Abstract
Tumors of the CNS represent one of the most complex groups of human cancer, with a vast number of different entities occurring across a spectrum of ages and anatomic locations. This heterogeneity makes accurate diagnosis challenging, with the current gold standard relying on multiple subjective elements. We recently proposed a classification algorithm based on tumor DNA methylation profiling as an objective way to assign samples to over 80 distinct molecular classes. Here we present a substantial update to our machine learning-based algorithm, with more than 170 molecular classes now being represented amongst the 5,915 samples in our reference cohort. These new classes include further subclassification of known groups such as medulloblastoma and ependymoma, as well as multiple new molecular entities described here for the first time. A further improvement is the introduction of a more rationally layered output, making use of ‘families’ of closely-related molecular classes to improve the compatibility with the current WHO classification of CNS tumors. This approach is designed to increase the clinical relevance of the primary output, while also retaining the full information content from the random forest-driven classification. Benchmarking our new algorithm by cross-validation and on an independent validation cohort indicates a retention of the excellent accuracy of diagnosis (error-rate < 4%), with a significant improvement in the proportion of confidently classifiable tumors compared with our previous tool. We believe that this approach, freely accessible through an online web portal, has the potential to enhance diagnostic precision and thereby support clinical care for brain tumor patients.