Learning Strand-Equivariant Tokenizations of DNA with MirrorDNA

MirrorDNA’s learned tokenizer enforces mirror-symmetric token boundaries across reverse complements, and the full 2.3M-parameter model preserves strand-consistent predictions when labels are strand-independent.

• Why it matters: Existing DNA language models tokenize one strand only, so reverse-complement inputs can yield inconsistent motif splits and prediction changes, undermining strand-invariant regulatory label tasks.
• What they did: The authors quantify that BPE and k-mer tokenizers split 92–100% of transcription-factor motifs differently, and NT-v2, DNABERT-2, and HyenaDNA change 25–31% of predictions under reverse-complement testing; they then introduce MirrorDNA with a tokenizer symmetry guarantee and prove it holds through the whole model.
• The result: Counting only cases correct on both strands, MirrorDNA more than doubles MCC of NT-v2, DNABERT-2, and HyenaDNA, matches Caduceus accuracy with 2.3M parameters, stays accurate on 16× longer inputs, and learns splice-donor signal AGGT as a single token.

AI summary of the abstract · may contain errors · not medical advice · read the paper.