2Cytosine methylation is essential for normal mammalian development, yet also provides a major mutagenic stimulus. Methylcytosine (5mC) is prone to spontaneous deamination, which introduces cytosine to thymine transition mutations (C>T) upon replication 1 . Cells endure hundreds of 5mC deamination events each day and an intricate repair network is engaged to restrict this damage. Central to this network are the DNA glycosylases MBD4 2 and TDG 3,4 , which recognise T:G mispairing and initiate base excision repair (BER). Here we describe a novel cancer predisposition syndrome resulting from germline biallelic inactivation of MBD4 that leads to the development of acute myeloid leukaemia (AML). These leukaemias have an extremely high burden of C>T mutations, specifically in the context of methylated CG dinucleotides (CG>TG). This dependence on 5mC as a source of mutations may explain the remarkable observation that MBD4-deficient AMLs share a common set of driver mutations, including biallelic mutations in DNMT3A and hotspot mutations in IDH1/IDH2. By assessing serial samples taken over the course of treatment, we highlight a critical interaction with somatic mutations in DNMT3A that accelerates leukaemogenesis and accounts for the conserved path to AML. MBD4-deficiency was also detected, rarely, in sporadic cancers, which display the same mutational signature. Collectively these cancers provide a model of 5mC-dependent hypermutation and reveal factors that shape its mutagenic influence.We identified three patients with AML, including two siblings, that were distinctive because of their early age of onset (all <35 years old) and an extremely high mutational burden (~33-fold above what is typical for AML) (Fig. 1a, Clinical Synopsis). Virtually all of the somatic mutations identified were C>T in the context of a CG dinucleotide (>95% of SNVs) (Fig. 1b, Extended Data Fig. 1). This differs markedly from the distribution of C>T mutations in AML generally and is more refined than the mutational signature ascribed to ageing, which includes a strong contribution from 5mC deamination 5 . All three cases carried rare germline loss-offunction variants in the gene encoding the DNA glycosylase MBD4 2 (Fig. 1c, Extended Data Table 1). Case EMC-AML-1 carried a homozygous MBD4 in-frame deletion of Histidine 567 (His567) in the glycosylase domain. An in vitro glycosylase assay confirmed that loss of His567 results in a catalytically inactive MBD4 protein (Fig. 1c). The siblings (WEHI-AML-1, WEHI-AML-2) were compound heterozygotes with a frameshift in exon 3 and a variant that disrupts the splice acceptor of exon 7 (Fig. 1c, Extended Data Table 1). Analysis of the MBD4 mRNA allowed for phasing of the variants to distinct alleles and confirmed aberrant splicing that excludes exon 7 and disrupts the glycosylase domain (Extended Data Fig. 2). MBD4 has not previously been associated with haematological malignancy, but somatic mutations have been detected in sporadic colon cancers with mismatch repair (MMR) deficiency 6,7 . Two patients...