Pretrained AI Classifies Nanopore Reads Without References

The preprint model, still unreviewed, could reduce reliance on reference genomes for sequencing research.

Updated on Nov. 6, 2026 in Biotech

Pretrained AI Classifies Nanopore Reads Without References

Researchers have presented Brieno, a self-supervised pretrained model for raw nanopore sequencing signals, described in a preprint published on bioRxiv on November 6, 2026. The work is at the preprint stage and has not been peer reviewed or released as a shipped product.

Why it matters

Raw nanopore signals are continuous and lack a native token vocabulary, which has left pretrained models underexplored for this data type. Brieno's results suggest pretraining on raw signals can transfer across classification and methylation tasks, a direction that could reduce dependence on reference genomes and labeled data.

The evidence comes from a bioRxiv preprint benchmarking a model pretrained on 118 million raw nanopore reads, which approaches Dorado's basecalling accuracy across six genomes and outperforms Dorado's 5-methylcytosine model and a dedicated 6-methyladenine caller. The authors do not fully characterize performance across every genome or task, and the results remain unreviewed.

The details

Brieno pairs a reconstruction-based tokenizer, which uses residual vector quantization to convert continuous signal into discrete units, with a Transformer, the neural architecture behind large language models, that learns contextual relationships across those units. The Transformer is pretrained with masked-token prediction, in which parts of the signal are hidden and the model learns to predict them from surrounding context. The resulting representations, used frozen or fine-tuned, support species identification without reference genomes and detection of CpG methylation with limited labeled data.

Timeline

  1. November 6, 2026: Preprint describing Brieno posted to bioRxiv

The Tech Race

Brieno is positioned alongside Dorado, the established basecaller for nanopore data, rather than as a replacement for it. Its results suggest a pretrained-model approach may complement incumbent tools across basecalling and methylation tasks.

Brieno is a preprint-stage research model and is not available to users today, and no release date or pricing has been announced. Its near-term relevance is to sequencing researchers who work with nanopore data and may eventually adopt pretrained models for classification and methylation analysis.

The takeaway

Pretrained representations of raw sequencing signals may reduce reliance on reference genomes and labeled data if independent groups reproduce these benchmark gains. Watch for peer-reviewed publication of the Brieno work and for independent evaluations comparing it with Dorado on additional genomes.

Further reading

Explore more developments in Biotech as researchers apply machine learning to sequencing data.