PATH-06. Explainable artificial intelligence of DNA methylation-based brain tumor diagnostics

We have previously developed a machine learning classifier that enables fast, accurate, and affordable classification of brain tumors based on genome-wide DNA methylation profiles that is widely employed in the clinic all over the world. Clinical and neuro-oncology research would benefit greatly from understanding the underlying artificial intelligence decision process, which currently remains largely unclear. Here, we describe an interpretable framework to explain the random forest classifier’s decisions and to uncover DNA methylation patterns distinguishing different tumor classes. We show that functional genomic regions of various sizes are predominantly employed to distinguish between different tumor classes, ranging from enhancers and CpG islands to large-scale heterochromatic domains. We detect a high degree of genomic redundancy, with many genes distinguishing individual tumor classes, explaining the robustness of the classifier and revealing potential targets for further therapeutic investigation. We anticipate that our resource will build up trust in machine learning in clinical settings, foster biomarker discovery and development of compact point-of-care assays, and enable further epigenome research of brain tumors. Our interpretable framework is accessible to the research community via an interactive web application (https://hovestadtlab.shinyapps.io/shinyMNP/).