DNA Model Learns From Genetic Variants and Methylation

Researchers' ModBERT model is trained on long-read genomes from 242 individuals rather than reference genomes alone.

Updated on Nov. 13, 2026 in Biotech

DNA Model Learns From Genetic Variants and Methylation

Researchers have developed ModBERT, a DNA language model pretrained on genetic variants and DNA methylation, rather than on reference genomes alone. The preprint was posted to bioRxiv on November 12, 2026.

Why it matters

Most DNA language models are pretrained almost exclusively on reference genomes, which cannot represent genetic variants or methylation. Modeling both could help prioritize candidate variants for experimental validation.

The model reaches an AUROC of 0.88 in predicting chromatin accessibility at held-out loci in held-out individuals, and its predicted variant effects agree with likely causal caQTLs at a Spearman correlation of 0.52. Results beyond these held-out tests have not been reported.

The details

ModBERT represents each genome position as a probability distribution over A, C, G, T and 5mC, allowing one individual's genotype and methylation to be encoded together. It uses a soft-target masked language modeling objective, outperforms models trained ab initio and models pretrained without methylation, and its variant predictions do not align with variants associated only through linkage.

Timeline

  1. - November 12, 2026: Preprint posted to bioRxiv (DOI 10.64898/2026.11.12.766258).

The Tech Race

ModBERT extends reference-genome DNA language models by drawing its pretraining data from the 1000 Genomes Long-Read Sequencing Consortium's population-scale cohort. This positions it as a step beyond earlier models that could not represent genetic variants or methylation.

For researchers, the model offers a way to prioritize candidate variants for experimental validation rather than testing them all. Its gains depend on population-scale long-read cohorts, which are still growing.

The takeaway

Models that account for both genetic variation and methylation may sharpen which DNA variants scientists test first. Their accuracy will depend on the size and diversity of the long-read data used to train them.

Further reading

Explore more Biotech coverage for context on genomic modeling advances.

Source note: This article includes information reported by Biorxiv.