MOSAIC: Methylation-Oriented Site Analysis and Information Classifier for Robust Epigenomic Classification of Acute Leukemia in Clinical Cohorts with Variable Tumor Purity

MOSAIC is a neural network developed for classifying acute leukemia based on DNA methylation patterns while maintaining accuracy even at lower tumor purity levels. Existing classifiers such as MARLIN and ALMA showed accuracy in only 7 out of 10 and 5 out of 10 cases respectively when analyzing low-purity tumor specimens. MOSAIC was trained on publicly available methylation array data supplemented with native methylation data from Oxford Nanopore sequencing. When tested on low-blast samples (all below 25% blasts, including one case at 1.4%), MOSAIC was concordant with expert pathology assessment in every case. Saliency analysis revealed that the network uses a partially distinct set of discriminative CpG probes when classifying low-blast specimens. The results demonstrate that augmenting training with clinically representative specimens improves the accuracy of leukemia classification across variable tumor purities encountered in clinical practice.