Towards a foundation model for raw nanopore signal through discrete tokenization

Nanopore sequencing deciphers nucleic acids by detecting alterations in ionic current that encode information beyond the base sequence. While self-supervised pretrained models have flourished in genomic sequence modeling and single-cell analysis, their application to raw nanopore signals, which are continuous and unstructured, remains relatively unexplored.

This research introduces Brieno, a methodology that fuses reconstruction-based tokenization with residual vector quantization and a Transformer to extract contextual representation.

The researchers trained the Transformer on masked-token prediction using raw nanopore signals from 118 million reads. With frozen or fine-tuned features, the model can identify species without relying on reference genomes. Brieno augments the Dorado tool, accurately classifying reads that remain uncategorized in native-barcode demultiplexing. Its performance surpasses Dorado's basecalling accuracy across six genomes, and outperforms both a dedicated 5-methylcytosine caller and a specific 6-methyladenine caller.

Furthermore, Brieno excels in precise CpG methylation detection even with minimal labeled data. Performance improvements across various tasks, when compared to non-pretrained baselines, show that the pretrained contextual representations prove advantageous. Moreover, these representations demonstrate a high degree of generalizability across different sequencing runs. Overall, these findings underscore the potential of self-supervised pretraining in crafting transferable representations for raw nanopore signals.