An anatomic transcriptional atlas of human glioblastoma

Abstract

Glioblastoma is an aggressive brain tumor that carries a poor prognosis. The tumor’s molecular and cellular landscapes are complex and their relationships to histologic features routinely used for diagnosis are unclear. Here we present the Ivy Glioblastoma Atlas, an anatomically-based transcriptional atlas of human glioblastoma that aligns individual histologic features with genomic and gene expression patterns, thus assigning a molecular significance to the most important morphologic hallmarks of glioblastoma. The atlas and its clinical and genomic database are freely accessible online data resources that will serve as a valuable platform for future investigations of glioblastoma pathogenesis, diagnosis, and treatment.

Glioblastoma is the most common and the most lethal malignant brain tumor (1). Even for patients receiving aggressive treatment, the median survival is 12–15 months (2). The tumors evolve rapidly as they acquire new mutations; the resultant increase in intratumor genomic heterogeneity leads to the development of drug resistance, which limits the long-term efficacy of therapies (3, 4). Two large-scale efforts aimed at characterizing the genomic alterations in human glioblastoma are The Cancer Genome Atlas (TCGA), which is a catalog of multi-omics data, including genomics, transcriptomic, DNA methylomics, proteomics, etc. (5, 6), and REpository for Molecular BRAin Neoplasia DaTa (REMBRANDT), which also includes multiple data domains (7). These efforts helped clarify the role of genomic alterations in the pathogenesis of glioblastoma but were not designed to address intratumor heterogeneity. Subsequent studies addressed heterogeneity spatially within bulk tumor or at the single cell level (4, 8–12), but we lack the systematic understanding of the molecular heterogeneity of this tumor as it relates to the anatomical heterogeneity framed by the variable combination of the classical histological features of glioblastoma, which include tumor infiltration, endothelial cell proliferation, and necrosis. This notion is underscored by the empirical pathology-guided selection of samples typically applied for molecular studies. Here, we report the Ivy Glioblastoma Atlas (http://glioblastoma.alleninstitute.org/), a comprehensive pathology-molecular map of glioblastoma, to guide the unbiased assignment of key molecular alterations to each of the known anatomical features of glioblastoma. By systematically determining the genomic alterations and gene expression profiles of each anatomic feature, we have generated a molecular-pathology encyclopedia of glioblastoma. The atlas will be invaluable for the accurate deconvolution of individual anatomical states within any new tumor, therefore providing unique information for the comprehensive diagnostic characterization of glioblastoma heterogeneity.

To create the atlas, we surveyed the anatomic features by in situ hybridization (ISH), analyzed their transcriptomes by laser microdissection (LMD) and RNA sequencing (RNA-Seq), and validated the feature specific, gene expression enrichment of newly-identified markers by ISH (Fig. 1). We created a clinical and genomic database (http://ivygap.org/) for the 41-patient cohort (table S1) whose tumors (n=42) were evaluated to create the atlas. We describe gene sets whose expression is enriched in the anatomic features, measurements of intra- and inter-tumor heterogeneity, and a molecular subtype classification of transcriptomic samples from our atlas and The Cancer Genome Atlas (TCGA). Together, these two on line resources constitute the Ivy Glioblastoma Atlas Project (Ivy GAP).

Fig. 1. Data generation, analysis, and presentation pipeline for the Ivy Glioblastoma Atlas Project.

(A) Clinical data were collected for the Ivy cohort of 41 patients. (B) Tissue preparation required en bloc resection and formation of tissue blocks with custom L bars. (C) Two studies, Anatomic Feature Based Profiling and Cancer Stem Cell Marker Based Profiling, provided a framework for the ISH surveys, LMD/RNA-Seq experiments, and ISH validations. (D) Informatics included image registration, ontology development, and anatomic feature prediction based on a novel machine learning (ML) analysis of histological data. Search tools support queries of the data set by tumor, tumor block, and gene expression filtered by anatomic feature, molecular subtype, and clinical information. Searchable manual labels delineating the laser microdissections for 270 RNA-Seq samples from the two studies overlay the histology images. The atlas is equipped with image viewers that resolve the histology at 0.5μm/pixel, a transcriptome browser, an application programming interface, and help documentation. The database has detailed longitudinal clinical information and MRI time courses (table S1). (E) This free resource is made available as part of the Ivy Glioblastoma Atlas Project (Ivy GAP). (http://glioblastoma.alleninstitute.org/) via the Allen Institute data portal (http://www.brain-map.org), the Ivy GAP Clinical and Genomic Database (http://ivygap.org/) via the Swedish Neuroscience Institute (http://www.swedish.org/services/neuroscience-institute), and The Cancer Imaging Archive (https://wiki.cancerimagingarchive.net).

To identify gene sets with enriched expression in each anatomic feature (fig. S1), we used LMD to isolate RNA from the leading edge (LE), infiltrating tumor (IT), cellular tumor (CT), pseudopalisading cells around necrosis (PAN), and microvascular proliferation (MVP). In total, we isolated 122 samples from 3 different blocks per tumor in 8–10 tumors. In consultation with a neuropathologist, we manually drew outlines (LMD guidelines) for each of the anatomic features on images of histologically-stained tissue sections. Three additional neuropathologists independently validated the LMD guide lines, and the results showed excellent concordance (table S2). Differential gene expression analysis revealed a total of 3627 genes that had enriched expression in LE, CT, PAN, and MVP samples (Fig. 2A, table S3). Multidimensional scaling demonstrated that samples from these four features were largely distinct, whereas IT appeared to fall on a continuum between LE and CT (Fig. 2B). Gene Ontology enrichment analysis of gene sets with enriched expression in anatomic features (Fig. 2C) confirmed and extended previous reports (13, 14). In general, samples from the same anatomic feature, whether derived from the same or different tumors, were more similar to each other than to other samples from the same tumor (Fig. 2D), and within a given anatomic feature, inter-tumor heterogeneity exceeded intra-tumor heterogeneity (fig. S2).

Fig. 2. Gene expression in anatomic features.

(A) Differential expression matrix based on genes identified in the 122 anatomic feature RNA-Seq samples isolated in triplicate from 8–10 tumors. Values are numbers of genes, of the total 3627, whose expression is enriched in the row feature relative to the column feature (FDR<0.01, fold change >2; P<0.1, BH corrected). Values on diagonal are numbers of genes with higher expression in one feature compared with all other features (i.e. top marker genes). (B) Multidimensional scaling of all genes reflects anatomic specificity. (C) Gene ontology enrichment analysis. LE and CT were enriched for gene ontology terms related to neuronal systems and glial cell differentiation, respectively, whereas PAN was associated with stress, hypoxia, and immune responses, and MVP was related to angiogenesis, immune regulation, and response to wounding. (D) Mean Euclidean distance within and between tumors based on hierarchical clustering of all genes in all 122 anatomic feature RNA-Seq samples grouped by anatomic feature (fig. S1; fig. S2). Cross Feature measures variance between anatomic features. (E-I) Representative marker genes showing RNA-Seq expression levels for features isolated by LMD, representative ISH, ML annotations for ISH and H&E, and H&E adjacent to ISH. LE (blue), IT (purple), CT (green), PNZ (light blue), PAN (turquoise), HBV (orange), MVP (red/magenta).

We selected 31 genes with enriched expression in anatomic features for further analysis by ISH, and found that 27 showed at least partial agreement and 22 showed good agreement between RNA-Seq and ISH assessments of enrichment in PAN, CT, or MVP (table S4) (Fig. 2E–I). Assessing enrichment of gene expression by ISH required that we calculate the overlap between the expression pattern and our machine learning (ML) annotations for each anatomic feature, which we validated using (i) ML-determined rates of accuracy and precision (table S5); (ii) an inter-neuropathologist test to establish agreement on definitions of anatomic features (fig. S1, tables S6, S7); and (iii) neuropathology concordance analyses (tables S8–table S11).

To characterize intra-tumor genetic heterogeneity across anatomic features, we assessed RNA-Seq derived copy number changes in the features and compared them to the DNA level copy number variations (CNVs) (12) from the corresponding bulk tumor (fig. S3; table S12). The CT and PAN samples consistently showed gene expression changes corresponding to the CNVs, whereas LE samples did not as LE samples by definition consist largely of non-neoplastic cells and hence would not harbor the CNVs. On the other hand, MVP samples showed some gene expression changes corresponding to the CNVs indicating a mixture of tumor and non-neoplastic cells. To evaluate the distribution of somatic mutations targeting key glioblastoma genes within the different anatomic features of this tumor, we used RNA-Seq to call Single Nucleotide Variants (SNVs) in eight genes (TP53, PTEN, EGFR, ATRX, IDH1, NF1, PIK3R1, PIK3CA) known to harbor recurrent and functionally important mutations in glioblastoma across anatomic features for tumors where there was at least one sample available from each of the LE, CT, PAN and MVP features (fig. S4; table S13). We detected somatically mutated alleles in RNA from CT, PAN and MVP samples, but only the wild-type variants in LE samples (fig. S4A). The ratio of mutant to wild-type expression was least for MVP relative to CT and PAN samples (fig. S4B). Some of the SNVs occurring across anatomic features were corroborated by ISH data (table S1). Together, the copy number and mutation analyses indicated that LE samples largely consist of non-neoplastic cells, CT and PAN samples comprise largely tumor cells and MVP samples have a mixture of tumor and non-neoplastic cells. The observed intra-tumor heterogeneity in copy number and mutation profiles is consistent with previous studies (8, 9). Only 3 tumors from our 41-patient cohort harbored the R132H mutation in isocitrate-dehydrogenase 1 (IDH1) (table S1); thus, there was insufficient statistical power for analysis of this mutation by anatomic feature. We did not identify any mutation associated with a particular anatomic feature that predicted overall survival better than the promoter methylation status of the MGMT gene in the bulk tumor (fig. S5A,B) (15).

Finally, we developed an admixture model using a 293 gene signature matrix (table S14) for computational decomposition of bulk tumor samples into four anatomic features (LE, CT, PAN, and MVP), and classified the 122 anatomic feature RNA-Seq samples on the basis of histology, admixture (table S14), molecular subtype (6) , and cell type gene expression signature (table S15) enrichment (fig. S6A–D; table S16). Several genes exhibited differential expression across known molecular subtypes of glioblastoma within each anatomic feature (fig. S7A–C). Enrichment of the cell type gene expression signatures in the anatomic features was consistent with Gene Ontology enrichment analyses (Fig. 2C). The correlation between the anatomic feature gene sets and molecular subtypes (table S16) is broadly consistent with results of previous studies (8, 9). When we applied our admixture model to 167 RNA-Seq samples of the TCGA data, we observed similar patterns (fig. S8A–C; table S16).

This atlas and the associated database for clinical and genomic data will serve as a valuable platform for developing and testing new hypotheses related to the pathogenesis, diagnosis, and treatment of glioblastoma. We note that investigators are already leveraging this resource (16–33). In one study, Miller et al. (22) used the atlas to prioritize potential druggable targets based on relationships to tumor microenvironment signatures. We envision use of the Ivy GAP dataset in preclinical studies where investigators identify the cells that drive tumor growth, and target their resident anatomic features for the preferred drug delivery route to maximize the therapeutic effect, as demonstrated by Yu et al.(31).