Abstract
Assessment of measurable residual disease (MRD) upon treatment of acute myeloid leukemia (AML) remains challenging. It is usually addressed by highly sensitive PCR- or sequencing-based screening of specific mutations, or by multiparametric flow cytometry. However, not all patients have suitable mutations and heterogeneity of surface markers hampers standardization in clinical routine. In this study, we propose an alternative approach to estimate MRD based on AML-associated DNA methylation (DNAm) patterns. We identified four CG dinucleotides (CpGs) that commonly reveal aberrant DNAm in AML and their combination could reliably discern healthy and AML samples. Interestingly, bisulfite amplicon sequencing demonstrated that aberrant DNAm patterns were symmetric on both alleles, indicating that there is epigenetic crosstalk between homologous chromosomes. We trained shallow-learning and deep-learning algorithms to identify anomalous DNAm patterns. The method was then tested on follow-up samples with and without MRD. Notably, even samples that were classified as MRD negative often revealed higher anomaly ratios than healthy controls, which may reflect clonal hematopoiesis. Our results demonstrate that targeted DNAm analysis facilitates reliable discrimination of malignant and healthy samples. However, since healthy samples also comprise few abnormal-classified DNAm reads the approach does not yet reliably discriminate MRD positive and negative samples.