Citations

Abstract Recurrent or refractory IDH-mutant (mt) gliomas represent a population with significant unmet clinical need. The Notch ligand delta-like 3 (DLL3) is a tumor-associated antigen that is highly and homogeneously expressed in IDH-mt gliomas but absent in normal brain tissue, making it an attractive therapeutic target. Tarlatamab is a half-life extended bispecific T-cell engager that targets DLL3 and CD3 to redirect cytotoxic T cells toward DLL3-expressing tumor cells. Given the specific expression of DLL3 in IDH-mt gliomas and promising activity of tarlatamab in other DLL3-positive tumors, this study investigates the safety, efficacy, and intracranial activity of tarlatamab in this patient population. TARGID is an open-label, phase II trial enrolling adult patients with histologically confirmed IDH1 or IDH2-mt astrocytoma or oligodendroglioma following up to two prior lines of systemic therapies. The study comprises two cohorts. Cohort 1 (n=10) includes patients with resectable disease who will receive up to three cycles of tarlatamab prior to surgery; postoperative continuation is permitted. The primary endpoint for this cohort is the change in CD8+ T-cell infiltration in tumor tissue and blood. Cohort 2 (n=34) includes patients with unresectable disease and uses a Simon’s optimal two-stage design (H0: P ≤ 0.1 versus H1: P ≥ 0.25). In the 1st stage, 13 patients will be accrued. If 2 or more responses are observed, 21 additional patients will be enrolled. Tarlatamab is administered intravenously at 10 mg every two weeks, following a 1 mg priming dose on Cycle 1 Day 1. The primary endpoint is objective response rate by RANO 2.0. Exploratory endpoints include DLL3 expression, circulating biomarkers, and immune correlatives. The trial is currently in activation, with first patient enrollment anticipated in Q3 2025. The study is supported by the provision of study drug from Amgen as an investigator-initiated trial.

Abstract Glioblastoma (GBM) progression and recurrence are often driven by therapy-resistant cellular states emerging within hypoxic tumor niches. While hypoxia is a hallmark of GBM, the spatial transcriptomic heterogeneity that underpins resistance in these regions remains poorly characterized. Here, we employ pimonidazole labeling in conjunction with high-resolution spatial and single cell transcriptomics across 29 primary, treatment-naive human GBM samples, syngeneic mouse models, and patient-derived glioma-initiating cells exposed to hypoxia to define the hypoxia-specific tumor microenvironment and its associated cellular programs. We reveal spatially organized sub-niches within hypoxic regions, each supporting diverse cellular programs yet coexisting tumor lineage states with minimal transitions between cell states. Phylogenetic reconstruction based on inferred copy number variations demonstrates consistent evolutionary trajectories from non-hypoxic to hypoxic tumor loci. Specific copy number alterations were enriched in hypoxic regions and associated with immune infiltration and poor prognosis in MGMT-methylated tumors. Cross-species analyses confirmed the conservation of hypoxia-associated gene programs and spatial architectures. Together, our study delineates the cellular, evolutionary, and spatial complexity of hypoxia in GBM, identifies hypoxia-induced genomic alterations that may promote tumor fitness, and offers insights into new therapeutic vulnerabilities within hypoxic niches.

Abstract Glioblastoma (GBM) is a highly aggressive brain tumor with poor survival outcomes and limited tools for accurate prediction of individual patient prognosis. While many prognostic models rely on advanced molecular testing, such methods are not routinely accessible across most clinical settings. In this study, we identify a practical and clinically applicable approach to estimate prognosis for GBM patients using three objective and readily accessible features: the patient’s age, tumor surface irregularity determined from MRI scans, and the expression of ICAM1 measured by standard immunohistochemistry. We show that higher ICAM1 levels are linked to worse survival and a more inflamed but immunosuppressive tumor microenvironment. These findings are functionally validated in ICAM1-wildtype and knockout mouse models using single-cell proteomics via cytometry by time-of-flight, and an immune-comprehensive panel of antibodies. By combining age, ICAM1, and MRI-derived shape features into a single Clinical-Imaging-Molecular (CIM) score, we developed a model that can reliably separate patients with a GBM into higher- and lower-risk groups. This score was validated across multiple international institutions and patient populations and does not require genomic sequencing or specialized technologies. The CIM model could serve as a supplement for care providers to better tailor treatment plans and follow-up strategies for individual patients, making personalized care more achievable even in resource-limited settings.

Abstract Computer vision now enables deep learning models to link histopathologic features to molecular features in a quantitative manner not previously achievable by human pathologists. However, human interpretability is often limited. We trained a deep neural network to predict gene expression from digital whole slide images (WSIs) using 848 central nervous system (CNS) tumors with paired WSI and RNA-sequencing data. We used inferred RNA expression levels as a surrogate for protein expression of the CNS tumor lineage markers GFAP, OLIG2, and SSTR2, as well as the proliferation marker MKI67 (Ki67). We established thresholds for categorizing the inferred expression levels as positive or negative based on levels observed in cross-validation testing in select tumor types to approximate routine clinical interpretation of immunohistochemical (IHC) staining. We tested the sensitivity and specificity of our ‘virtual’ IHC on an independent, multi-institutional cohort of over 2,000 CNS tumors with objective diagnostic labels derived from DNA methylation-based tumor classification. As a dichotomous variable, inferred GFAP expression showed a sensitivity of 74% and a specificity of 99% as a glial marker in a cohort of gliomas and meningiomas. In the same cohort, OLIG2 showed a sensitivity of 75% and a specificity of 97% as a glial marker, while SSTR2 showed a sensitivity of 83% and a specificity of 97% as a marker of meningioma. In a cohort of ependymomas and gliomas, virtual IHC for OLIG2 showed a sensitivity of 75% and a specificity of 82% in distinguishing ependymomas from gliomas. Finally, in a cohort of gliomas, inferred expression of MKI67 demonstrated a trend consistent with what would be expected by actual IHC, with increasing MKI67 expression with increasing tumor grade. Our model provides the basis for a human interpretable and clinically applicable deep neural network to aid human pathologists in the diagnosis and grading of CNS tumors.

Abstract Molecular testing has become critical components in the diagnostic classification of the central nervous system (CNS) tumors. However, these methods require substantial resources, limiting accessibility for many patients. Recent advances in artificial intelligence (AI) and computer vision empower deep learning models to infer molecular features from histopathology images, which may often be more readily available than molecular testing, to classify CNS tumors. We trained pan-CNS tumor models to predict DNA methylation and gene expression from whole-slide images (WSIs). These molecular predictions were then used in a hierarchical machine learning framework, termed Neuropath-AI, to predict nine broad tumor families and 52 tumor types using a large diverse cohort of 5,715 histopathology samples, of which 2,988 had paired DNA methylation profiling and 848 had paired RNA-sequencing. We evaluated Neuropath-AI on another large independent multi-institutional cohort of 5,516 CNS tumors, for which Neuropath-AI predicted tumor class with an associated confidence score. Neuropath-AI designated 46% of test samples as predictable with high-confidence, which it then accuracy classified in 97% of samples. It made moderate-confidence (and above) predictions for 87% of test samples, for which it achieved a diagnostic accuracy of 80% for the top-1 prediction and 86% accuracy for a top-2 predictions. A comparison with human neuropathologists showed that Neuropath-AI is comparable at classifying tumors. A second test showed that human neuropathologists were more accurate when the Neuropath-AI classifier results were provided to them, testifying to the potential translational benefit of integrating AI classification tools into the pathologist’s workflow. Our model provides the basis for a clinically applicable deep learning assistant to improve human efficiency and accuracy of CNS tumor diagnosis. The model will be made publicly available and can be readily implemented to assist human pathologists in clinical workflows, in further expanded prospective studies.

Abstract Pediatric, adolescent and young adult patients (Ped/AYA; 0–39 years) with meningiomas are underrepresented in neuro-oncology research, despite meningiomas being the second most common primary brain tumor in this age group. While DNA methylation-based models have advanced classification and prognostication in adults, their utility in younger populations remains uncertain. To address this gap, we aimed to evaluate the performance of current DNA methylation-based classification and recurrence prediction models in Ped/AYA patients with meningiomas and identify age-associated molecular differences. As part of this study, we analyzed 1,568 meningioma samples with DNA methylation and clinical data, including subsets with whole-exome and RNA sequencing. Patients were grouped into Ped/AYA (≤39 years, n = 213) and adult (>39 years, n = 1,355) cohorts. We assessed classifier performance, performed subgroup-specific model retraining, and compared molecular profiles across age groups. Our results showed that molecular classification differed significantly between age groups, with Ped/AYA tumors less frequently belonging to the most aggressive Proliferative Molecular Group. The DNA methylation-based recurrence model performed well in adults (AUC = 0.83) but poorly in the Ped/AYA cohort (AUC = 0.57). However, retraining the model on only younger patients improved performance (AUC = 0.75). In addition, Ped/AYA tumors exhibited significantly fewer chromosomal alterations, including lower frequencies of 1p, 6q, and 14q loss (p<0.05). Taken together, these findings indicate that predominantly adult-trained methylation models underperform in younger meningioma patients due to both biological divergence and exclusion from training datasets. Our findings underscore the need for age-specific molecular tools to improve prognostication and support personalized care across the lifespan in neuro-oncology.

Abstract The WHO 2021 classification of central nervous system tumors incorporated molecular markers to better define diffuse glioma subgroups. We analyzed array-based copy number, targeted next-generation sequencing, germline genotyping, and Illumina EPIC array methylation data from 404 primary IDH-mutant astrocytomas to further understand the relationship between copy number and outcomes. Loss of chromosome arm 13q was associated with poor overall survival after adjusting for age, sex, tumor grade, and treatment status (hazard ratio [HR]=2.04 (95% confidence interval: 1.25-3.33), p=0.004). Chromosome 13q loss was also significantly associated with the total number of copy number alterations (p<0.001). Loss of arm 11p was associated with poor survival in the adjusted analysis (HR=2.29 (1.38-3.79), p=0.001). Unsupervised clustering of the raw copy number data identified 12 clusters, where each cluster was defined by distinct copy number alteration(s). Clusters that contained 13q loss exhibited worse overall survival compared to other clusters. The clusters were significantly associated with age at diagnosis (p=0.015), tumor grade (p<0.001) and copy number complexity (p<0.001). The clusters with a high proportion of grade 4 tumors and with the highest complexity had the poorest overall survival. These findings highlight the prognostic relevance of chromosome 13q and 11p losses in IDH-mutant astrocytomas and support the integration of additional copy number alterations into molecular classification frameworks of IDH-mutant astrocytomas to improve risk stratification and patient management.

Abstract Brain metastases (BMs) are a common and often fatal progression of systemic cancers, affecting up to 25% of patients. Despite advances in surgical resection and radiotherapy, the median survival remains limited to 10–16 months. While genomic profiling has enabled the development of targeted therapies, there remains a paucity of prognostic tools for clinical use. To address this, we explored the utility of DNA methylation profiling—an epigenetic marker increasingly recognized for its diagnostic and prognostic value. We profiled fresh-frozen tissue from 327 surgically resected BM samples of lung, breast, melanoma, and gastrointestinal (GI) origin using the Illumina Infinium MethylationEPIC BeadChip array. Unsupervised analysis via partitioning around medoids (PAM) clustering identified five robust groups. Four of these correlated closely with primary tumor origin. However, a distinct “Poly-Origin Cluster” emerged, comprising tumors from multiple primary sites with a shared epigenetic signature. Poly-Origin tumors showed improved survival outcomes compared to origin-aligned clusters (20.2 vs. 10.1 months, p=0.0018), independent of clinical covariates. Notably, these tumors were enriched for immune cell infiltration—including CD14+ macrophages, CD19+ B-cells, and CD56+ NK cells—based on deconvolution analysis, and exhibited reduced genomic instability as measured by copy number variation. These tumors were enriched for pathways related to cell signaling. We validated the prognostic significance of the Poly-Origin methylation signature in a publicly available independent cohort (n=96), and found that high signature concordance correlated with extended survival. Finally, we demonstrated the feasibility of detecting this signature in plasma-derived cell-free DNA using cfMeDIP-seq, with high classification accuracy (AUC = 0.98). These findings reveal a clinically meaningful subtypes of BM not captured by tissue origin or mutation status alone. Our work suggests that DNA methylation profiling may be a powerful tissue- and liquid biopsy–based tool that holds promise as a non-invasive prognostic biomarker in the management of brain metastasis.

Abstract A major barrier in understanding adult diffuse gliomas lies in the inability to track tumor evolution in real time. To address this, we employed a novel method known as continuous lineage tracing, which integrates CRISPR/Cas9-based expressed DNA barcoding with single-cell RNA sequencing, enabling a phylogenetic approach to studying tumor development. Patient-derived glioma-initiating cell lines were engineered with expressed barcodes targeted by CRISPR/Cas9 and implanted into mice to create intracranial xenografts. Tumors underwent single-cell RNA sequencing; expressed barcodes were used to infer clonal relationships, and transcriptomic profiles enabled cell state classification. Phylogenetic lineage trees were reconstructed using lineage inference algorithms to characterize cell fitness, expansion, and plasticity. Our analysis uncovered extensive intra-clonal cell state heterogeneity, indicating active phenotype switching prior to therapy. We identified consistent transcriptional programs associated with tumor engraftment and in vivo clonal advantage. Lineage tracing revealed gene expression signatures linked to key phenotypes: fitness, enriched for neural-mesenchymal and injury-response pathways; expansion, associated with RNA splicing; and plasticity, correlated with cell cycle and DNA repair programs. Glioma stem cells appeared to span a transcriptional continuum from undifferentiated, high-fitness states to more differentiated, low-fitness states, with high-fitness cells potentially representing transitional phenotypes. We then validated these findings in a cohort of 185 surgically resected gliomas with matched bulk RNA and proteomic data. Phylogenetic gene signatures differed markedly between IDH-mutant and IDH-wildtype tumors. When stratified by tumor type, both fitness and expansion signatures were significantly associated with overall survival in GBM (median 12.6 months, p=0.041) and oligodendrogliomas (median 66 months, p=0.027). This study demonstrates the utility of continuous lineage tracing to reconstruct tumor evolution and identify transcriptional programs linked to tumor growth and prognosis. Our approach provides a powerful framework for dissecting glioma biology and identifying potential therapeutic vulnerabilities.

Abstract INTRODUCTION Brain metastases (BM) are the most common intracranial tumor, with a poor prognosis of 10-16 months, and lung adenocarcinoma (LUAD) is the most common source. Currently, we cannot reliably predict which patients will develop BMs clinically, so most are detected after they develop and growing to cause deficits. Here, we use tumor DNA methylomes to predict which patients are likely to develop BM through the course of disease and we use both cerebrospinal fluid (CSF) and plasma methylomes for BM liquid biopsy, enabling their early detection and treatment. METHODS DNA methylomes were acquired from 419 tumor and plasma samples from 363 LUAD and BM patients. We built DNA methylation-based models that predict BM development and that also identify BMs non-invasively from CSF and plasma in discovery data subsets. A nomogram was built that predicts BM risk using methylation scores and clinical stage together. Models were evaluated in independent validation sets. RESULTS Our model predicting BM development was accurate (Univariable Cox: HR=5.7, 95% CI=1.9-17.2), p=0.0023), independent of clinical factors (Multivariable Cox: HR=8.9, 95% CI=2.0-40.5, p=0.0046), and more predictive than cancer staging used currently clinically (AUC=0.81 versus 0.65). Our combinatorial nomogram using methylation and staging data together showed enhanced BM prediction (Univariable Cox: HR=17.2, 95% CI=4.1-71.3, p<0.0001; AUROC=0.82). Our models identifying BM through biofluid methylomes accurately distinguished BM from differential diagnoses of gliomas and lymphomas in plasma (AUROC=0.80, 95% CI=0.68-0.93) and with greater accuracy in CSF (AUROC=0.93, 95% CI=0.71-1.0). CONCLUSION Overall, we show the first approaches for robust molecular prediction of BM development and for reliable liquid biopsy of BM. This work is expected to transform cancer management by identifying high-risk lung cancer patients after initial treatment for inclusion in screening protocols using neuroimaging and liquid biopsy, to identify BM development early with less advanced disease and more optimal therapeutic outcomes.