Sebastian Brandner

Deep learning applied to whole-slide histopathology images (WSIs) has the potential to enhance precision oncology and alleviate the workload of experts. However, developing these models necessitates large amounts of data with ground truth labels, which can be both time-consuming and expensive to obtain. Pathology reports are typically unstructured or poorly structured texts, and efforts to implement structured reporting templates have been unsuccessful, as these efforts lead to perceived extra workload. In this study, we hypothesised that large language models (LLMs), such as the generative pre-trained transformer 4 (GPT-4), can extract structured data from unstructured plain language reports using a zero-shot approach without requiring any re-training. We tested this hypothesis by utilising GPT-4 to extract information from histopathological reports, focusing on two extensive sets of pathology reports for colorectal cancer and glioblastoma. We found a high concordance between LLM-generated structured data and human-generated structured data. Consequently, LLMs could potentially be employed routinely to extract ground truth data for machine learning from unstructured pathology reports in the future. © 2023 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Understanding the exact molecular mechanisms involved in the aetiology of epileptogenic pathologies with or without tumour activity is essential for improving treatment of drug-resistant focal epilepsy. Here, we characterize the landscape of somatic genetic variants in resected brain specimens from 474 individuals with drug-resistant focal epilepsy using deep whole-exome sequencing (>350×) and whole-genome genotyping. Across the exome, we observe a greater number of somatic single-nucleotide variants in low-grade epilepsy-associated tumours (7.92 ± 5.65 single-nucleotide variants) than in brain tissue from malformations of cortical development (6.11 ± 4 single-nucleotide variants) or hippocampal sclerosis (5.1 ± 3.04 single-nucleotide variants). Tumour tissues also had the largest number of likely pathogenic variant carrying cells. low-grade epilepsy-associated tumours had the highest proportion of samples with one or more somatic copy-number variants (24.7%), followed by malformations of cortical development (5.4%) and hippocampal sclerosis (4.1%). Recurring somatic whole chromosome duplications affecting Chromosome 7 (16.8%), chromosome 5 (10.9%), and chromosome 20 (9.9%) were observed among low-grade epilepsy-associated tumours. For germline variant-associated malformations of cortical development genes such as TSC2, DEPDC5 and PTEN, germline single-nucleotide variants were frequently identified within large loss of heterozygosity regions, supporting the recently proposed 'second hit' disease mechanism in these genes. We detect somatic variants in 12 established lesional epilepsy genes and demonstrate exome-wide statistical support for three of these in the aetiology of low-grade epilepsy-associated tumours (e.g. BRAF) and malformations of cortical development (e.g. SLC35A2 and MTOR). We also identify novel significant associations for PTPN11 with low-grade epilepsy-associated tumours and NRAS Q61 mutated protein with a complex malformation of cortical development characterized by polymicrogyria and nodular heterotopia. The variants identified in NRAS are known from cancer studies to lead to hyperactivation of NRAS, which can be targeted pharmacologically. We identify large recurrent 1q21-q44 duplication including AKT3 in association with focal cortical dysplasia type 2a with hyaline astrocytic inclusions, another rare and possibly under-recognized brain lesion. The clinical-genetic analyses showed that the numbers of somatic single-nucleotide variant across the exome and the fraction of affected cells were positively correlated with the age at seizure onset and surgery in individuals with low-grade epilepsy-associated tumours. In summary, our comprehensive genetic screen sheds light on the genome-scale landscape of genetic variants in epileptic brain lesions, informs the design of gene panels for clinical diagnostic screening and guides future directions for clinical implementation of epilepsy surgery genetics.

BACKGROUND: Glioblastoma (GB), particularly IDH-wild type, is the most aggressive brain malignancy with a dismal prognosis. Despite advances in molecular profiling, the complexity of its tumor microenvironment and spatial organization remains poorly understood. This study aimed to create a comprehensive single-cell and spatial atlas of GB to unravel its cellular heterogeneity, spatial architecture, and clinical relevance. METHODS: We integrated single-cell RNA sequencing data from 26 datasets, encompassing over 1.1 million cells from 240 patients, to construct GBmap, a harmonized single-cell atlas. High-resolution spatial transcriptomics was employed to map the spatial organization of GB tissues. We developed the Tumor Structure Score (TSS) to quantify tumor organization and correlated it with patient outcomes. RESULTS: We showcase the applications of GBmap for reference mapping, transfer learning, and biological discoveries. GBmap revealed extensive cellular heterogeneity, identifying rare populations such as tumor-associated neutrophils and homeostatic microglia. Spatial analysis uncovered 7 distinct tumor niches, with hypoxia-dependent niches strongly associated with poor prognosis. The TSS demonstrated that highly organized tumors, characterized by well-defined vasculature and hypoxic niches, correlated with worse survival outcomes. CONCLUSIONS: This study provides a comprehensive resource for understanding glioblastoma heterogeneity and spatial organization. GBmap and the TSS provide an integrative view of tumor architecture in GB, highlighting hypoxia-driven niches that may represent avenues for further investigation. Our resource can facilitate exploratory analyses and hypothesis generation to better understand disease progression.