Single-cell DNA methylation analysis of human cancer

Division Of Basic Sciences - Nci

Investigators

Linked publications & trials

Abstract

We have used techniques in machine learning and deconvolution of methylation data to elucidate genomic profiles of several CNS cancers. Recent advances in artificial intelligence (AI) and computer vision empower deep learning models to infer molecular features from histopathology images to classify central nervous system (CNS) tumors. We trained deep neural networks to predict DNA methylation and gene expression from whole slide images (WSIs). Using these predictions, we trained a hierarchical machine learning framework, termed Neuropath-AI, to predict 52 tumor types using a large, diverse cohort of hematoxylin and eosin (HandE)-stained WSIs from 5,835 patient samples of a variety of CNS tumors. We tested our model on a large, independent, multi-institutional cohort of 5,516 patient samples, for which Neuropath-AI predicted tumor class with an associated confidence level. High-confidence predictions were achieved in 46% of the test samples and showed an accuracy of 97% (balanced accuracy 83%) for the highest-scoring (top-1) prediction, which increased to 98% (92%) when the two highest-scoring (top-2) predictions were considered. Predictions with at least moderate confidence were achieved in 86% of the test samples, with top-1 accuracy of 80% (66%) and top-2 accuracy of 86% (75%). In a test comparing diagnostic accuracy of Neuropath-AI to four human pathologists across a range of experience using single WSIs of HandE-stains, Neuropath-AI accuracy was higher than that of the pathologists. When pathologists then re-evaluated the same WSIs with access to Neuropath-AI predictions, they demonstrated substantial improvements in accuracy, testifying to the potential translational benefit of integrating AI classification tools into the pathologist workflow. This data resulted from our investigations into DNA methylation studies.