Abstract
Despite producing a panoply of potential cancer-specific targets, the proteogenomic characterization of human tumors has yet to demonstrate value for precision cancer medicine. Integrative multi-omics using a machine-learning network identified master kinases responsible for effecting phenotypic hallmarks of functional glioblastoma subtypes. In subtype-matched patient-derived models, we validated PKCδ and DNA-PK as master kinases of glycolytic/plurimetabolic and proliferative/progenitor subtypes, respectively, and qualified the kinases as potent and actionable glioblastoma subtype-specific therapeutic targets. Glioblastoma subtypes were associated with clinical and radiomics features, orthogonally validated by proteomics, phospho-proteomics, metabolomics, lipidomics and acetylomics analyses, and recapitulated in pediatric glioma, breast and lung squamous cell carcinoma, including subtype specificity of PKCδ and DNA-PK activity. We developed a probabilistic classification tool that performs optimally with RNA from frozen and paraffin-embedded tissues, which can be used to evaluate the association of therapeutic response with glioblastoma subtypes and to inform patient selection in prospective clinical trials.
Main
The classification systems of malignant tumors have evolved in the past 15 years under the pressure of mounting molecular and genetic data and remain an active area of cancer research. The need for more accurate classifiers derives from the urgency of precision oncology and drug development targeting homogeneous tumor subsets1. Whereas genomics offers a comprehensive view of the genetic makeup of individual tumors, the integration of genomics, protein profiling and post-translational regulation delivers a deeper understanding of tumor biology and recognizes similarity patterns within individual tumor types, and possibly across multiple types of tumors that can fine-tune targeted therapeutics2.
Cancer proteomics consortia have recently provided proteogenomic data and the initial framework for analysis of the proteomic platforms and integration with genomic data3,4.
Here, we reconstructed four functional subtypes of glioblastoma (GBM)5 using proteomics, phospho-proteomics, acetylomics, metabolomics and lipidomics data using the GBM dataset from the Clinical Proteomic Tumor Analysis Consortium (CPTAC)6. We developed a computational approach, Substrate PHosphosite-based Inference for Network of KinaseS (SPHINKS) to generate unbiased kinome-phosphosite networks and extract the master kinases (MKs) driving GBM subtypes. We experimentally validated protein kinase Cδ (PKCδ) and DNA-dependent protein kinase catalytic subunit (DNA-PKcs) as the MKs that sustain cell growth and tumor cell identity of the glycolytic/plurimetabolic (GPM) and proliferative/progenitor (PPR) functional GBM subtypes, respectively. We confirmed PKCδ and DNA-PKcs as MKs in GPM and PPR tumors from pediatric glioma (PG), breast carcinoma (BRCA) and lung squamous cell carcinoma (LSCC) cohorts classified according to the four functional classes that recapitulate metabolic and proliferation tumor cell states. Finally, we developed a probabilistic classification tool for GBM that exhibits optimal performance in both frozen and formalin-fixed, paraffin-embedded (FFPE) tumor tissue for application in cancer clinical pathology.
Proteogenomic analysis captures functional subtypes of GBM
We recently reported a single-cell-guided, pathway-based classification of isocitrate dehydrogenase (IDH) wild-type GBM that consists of four subtypes within two functional branches: neurodevelopment (PPR and neuronal, or NEU) and metabolism (GPM and mitochondrial, or MTC)5. Here, we used the proteogenomic data of 92 IDH wild-type GBM from the CPTAC cohort that was profiled by genomics, transcriptomics, proteomics, phospho-proteomics, metabolomics, acetylomics and lipidomics to explore the biology associated with the multi-omics taxonomy and uncover therapeutic opportunities (Extended Data Fig. 1a)6. As functional copy-number variations (fCNVs), the CNVs of genes associated with coherent transcriptomic changes in cis and gene expression were the primary data sources for the pathway-based classifier of GBM5, we selected validated fCNVs and transcripts as input features of similarity network fusion (SNF)7 and obtained four stable clusters (Extended Data Fig. 1b). Using 52 GBM classified according to the highest transcriptomic simplicity score as anchors, we classified 33 of the 40 remaining tumors by the SNF distance matrix (Supplementary Table 1a). Genes differentially expressed by each SNF cluster were enriched with biological activities previously assigned to GPM, MTC, PPR and NEU GBM subtypes (Supplementary Table 2a–c)5. Inspection of proteome revealed that the most differentially abundant proteins and enriched pathways coincided with activities biologically congruent with fCNV and gene expression-guided functions and recapitulated the predominant biology assigned to each subtype by SNF clustering (Fig. 1a,b and Supplementary Table 2d,e).
To ask whether fCNVs impact protein abundance in cis, we integrated genomics, transcriptomics and proteomics data to identify genes for which gain or loss correspondingly changed messenger RNA and protein expression (fCNVprot). Overall, 2,205 genes with fCNV gain and 2,837 genes with fCNV loss had concordant changes in protein abundance when compared to copy-number neutral samples (Supplementary Table 2f). Among them, 553 (25.08%) fCNVprot gains and 415 (14.63%) fCNVprot losses segregated with one subtype (Fig. 1c and Supplementary Table 2g–j). fCNVprot contributed directly to activation/deactivation of the subtype-specific biological hallmarks (Extended Data Fig. 1c and Supplementary Table 2k).
To understand the relationship between pathway-based classification (GPM, MTC, PPR and NEU) and previously proposed transcriptional (TCGA: proneural, classical and mesenchymal)8 and epigenetic (MolecularNeuroPathology (MNP): mesenchymal, RTK I, RTK II, RTK III, MID, MYCN and G34)9 subtypes of GBM, we selected 199 and 83 IDH wild-type GBM profiled by both RNA-seq and DNA methylation arrays from TCGA and CPTAC, respectively. We performed a three-way comparison. The GPM subtype exhibited clear association with the mesenchymal subtypes of TCGA and MNP classifiers. Conversely, MTC tumors were mapped to all TCGA and MNP subtypes, with slight preference for RTK II and mesenchymal subtype in the TCGA and CPTAC dataset, respectively (Extended Data Fig. 1d–f and Supplementary Table 1a,b). PPR and NEU had limited overlap with the TCGA and MNP classes, with proneural and RTK I contributing to most PPR and NEU tumors (Extended Data Fig. 1d,e and Supplementary Table 1a,b). Although the epigenetic RTK III, MID, MYCN and G34 subtypes were only minimally represented in TCGA and CPTAC datasets (4.5% and 1.2%, respectively), six of nine tumors were classified as PPR (Extended Data Fig. 1d,e). We also compared functional subtypes with proneural-like, classical-like and mesenchymal-like subtypes reported by CPTAC6. GPM tumors were mainly CPTAC mesenchymal-like; however, the mesenchymal-like group also included a significant fraction of MTC cases (Extended Data Fig. 1f), indicating that our classification uniquely discriminates tumors exhibiting alternative metabolic fluxes (MTC and GPM) and clinical characteristics5. The CPTAC proneural-like subtype included similar fractions of PPR and NEU, whereas the classical-like subtype was preferentially enriched with PPR tumors.
The analysis confirmed orthogonal distribution of MTC GBM and indicated that, with the description of PPR and NEU subtypes, the pathway-based classifier more accurately captures the neurogenesis stages than the vague definition of proneural state.
Proteogenomics enables integrative modules of GBM subtypes
To understand whether each functional subtype of GBM reflects a unique configuration of elements that compose a distinct functional module, from genetic drivers to clinical characteristics such as age, sex and location of the tumor in the brain or radiological features that are obtained at diagnosis by magnetic resonance imaging (MRI), we applied a univariate logistic regression that determined the association of mutations and fCNV5 with each subtype. In an independent model we asked whether proteins encoded by GBM driver genes provide orthogonal validation to the genetic associations (Extended Data Fig. 2). We found that PPR activity predominantly associated with fCNV amplification/mutation/high protein abundance of GBM oncogenes (CDK6, EZH2, MDM4 and EGFR) and fCNV deletion/mutation/protein depletion of CDKN2A, all connected to PPR hallmarks. GPM activity was associated with MET fCNV amplification/high protein abundance and NF1 fCNV deletion/mutation/protein depletion (Extended Data Fig. 2a,c). Confirming our previous findings10, the MTC subtype was associated with FGFR3-TACC3 fusion-positive tumors in the cohort of 178 GBM that we used to validate the probabilistic classifier (see below and Extended Data Fig. 2b)11. fCNV deletion of RERE and SLC45A1 genes located in the ‘metabolic’ region of chromosome 1p36.23 previously identified as a driver of the MTC subtype5 was associated with increased MTC activity. The positive correlation between low RERE protein abundance independently supported the association whereas the SLC45A1 protein was not detected in the CPTAC proteome (Extended Data Fig. 2c). With the limitation of the small number of CPTAC samples, the overall analysis indicated that protein abundance was generally a better indicator of subtype activity than CNV and mutations, a finding that likely reflects control of oncogenic protein abundance by non-genetic factors.
Next, we analyzed the correlation between clinical characteristics and subtype transcriptomic activity. GPM activity showed significant association with male sex and age between 40 and 65 years. When aggregated, PPR and NEU activities approached significance in association with female sex (Fig. 2a). GPM tumors were more frequently found in the frontal and parietal lobes but were excluded from the temporal region. Conversely, MTC tumors were more frequent in the temporal lobe and were excluded from the parietal lobe, indicating a reciprocal brain location pattern for the metabolic subtypes (Fig. 2b)….
