L-MAP

Whole-genome bisulfite sequencing only reports a methylation level where reads pile up deep enough to call one, and cytosines below the coverage threshold are dropped. Deciding whether a given cytosine's methylation actually fluctuates — rather than sitting stably methylated for the life of the cell — then requires comparing a wild-type sample against a knockout of the enzymes that write and erase the mark, and those knockout lines are expensive enough that only a handful exist.

L-MAP — Language model-based Methyltransferases Activity Predictor — asks whether the sequence itself already carries the answer. Saleh Sereshki and Stefano Lonardi at the University of California, Riverside take the 512 bp of genomic sequence centred on a cytosine, tokenize it into overlapping 6-mers, and push it through DNABERT into a small fully connected head that emits a two-class decision: differentially methylated in a DNMT3 or TET knockout, or not. The backbone is fine-tuned end to end rather than frozen, so the resulting checkpoint is a sequence model specialized to one enzymatic context. Nothing about the cytosine's own methylation is supplied at inference time.

The work was posted to bioRxiv in May 2024 and published in Briefings in Bioinformatics in March 2025, where the model acquired its name. It belongs to the same family as DNABERT-Epi and DNABERT-Enhancer — a single-task classifier built on a released genomic language model — and the authors distribute one trained checkpoint per knockout dataset rather than one general predictor.

#Key Features

• Sequence-only inference: A prediction needs the cytosine's coordinates and a FASTA assembly, nothing else, which is what makes the model usable on positions that had too little read coverage to be called experimentally.
• Seven released checkpoints: One per knockout dataset — four human embryonic stem cell lines and three mouse lines — so a user selects the enzymatic context matching their experiment instead of fine-tuning their own.
• Attention-derived motif discovery: High-attention windows are extracted through DNABERT's motif tool, passed to STREME for enrichment, and matched against JASPAR, turning the classifier into an instrument for finding sequence determinants of enzyme activity.
• Transfer experiments across species and enzymes: The authors train on one knockout and test on another, and train on human and test on mouse, reporting these explicitly as transfer results rather than as the model's intended operating mode.
• Optional methylation-conditioned variant: A second architecture concatenates the methylation levels of neighbouring cytosines onto the sequence embedding before the head.

#Technical Details

The backbone is the 6-mer DNABERT checkpoint, chosen after a comparison in which it tied with Nucleotide Transformer on the TET knockout data and won on the availability of a motif-finding module. Its pooled output feeds three linear layers of 128, 24, and 2 units with 0.5 dropout and ReLU activations, closing on a softmax. Training uses Adam at a learning rate of 1e-5 for five epochs, roughly 40 minutes per epoch on an NVIDIA RTX 3090 and about 24 GB of memory. A 512 bp window is DNABERT's maximum input and also the best-performing choice among the sizes the authors tested.

Labels come from three published bisulfite studies: HUES8 human embryonic stem cells carrying DNMT3, TET, quadruple and quintuple knockouts; mouse embryonic stem cells with DNMT3A and DNMT3B knockouts; and mouse intestinal stem cells lacking TET2 and TET3. A cytosine counts as differentially methylated when its level shifts by at least 0.6 between wild-type and knockout, with at least 10-fold coverage in both. Each dataset contributes a balanced 90,000-cytosine training set and a 10,000-cytosine test set held at least 256 bp away from any training window. Accuracy peaks on the human triple-TET knockout at 0.89 AUC and 0.79 accuracy, and on the mouse TET2/TET3 knockout at 0.81 AUC and 0.73 accuracy; the multi-enzyme quadruple and quintuple knockouts score lower. Fine-tuning the backbone beat random forest and SVM classifiers trained on frozen DNABERT embeddings.

#Applications

The stated use is imputation: filling in differential-methylation calls for cytosines a bisulfite experiment covered too thinly to score, so they can re-enter a downstream analysis. The second use is interpretive — the motifs recovered from attention all matched C2H2 zinc finger factors, including PRDM9, ZBTB14, GLIS2 and KLF17 binding sites, and CTCF recurred most often among the top TET-associated hits, nominating transcription factors that may partner with these enzymes in pluripotent cells.

#Impact

This is the first published attempt at predicting which cytosines DNMT3 and TET act on from sequence context alone, and its clearest finding is asymmetric: TET activity is more sequence-predictable than DNMT3 activity in both species, while cross-species TET transfer degrades sharply. The scope is narrow and the authors say so — seven knockout datasets in two species and two cell types, and supplying neighbouring methylation levels helped in only one of seven cases. The repository provides the training, prediction, and motif-finding scripts, and the seven checkpoints are distributed through a Google Drive folder because they exceed GitHub's file size limit; no license file accompanies either.