glioma tissue, with a focus on pathways related to synaptic
   synthesis, neuroinflammation, and PD-1 pathways. These pathways are
   closely associated with tumor metabolism and immune inflammation.
   Furthermore, we categorized WHO grade 2–4 gliomas into two subtypes
   based on proteomic differences, revealing significant disparities in
   patient survival between these subtypes. Notably, this proteomic
   classification remained robust across varying tumor WHO grade and IDH
   status. To further explore the underlying prognostic mechanisms, we
   selected differentially expressed proteins associated with metabolism
   for subsequent analysis and investigation.

   Metabolic reprogramming stands as a well-established hallmark of
   cancer, and the growing understanding of the cross-talk between
   metabolism and cancer epigenetics expands our insights into the
   contributions of metabolic derangements to tumor malignancy. Within
   this context, pyrimidine metabolism has emerged as a promising
   vulnerability in IDH mutant glioma. For example, BAY 2402234, an
   inhibitor targeting dihydroorotate dehydrogenase in pyrimidine
   synthesis, has exhibited the capacity to sensitize DNA damage in
   response to nucleotide pool imbalance, thus influencing clinical
   outcomes in patients^[206]53,[207]54. However, the maintenance of a
   dynamic DNA metabolism pool requires not only enzymes involved in
   anabolism but also those participating in catabolism. In our
   investigation, we unveiled the critical role played by pyrimidine
   catabolism enzymes in glioma, and notably, their association with
   clinical prognosis, which is consistent with the current prevailing
   understanding in the field. Our study provides a valuable contribution
   to the field of cancer research by highlighting the critical importance
   of nucleotide metabolism in cancer cells and the potential of targeting
   this metabolic dependency as an effective cancer therapy strategy.
   DPYD, which serves as the initial committed step in pyrimidine
   degradation, has been extensively investigated, particularly in the
   context of 5-FU metabolism, a widely used chemotherapy drug. Altered
   DPYD activity in the context of tumorigenesis may disrupt the balance
   of nucleotide pools within cells, potentially leading to genomic
   instability and contributing to cancer development. Notably, in glioma
   research, studies have demonstrated elevated DPYD expression in diffuse
   midline gliomas compared to normal astrocytes, indicating resistance to
   5-FU treatment and underscoring the malignant phenotype of glioma
   biology^[208]55. Furthermore, in pancreatic cancer, a disease
   exhibiting similar aggressiveness like glioma, the inhibition of
   STAT3-DYPD expression by chemopreventive agent has demonstrated the
   capacity to restrain tumor proliferation^[209]56. Beyond its influence
   on tumor cell proliferation, DPYD has been implicated in maintaining a
   mesenchymal-like state in epithelial tumors by directing pyrimidine
   metabolites towards the catabolic pathway^[210]57. TYMP encodes the
   enzyme thymidine phosphorylase, an enzyme that exhibits overexpression
   in various cancer types, with established roles in tumor growth,
   invasion, and chemotherapy resistance^[211]58–[212]60. Beyond its
   primary role in thymidine metabolism, TYMP serve as a potent angiogenic
   factor and functions as a cytokine that stimulates endothelial cell
   proliferation and migration, thereby promoting the formation of new
   blood vessels^[213]61. This angiogenic activity of TYMP holds
   particularly relevant in the context of glioma, given that tumor growth
   and invasion critically depend on the development of a robust blood
   supply. Elevated TYMP levels have been associated with increased
   vascularity and more aggressive tumor behavior. In general, our study
   provided experimental validation of the roles of DPYD and TYMP, which
   have traditionally been investigated primarily in the context of
   chemotherapy response within the malignant progression of glioma. These
   findings demonstrated that the downregulation of these two enzymes
   significantly reduces the self-renewal capacity of GSCs, ultimately
   leading to DNA damage, apoptosis, and impairments of in vivo
   tumorigenesis, which strongly suggest that DPYD and TYMP might
   contribute to the genesis of glioma through mechanisms distinct from
   their known involvement in drug metabolism. Moreover, considering the
   widespread use of drugs targeting these two enzymes in preclinical
   gastrointestinal and urinary tumor models^[214]62–[215]65, conducting
   clinical trials to evaluate the efficacy of these drugs in glioma
   patients represents a promising direction for future research.

   Our study has some limitations that should be acknowledged. First, our
   analysis only measures protein expression, and does not account for
   post-translational modifications, such as phosphorylation and
   glycosylation, that may also affect glioma pathogenesis. Second, our
   proteomic subtyping approach needs to be validated in larger cohorts to
   increase its robustness. Third, more research is needed to elucidate
   the exact molecular pathways and downstream effectors of DPYD and TYMP
   in glioma development.

   In conclusion, our study provides insights into the molecular
   mechanisms of glioma and identifies potential biomarkers and
   therapeutic targets. Proteomic subtyping emerges as a more accurate
   predictor of patient prognosis in glioma. Notably, enzymes involved in
   pyrimidine catabolism, specifically DPYD and TYMP, emerge as pivotal
   players in glioma pathogenesis. Further research is warranted to
   validate and expand upon these findings, with the aim of improving the
   diagnosis and treatment of this devastating cancer.

Methods

   This study was approved by the Ethics Committee of Huashan Hospital
   (KY2021-064). Patients provided informed consent, agreeing to donate
   tumor samples and to the publication of clinical information for
   research purposes.

Clinical samples collection and preparation

   Formalin-fixed and paraffin-embedded (FFPE) human brain glioma samples
   were obtained from the Department of Neurosurgery, Huashan Hospital of
   Fudan University during January to December 2018. The specimens were
   diagnosed and graded according to WHO guidelines by two
   neuropathologists^[216]3. Patients provided informed consent
   (KY2015-256) before surgery, agreeing to donate tumor samples and
   personal information for research purposes. The acquisition and use of
   human glioma samples were approved by the Ethics Committee of Huashan
   Hospital (KY2021-064) and adhered to ethical principles outlined in the
   Helsinki Declaration (1964, amended in 2013) and the guidelines of the
   World Medical Association^[217]66. A total of 188 glioma patients were
   included in the study, with a male-to-female ratio of 1.76:1 (120
   males, 68 females).

   For the subsequent proteomic analysis, two neuroscientists selected
   representative histological area of either the tumor or
   normal-appearing tissue based on H&E staining slides, ensuring sample
   purity for punching. Areas displaying hemorrhagic, necrotic, or
   degenerative changes, such as mucous or cystic changes, were
   deliberately excluded from the punching process. Normal-appearing brain
   tissue samples (53 samples from 32 cases) were included as controls.
   All samples were obtained using a 1 mm biopsy punch (Kai medical,
   #BP-10F) for MS analysis. To minimize potential bias from tumor
   heterogeneity, portions of FFPE samples were punched at two independent
   sites.

Hematoxylin and eosin (H&E) staining

   The FFPE tissues were sliced into 4 µm sections, and H&E staining was
   performed following standard protocols using the fully automated
   BOND-III stainer system (Leica Biosystems, USA).

Peptide sample preparation

   Protein extraction and peptide digestion were performed as described
   previously^[218]67,[219]68. Briefly, the FFPE tissue samples were
   dewaxed by heptane and hydrated with a gradient ethanol solution.
   Proteins were extracted from the tissues and then enzymatically
   digested into peptide samples using trypsin (Hualishi Scientific,
   China) and LysC (Hualishi Scientific, China) relying on PCT (Pressure
   Biosciences Inc., MA, USA)^[220]67,[221]68.

Liquid chromatography-tandem mass spectrometry (LC–MS/MS)

   Clean peptide samples were separated with a nano Elute system (Bruker
   Daltonics, Germany) at a flow rate of 300 nL/min, at 217.5 bar. The
   linear liquid phase gradient of buffer B was 5–27% for 90 min during
   the data-dependent acquisition (DDA) mode and then for 60 min during
   the PulseDIA acquisition. The mobile phase was mixed with buffer A (2%
   ACN, 0.1% formic acid) and buffer B (98% ACN, 0.1% formic acid).
   Peptides were scanned by a Captive Spray nanoelectrospray ion source on
   a hybrid trapped ion mobility spectrometer (TIMS) quadrupole
   time-of-flight mass spectrometer (TIMSTOF Pro, Germany).

   To generate a library of ion mobility-enhanced spectra, we performed
   the DDA of the Parallel Cumulative Sequential Fragmentation (PASEF)
   mode with 10 PASEF scans every top-N acquisition cycle. A total cycle
   time of 1.17 s was achieved with an accumulation and ramp time of
   100 ms each for the dual TIMS analyzer. The ion mobility was scanned
   from 0.6 to 1.6 V/cm^2. The MS1 and MS2 acquisitions were performed in
   the m/z range of 100–1700 Th. Precursors reaching a target value of
   20,000 arbitrary units were dynamically excluded for 0.4 min, and
   singly charged precursors were excluded at positions in the m/z-plane
   of ion mobility.

   The peptide samples were analyzed using the PulseDIA mode^[222]69 of
   the diaPASEF^[223]70. Ion mobility was scanned from 0.7 to 1.3 V/cm^2.
   MS1 and MS2 acquisitions were performed in the m/z range of
   100–1700 Th. We defined two sets of complementary isolation windows and
   applied them to the two MS methods for two injections. We set a window
   with a width of 25, and the rest of the parameters were the same as for
   the DDA.

High-pH reversed-phase chromatography fractionation

   Approximately 100 μg of peptide mix of glioma tumor and
   normal-appearing brain were fractioned on a chromatographic column (BEH
   C18, 300 Å, 5 μm, 4.6 mm × 250 mm) coupled to a Thermo Dionex Ultimate
   3000 (Thermo Fisher Scientific, USA). A linear gradient of 5–35% buffer
   B (98% ACN, 0.6% ammonia, pH = 10) was kept for 68 min to generate 60
   fractions. The fractions were next combined into 20 samples according
   to a published strategy^[224]71. Finally, all samples were dried by a
   SpeedVac (Thermo Fisher Scientific, USA), reconstituted in 0.1% formic
   acid, and spiked with standard peptides (iRT, Switzerland).

Human glioma-specific spectral library generation and proteomic data analysis

   The 20 glioma DDA data files and 88 normal brain DDA data files were
   analyzed using Spectronaut (version 14.5). The background FASTA file
   was downloaded from SwissProt on 22 January 2020 and contained 20,365
   human protein entries. The tolerance for fragment and precursor was set
   to dynamic, which depends on each file. The digestion enzyme was
   trypsin, cutting after “KR” but not before “P.” False discovery rates
   (FDRs) of peptide-spectrum matches, precursors, and proteins were set
   to 1%. A total of 17,0445 precursors and 10,067 proteins were
   identified in this spectral library. The self-constructed library was
   further used for the PulseDIA data analysis by Spectronaut (version
   14.5) without cross-run normalization. FDRs of precursors and proteins
   were set to 1%. Other settings were used as default parameters. Missing
   values were imputed by 0.8 times the minimum intensity found in the
   matrix of 8561 proteins.

Proteomic data quality control analysis

   To minimize the batch effect, the 343 tumor and 53 normal-appearing
   brain samples were randomly separated into 27 batches each
   (Supplementary Data [225]2) for proteomic data acquisition. During
   sample preparation, from batches 1 to 25, each batch contained 13 tumor
   samples, two normal-appearing brain samples, and one mouse liver sample
   for PCT quality control; batch 26 included ten tumor samples, two
   normal-appearing brain samples, and one mouse liver sample; batch 27
   included nine samples contaminated by blood and one mouse liver sample.
   Additionally, for the LC–MS/MS acquisition, each batch contained one
   pooled peptide sample and one randomly selected peptide sample as a
   replicate for the LC–MS/MS technical quality control.

Differentially expressed analysis

   A two-sided unpaired Welch’s t-test was used to compare the two groups.
   And p values were adjusted using the B–H method. The differentially
   expressed proteins between two groups were filtered with fold change
   and B–H adjusted p value, while among over two groups were filtered
   with one-way ANOVA p value. The pathway enrichment analysis was
   performed using differentially expressed proteins among different
   groups. The activation degree score of the pathway is calculated in
   IPA.

Proteomic-based clustering analysis

   To determine the optimal number of subtypes, we utilized K-means
   clustering (implemented through the “kmeans” function in R) and
   consensus clustering (implemented through the “consensusClusterPlus”
   package in R). In order to cluster the samples based on the constituent
   pattern of each pathway, we scaled each sample. Consensus clustering
   was then employed to evaluate the robustness of the K-means clustering,
   with 1000 iterations and 80% resampling. Finally, a silhouette analysis
   was performed to confirm the stability of the clustering. The best
   method was used for clustering all samples.

Cell culture and reagents

   Patient-derived GSCs (T4121 and Mes28) were derived by our laboratory
   and obtained via a material transfer agreement with Duke University.
   These GSCs were cultured in Neurobasal media (Thermo Fisher Scientific,
   #21103049) supplemented with 2% B27 supplement without vitamin A
   (Thermo Fisher Scientific, # 12587010), EGF and bFGF (20 ng/mL each;
   R&D Systems, #236-EG and # 3718-FB), 1% sodium pyruvate (Thermo Fisher
   Scientific, #11360070), and 1% GlutaMAX (Thermo Fisher Scientific,
   #35050079). In addition, 293T cells (sourced from ATCC) were cultured
   in Dulbecco’s modified Eagle’s medium (Thermo Fisher Scientific,
   #11965092) with 10% fetal bovine serum (Thermo Fisher Scientific,
   #A4766801).

Plasmids and lentiviral transduction

   Two non-overlapping shRNAs targeting human DPYD (target sequences:
   GCCGTATGATGTAGTGAATTT and GCAATTTGCTACTGAGGTATT), TYMP (target
   sequences: GCCTCCATTCTCAGTAAGAAA and GCTGGAGTCTATTCCTGGATT), and a
   non-targeting control shRNA (shCont, CCTAAGGTTAAGTCGCCCTCG) were
   construed into a PLKO.1 plasmid. To generate lentiviral particles, we
   employed 293T cells and conducted co-transfection with the packaging
   vectors psPAX2 (Addgene, #12259) and pMD2.G (Addgene, #12259) using the
   polyethylenimine transduction method in the OPTI-MEM medium (Thermo
   Fisher Scientific, #31985062). After 8 h, the medium was changed. The
   supernatants containing the viruses were collected in a 50 mL tube at
   24, 48, and 72 h after transfection, filtered and concentrated by
   PEG8000 (Sigma, # 89510) precipitation. Cells were then infected with
   lentiviruses for further experiments with 8 μg/mL polybrene which
   enhances infection efficiency. Stable cells were selected with media
   containing 2 μg/mL puromycin. The stable cells were detected by qRT-PCR
   and immunoblotting.

Quantitative RT-PCR

   An RNA purification kit (EZB, #B0004DP) was used to isolate total
   cellular RNA from cell pellets. The 4× EZscript Reverse Transcription
   Mix II with gDNA remover (EZB, RT2GQ) was utilized to reverse cDNA
   transcription. Quantitative real-time PCR was performed on the
   QuantStudio 6 Flex Real-Time PCR System (Thermo Fisher Scientific) with
   a 2× S6 universal SYBR qPCR mix (NovaBio, #Q204). The primer pairs for
   qPCR in this study were as follows: human DPYD forward
   5′-GCTGTCCCTGAGGAGATGGA-3′ and reverse 5′- GTCCGAACAAACTGCATAGCAA-3′;
   TYMP forward 5′-CACCTTGGATAAGCTGGAGTC-3′ and reverse
   5′-GGCTGCATATAGGATTCCGTC-3′; ACTIN forward 5′-TCCCTGGAGAAGAGCTACG-3′
   and reverse 5′-GTAGTTTCGTGGATGCCACA-3′.

Western blot

   Equal cell amounts per group were collected and boiled in sodium
   dodecyl-sulfate polyacrylamide gel electrophoresis loading buffer.
   Electrophoresis was performed using 12% protein gels, followed by
   transfer onto polyvinylidene fluoride membrane. TBST supplemented with
   3% bovine serum albumin (BSA) was used for blocking for 1 h at room
   temperature. Blotting with primary antibody was performed at 4 °C
   overnight. The following antibodies were used: DPYD (1:1000, Abcam,
   ab134922), anti-Cleaved Caspase-3 (1:1000, CST, #9661),
   anti-phospho-γH2A.X (1:1000, CST, #9718), and anti-α-Tubulin (1:5000,
   Proteintech, #66031-1-lg). A secondary antibody (1:2000, Proteintech,
   #SA00001-2 and #SA00001-1) was applied to bind with the primary
   antibody for 2 h. Finally, membranes were visualized under an enhanced
   chemiluminescence detection system (Typhoon FLA 9500, GE Healthcare).

In vitro limiting dilution assay

   In vitro limiting dilution was performed following established
   protocols^[226]72,[227]73. Briefly, decreasing numbers of cells per
   well (50, 20, 10, 5, and 1) were plated into 96-well plates. Seven days
   after plating, the presence and count of neurospheres in each well were
   recorded. An extreme limiting dilution analysis was performed using
   software available at
   [228]http://bioinf.wehi.edu.au/software/elda^[229]72,[230]73.

Cell proliferation and neurosphere formation assay

   Cell proliferation experiments were conducted by seeding the cells of
   interest at a density of 10,000 cells/well in 12-well plates, with four
   replicates for each condition. Cell counts were performed on day 3 and
   day 5 to assess cell proliferation. The neurosphere formation was
   measured by seeding 5000 cells into a 12-well plate with eight
   replicates for each group. The presence and quantity of tumor spheres
   in each well were recorded 7 days after plating.

Immunohistochemistry (IHC) staining

   The IHC staining procedures followed established protocols^[231]74. In
   brief, 3% H[2]O[2] was used to quench endogenous peroxidases, followed
   by antigen retrieval using the citrate buffer. Primary antibodies
   against DPYD (1:500, Proteintech, #27662-1-AP) and TYMP (1:200,
   Proteintech, #12383-1-AP) were incubated overnight at 4 °C. Staining
   was carried out using an LSAB IHC kit (DAKO Cytomation, #VB-6017D),
   with 3,3-diaminobenzidine as the chromogen and Mayer’s hematoxylin for
   nuclear staining.

   For semi-quantitative analysis of the IHC staining, three randomly
   selected fields (100 μm^2 each) were chosen from each sample, and
   images were captured using a charge-coupled device camera. Staining
   intensity was analyzed using IMT i-Solution software version 10.1 (IMT
   i-Solution, Inc., Canada), as previously described^[232]75. The results
   of IHC image analysis were presented as mean ± SEM. Based on the IHC
   staining density, cases were categorized into a high-expression group
   with staining intensity above the average value and a low-expression
   group with staining intensity below the average value.

Immunofluorescence staining

   For immunofluorescence microscopy, cells were digested into single
   cells and then collected. Cells were fixed with 4% paraformaldehyde
   twice for 30 min at room temperature, followed by permeabilization
   using 0.01% Triton X-100 (Thermo Fisher Scientific, # HFH10) for
   10 min, then blocked in 3% BSA at room temperature for 30 min. Next,
   cells were incubated with anti-phospho-γH2A.X (1:500, CST, #9718)
   antibodies at 4 °C overnight. After washing with PBS, cells were
   stained with the second antibody conjugated with Alexa Fluor 647
   (1:1000, Invitrogen, #A32733) at 37 °C for 1 h. Nuclei were stained
   with DAPI (1:500, Sigma, MBD0015). After adding the Fluoromount-G
   (SourthenBiotech, #0100), the cells were examined using a Zeiss LSM900
   confocal microscope. All imaging results were processed using the
   ImageJ software. Phospho-γH2A.X foci were quantified in individual
   nuclei, with least 50 nuclei counted for foci quantification in each
   group.

In vivo tumorigenesis

   In the knockdown experiment, GSCs were transduced with lentiviral
   vectors expressing shDPYD, shTYMP, and a non-targeting control shRNA
   (shCont). After 48 h of infection, cells were assessed by qPCR and
   western blot to confirm successful knockdown. Then the cells were
   counted and randomly implanted into the right cerebral cortex of
   healthy, 4–6 weeks old male mice of Balb/c nude background purchased
   from the GemPharmatech Company in Jiangsu Province. All mice
   experiments were performed following an animal protocol approved by the
   Institutional Animal Care and Use Committee of the Shanghai Medical
   College, Fudan University. All experimental mice were housed in a
   specific pathogen-free environment with ambient temperature of 20–26 °C
   and humidity of 30–70% in a 12 h light/12 h dark cycle, with no more
   than 5 mice per cage. The animals were sacrificed at the point that
   they displayed neurological symptoms, in accordance with the protocol
   outlined by the NIH Office of Animal Care and Use. The brains were
   collected, fixed in 4% formaldehyde, paraffin-embedded, and then
   sectioned. These sections were stained with H&E for histological
   analysis. In parallel survival experiments, animals were monitored
   until they developed neurological signs.

Statistical analysis

   All analyses were performed using R statistical software (R-4.1.1) and
   SPSS (version 19.0; SPSS Inc., USA). Unless otherwise specified, all
   tests were two-tailed, and P values less than 5% were considered
   statistically significant.

Reporting summary

   Further information on research design is available in the [233]Nature
   Portfolio Reporting Summary linked to this article.

Supplementary information

   [234]Supplementary Information^ (95MB, pdf)
   [235]Supplementary Data 1^ (22.9KB, xlsx)
   [236]Supplementary Data 2^ (17.5KB, xlsx)
   [237]Supplementary Data 3^ (14.9KB, xlsx)
   [238]Supplementary Data 4^ (32.6KB, xlsx)
   [239]Supplementary Data 5^ (11.3KB, xlsx)
   [240]Supplementary Data 6^ (12.3KB, xlsx)
   [241]Reporting Summary^ (2.6MB, pdf)
   [242]41467_2024_54352_MOESM9_ESM.pdf^ (100.2KB, pdf)

   Description of Additional Supplementary Files
   [243]Transparent Peer Review file^ (5.7MB, pdf)

Source data

   [244]Source Data^ (78.2MB, zip)

Acknowledgements