Document Type

Conference Proceeding

Publication Date


Publication Title

Journal of Neuropathology and Experimental Neurology


Whole genome methylation profiling is proving to be an incredibly powerful tool to diagnose tumors that are otherwise difficult to classify by histopathology and sequencing. The reference library of tumor DNA methylation data, using a machine learning-based tumor classifier, was developed by the German Cancer Research Network (DKFZ), and is available publicly on a research basis. However, medical institutions desiring to implement that classifier for routine clinical care need to import, adapt, and validate the classifier internally. We did this for our own DNA methylation-based classifier of central nervous system (CNS) tumors, using the same training and validation datasets as the DKFZ group. In addition, we validated additional samples from our own hospital, and compared the performance of both the original DKFZ classifier and our internal modified classifier. Using the validation data set, our classifier’s performance showed high concordance (92%) and comparable accuracy (specificity 94.0% v. 84.9% 29 for DKFZ, sensitivity 88.6% v. 94.7% for DKFZ). Receiver operator curve showed areas under the curve of 0.964 v. 0.966 for NM and DKFZ classifiers, respectively. Our classifier performed similarly well with samples tested in our own laboratory, and is now being used on a routine clinical basis for CNS tumors. Herein, we describe that process of importation, adaptation, and validation of the DKFZ dataset, with the objective of providing a template by which other institutions can do likewise.





First Page


Last Page