Brieno Model Decodes Raw Nanopore Signals
A self-supervised Transformer pretrained on raw nanopore sequencing signals can classify species and detect DNA methylation without reference genomes.
Updated on Nov. 6, 2026 in Biotech
Researchers introduced Brieno, a self-supervised pretrained model that learns from raw nanopore signals using masked-token prediction on data from 118 million reads. The model is described in a preprint published on bioRxiv on November 6, 2026.
Why it matters
Raw nanopore signals are continuous and lack a native token vocabulary, which has left pretrained models underexplored for this type of sequencing data. Brieno suggests that contextual representations learned directly from these signals can improve downstream genomic tasks.
The preprint reports pretraining on 118 million raw nanopore reads, with basecalling performance approaching Dorado across six genomes and exceeding both Dorado's 5-methylcytosine model and a dedicated 6-methyladenine caller. The extraction does not indicate how the model performs on genomes beyond those six, or whether results have been independently reproduced.
The players
Brieno
Brieno is a self-supervised pretrained model for raw nanopore sequencing signals that pairs residual vector quantization tokenization with a Transformer.
Dorado
Dorado is the basecaller developed for Oxford Nanopore sequencing data, against which Brieno's basecalling and methylation performance is benchmarked.
The details
Brieno combines reconstruction-based tokenization through residual vector quantization with a Transformer that learns contextual representations. With frozen or fine-tuned features, the model identifies species without reference genomes and classifies reads left unclassified during native-barcode demultiplexing in Dorado, while also supporting CpG methylation detection with limited labeled data.
Timeline
• November 6, 2026: Preprint describing Brieno published on bioRxiv.
Big Picture
Brieno is positioned as a complement to Oxford Nanopore's Dorado rather than a full replacement, outperforming Dorado's 5-methylcytosine model while approaching its basecalling accuracy. The approach reflects a broader move toward pretrained representation learning applied directly to raw sequencing signals.
For researchers running nanopore sequencing, Brieno offers a way to classify unassigned reads and detect methylation without building a separate specialized caller for each task. Labs with limited labeled data may find the methylation detection approach practical for CpG studies.
The takeaway
Pretrained models built on raw instrument signals may reduce the reliance on reference genomes and task-specific labeled datasets in sequencing research. Researchers evaluating nanopore workflows can watch for independent replications of Brieno's benchmark results.
Further reading
Explore more in Biotech coverage of sequencing and genomic tools.
