Learning Strand-Equivariant Tokenizations of DNA with MirrorDNA
- 1Department of Computer Science and Engineering, Bangladesh University of Engineering and Technology, Dhaka, Bangladesh
Abstract
DNA has two strands, and for most regulatory label prediction tasks a sequence and its reverse complement should get same answer. Yet DNA language models are tokenized, trained and tested on one strand at a time. We show that this makes them inconsistent. BPE and k-mer tokenizers split 92--100\% of transcription-factor motifs differently on the two strands, and NT-v2, DNABERT-2 and HyenaDNA change 25-31\% of their predictions when the test sequences are reverse-complemented. We introduce {\model}, a compact DNA language model with a learned tokenizer that guarantees mirror-symmetric token boundaries across reverse-complement sequences, regardless of its learned parameters. We prove that this symmetry holds through the whole model. As a result, the same model can give identical answers on both strands when the label does not depend on the strand, and can still tell the strands apart when it does, which no strand-blind model can learn. Counting only predictions that are correct on both strands, \model\ more than doubles the MCC of NT-v2, DNABERT-2 and HyenaDNA. With 2.3M parameters, it is more accurate than the Caduceus models trained on the same data, and not significantly different from NT-v2 (98M parameters) on Genomic Benchmarks or from DNABERT-2 (117M) on the Nucleotide Transformer tasks. It also stays accurate on inputs sixteen times longer than those it was trained on. The same model trained with reverse-complement augmentation instead still changes about one prediction in ten with the strand, unless its outputs are averaged over both strands. Without any annotation, the tokenizer learns to keep the splice-donor signal \texttt{AGGT} in a single token. Code: https://github.com/NIRJHOR-DATTA/MirrorDNA.
Related
- OmniScore: Universal Scoring of Diverse Biomolecular Complexes via Equivariant Geometry-Aware Discrete Representation Learning
- Protocol for mapping sister chromatid exchanges and structural variations using sci-L3-strand-seq.
- Single-strand break repair pathway disruption impacts radiosensitivity to both low- and high-LET radiation
- The hidden diversity: Replication-strand asymmetry shapes short-term nucleotide and structural variation in an RNA virus
- The nuclear actin cytoskeleton supports DNA double-strand break repair via VCP-mediated extraction of the KU70/80 complex from damaged chromatin
BibTeX
@misc{datta2026learning,
title = {{Learning Strand-Equivariant Tokenizations of DNA with MirrorDNA}},
author = {N. Datta and S. Shatabda and M. S. Bayzid and M. S. Rahman},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.04.764207},
url = {https://www.biorxiv.org/content/10.64898/2026.11.04.764207v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Learning Strand-Equivariant Tokenizations of DNA with MirrorDNA
AU - N. Datta
AU - S. Shatabda
AU - M. S. Bayzid
AU - M. S. Rahman
PY - 2026
DA - 2026/11/04
PB - bioRxiv
DO - 10.64898/2026.11.04.764207
UR - https://www.biorxiv.org/content/10.64898/2026.11.04.764207v1
AB - DNA has two strands, and for most regulatory label prediction tasks a sequence and its reverse complement should get same answer. Yet DNA language models are tokenized, trained and tested on one strand at a time. We show that this makes them inconsistent. BPE and k-mer tokenizers split 92--100\% of transcription-factor motifs differently on the two strands, and NT-v2, DNABERT-2 and HyenaDNA change 25-31\% of their predictions when the test sequences are reverse-complemented. We introduce {\model}, a compact DNA language model with a learned tokenizer that guarantees mirror-symmetric token boundaries across reverse-complement sequences, regardless of its learned parameters. We prove that this symmetry holds through the whole model. As a result, the same model can give identical answers on both strands when the label does not depend on the strand, and can still tell the strands apart when it does, which no strand-blind model can learn. Counting only predictions that are correct on both strands, \model\ more than doubles the MCC of NT-v2, DNABERT-2 and HyenaDNA. With 2.3M parameters, it is more accurate than the Caduceus models trained on the same data, and not significantly different from NT-v2 (98M parameters) on Genomic Benchmarks or from DNABERT-2 (117M) on the Nucleotide Transformer tasks. It also stays accurate on inputs sixteen times longer than those it was trained on. The same model trained with reverse-complement augmentation instead still changes about one prediction in ten with the strand, unless its outputs are averaged over both strands. Without any annotation, the tokenizer learns to keep the splice-donor signal \texttt{AGGT} in a single token. Code: https://github.com/NIRJHOR-DATTA/MirrorDNA.
ER -