Towards a foundation model for raw nanopore signal through discrete tokenization

Brieno, a Transformer pretrained on raw nanopore signals from 118 million reads using residual vector quantization, learns contextual representations that enable species identification without reference genomes.

• Why it matters: Raw nanopore signals are continuous and lack a native token vocabulary, leaving self-supervised pretrained models underexplored for tasks like classification and methylation calling.
• What they did: The method tokenizes signals via reconstruction-based residual vector quantization and trains with masked-token prediction, then evaluates frozen or fine-tuned features on species ID, demultiplexing, and methylation tasks.
• The result: Brieno identifies species without reference genomes, classifies reads left unclassified by native-barcode demultiplexing, approaches Dorado across six genomes, exceeds Dorado’s 5-methylcytosine model and a 6-methyladenine caller, and improves CpG methylation detection with limited labels.

AI summary of the abstract · may contain errors · not medical advice · read the paper.