XNA Basecaller

The XNA Basecaller is a nanopore basecalling model that reads non-canonical bases (NCBs, also called unnatural bases) directly from raw electrical signal, alongside the four canonical DNA bases. It was developed by Mauricio Perez, Michiko Kimoto, Priscilla Rajakumar and colleagues at the Genome Institute of Singapore and the Institute of Bioengineering and Bioimaging, with Ichiro Hirao's group and Xenolis, posted as a preprint in December 2024 and published in Nature Communications in July 2025.

Synthetic genetic alphabets built on unnatural base pairs underpin xenobiology, aptamer selection and DNA data storage, but until now there was no way to sequence them at scale: standard basecallers only emit A, T, C and G, so an unnatural base is silently miscalled as whichever canonical base its signal most resembles. The team showed that xeno-nucleic acids containing the Ds–Px unnatural base pair run on an Oxford Nanopore MinION without disrupting throughput — more than 2 million reads per flowcell — and that their raw current traces differ measurably from canonical DNA (median fold-change above 6×). What was missing was a decoder.

Getting one was a training-data problem more than an architecture problem. Nanopore signal depends on roughly six bases in the pore at once, so a model must see each unnatural base in every 6-mer context. The authors synthesized a library of 1,024 oligonucleotides in which an NCB is flanked by matched 5-mers, covering all 6,144 single-NCB 6-mer contexts, then combined bootstrapped relabelling with read-splicing augmentation to stretch it further.

#Key Features

• Six-letter decoding: The basecaller's decoding states are expanded beyond A, T, C, G to emit the non-canonical bases X (Px) and Y (Ds), so an unnatural base is reported as itself rather than substituted.
• Bootstrapped label refinement: Training proceeds in rounds, each using the previous round's model to extract cleaner signal-to-sequence chunks — a way to get accurate ground truth for signal no existing model can read.
• Spliced-read augmentation: Real unnatural-base signal slices, located by a dynamic-time-warping-refined alignment, are grafted into canonical DNA signal chunks from the Bonito training set, generating reads with far greater sequence-context diversity than the physical library alone provides.
• High specificity: Canonical bases are almost never miscalled as unnatural ones (99% specificity); the dominant residual error is a missed NCB, i.e. a deletion, the same failure mode nanopore basecallers show on canonical DNA.
• Partial-freezing domain adaptation: Freezing the first six of nine network layers during fine-tuning raises NCB accuracy while leaving canonical DNA accuracy essentially intact.

#Technical Details

The model uses Bonito v0.5.0 Super Accuracy — Oxford Nanopore's convolutional research basecaller — as its backbone, adapted to a six-letter alphabet, with a DNA-only checkpoint (xna_r9.4.1_e8_sup@v3.3) as the initialization for fine-tuning. Signal chunks of 3,600 samples are SMAD-normalized to match Bonito's preprocessing. Roughly 316,000 reads from the complex library were split 297,000 for training and 19,000 for validation; fine-tuning ran for 5 epochs at a learning rate of 5e-4 with 50% dropout on the top layer and 5% elsewhere. On the held-out proof-of-concept library, single-base models reach 77% accuracy for X and 81% for Y with 91–92% canonical DNA accuracy, while a joint model calling both bases reaches 71%; median per-template accuracy on the complex library is 86% and 82% for the single-base models and 73% for the joint model. Code, the DNA-only baseline checkpoint and preprocessed training data are public under MIT, and the raw nanopore runs are deposited at ENA under PRJEB82716. The trained X, Y and XY checkpoints are not distributed as downloads — the released train_and_eval.sh pipeline reproduces them in a few hours from the provided data.

#Applications

The immediate beneficiaries are groups building or characterizing xeno-nucleic acids: aptamer engineers using genetic-alphabet expansion to raise binding affinity, synthetic biologists incorporating unnatural base pairs into engineered constructs, and DNA data-storage efforts that use extra letters to increase information density. Where quality control previously relied on indirect assays or specialized Sanger chemistry, a MinION run plus this basecaller returns nucleotide-level identity across a whole pool, making it practical to verify NCB incorporation position by position across a library rather than one construct at a time.

#Impact

This is the first demonstration that non-canonical bases can be deconvolved directly and at high throughput from nanopore signal, and the methodological contribution generalizes beyond the specific chemistry: bootstrapped relabelling and signal splicing are strategies for any expanded alphabet where physically synthesizing every sequence context is infeasible, a constraint that tightens rapidly as more letters are added. The limitations are concrete: accuracy for unnatural bases remains well below the ~99% routine for canonical DNA, and the models are tied to R9.4.1 chemistry and the Ds–Px pair, with generalization to other unnatural base pairs or newer flowcells untested. The authors point to duplex basecalling and richer signal measurements as the route to closing the gap.