Wreath predicts DNA methylation across more than 90 tissues and cell types

A newly published genomic AI system is designed to answer a question that ordinary DNA sequencing does not resolve: if a particular letter of someone's DNA changes, what might happen to the chemical marks that regulate gene activity in a specific tissue? Researchers reporting in Nature Genetics on November 3 describe Wreath, a model that predicts DNA methylation across roughly 28 million CpG sites, drawing on 296 experimental profiles spanning more than 90 tissues and cell types. Its real contribution is to place the same genetic variant into different biological contexts. A mutation may matter in blood without having the same effect in brain or muscle; a model that averages across them can miss precisely that distinction.[1]

DNA methylation commonly refers to a chemical modification of cytosine at CpG sites, positions where cytosine is followed by guanine. It helps regulate gene activity and is involved in development, cell identity and disease. The paper says earlier sequence-based prediction models covered fewer than 3% of the roughly 28 million CpG positions in the human genome. Wreath draws on genome-wide experimental measurements from different tissues, which first had to be harmonised into a consistent training resource.[1]

The dataset was as important as the neural network

The team assembled genome-wide measurements representing different tissues and developed a set of methylation classes describing characteristic regions and patterns. It then trained a convolutional sequence model, Wreath, to examine a 2,048-base-pair window around a CpG site and predict continuous methylation levels for 296 profiles simultaneously. The neural network was initialised from a model previously trained on chromatin measurements, rather than entirely from scratch. That transfer matters because methylation and chromatin organisation share some sequence determinants. It does not mean the model is observing an individual's live molecular state: it predicts patterns from sequence using learned relationships in the training data.[1]

On a held-out chromosome-based test, Wreath achieved an average Pearson correlation of 0.825 and Spearman rank correlation of 0.640 across the methylation profiles. Those are two different measures of association, not accuracy percentages. Pearson evaluates alignment of numerical values; Spearman evaluates rank ordering. Strong average correspondence can coexist with poor predictions for particular rare sites, unusual cells or variants outside the training distribution. The chromosomal separation is a meaningful design choice because allowing nearby genomic sequences to appear in both training and test partitions can inflate apparent generalisation.[1]

The paper then tested predictions against independent methylation quantitative trait locus data, which connect genetic variants to measured methylation differences. Among variants in the top 20% by model-predicted effect magnitude, Wreath identified the direction of change correctly more than 90% of the time in the reported analyses; the blood result approached 98%. These results concern selected, strongly predicted methylation effects, so they should not be read as accuracy across all variants or as disease-diagnosis performance.[1]

A new way to generate biological hypotheses

The authors also grouped methylation sites into classes and explored their associations with complex traits through genetic analyses. They report that 17 of 23 classes were significantly enriched for 51 of 60 traits tested at a false-discovery-rate threshold below 0.1. Such enrichment highlights regions that may be relevant to disease biology; it does not establish that a particular methylation mark causes disease or that changing it will improve treatment. The model's outputs can help researchers choose variants and tissues for laboratory follow-up, an efficiency gain in the question-selection stage of science rather than a replacement for experiments.[1]

Wreath is not the only attempt to fill methylation gaps. Earlier work on DeepMethylation, published in PLOS Computational Biology, also used machine learning for tissue-specific methylation and variant effects. However, its data, modelling targets and headline metric differ from Wreath's genome-wide continuous prediction and tissue-wide framework. Comparing an AUROC from a classification formulation with Wreath's correlation coefficients would be invalid. The new contribution is best understood through its harmonised data coverage, its tissue-dependent sequence predictions and its tests against measured variant effects, rather than an invented leaderboard of unrelated scores.[1][2]

The released framework gives researchers a way to prioritise variants and tissues for experimental follow-up. Testing those predictions in biological settings poorly represented in training will help establish where it is most useful. A clinical application would additionally need disease-specific validation and a clearly defined decision use.[1]

Report an error or suggest a correction