PATH-26. COMPLEMENTARY CLASSIFICATION APPROACH TO RESOLVE LOW-SCORE RESULTS FROM THE DKFZ METHYLATION CLASSIFIER FOR CNS TUMOR DIAGNOSTICS

Abstract

Abstract DNA methylation-classification using the DKFZ CNS tumor classifier is highly useful, but experience indicates that approximately one-third of clinical cases do not reach the confidence threshold (0.84) for a definitive match. A low-confidence score can be observed for a variety of reasons, and addressing such instances is not always clear. We developed a complementary classification approach that differs from the DKFZ classifier in several important ways. First, it utilizes a new reference set of several thousand samples, utilizing the entire ~850k probe set on the EPIC array. Second an alternative classification approach, it establishes multiple (50) classifiers by dividing the probes into 50 independent bins, allowing dual measurements of both an overall confidence score, as well as a consistency score. To evaluate this approach, we first evaluated a set of well-classified samples (N=521, DKFZ family score > 0.84) as ground-truth, showing that our bin-based approach was equivalent (100% accuracy) in well-classified samples. Second, we then evaluated a set (n=329) of samples that did not match to the 0.84 threshold on the DKFZ classifier. Using cutoff of 0.8 consistency score and 0.7 confidence score, we identified 99/329 cases with a match. To evaluate the potential ground-truth of our findings, orthogonal data was used to validate matches in our classifier. For example, many of our matches to GBM showed the expected abnormalities (combined +7/-10 CNV and TERT promoter mutation). Two cases that received low-score IDH-glioma matches by DKFZ matched to K27-altered-DMG in our classifier and were later shown to harbor the H3-K27M mutation. One case that received a low-score suggestion of GBM-MES on DKFZ matched in our classifier to PXA; this was then shown to harbor the expected BRAFV600E mutation. Overall, our method may add to and complement the DKFZ classifier as a means to resolve a subset of otherwise difficult-to-classify CNS tumors.