Human PMS1-dependent non-canonical mismatch repair engages with MBD4 to repair methylated CpG deamination

Abstract

CpG dinucleotides are mutational hotspots due to spontaneous deamination of 5-methylcytosine (5mC), resulting in T:G mismatches that can lead to CpG>TpG transitions. These mutations are a hallmark of aging and cancer and play a central role in the evolution of vertebrate genomes. We have previously uncovered MBD4 as the primary base excision repair (BER) glycosylase responsible for 5mC deamination repair. Here, we employ an APOBEC1 deaminase fused to a catalytically dead Cas9 to induce targeted 5mC deamination independently of DNA replication and track its repair in human cells. This approach reveals that MBD4 elicits a coordinated repair response with a non-canonical branch of mismatch repair (MMR) involving complexes MutLβ (MLH1–PMS1) and MutSα (MSH2–MSH6). We uncover the physical interaction between MBD4 and MutLβ and demonstrate that MBD4-mediated repair requires MLH1. We show that PMS1 deficiency phenocopies the CpG>TpG hypermutation signature characteristic of MBD4 loss, establishing 5mC deamination repair as a key function of human PMS1. In alignment with our experimental data, we show that the CpG>TpG mutational burden in MMR-deficient tumors is partly explained by replication-independent processes. Altogether, we uncover a novel function of non-canonical MMR that underscores its interplay with BER in safeguarding genomic integrity against damage to methylated DNA.

Graphical Abstract

Graphical Abstract.

Introduction

DNA methylation is a common epigenetic modification, in which a methyl group is added to the 5-position of a cytosine residue to form 5-methylcytosine (5mC). In vertebrate genomes, most DNA methylation occurs at CpG dinucleotides [1]. CpG dinucleotides are genomic hotspots for mutation, primarily due to the spontaneous deamination of 5mC into thymine. This endogenous DNA damage results in T:G mismatches in double-stranded DNA, leading to CpG>TpG transitions if left unrepaired before DNA replication [2]. CpG>TpG mutagenesis is a major force shaping vertebrate genome evolution [3, 4], a primary source of germline single-nucleotide polymorphisms (SNPs) in humans, and a frequent cause of genetic diseases [5, 6]. Moreover, somatic CpG>TpG variants accumulate during life [7], and this hallmark of aging accounts for a large portion of the mutational burden in cancer genomes [8, 9]. Collectively, these observations underscore 5mC deamination as a ubiquitous mutagenic process with widespread biological consequences.

The base excision repair (BER) pathway is involved in the repair of deaminated 5mC through the activity of specialized DNA glycosylases, which recognize T:G mismatches as a substrate and direct excision of the mispaired thymine [10]. We uncovered MBD4 (methyl-CpG binding domain protein 4) as the primary BER glycosylase responsible for 5mC deamination repair [11]. Furthermore, our group and others have shown that loss of MBD4 in cancer is associated with a CpG>TpG hypermutation phenotype and an increased response to immune checkpoint inhibitor therapies [12–15].

DNA mismatch repair (MMR) is a highly conserved biological pathway that plays a key role in maintaining genomic stability. One of its primary functions is to excise mismatches generated by nucleotide misincorporation during S-phase DNA replication, often referred to as canonical post-replicative MMR. During this process, base–base mismatches are generally recognized by the MutSα (MSH2–MSH6) complex, which then recruits the MutLα complex (MLH1–PMS2). This interaction activates the latent endonuclease activity of MutLα in a PCNA-dependent manner to generate nicks in the daughter DNA strand and initiate repair [16–19]. The MutLβ complex (MLH1–PMS1) lacks the motif required for endonuclease activity [18], is dispensable for canonical MMR [20], and its role in the repair of mismatches has remained elusive. Notably, the role of MMR in G₀/G₁, often referred to as non-canonical MMR [21], remains poorly understood.

Growing evidence suggests that BER and MMR pathways may converge in certain instances of DNA repair [22, 23]. MBD4 was first identified due to its interaction with MLH1 [24], which enhances efficient recognition of T:G mismatches by MBD4 within methylated CpGs [25]. Moreover, germline loss-of-function mutations in either MBD4 or MMR genes predispose to a partially overlapping spectrum of cancers that includes uveal melanomas and colorectal polyposis [26–28]. However, the cooperative roles of BER and MMR pathways in preventing genomic instability remain largely unexplored.

In this study, we uncover that 5mC deamination repair relies on the coordinated activities of the BER glycosylase MBD4 and a non-canonical MMR axis built around MutLβ and MutSα. Using a Cas9-based targeted deamination system, in addition to genome-wide mutational profiling of engineered cell lines and human tumors, we show that MMR is central to limiting CpG>TpG mutagenesis independently of its canonical post-replicative roles. Moreover, we uncover that MBD4 physically interacts with MutLβ and demonstrate that MBD4 activity against 5mC deamination requires MLH1. Overall, our data support a model in which T:G mismatches arising from 5mC deamination are recognized by MutSα, leading to the recruitment of the MBD4–MutLβ complex to elicit repair of these endogenous lesions and maintain genomic stability.

Materials and methods

Isogenic cell models

HAP1 cells were cultured in complete media [IMDM (Gibco, 12440053), bovine fetal serum 10% (BioSera, FB-1003), Penicillin-Streptomycin 100 U/ml (Gibco, 15140122)] at 37°C in 5% CO₂. Diploidized clones of HAP1 wild-type or MBD4 knockout cells (Horizon Discovery, C631 and HZGHC000921c002, respectively) were transfected using jetOPTIMUS (Polyplus, 101000051) with all-in-one CRISPR–Cas9 vectors (pSpCas9(BB)—2A-GFP PX458; Addgene, #48138) with single-guide RNA (sgRNA) target sequences for UNG (AAAGCCCACGGGCACGTTGC), MLH1 (AAGACAATGGCACCGGGATC), PMS1 (GTATCCTTAAACCTGACTTA), PMS2 (ATGCTGTCTTCTAGCACTTC), or MSH6 (GGAACATTCATCCGCGAGAA). GFP-positive cells were single-cell sorted using an SH800S Cell Sorter (Sony Biotechnology) and expanded for ~14 days. Clones successfully knocked out for each gene were identified by western blotting, as described below. MBD4 knockout in the parental diploidized clone of HAP1 used in this study was previously confirmed with antibodies raised against the N-terminal (Abcam, ab224809) and C-terminal (Abcam, ab12187) of MBD4 [11]. The identities of HAP1 cell models were confirmed by whole genome sequencing (WGS), as described below. All HAP1 cell models tested negative for mycoplasma with the Eurofins Mycoplasmacheck service, both before and after long-term culturing.

Cytosine targets selection and bisulfite sequencing

Target cytosines were selected based on pseudo-bulk analysis of whole-genome bisulfite sequencing data of KBM-7 cells (GEO accession number GSE65196) [29], as previously described [11]. We opted to target CpGs found within either highly methylated (>99%) or unmethylated (<1%) genomic regions based on average CpG methylation levels across 2 kb genomic bins, and within open-chromatin regions based on chromatin state annotations from ENCODE (dataset ENCFF300WNA). We further selected isolated CpGs not flanked by other cytosines in the same DNA strand within a ±7-bp window from the target CpG. All four selected targets overlapped intronic regions of transcribed genes, as confirmed using normalized transcripts per million (nTPM) values obtained from the Human Protein Atlas in HAP1 cells. These comprised 5mC target #1 in DNAJC16 (nTPM = 7.6), 5mC target #2 in LRP8 (nTPM = 12.2), C target #1 in STK38 (nTPM = 24.7), and C target #2 in HNRNPDL (nTPM = 534.8). To experimentally confirm the DNA methylation status of the selected target cytosines in HAP1 cells, 500 ng of genomic DNA (gDNA) was bisulfite-converted with EZ DNA Methylation-Gold Kit (Zymo, ZD5005). Polymerase chain reaction (PCR) on 50 ng bisulfite-converted DNA per reaction was performed with ZymoTaq Polymerase (ZE2001) using the following primers: 5mC target #1 (fwd GGGATTTTATTTTTTTATGTTTTAGTTTT; rev TCATACCTATAATACCAATACTTTAAAAAA), 5mC target #2 (fwd TATTATGTGTTAGATATTGTTAGGAGTT; rev TTTAATTTTTTAACCTAACACACAAAAC), C target #1 (fwd AAAAATGTGGAGAGGGTTAATAAG; rev CTACCTCTCCCTAAACTCTAATAAATTA), C target #2 (fwd AATAATTTAAATGATTTTTTAGTTTTGG; rev ACAAAACCTTTACTATATTTTATTAATCTCT). Amplified DNA was either directly analyzed by Sanger sequencing or cloned with CloneJET PCR Cloning Kit (Thermo Fischer Scientific, K1231) followed by Sanger sequencing on 10 clones per target. High efficiency of bisulfite conversion was observed in all samples analyzed by sequencing.

dCas9-APOBEC1 targeted deamination

The cytosine base editor (CBE) variant CMV-YE1-BE3-FNLS-CMV-mCherry [30] (Addgene, #154005) was reengineered with the Q5 Site-Directed Mutagenesis Kit (NEB, E0554S). First, we removed its uracil glycosylase inhibitor, which was originally intended to inhibit BER and enhance base-editing efficiencies [31]. Second, we mutated the D10A nickase Cas9 (nCas9) with the H840 variant to generate a catalytically dead Cas9 (dCas9). We refer to the resulting construct as dCas9-APOBEC1. sgRNA sequences were cloned into the vector pGL3-U6-sgRNA-EGFP [32] (Addgene, 107721), including for 5mC target #1 in DNAJC16 (AGATACGAAAGAGATGGTGA), 5mC target #2 in LRP8 (TGAATCGAGGGGTGGGGAGA), C target #1 in S TK38 (GAGTCGGAGGTTGGGGAAGG), and C target #2 in HNRNPDL (TAAATCGTGAAATAAAAACA). HAP1 cells were plated in complete media 24 h before transfection (1 million cells per well of six-well plate), and media was replaced with fresh complete media supplemented with 1 μM Palbociclib (CST, 47284S) immediately before co-transfection with 1.4 μg CBE vector and 150 ng of each sgRNA vector per well with jetOPTIMUS (Polyplus, 101000051) at a 1:1 ratio. At 24 h post-transfection, media was replaced with fresh complete media supplemented with 1 μM Palbociclib. At 48 h post-transfection, cells were harvested, resuspended in FACS buffer containing 1 μM DAPI, and live cells were sorted by predefined narrow bands of EGFP and/or mCherry with a FACSAria Fusion Cell Sorter (BD Biosciences). For the experiment involving base-editing detection by Sanger sequencing, Palbociclib treatment was omitted, cells were transfected with 1.4 μg CBE vector and 600 ng of sgRNA vector targeting either of the two 5mC sites, and cells with high EGFP and mCherry brightness were sorted. For experiments involving base-editing detection by droplet digital PCR (ddPCR), we fine-tuned EGFP and mCherry brightness levels used during sorting to maximize the quantitative range of the assay across HAP1 models and CpG targets while avoiding saturating levels of deamination. For MBD4 rescue experiments, only mCherry was used for sorting, as the MBD4 overexpression vector also expressed EGFP. gDNA was extracted from sorted cells with Quick-DNA Microprep Kit (Zymo, D3021). Custom TaqMan SNP Genotyping Assays (Thermo Fischer Scientific, 4 331349) were used to perform ddPCR quantitative genotyping using the QX600 ddPCR System (Bio-Rad), separately for each cytosine target. The amount of 10–30 ng DNA was used per reaction with the ddPCR Supermix for Probes no dUTP (Bio-Rad, #1863024). C>T fractional abundances were determined by Poisson distribution analysis using QX Software (Bio-Rad). Calculations were based on the number of droplets that were either positive for the T-specific probe (indicating the presence of at least one DNA molecule derived from unrepaired deamination, including T:G, U:G, T:A, or U:A at the target position) or exclusively positive for the C-specific probe (only DNA molecules with C:G at the target position). Therefore, under non-saturating DNA amounts per ddPCR reaction, the theoretical upper detection limit of the assay is 100% C>T fractional abundance, where all DNA molecules in in the sample are the result of unrepaired deamination events. Based on a minimum of 1069 droplets positive for target DNA analyzed per ddPCR replicate, the theoretical sensitivity of the assay was estimated to be <0.1% C>T fractional abundance across experiments and CpG targets. A negative control reaction using gDNA of untransfected cells was systematically included in all ddPCR runs, thus enabling estimation of assay specificity per run.

Isogenic cell models WGS

HAP1 cells were split every 3–4 days and kept at low density (<50% confluency) for ~60–90 days following knockout clone selection and initial expansion. A second clonal step was performed by morphology-based single-cell sorting with SH800S Cell Sorter. Expanded subclones had their DNA extracted with DNeasy Blood & Tissue Kit (Qiagen, 69504) and quantified using the Qubit dsDNA BR Assay Kit (Invitrogen, Q32850). One microgram DNA was used to prepare WGS libraries using the Kapa HyperPrep kit (Roche, 07962363001). Paired-end libraries (2 × 150 bp) were sequenced on a NovaSeq 6000 instrument. WGS coverage depth was set to 25×. Sequencing data QC and adapter trimming on all samples was performed with FastQC v0.11.9 and TrimGalore v0.6.10. Mapping was performed with BWA-MEM v0.7.17 (https://github.com/lh3/bwa) to GRCh38 (Primary assembly + ALT contigs). Duplicates were removed with Sambamba v1.0 (https://github.com/biod/sambamba). All samples had >90% mapping rate and average fragment size >300 bp. Single-nucleotide variant (SNV) and insertion/deletion (InDel) discovery was performed with GATK4 v4.4.0.0 (https://github.com/broadinstitute/gatk) Mutect2 in “tumor” versus “normal” mode, using sequencing data from HAP1 wild-type parental cells as the normal reference. Variants with a PASS call were selected using the FilterMutectCalls function and annotated with VEP v104.3 (https://github.com/Ensembl/ensembl-vep). Variants with less than four alternative allele reads or variant allele frequency <0.25 were excluded. Variants shared among subclones of each genotype were further excluded, as they represented variants most likely acquired before knockout first clonal selection step. As the mutational background due to culturing is expected to accumulate at similar rates across HAP1 isogenic models [33], it can serve as an internal normalization factor to estimate mutational burdens arising from DNA repair deficiencies independently of culturing time. Estimation of SNV burdens due to DNA repair deficiencies or background mutagenesis was performed with single base substitution (SBS) signature analysis with R 4.4.3 package signature.tools.lib v2.4.4 (https://github.com/Nik-Zainal-Group/signature.tools.lib) with the FitMS function in constrained Fit multi-step mode with threshold percent and minCosSimRareSig turned off. Estimation of background InDel burden was performed with R 4.5.1 package indelsig.tools.lib v1 (https://github.com/Nik-Zainal-Group/indelsig.tools.lib). InDels from Mutect2 were formatted to fit the package requirements by adding the nucleotide 5′ of a deletion in the REF column or the nucleotide 5′ of an insertion in the ALT column. InDels were then filtered with the indel_highspecific function to remove variants in highly repeated regions (≥10 repeats) and longer than 100 base pairs. As for SBS, InDel burden due to culturing was performed with the FitMS function in constrained Fit multi-step mode. The SBS and InDel background mutational signatures were considered as the average of the mutational profile of HAP1 wild-type subclones, two from this study and four from Silveira et al. [11], which were assigned as common signatures for fitting. MMR gene knockout signatures previously obtained in human iPSC cells [33] and MBD4 deficiency signature in human tumors (SBS96) [34] were assigned as rare SBS signatures. MMR deficiency signature in human tumors (InD7) [35] was assigned as rare InDel signature. This allowed us to estimate the proportion of the mutational burden in each subclone attributable to background mutagenesis or to mutagenesis associated with the knockout of specific DNA repair genes, independently of duplication rates and time of long-term culturing.

Human tumors WGS analysis

WGS data from publicly available MBD4-deficient tumors [11, 36, 37] and from the National Genomic Research Library (NGRL) available through Genomics England [38] data release v17 (2023–03-30) were selected, filtered, and processed as previously described [11]. SBS mutational signature analysis of tumor samples was performed on filtered somatic variants with signature.tools.lib v2.4.1, as previously described [11]. MMR-deficient samples were kept for downstream analyses if they showed (i) any contribution of MMR-deficiency rare signatures SBS6, SBS15, SBS26, SBS44. and SBS97; (ii) no contribution of other rare signatures, including signatures linked to MMR-deficiency in combination with POLE or POLD dysfunctions; and (iii) germline or biallelic somatic loss-of-function mutations in a single of the MMR genes PMS2, MLH1, MSH6, or MSH2. Replication strand asymmetry analysis was performed with R 4.2.1 package MutationalPatterns v3.8.1 (https://github.com/UMCUGenetics/MutationalPatterns).

Epitope-tagged MBD4 and MLH1 expression

HAP1 cells stably overexpressing N- or C-terminally FLAG-tagged MBD4 were obtained as previously described [11]. Additional synthetic lentiviral vectors were obtained from VectorBuilder for expression of 3xFLAG N-terminally tagged wild-type MBD4 (NM_003925.3) or 3xHA N-terminally tagged wild-type MLH1 (NM_000249.4) under the control of the hPGK promoter, with open reading frames followed by IRES and EGFP gene. These vectors were further modified with the Q5 Site-Directed Mutagenesis Kit (NEB, E0554) to generate the Y417A-F418A mutant of MBD4 and Y625E mutant of MLH1. Lentiviral particles were generated in HEK-293 and used to transduce HAP1 cells. Cells expressing each construct within a narrow band of EGFP expression were sorted with a SH800S Cell Sorter. Identity of the constructs expressed by each isogenic cell model was confirmed by PCR on gDNA followed by Sanger sequencing.

Co-immunoprecipitation

For co-immunoprecipitation (co-IP) experiments involving G₁-cell cycle arrest, HAP1 cells stably overexpressing N- or C-terminally FLAG-tagged MBD4 were cultured in complete media supplemented with 1 μM Palbociclib or vehicle 0.1% dimethyl sulfoxide (DMSO) for 17 h. For other co-IP experiments, HAP1 cells were cultured in complete media without further treatment. Total cell lysates were obtained from ~20 million cells in 300 μl lysis buffer [Tris–HCl, pH 8, 50 mM; NP-40, 1%; ethylenediaminetetraacetic acid (EDTA), pH 8, 0.2 mM; 1× protease inhibitor cOmplete EDTA-free (Roche, 11836170001)] at 300 mM NaCl and supplemented with benzonase 250 U/ml, followed by 15 min incubation at room temperature and 15 min at 4°C under agitation. Soluble fractions were obtained by centrifugation at 16 000 × g for 15 min at 4°C and collection of the supernatants. Protein concentrations were measured with the Pierce BCA Protein Assay Kit (Pierce, 23227). Lysates were diluted in lysis buffer to reach final concentrations of 1 μg/μl protein and 150 mM NaCl (input). For co-IP reactions using FLAG-tagged MBD4 as bait, 500 μl diluted lysates and 50 μl pre-washed Anti-DYKDDDK magnetic beads (Pierce, A36798) were incubated for 2 h at 4°C under agitation. Beads were washed five times with 1 ml lysis buffer 150 mM NaCl, and samples were eluted in 2× Laemmli Sample Buffer (Bio-Rad #1610737) with 50 mM dithiothreitol (DTT) for 5 min at 95°C. For co-IP reactions using HA-tagged MLH1 as bait, 500 μl diluted lysates and 25 μl pre-washed Anti-HA Magnetic Beads (Pierce, 88836) were incubated for 30 min at room temperature under agitation. Beads were washed five times with 1 ml lysis buffer 150 mM NaCl, samples were eluted in 60 μl HA peptide (Thermo Fischer Scientific, 26184) at 2 mg/ml in TBS for 10 min at 37°C, and supernatants were collected and denatured in an equal volume of 2× Laemmli Sample Buffer with 50 mM DTT for 5 min at 95°C. For co-IP reactions using endogenous proteins as bait, 500 μl diluted lysates were incubated overnight at 4°C under agitation with antibodies against MLH1 (2 μg; CST, 4256S), PMS1 (10 μg; Invitrogen, PA5-35952), or equivalent amount of IgG control (CST, 3900S), followed by 1 h incubation at room temperature under agitation with 25 μl pre-washed Protein A/G Magnetic Beads (Pierce, 88802). Beads were washed five times with 1 ml lysis buffer 150 mM NaCl and 0.1% NP-40, samples were eluted in 2× Laemmli Sample Buffer with 50 mM DTT for 5 min at room temperature, and supernatants were incubated for 5 min at 95°C. Inputs were denatured in an equal volume of 2× Laemmli Sample Buffer with 50 mM DTT for 5 min at 95°C. Samples were analyzed by western blotting, as described below.

Western blotting

For knockout confirmation, HAP1 total cell lysates were extracted in RIPA buffer supplemented with 1× protease inhibitor cOmplete (Roche, 11697498001) and sonicated with Bioruptor Pico (Diagenode) for 10 cycles of 30 s on and 30 s off at 4°C. Soluble fractions were obtained by centrifugation at 16 000 × g for 15 min at 4°C and collection of the supernatants. Protein concentrations were measured with the Qubit Protein Broad Range Assay (Invitrogen, A50668). Samples were run in NuPAGE 4%–12% Bis–Tris gel (Invitrogen, WG1402BOX) or NuPAGE 3%–8%, Tris–acetate gel (Invitrogen, WG1602BOX) and transferred to 0.45  µm nitrocellulose membranes. Membranes were blocked with 5% non-fat milk in TBS-T (Tris-buffered saline, 0.1% Tween-20), and incubated with primary antibodies in 5% bovine serum albumin in TBS-T overnight at 4°C. Antibodies against UNG (1:2000; ProteinTech, 12394-1-AP), MBD4 (1:1000; Abcam, ab224809), FLAG (1:2000; CST, 14793S), MLH1 (1:1000; Sigma, HPA052707), PMS1 (1:2000; Invitrogen, PA5-35952), PMS2 (1:1000; BD, 556415), MSH6 (1:1000; CST, 5424S), Cyclin A (1:1000; Santa Cruz, sc-56299), HA (1:2000; CST, 3724S), GFP (1:2000; Abcam, ab290), and Tubulin (1:2500; Invitrogen, 14-4502-82) were incubated overnight at 4°C. Secondary antibodies used were anti-rabbit HRP-linked (1:10 000–20 000; CST, 7074), anti-mouse HRP-linked (1:10 000–20 000; CST, 7076), anti-rabbit TrueBlot HRP-linked (1:1000; Rockland,18-8816-3), anti-rabbit IRDye 800CW (1:20 000; LI-COR, 926-32213), or anti-mouse IRDye 680RD (1:20 000; LI-COR, 926-68070). Blots were imaged with Odyssey Imaging System (LI-COR) or ChemiDoc Imaging System (Bio-Rad).

Structural modeling

The structural modeling of MLH1 and MBD4 complex was performed using AlphaFold2 [39] with a fragmentation strategy shown to increase the sensitivity of detection [40]. Sequences of MBD4 (O95243) and MLH1 (P40692) were retrieved from the UniProt database and were submitted to three iterations of MMseqs2 [41] against the uniref30_2202 database. The resulting alignments were filtered using hhfilter [42] using the parameters (‘id’=100, ‘qid’=25, ‘cov’=50) and the taxonomy assigned to every sequence, keeping only one sequence per species. Full-length sequences in alignments were then retrieved, and these sequences were realigned using MAFFT [43] with the default FFT-NS-2 protocol. MLH1 and MBD4 were fragmented into 13 and 30 overlapping fragments, respectively, each of ~30 amino acids long in the disordered regions. To build the mixed co-alignments, sequences in the alignment of individual partners were paired according to their assigned species and left unpaired in case no common species were found. The concatenated multiple sequence alignments of tested partners were used as input of one run of the AlphaFold2 algorithm with three recycles each and five generated models using ColabFold v1.5.2 [44] with the Multimer v2.3 model parameters [45]. Five scores were used to rate the quality of the models, the pLDDT, the pTMscore, the ipTMscore, and the model confidence score (weighted combination of pTM and ipTM scores with a 20:80 ratio) together with the actifpTM score [46]. 3D structures were represented using ChimeraX [47] and multiple sequence alignments were displayed using Jalview [48].

Statistical analysis

Unpaired two-sided Wilcoxon tests and unpaired two-sided t-tests with Benjamini–Hochberg (BH) multiple comparisons adjustment for controlling the false discovery rate were performed with R 4.4.3 package ggpubr v0.6.0 (https://github.com/kassambara/ggpubr).

Results

A Cas9-based system to induce targeted cytosine deamination

Quantifying mutagenesis that arises from 5mC deamination has remained challenging, largely because spontaneous deamination occurs at low rates and is distributed across a pool of ∼28 million CpG dinucleotides in the human genome [11, 49]. We thus developed a system that leverages a CBE to induce site-specific cytosine deamination in human cells. CBEs typically consist of a catalytically impaired Cas9 variant fused to a cytidine deaminase of the APOBEC family, which is guided to a specific locus by an sgRNA. Upon binding, the deaminase converts a target 5mC to T, or a target unmethylated C to U, within an exposed single-stranded editing window [31]. We redesigned the CBE variant YE1-BE3-FNLS, which shows high on-target editing efficiency and extremely low off-target and bystander edits [30], to encode a dCas9 fused to its engineered APOBEC1. Hence, deamination can occur without generating flanking DNA nicks (Fig. 1A).

Figure 1.

Targeted cytosine deamination induced by a Cas9-based system retrieves the specialized repair roles of BER glycosylases. (A) Scheme of the CBE used in this study, based on a dCas9 fused to an engineered rat APOBEC1 cytidine deaminase. The tool induces strand-specific cytosine deamination to produce T:G or U:G mismatches within an exposed single-stranded editing window of the target sequence. (B) Sanger sequencing on amplified gDNA of HAP1 cells exposed to the CBE, with sgRNA targeting either of two independent methylated cytosine (5mC) loci within CpGs. (C) Experimental framework to quantify efficiency of repair following targeted cytosine deamination, based on transfection of HAP1 cells exposed to CDK4/6 inhibitor Palbociclib with CBE and sgRNA vectors, sorting of cells of a narrow range of EGFP (co-expressed in sgRNA vectors) and mCherry (co-expressed in CBE vector), and ddPCR quantitative genotyping on gDNA. (D) Western blotting on total cell extracts of HAP1 cells wild-type or knockout (^KO) for BER glycosylases MBD4 or UNG. Arrows indicate specific bands of UNG1 (mitochondrial) and UNG2 (nuclear). Positions of molecular weight markers in kDa are indicated. (E) ddPCR quantification of CpG>TpG frequency at two 5mC loci and two unmethylated C loci within CpGs, following targeted deamination in HAP1 cells. Higher CpG>TpG frequencies than wild-type cells indicate disruption of DNA repair. Values are shown relative to the mean CpG>TpG fractional abundance of wild-type replicates, indicated by the percentages and the dark gray dashed line. The theoretical lower (0%) fractional abundance limit is indicated as a light gay dashed line. The theoretical upper (100%) fractional abundance limit falls outside the plotted scales. Five biological replicates per genotype from a single experiment are shown. The single replicate negative control (neg) represents HAP1 cells not exposed to the CBE. Two-sided unpaired t-test P-values with BH multiple comparisons adjustment across facets are indicated following the notation: *P ≤ .05, **P ≤ .01, ^****P ≤ .0001. Non-significant P-values are not shown. Boxes indicate the median, 25th, and 75th percentiles. Whiskers extend to the largest or lowest value up to 1.5 times the inter-quartile range from the hinges.

We transiently co-transfected HAP1 cells with our dCas9-APOBEC1 and sgRNAs targeting two methylated CpG loci, as confirmed by bisulfite sequencing (Supplementary Fig. S1). Sanger sequencing 48 h post-transfection confirmed CpG>TpG mutagenesis specifically at the target position harboring a cytosine on the Watson strand, with no detectable mutagenesis at the adjacent nucleotide position harboring a cytosine on the Crick strand (Fig. 1B). Overall, our system enables highly efficient strand-specific CpG deamination to introduce mismatches in the genome independently of nucleotide misincorporation during DNA replication, thereby mimicking spontaneous deamination under physiological conditions.

Targeted cytosine deamination identifies the specialized roles of BER glycosylases

We next sought to determine whether our dCas9-APOBEC1 system could be used to study cytosine deamination repair. We targeted two methylated and two unmethylated CpGs in isogenic HAP1 cell models (Supplementary Figs S1 and S2) to generate T:G and U:G mismatches, respectively. Transfection was performed in cells concomitantly exposed to the cyclin-dependent kinases CDK4/6 inhibitor Palbociclib [50], thereby leading to G₁-arrest (Supplementary Fig. S3) and limiting fixation of mismatches as C>T transitions during S-phase. To distinguish DNA molecules containing a T/U or C at each targeted CpG, we used ddPCR quantitative genotyping on gDNA of transfected cells. Notably, U has the base-pairing properties of T and is recognized as such following the end-point ddPCR reaction (Fig. 1C). In our experimental workflow, higher C>T frequencies relative to wild-type cells indicate ineffective repair of these DNA lesions.

In a proof-of-concept experiment, we applied our workflow in HAP1 cell lines knocked out for MBD4 or UNG, the main BER glycosylases responsible for 5mC or unmethylated C deamination repair, respectively [11, 51] (Fig. 1D and Supplementary Figs S1 and S2). MBD4 knockout (^KO) led to a 2.2–2.9-fold increase in C>T frequency at 5mC loci (Fig. 1E), consistent with previous reports relying on the endogenous rates of 5mC deamination in mammalian cells [11, 36]. Surprisingly, MBD4^KO led to a small but significant decrease in the C>T frequency at unmethylated C loci (Fig. 1E). MBD4 also recognizes U:G mismatches as substrates, and its loss may increase mismatch accessibility for higher-efficiency repair by UNG or SMUG1 [52]. UNG^KO led to a 1.9–5.1-fold increase in C>T frequency specifically at unmethylated C loci, with no effect on the repair of deaminated 5mC loci (Fig. 1E). Overall, our dCas9-APOBEC1 system successfully teased out the known specialized roles of these BER glycosylases, thus enabling the study of cytosine deamination repair of endogenous CpG loci at their native chromatin state.

Non-canonical MMR is central to 5mC deamination repair in coordination with MBD4

To explore the interplay of BER and MMR in 5mC deamination repair, we disrupted different MMR complexes by inactivating MMR genes in wild-type or MBD4^KO HAP1 cells. We targeted the complexes MutLα (MLH1^KO or PMS2^KO), MutLβ (MLH1^KO or PMS1^KO), or MutSα (MSH6^KO) (Fig. 2A and Supplementary Fig. S2). We then applied our dCas9-APOBEC1 system to assess whether these DNA repair deficiencies would disrupt deamination repair at 5mC or unmethylated C loci (Supplementary Fig. S1).

Figure 2.

MMR complexes MutLβ and MutSα are involved in 5mC deamination repair in coordination with BER glycosylase MBD4. (A) Western blotting on total cell extracts of HAP1 cells wild-type or knockout (^KO) for MMR members, in combination or not with MBD4 double-knockout (^dKO). MLH1^KO disrupts the stability of PMS1 and PMS2 proteins. PMS1^KO partially disrupts the stability of MLH1 protein. Positions of molecular weight markers in kDa are indicated. (B) ddPCR-based quantification of CpG>TpG frequency at two 5mC and two unmethylated C loci within CpGs, following targeted deamination in HAP1 cells. Values are shown relative to the mean CpG>TpG fractional abundance of wild-type replicates, indicated by the percentages and the dark gray dashed line. The theoretical lower (0%) and upper (100%) fractional abundance limits are indicated as light gray and red dashed lines, respectively, unless they fall outside the plotted scales. Three biological replicates per genotype per experiment from two independent experiments are shown. The single replicate negative control (neg) per experiment represents HAP1 cells not exposed to the CBE. Two-sided unpaired t-test P-values with BH multiple comparisons adjustment across facets are indicated following the notation: *P ≤ .05, **P ≤ .01, ***P ≤ .001, ^****P ≤ .0001. Non-significant P-values are not shown. Boxes indicate the median, 25th, and 75th percentiles. Whiskers extend to the largest or lowest value up to 1.5 times the inter-quartile range from the hinges. (C) Working model of 5mC deamination repair, highlighting the cooperative repair roles of MBD4 and MMR complexes MutLβ and MutSα.

First, MSH6^KO led to the strongest phenotype at 5mC loci, with a 6.9–7.2-fold increase in C>T frequency relative to wild-type cells (Fig. 2B). This suggests the importance of upstream mismatch recognition by the MMR complex MutSα during 5mC deamination repair. Next, we delineated the downstream MMR complexes involved. Single MLH1^KO, which also disrupts the stability of PMS1 and PMS2 proteins (Fig. 2A), led to a 4.7–5.1-fold increase in C>T frequency at 5mC loci. Surprisingly, single PMS1^KO led to a comparable effect to single MLH1^KO, with a 3.8–4.5-fold increase in C>T frequency at 5mC loci. In contrast, single PMS2^KO did not lead to a significant increase in C>T frequency at 5mC loci (Fig. 2B), indicating that 5mC deamination repair is largely independent of MutLα-mediated post-replicative MMR. Furthermore, none of the MMR knockouts resulted in a consistent increase in C>T frequencies at unmethylated C loci compared with cells matched by their MBD4 status (Fig. 2B), suggesting that MMR plays a central role in deamination repair specifically at methylated CpGs. Of note, our experimental system interrogates deamination repair independently of CpG mutagenesis linked to replicative errors [53, 54]. Hence, our results support the hypothesis that non-canonical MMR involving the complexes MutLβ and MutSα plays a central role in the recognition and repair of deaminated 5mC.

Next, we investigated the crosstalk between MMR and BER glycosylase MBD4. Remarkably, MBD4^KO in combination with MLH1^KO, PMS1^KO, or MSH6  ^KO led to no additive increase in C>T frequencies at 5mC loci as compared to their respective MMR gene single knockouts (Fig. 2B). This indicates that MBD4 activity is dependent on the MMR complexes MutLβ and MutSα. Strikingly, loss of MBD4 only partially reproduced the phenotype observed upon disruption of MutLβ or MutSα, suggesting these MMR complexes can elicit some level of 5mC deamination repair independently of MBD4. Overall, our data reveal a previously unrecognized function of non-canonical MMR involving MutLβ and MutSα complexes, highlighting its interplay with the BER glycosylase MBD4 in 5mC deamination repair (Fig. 2C).

PMS1 loss phenocopies the CpG>TpG hypermutation linked to MBD4 deficiency

In the absence of induced DNA damage, inactivating a gene that is critical to maintaining genome stability from endogenous DNA damage should produce a specific signature. This would manifest as an altered mutation profile and/or an increased mutation burden not explained by the mutational background. To orthogonally confirm the results obtained with our dCas9-APOBEC1 system, we took a genome-wide approach relying on the endogenous cytosine deamination rates. Clonal knockout lines were long-term cultured to permit mutation accumulation, followed by a second single-cell cloning step, expansion and WGS of 3–4 individual subclones per genotype (Fig. 3A). We excluded variants present in parental HAP1 cells or shared among subclones, thereby focusing on mutations that occurred after isolation of the initial knockout clones. This dataset was complemented by mutational profiles we previously obtained using a similar strategy in MBD4^KO versus wild-type HAP1 cells [11]. Data from wild-type HAP1 subclones were used as a reference for the background mutagenesis arising from in vitro culturing [55].

Figure 3.

Disruption of MutLβ and MutSα lead to genome-wide CpG>TpG hypermutagenesis. (A) Workflow used to quantify mutational burdens of clonally expanded isogenic cell models. Subclone-specific mutational burdens represent a combination of knockout-related mutations due to specific DNA repair deficiencies and background mutations due to in vitro culturing. (B) Mutational signature analysis of subclones to linearly estimate the mutational burdens due to knockouts of DNA repair genes or background. The background mutational signature was obtained from wild-type subclones. Data for all SNV classes are shown. (C) Distributions of CpG>TpG substitution frequencies among all SNVs in HAP1 subclones, relative to the mean of wild-type subclones (shown as a dashed line). Two-sided unpaired t-test P-values with BH multiple comparisons adjustment, applied separately in panels B and C, are indicated following the notation: *P ≤ .05, ***P ≤ .001, ^****P ≤ .0001. Non-significant P-values are not shown. Boxes indicate the median, 25th, and 75th percentiles. Whiskers extend to the largest or lowest value up to 1.5 times the inter-quartile range from the hinges. (D) C>T substitution frequencies by trinucleotide context among all SNVs, relative to the mean of wild-type subclones (shown as a dashed line). Bars represent the means of subclones, shown as dots.

We first applied mutational signature analysis to linearly estimate mutational burdens specifically linked to each gene knockout [33–35]. PMS2^KO, MLH1^KO, and MSH6^KO were associated with strong SNV and InDel hypermutated profiles, as expected following disruption of post-replicated MMR in highly proliferative HAP1 cells (Fig. 3B and Supplementary Fig. S4). In contrast, MBD4^KO and PMS1^KO led to a more focused phenotype, with a low overall contribution of InDels and an SNV burden highly enriched in CpG>TpG mutations (Fig. 3B and C; and Supplementary Figs S4 and S5). Consistent with data using our dCas9-APOBEC1 system, MBD4^KO was associated with a 2-fold increase in CpG>TpG frequency, while PMS1^KO was associated with a slightly higher effect of 2.3-fold (two-sided unpaired t-test; P = .021). Remarkably, CpG>TpG mutations in PMS1^KO cells showed similar relative frequencies by trinucleotide context to MBD4^KO cells (Fig. 3D), with the pattern ACG>CCG>GCG>TCG characteristic of the SBS96 mutational signature of MBD4-deficient human tumors [11, 34]. Altogether, we show that PMS1 loss phenocopies the preferential CpG>TpG hypermutator profile of MBD4-deficient cells, indicating that 5mC deamination repair may represent one of the main functions of human PMS1.

The relative frequency of CpG>TpG mutations in PMS2^KO cells was comparable to wild-type cells, indicating that disruption of post-replicative MMR does not lead to preferential CpG>TpG mutagenesis (Fig. 3C). In contrast, both MLH1^KO and MSH6^KO led to a robust increase in CpG>TpG frequency (2.7- and 2.9-fold, respectively), although C>T transitions in other trinucleotide contexts were also highly enriched (Fig. 3C and D). Notably, MLH1^KO and MSH6^KO shared features of both 5mC deamination repair disruption (the highest CpG>TpG peak in ACG context) and post-replicative MMR disruption (the highest C>T peak in GCC context and a large contribution of T>C mutations) (Fig. 3D and Supplementary Fig. S5). Overall, our genome-wide data support the hypothesis that MutLβ and MutSα complexes play central roles in both post-replicative MMR and 5mC deamination repair, while the MutLα complex is primarily essential for post-replicative MMR.

5mC deamination contributes to the CpG>TpG burden of tumors deficient for MLH1 and MSH6/MSH2

We next wondered whether the mutational landscape of MMR-deficient human tumors would support the involvement of MMR factors in 5mC deamination repair. We reanalyzed whole genomes from the Genomics England (GEL) pan-cancer series (release data v17 with 12 726 high-quality tumor samples) [38], which includes 12 MBD4-deficient tumor samples. We complemented this series with six previously described MBD4-deficient tumors [11, 36, 37]. We identified MMR-deficient tumors in the GEL series by performing mutational signature analysis [34], excluding samples with overlapping DNA polymerase defects, and focusing on those harboring germline or biallelic somatic loss-of-function mutations in PMS2, MLH1, MSH6, or MSH2. To note, PMS1 loss in human cancer remains anecdotal, and previously reported cases lack WGS data on tumor samples [56, 57].

Preferential CpG>TpG mutagenesis at the ACG context, indicative of 5mC deamination repair deficiency (e.g. MBD4 and PMS1 loss), was minimally present in PMS2-inactivated non-colorectal tumors, but was increasingly prominent in MLH1-inactivated and MSH6/MSH2-inactivated non-colorectal tumors, respectively (Fig. 4A and Supplementary Fig. S6). Furthermore, we observed the lowest CpG>TpG percentages and absolute burden in PMS2-inactivated tumors (6.0 ± 0.6%; 2219 ± 255), followed by MLH1-inactivated tumors (16.2 ± 5.0%; 6229 ± 3643), and MSH6/MSH2-inactivated tumors (27.7 ± 10.3%; 17 816 ± 7380). The same overall pattern was observed for tumors of colorectal origin, which are generally associated with a higher burden of all SNVs, including CpG>TpG transitions (Fig. 4B and Supplementary Fig. S7). Overall, we show that the CpG>TpG mutational landscape of MMR-deficient tumors closely mirrored those observed in our HAP1 isogenic models beyond background levels.

Figure 4.

Genome-wide CpG>TpG hypermutagenesis of MMR-deficient tumors is partly linked to 5mC deamination. (A) C>T substitution percentages by trinucleotide context among all SNVs of human tumor samples of non-colorectal origin. Bars represent the means among cases harboring germline or somatic biallelic mutations in MBD4 (n = 18), PMS2 (n = 4), MLH1 (n = 9), or MSH2/6 (n = 24). (B) Distributions of CpG>TpG percentages among all SNVs or absolute CpG>TpG burdens of human tumor samples of either colorectal or non-colorectal origins. Two-sided Wilcoxon test P-values with BH multiple comparisons adjustment across facets are indicated. (C) Replication strand asymmetry bias in the leading versus the lagging strands among C>T mutations. Tumor samples of both colorectal and non-colorectal origins were grouped. No bias for either strand is indicated as a red line. Two-sided unpaired t-test P-values with BH multiple comparisons adjustment across facets are indicated. Statistics shown throughout the figure follow the notation: *P ≤ .05, **P ≤ .01, ***P ≤ .001, ^****P ≤ .0001. Non-significant P-values are not shown. Boxes indicate the median, 25th, and 75th percentiles. Whiskers extend to the largest or lowest value up to 1.5 times the inter-quartile range from the hinges.

Previous studies have shown that post-replicative repair defects cause asymmetric mutational patterns arising during leading- and lagging-strand DNA synthesis [58]. In contrast, CpG>TpG mutagenesis driven by spontaneous 5mC deamination in MBD4-deficient tumors does not exhibit replication-strand bias (Fig. 4C). Strikingly, tumors inactivated for MLH1, MSH6, or MSH2 exhibited a weaker replication-strand bias toward the leading strand for CpG>TpG mutations compared with other C>T transitions (Fig. 4C). This pattern suggests that the elevated CpG>TpG load in these tumors is at least partly due to failed repair of 5mC deamination, rather than post-replicative MMR disruption. Notably, PMS2-mutated tumors showed similar levels of replication-strand bias for CpG>TpG and other C>T transitions (Fig. 4C). Overall, the mutational landscape of MMR-deficient tumors aligns with our experimental results, showing that the MutLβ and MutSα complexes, but not the MutLα complex, play major roles in 5mC deamination repair.

MBD4 engages with MutL complexes through its physical interaction with MLH1

MBD4 has been shown to form a stable complex with MutLα, although experiments using a co-expression baculovirus system indicated a direct physical interaction only between MBD4 and MLH1, and not between MBD4 and PMS2 [25]. Given that our data do not support a role for PMS2 in 5mC deamination repair, we next considered whether MBD4 might also assemble into a multimeric complex with MutLβ. To test this hypothesis, we generated HAP1 cell lines stably overexpressing MBD4 constructs bearing either C- or N-terminal FLAG tags and performed co-IP assays. We successfully co-immunoprecipitated MLH1, PMS1, and PMS2 with both MBD4 constructs as baits. Comparable interaction levels were detected in asynchronous versus G₁-arrested cells, indicating that MBD4 association with these MMR components is largely independent of cell cycle phase regulation (Fig. 5A). We further confirmed the interaction between MBD4 and the MutLβ complex by immunoprecipitating endogenous MBD4 from wild-type HAP1 cell extracts using PMS1- and MLH1-specific antibodies (Fig. 5B).

Figure 5.

MBD4 physically interacts with MutL complexes in a MLH1-dependent manner. (A) Western blotting of anti-FLAG immunoprecipitated HAP1 total cell extracts, and respective inputs. Parental cells or stably overexpressing C- or N-terminally FLAG-tagged MBD4 (MBD4^FLAG) cultured with vehicle (asynchronous) or with 1 µM Palbociclib (G₁-arrested). Cyclin A was used as a marker of cycling cells. (B) Western blotting of immunoprecipitated total cell extracts of non-tagged parental HAP1 cells, using antibodies specific for MLH1, PMS1, or isotype IgG control. Endogenous (endog.) MBD4 bands are indicated. (C) Evolutionary conservation of the MLH1-interacting region in MBD4. Schematic representation of human MBD4 highlighting the methyl-CpG binding domain (MBD) and the DNA glycosylase domain, together with AlphaFold confidence coloring along the protein sequence. A multiple sequence alignment of the MBD4 C-terminal region (around residues 400–450) from representative metazoan species is shown below. The conserved aromatic residues Y417 and F418 (in human), which mediate interaction with the MLH1 C-terminal domain, are highlighted. This region shows strong conservation across vertebrates and invertebrates, supporting an evolutionarily conserved MLH1–MBD4 interaction interface. (D) AlphaFold2 model of the MLH1 C-terminal domain in complex with MBD4. Ribbon representation of the MLH1 C-terminal domain (residues 500–756) shown in blue, in complex with a C-terminal fragment of MBD4 (residues 410–430) shown in pink. The interaction interface is highlighted, with MBD4 residues Y417 and F418 displayed as sticks and indicated by arrows. These conserved aromatic residues form a key contact surface with MLH1. The model was generated using AlphaFold2, with an overall confidence score of 0.81, an ipTM score of 0.79, and an actual interface pTM (actifpTM) score of 0.94, indicating a high-confidence interface prediction. (E) Western blotting of anti-FLAG immunoprecipitated total cell extracts and respective inputs of parental HAP1 wild-type cells or stably overexpressing MBD4^FLAG wild-type (MBD4^wt) or Y417A-F418A mutant (MBD4^mut). (F) Western blotting of anti-HA immunoprecipitated total cell extracts and respective inputs of HAP1 wild-type cell, and MLH1^KO cells stably expressing HA-tagged MLH1 wild-type (MLH1^wt) or Y625E mutant (MLH1^mut). Positions of molecular weight markers in kDa are indicated in western blotting panels.

Next, we used a scanning protocol based on AlphaFold2 Multimer [39, 40] to generate a structural model of the direct interaction between MBD4 and MLH1. AlphaFold2 predicted a high-confidence interaction between the C-terminal region of MLH1 encompassing residue Y625 and a highly conserved motif in MBD4 encompassing residues Y417 and F418 (confidence score of 0.81) (Fig. 5C and D, and Supplementary Fig. S8). This short MBD4 motif lies immediately adjacent to the DNA glycosylase domain but is predicted to be intrinsically disordered in the absence of MLH1 (Fig. 5C).

Based on the predicted 3D interaction map, we examined the contact interface between MBD4 and MLH1 by introducing point mutations into residues predicted to contribute to complex formation. First, we generated HAP1 cell lines stably expressing N-terminally FLAG-tagged MBD4, either wild-type (MBD4^WT) or Y417A/F418A mutant (MBD4^Y417A/F418A). Co-IP assays showed that these mutations resulted in near-complete loss of interaction not only to MLH1, but also to PMS1 and PMS2 (Fig. 5E). Second, we generated HAP1 cell lines stably expressing N-terminally HA-tagged MLH1, either wild-type (MLH1^WT) or Y625E mutant (MLH1^Y625E) in a MLH1^KO background. Consistent with previous reports describing the pleiotropic effects of disrupting this MLH1 residue [59, 60], MLH1^Y625E exhibited reduced protein stability relative to MLH1^WT. To compensate this difference, we selected HAP1 cell populations expressing comparable levels of either MLH1 construct (Fig. 5F). Co-IP assays showed complete loss of MLH1^Y625E interaction with MBD4, whereas interactions with PMS1 and PMS2 were largely retained. Moreover, the near-complete loss of PMS1 and PMS2 protein levels in MLH1^KO cells was partially reverted upon MLH1^Y625E expression (Fig. 5F). These findings indicate that MLH1^Y625E retains its capacity to serve as a central scaffold for the assembly of the MutLα and MutLβ complexes, while selectively disrupting its interaction with MBD4. Collectively, these results support a direct interaction between MLH1 and MBD4 and suggest that MLH1 serves as the molecular bridge linking MBD4 to PMS1- and PMS2-containing MutL complexes.

MBD4-mediated 5mC deamination repair is dependent on MLH1

To determine the functional consequences of disrupting the MBD4-MLH1 interaction, we overexpressed either MBD4^WT or the interaction-deficient MBD4^Y417A/F418A mutant in MBD4^KO cells (Fig. 6A). Using our dCas9-APOBEC1 system, we observed complete restoration of 5mC deamination repair proficiency with MBD4^WT, whereas the MBD4^Y417A/F418A mutant failed to rescue repair activity (Fig. 6B). Although we cannot exclude the possibility that the MBD4^Y417A/F418A mutant shows impaired glycosylase activity, our data are consistent with the interpretation that MBD4 is functionally dependent on its physical interaction with MLH1.

Figure 6.

Rescue of MBD4-mediated 5mC deamination repair is dependent on MLH1. (A,C) Western blotting on total cell extracts of HAP1 cells wild-type, MBD4^KO (pink box) or MLH1^KO (blue box), stably overexpressing or not MBD4^FLAG wild-type (wt) or Y417A-F418A mutant (mut). Positions of endogenous and exogenous MBD4 bands are indicated. Positions of molecular weight markers in kDa are indicated. (B,D) ddPCR-based quantification of CpG>TpG frequency at two 5mC and two unmethylated C loci within CpGs, following targeted deamination in HAP1 cells. Values are shown relative to the mean CpG>TpG fractional abundance of wild-type replicates, indicated by the percentages and the dark grey dashed line. The theoretical lower (0%) fractional abundance limit is indicated as a light grey dashed line. The theoretical upper (100%) fractional abundance limit falls outside the plotted scales. Five biological replicates per genotype per experiment are shown. Single independent experiments are shown in (B) and (D). The single replicate negative control (neg) per experiment represents HAP1 cells not exposed to the CBE. Two-sided unpaired t-test P-values with BH multiple comparisons adjustment across facets, applied separately in B and D, are indicated following the notation: *P ≤ .05, **P ≤ .01, ***P ≤ .001, ^****P ≤ .0001. Non-significant P-values are not shown. Boxes indicate the median, 25th, and 75th percentiles. Whiskers extend to the largest or lowest value up to 1.5 times the inter-quartile range from the hinges.

To confirm this hypothesis, we re-expressed MBD4^WT in different MLH1 backgrounds (Fig. 6C). We observed that MBD4^WT restored 5mC deamination repair only in an MLH1-proficient background but not in MLH1^KO cells (Fig. 6D). These findings are fully consistent with our earlier observation that MBD4 loss does not further increase C>T mutation frequencies at 5mC sites in MLH1^KO and PMS1^KO backgrounds (Fig. 2B). Altogether, we demonstrate that MBD4-mediated repair of deaminated 5mC requires MLH1. Lastly, our data provide a mechanistic rationale for the crosstalk between BER and non-canonical MMR in the repair of methylated DNA damage.

Discussion

Spontaneous deamination of 5mC is a major endogenous source of CpG>TpG mutations involved in vertebrate genomic evolution, aging, and cancer. In this study, by redesigning a high-fidelity Cas9-based deamination system to induce T:G and U:G mismatches with high efficiency in native chromatin, we unlocked the potential to study deamination repair in near-physiological conditions. We validated the sensitivity and specificity of our dCas9-APOBEC1 deamination system, disentangling parallel repair activities by specialized BER glycosylases. We extended on these findings by showing that T:G mismatches arising from 5mC deamination are processed by a pathway that couples the BER glycosylase MBD4 to non-canonical MMR involving MutSα and MutLβ.

Strikingly, disruption of 5mC deamination repair was most pronounced upon MutSα loss, intermediate upon MutLβ loss, and less pronounced upon MBD4 loss. Furthermore, we show evidence that MBD4 acts downstream of MutSα and MutLβ, and we demonstrate that MBD4 activity against 5mC deamination requires MLH1. Altogether, our data suggest a stepwise repair hierarchy in which T:G mismatches are most likely recognized by MutSα, leading to the recruitment of downstream repair effectors such as MutLβ-MBD4 and potentially other DNA repair complexes yet to be defined. We provide a mechanistic rationale for the interplay between BER and MMR, showing that an intrinsically disordered domain in MBD4 enables its physical coupling to MutL complexes through an interaction with the MLH1 C-terminus. Interestingly, this MBD4 domain was omitted from previous crystal structures of its catalytic domain, which resolved only the folded glycosylase core [61, 62]. While we demonstrate that MBD4 acts in coordination with MutLβ during 5mC deamination repair, the functional significance of its interaction with MutLα remains unclear. Given that MutLα is the primary MutL complex in canonical MMR and also interacts with MBD4, the basis for the preferential engagement of MutLβ in this pathway remains to be elucidated. Despite these unresolved questions, our findings challenge the long-held view that 5mC deamination is repaired primarily through the classical BER pathway.

Notably, our genome-wide mutational analyses confirmed previous observations that loss of MutSα results in an increased CpG>TpG burden with replication-independent features [49, 63]. Although genomic data alone cannot fully disentangle CpG>TpG mutations arising from defective repair of 5mC deamination from those associated with DNA replication, our findings are consistent with the patterns observed using the dCas9-APOBEC1 deamination system, where mismatches are generated independently of replication. Furthermore, we found that the genome-wide mutational consequences of PMS1 loss closely recapitulated those observed in MBD4-deficient cells. These findings raise the possibility that, akin to MBD4, the repair of 5mC deamination lesions may constitute one of the major biological functions of human PMS1.

While recurrent homozygous and heterozygous loss-of-function mutations in MBD4 are associated with cancer predisposition syndromes [27, 28], PMS1 mutations have not been linked to Lynch syndrome, and Pms1 deficiency does not promote tumor formation in mice [64]. Although recent studies have implicated PMS1 in repeat expansion [65, 66] and our findings identify a role for PMS1 in 5mC deamination repair, evidence supporting a contribution of PMS1 to human cancer remains limited [56, 57]. The reasons underlying this apparent discrepancy remain unclear and warrant further investigation.

Because both nucleotides within a T:G mismatch are natural to DNA, the repair machinery requires strand-specific signals in cis to guarantee repair of the damaged strand. During post-replicative MMR, this is elicited by transient replication-associated signals that distinguish the daughter and parental strands. Although the precise mechanism of strand discrimination in eukaryotes remains unclear, it may involve pre-existing strand breaks (such as the ends of Okazaki fragments) or the directional loading of PCNA to direct MutLα endonuclease activity to the daughter strand [16–19]. In contrast, non-canonical MMR has been proposed to lack strand specificity and promote the recruitment of error-prone DNA polymerases [22, 23, 67]. In the context of 5mC deamination, we propose an alternative model in which MBD4 guides non-canonical MMR toward the damaged DNA strand, relying on its MBD domain to asymmetrically bind to a T:G mismatch within a methylated CpG [68] and its specific glycosylase activity toward the thymine [61]. This model would be consistent with the lack of endonuclease activity of the MutLβ complex [18]. While our work provides a foundation for re-evaluating 5mC deamination repair, the precise mechanistic basis for MBD4 engagement with MMR factors may be addressed in future studies.

In summary, our study delineates a previously unrecognized MutSα-MutLβ-MBD4-driven repair pathway that preserves the integrity of methylated cytosines, establishing non-canonical MMR as a fundamental safeguard of the human genome against CpG mutagenesis. Our work may redefine current models of mutagenesis and inform therapeutic strategies targeting DNA repair pathways.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

M.R. reports non-financial support from AstraZeneca and Merck Sharp and Dohme, grants from Daiichi Sankyo, personal fees from AstraZeneca, Immunocore, Merck Sharp and Dohme, and GlaxoSmithKline, all outside the submitted work. M.-H.S. reports grants from Immunocore and Bionano, and royalties from Myriad Genetics, all outside the submitted work. The remaining authors have no conflict of interest to declare.

Funding

This work was supported by funding from the French National Cancer Institute (INCa) and Organisation de la direction générale de l’offre de soins (DGOS) Programme de Recherche Translationnelle en Cancérologie (PRT-K19-51); the INCa Projets libres de recherche Biologie et Sciences du Cancer (PLBIO21-012); the Agence Nationale de la Recherche (ANR) Equipements d’Excellence (EQUIPEX) Investissements d’Avenir program (ANR10-INBS-09-08 to the Institut Curie ICGex NGS platform); the Institut National de la Santé et de la Recherche Médicale (INSERM) Interface program (to M.R.); the Institut du développement et des ressources en informatique scientifique (IDRIS) and Grand équipement national de calcul intensif (GENCI) for HPC ressources access (2025-AD010314343R2 to R.G.); and the Assistant Secretary of Defense for Health Affairs endorsed by the Department of Defense, in the amount of $1 907 499.00, through the Melanoma Research Program Focused Program Award—Rare Melanomas FY22 (Award No. HT94252310722). Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the Assistant Secretary of Defense for Health Affairs or the Department of Defense. Funding to pay the Open Access publication charges for this article was provided by the United States Department of Defense (Award No. HT94252310722).

Data availability

New molecular models based on computational modeling are deposited in ModelArchive under the link https://modelarchive.org/doi/10.5452/ma-8z967. Newly generated cell line WGS are deposited in the European Genome-phenome Archive (EGA) database under accession EGAS50000001662. Previously published WGS are deposited in the EGA database under accessions EGAS50000000536, EGAD00001003568, and EGAD00001005454. Data from the National Genomic Research Library (NGRL) used in this research are available within the secure Genomics England Research Environment. Access to NGRL data is restricted to adhere to consent requirements and protect participant privacy. Data used in this research include tumor-normal WGS pairs from the NGRL cancer series, with derived datasets available within the Research Environment at location /re_gecip/cancer_pan/abortoli. Access to NGRL data is provided to approved researchers who are members of the Genomics England Research Network, subject to institutional access agreements and research project approval under participant-led governance. For more information on data access, visit: https://www.genomicsengland.co.uk/research.