Flip the strand and DNA language models change their answer; MirrorDNA cuts both strands at mirrored points to give one prediction
Synopsis
DNA language models usually tokenize, train and test on one strand, so a sequence and its reverse complement can get different predictions. MirrorDNA uses a symmetric router that cuts both strands at mirrored positions and proves the whole model is equivariant to the reverse complement. On 12 strand-symmetric Nucleotide Transformer tasks it reaches an MCC of 0.491 on both strands, against 0.152, 0.187 and 0.080 for NT-v2, DNABERT-2 and HyenaDNA.
Interpretation
DNA language models give different answers for the two strands of the same locus, and that inconsistency comes from how the sequence is cut into tokens. BPE and 6-mer tokenizers split 92% to 100% of transcription-factor motifs differently on the two strands, and NT-v2, DNABERT-2 and HyenaDNA change 25% to 31% of their predictions when test sequences are reverse-complemented. On 12 strand-symmetric Nucleotide Transformer tasks, counting only sequences predicted correctly on both strands, these three models score MCC 0.152, 0.187 and 0.080, against 0.51 to 0.54 on the forward strand.
MirrorDNA breaks both strands at mirrored positions, so the same DNA yields chunks that are reverse complements of each other on the two strands. It averages the adjacent-base difference score with its mirrored version to get a symmetric cut score, then reads each strand in its own 5′ to 3′ direction with a shared-weight encoder. Over 2,000 random 1 kb windows the Jaccard index of the two strands' cut sets is 1.00, and 0% of JASPAR motif occurrences are tokenized differently on the two strands.
The symmetry holds through the whole model, so strand-symmetric labels get identical outputs on both strands while strand-dependent labels remain distinguishable. The chunk transformer and decoder are group-averaged, and the backbone carries an invariant head and an orientation-aware head, the first giving the same answer on both strands and the second keeping direction information. On the 12 strand-symmetric NT tasks MirrorDNA scores MCC 0.491 on both strands, above every non-equivariant model (Wilcoxon p<0.001) and not significantly different from the equally equivariant Caduceus-PS-1k at 0.487 (p=0.15).
MirrorDNA, with 2.3M parameters, comes close to models with far more parameters on several benchmarks. MirrorDNA uses the same pretraining data and token budget as the 1 kb Caduceus models, with 2.3M parameters. It scores 0.751 on Genomic Benchmarks, not significantly different from the 98M-parameter NT-v2 at 0.741 (p=0.36), and 0.618 on the NT tasks, not significantly different from the 117M-parameter DNABERT-2 at 0.631 (p=0.42).
Perspective
The demonstrated setting is a 2.3M-parameter model pretrained on 1 kb windows of human DNA, evaluated on 18 Nucleotide Transformer tasks, 9 Genomic Benchmarks, 2 RefSeq gene-strand tasks and DeepSTARR enhancer-activity regression, with results averaged over three seeds.
The authors note that MirrorDNA is small and pretrained on human DNA only, still below NT-v2 on the NT tasks, below single-nucleotide models on Ensembl regulatory elements and slightly below Caduceus on DeepSTARR; scaling, multi-species data, variant effects and long-range tasks remain to be tested. The gene-strand tasks satisfy the condition under which an invariant head cannot read orientation only approximately, not by construction.