Brieno turns raw nanopore current into discrete tokens, then pretrains a Transformer that reads methylation with few labels

Raw nanopore current can now be discretized and self-supervised pretrained much as speech is, yielding representations reused across tasks instead of one model per task. Modification calling, barcode demultiplexing and species classification previously relied on task-specific models or reference-based mapping, with little representation shared across tasks on the same raw signal. On two held-out ZymoBIOMICS mock-community runs, fine-tuned Brieno reached read-level species accuracy of 0.9887 and 0.9965 and frozen representations 0.9728 and 0.9916, both above the reference-based RawHash2 at 0.9564 and 0.9891.

The recipe has two steps: turn the current signal into discrete tokens from a finite vocabulary, then let a Transformer learn to fill in masked tokens from context. Residual vector quantization approximates a continuous vector by successively subtracting the nearest codeword across several codebooks; here each frame covers eight samples and carries four tokens. Pretraining used 16 public datasets totaling 118 million reads from ONT R9.4.1 flow cells, spanning bacteria, fungi, plants, humans, synthetic oligonucleotides and two mock communities across three sequencing platforms and five library kits.

Pretraining itself contributes a measurable gain, not just a different architecture. Against a Transformer of identical architecture but random initialization, the pretrained model scores higher on methylation detection. Under the same training data and optimization step budget, fine-tuned Brieno reached a methylation-detection AUROC of 0.978 against 0.921 for the randomly initialized control.

The demonstrated scope is public reads from ONT R9.4.1 flow cells, evaluated on species classification, barcode demultiplexing, basecalling and 5mC and 6mA detection, all on held-out runs or flow cells; R10.4.1 data received only an initial species-classification adaptation.

On basecalling, fine-tuned Brieno only approaches Dorado super accuracy, at 0.9546 versus 0.9557 on lambda and 0.9446 versus 0.9456 on E. coli, with larger gaps on K. pneumoniae and C. reinhardtii; in barcode demultiplexing Brieno's advantage is coverage rather than per-read precision, since at matched coverage Dorado is more precise on the reads it calls, 0.9997 against 0.9982; unseen-species detection reached an overall AUROC of 0.72, ranging from 0.54 for S. enterica to 0.88 for C. neoformans; and 6mA cross-species transfer was shown only within the GATC sequence context.