Explicit dynamic cross-strand interactions for DNA sequence language modelling

Abstract

Large language models are transforming the functional interpretation of genomic sequences. Most DNA sequence language models either treat DNA as unidirectional text or enforce reverse-complement symmetry through data augmentation or equivariance, yet both approximate double-strand interactions implicitly and statically, limiting the capture of context-dependent cross-strand information exchange during representation learning. In biological systems, double-strand coordination is shaped by physical coupling, functional synergy and information transfer, and is fundamental to genomic function. Here we present CrossDNA, an explicit and dynamic language model for DNA cross-strand modelling. CrossDNA uses dual branches with alternating exposure to forward and reverse-complement sequence segments as a duplex-inspired training strategy, and establishes interstrand communication through a lightweight cross-strand communication module. To support long-context modelling, CrossDNA couples a recurrent long-context backbone with sliding-window attention. Across diverse genomics tasks, CrossDNA delivers consistent performance gains and improves robustness to sequence orientation, particularly in enhancer prediction. CrossDNA uses only million-scale parameters yet achieves comparable or superior performance to DNA foundation models containing hundreds of millions of parameters, demonstrating markedly greater parameter efficiency. These results advance DNA representation from the implicit, static approximation of double-strand relationships to explicit, dynamic cross-strand modelling, enabling regulatory logic interpretation, discovery of novel regulatory elements and prioritization of disease-associated non-coding variants.