METB-10. EXPLAINABLE ARTIFICIAL INTELLIGENCE REVEALS DNA METHYLATION PATTERNS UNDERLYING BRAIN TUMOR CLASSIFICATION

Abstract

Abstract Over 100 different molecular classes of brain tumors are recognized across histopathological grades and age groups. Their precise diagnosis is crucial for prognostication and appropriate treatment decisions, particularly in pediatric patients. Recently, a random forest-based classifier trained on genome-wide DNA methylation profiles has been developed to guide the accurate classification of these tumors. Over recent years, this classifier has been applied to tens-of-thousands of patient tumors all over the world. Clinical and basic neuro-oncology research would benefit greatly from a detailed understanding of the underlying artificial intelligence decision process. However, why the algorithm arrives at a certain prediction remains unclear. Here, we develop an interpretable framework to explain the model’s decisions and to uncover DNA methylation patterns distinguishing different tumor classes. Using this approach, we demonstrate both global and local differences across epigenetic landscapes. For instance, the model frequently uses DNA methylation probes in CpG islands to distinguish IDH mutant gliomas from other tumor classes. Tumor classes associated with global hypomethylation are distinguished by probes located in broad heterochromatic domains. At the level of individual genes, we find hundreds of promoter and gene body regions specifically methylated or unmethylated in specific classes. In general, we detect a high degree of redundant information in the dataset, with multiple genes across the genome distinguishing individual tumor classes, and many CpG probes representing each gene. These findings explain the technical robustness of the array-based classifier, and provide a starting point for the design of more compact point of care assays. The results of our approach are provided to the research community via an interactive web interface. We hope that this resource will help to build up trust in the machine learning-based classifier in clinical settings, foster biomarker discovery, and enable cancer epigenome research in the context of the central nervous system tumors.