SeqTagger

A nanopore flowcell is an expensive way to sequence one sample. Oxford Nanopore sells barcoding kits for DNA and cDNA, but for direct RNA sequencing (DRS) — the mode that reads native transcripts with their modifications and poly(A) tails intact — no commercial kit exists. The community's answer, DeePlexiCon, turned raw ionic-current traces into 2D images and classified them with a convolutional network. It worked, but it capped out at four barcodes, would not run on GPUs newer than 2019, did not support the RNA004 chemistry, and consumed half or more of a DRS pipeline's preprocessing compute.

SeqTagger changes what the network predicts. A barcode is not an abstract class here — it is a stretch of DNA inside the reverse-transcription adapter ligated to every molecule during library prep. So instead of classifying a picture of the signal, SeqTagger basecalls that DNA into letters, aligns the result against a FASTA of expected barcodes with minimap2, and drops reads whose median base quality falls below a cutoff. Because the output space is nucleotides rather than an N-way label, going from four barcodes to ninety-six changes the reference file more than the model.

An off-the-shelf DNA basecaller cannot read this DNA: RNA threads through the pore 3'→5' and its helicase moves at 70–130 bases per second against 400–450 for DNA, so the same barcode gives a different trace under the two kits. SeqTagger comes from Leszek Pryszcz, Gregor Diensthuber and Eva Maria Novoa's group at the Centre for Genomic Regulation with Universitat Pompeu Fabra, published in Genome Research in 2025.

#Key Features

• Barcode read as sequence: The network emits nucleotides over an N/A/C/G/T alphabet and barcode identity is settled downstream by alignment, letting one architecture cover barcode sets of very different size.
• Both chemistries, both devices: Checkpoints cover the current SQK-RNA004 kit and the deprecated SQK-RNA002/R9.4, on MinION and PromethION, reading pod5 or fast5.
• 96-way multiplexing: A 24-fold rise in barcode count cost little accuracy — 98.8% precision on validation — and a dropout test omitting three barcodes showed cross-contamination below 0.00002% of reads per barcode.
• Libraries without poly(A) tails: v1 finds the barcode from the poly(A) signal; v2 instead basecalls each read across a rolling window and keeps the highest-quality one, making Nano-tRNAseq libraries tractable at 98.3% precision.
• Filtering on base quality: The quality cutoff removed 96–97% of misassigned barcodes while retaining 93–95% of reads.

#Technical Details

Every checkpoint is a CTC-CRF basecaller trained with bonito v0.7.2 from Oxford Nanopore's fast configuration: three convolutional layers feeding five stacked LSTM layers of 96 features, then a linear conditional-random-field head over the four bases plus a blank, with a state length of 4 — 519,880 parameters. Larger 384- and 768-feature variants (6.5M and 24.8M parameters) gave only marginal gains, so the smallest shipped. Training used a 31-sample window at stride 10, with 120,000 reads per barcode for the four-barcode models and 40,000 for the ninety-six-barcode ones. On an independent RNA002 test set SeqTagger reached 99.0% precision at 94.9% recall, against DeePlexiCon's 92.8%/88.5% in high-recovery mode and 96.9%/74.0% in high-accuracy mode, and ran roughly 9× faster; on RNA004 it exceeded 99% precision at 97% recall, and a million reads demultiplex in about 5 minutes from a local SSD. The zoo is indexed by chemistry, barcode count and RNA biotype rather than by one flagship checkpoint: four RNA004 models and three RNA002 models, mutually back-compatible from SeqTagger v2 onward.

#Applications

The immediate use is cost: pooling DRS libraries on one flowcell brings low-input samples and pilot experiments within reach of a technology that otherwise charges a flowcell per condition. Because DRS is chosen precisely when native RNA modifications, poly(A) tail lengths or full-length isoforms matter, the demultiplexer must leave those readouts undisturbed — the authors recommend b13_RNA004_mRNA over the 96-barcode model when tail-length estimation matters, since some 37-nt barcodes interfere with those calls. Deployment is a Docker image needing an NVIDIA GPU with float16 support, or a preprocessing option inside the lab's MasterOfPores NextFlow workflow.

#Impact

By replacing signal-to-image classification with direct basecalling, SeqTagger lifted both the accuracy ceiling and the compute bottleneck that had kept DRS effectively single-sample, and it did so for the chemistry the field was migrating to. Its backbone is the same bonito CTC-CRF lineage that XNA Basecaller retargets to non-canonical bases. Two constraints shape reuse. Code, checkpoints and analysis notebooks are released under CC BY-NC-ND 4.0 — non-commercial, no derivatives — with the vendored bonito module retaining Oxford Nanopore's own licence, and the authors have filed European patent applications (EP24382340 and EP24383144) on the method. And each barcode set ships as its own checkpoint rather than a swappable reference, so a custom design means training a model; the raw signal behind the published ones sits at the European Nucleotide Archive under PRJEB78482.