Methven

A methylation quantitative trait locus records something sequence models have trouble representing: a single-nucleotide substitution that shifts the methylation level of a CpG site tens of kilobases away. CpGenie, the tool built for exactly this prediction, reads a 500 bp window around the CpG — enough when the variant sits on the site itself, and blind to the majority of meQTLs, where SNP and CpG are separated by far more sequence than the convolution can see.

Methven takes a different route. It cuts the locus around the CpG site into 500 bp segments, embeds each segment independently with a frozen DNA language model, and hands the resulting ordered series of segment embeddings to a small recurrent network. Long-range dependency is carried by recurrence over segments rather than by any one filter's receptive field, which keeps the trainable model under a million parameters while spanning 100 kbp either side of the variant. Two copies of the locus go in — the reference sequence and the same sequence carrying the alternate allele — so the network scores a difference rather than an absolute methylation level.

The model came from Guan Ning Lin's group at Shanghai Mental Health Center and Shanghai Jiao Tong University, first as a preprint in September 2024 and then in Advanced Science in January 2025.

#Key Features

• Positional-wise segmentation: The window centred on the CpG is cut into a 501 bp central segment carrying the CpG, uniform 500 bp segments either side and a 250 bp flank at each end — 41 segments for the short-range model and 401 for the long-range one, each short enough for the language model's input limit.
• Paired pre- and post-mutation input: Reference and alternate sequences are embedded through the same backbone and concatenated, so the head sees both states of the locus.
• Chromatin accessibility channel: An ATAC-seq track for the cell type of interest is average-pooled over the same segmentation and concatenated with the two embedding streams, letting one variant be scored under different chromatin states.
• Divide-and-conquer by distance: Separate weights handle SNP–CpG distances below 10 kbp and between 10 and 100 kbp, dispatched automatically from the distance.
• Direction and magnitude as separate tasks: A classification head predicts up- or down-regulation and an independent regression head estimates the meQTL slope, avoiding interference between the two objectives.

#Technical Details

The backbone is DNABERT-2 (117M parameters), frozen and used only to emit a 768-dimensional vector per segment. The trainable module is two stacked bidirectional GRU layers of 64 units followed by batch normalisation and a dense stack, ending in a two-way softmax for classification or a scalar for slope regression: 722,786 parameters for the short-range model and 814,946 for the long-range one. Training data came from 244,491 CD4+ T cell cis-meQTLs in the EPIC meQTL Database with matched ATAC-seq from EpiMap, annotated against hg19 into 50,190 SNP–CpG pairs (19,874 short-range, 30,316 long-range) and class-balanced by down-sampling before an 80/10/10 split. Training ran in TensorFlow on a single NVIDIA 3090. On the held-out test sets, classification reached ACC 0.920 and AUC 0.969 over 1,988 short-range pairs and ACC 0.837 and AUC 0.918 over 3,032 long-range pairs; slope regression reached RMSE 1.59 with PCC 0.87 and RMSE 2.10 with PCC 0.81 respectively. Under ten-fold cross-validation, CpGenie extended to the same distances reached mean ACC 0.782 (short-range) and 0.706 (long-range), and functional annotations from Enformer also trailed the learned embeddings.

#Applications

The practical target is GWAS follow-up: given a non-coding risk variant and matched ATAC-seq, Methven ranks the CpG sites within 100 kbp by how strongly the variant is predicted to move their methylation, narrowing a locus to candidate regulatory sites without new assays. The authors demonstrate this on rheumatoid arthritis, scoring eight risk SNPs against CD4+ T cell ATAC-seq collected unstimulated and after 24 h of anti-CD3/CD28 stimulation and classifying affected CpGs as up-enhanced, down-enhanced, reduced or unaffected by disease state; the rs968567–cg06781209 pair in the FADS2 promoter, an SREBF2 binding site, reproduces a published mechanism.

#Impact

Methven shows that a frozen genomic language model plus a few hundred thousand trainable parameters can cover a range purpose-built convolutional predictors could not, and that pairing sequence with an accessibility track makes the prediction conditional on cellular state rather than fixed per genome. Its scope is correspondingly narrow: the released weights are trained on CD4+ T cell meQTLs, accept only single-nucleotide substitutions on hg19, and require a cell-type ATAC-seq track as input. Transfer to monocytes reached classification AUCs of 0.898 (short-range) and 0.770 (long-range) trained end to end and 0.939 and 0.823 after fine-tuning from the CD4+ T cell weights — evidence the approach moves to a new cell type when matched accessibility data exist, and a reminder that predictions beyond those two cell types are extrapolation.