Towards a foundation model for raw nanopore signal through discrete tokenization

Abstract

Nanopore sequencing reads nucleic acids as changes in ionic current that carry information beyond the base sequence. Self-supervised pretrained models have advanced genomic sequence modeling and single-cell analysis, but remain underexplored for raw nanopore signals, which are continuous and lack a native token vocabulary. Here we present Brieno, an approach combining reconstruction-based tokenization through residual vector quantization with a Transformer for contextual representation learning. We pretrain the Transformer using masked-token prediction on raw nanopore signals from 118 million reads. With frozen or fine-tuned features, the model identifies species without requiring reference genomes. It complements Dorado by accurately classifying reads left unclassified in native-barcode demultiplexing, approaches Dorado's basecalling performance across six genomes, and exceeds Dorado's specialized 5-methylcytosine model as well as a dedicated 6-methyladenine caller. Brieno also supports accurate CpG methylation detection with limited labeled data. Across tasks, pretrained contextual representations improve downstream performance over non-pretrained baselines and generalize across sequencing runs. Together, these findings support self-supervised pretraining as a basis for transferable representations of raw nanopore signals.