Explicit dynamic cross-strand interactions for DNA sequence language modelling

2026-06-04 · Source: Nature Machine Intelligence · Field: Science & Research — Life Sciences & Biology, Mathematics & Computational Sciences, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

CrossDNA is a novel DNA sequence language model designed for explicit and dynamic cross-strand interactions, addressing limitations of existing models that implicitly approximate double-strand relationships. It employs dual branches with alternating exposure to forward and reverse-complement DNA segments, coupled with a lightweight cross-strand communication module. To handle long contexts, CrossDNA integrates a recurrent long-context backbone and sliding-window attention. The model demonstrates consistent performance gains and enhanced robustness to sequence orientation across diverse genomics tasks, notably in enhancer prediction. With only million-scale parameters, CrossDNA achieves performance comparable to or superior to DNA foundation models containing hundreds of millions of parameters, showcasing significant parameter efficiency. This advancement enables improved regulatory logic interpretation and discovery of novel regulatory elements.

Key takeaway

For AI Scientists and Machine Learning Engineers developing genomic models, you should consider integrating explicit double-strand interaction mechanisms like CrossDNA's to improve model robustness and parameter efficiency. This approach offers superior performance in tasks such as enhancer prediction and variant prioritization, potentially accelerating the discovery of novel regulatory elements and enhancing biological interpretability. Evaluate CrossDNA's open-source implementation for your long-context DNA sequencing projects.

Key insights

Explicitly modeling DNA's double-strand interactions improves genomic sequence language model performance and efficiency.

Principles

• DNA's double-strand nature is crucial for genomic function.
• Explicit cross-strand modeling enhances context capture.
• Parameter efficiency is achievable with targeted architectural design.

Method

CrossDNA uses dual branches for forward/reverse-complement segments, a cross-strand communication module, and a recurrent long-context backbone with sliding-window attention for long sequences.

In practice

• Apply CrossDNA for robust enhancer activity prediction.
• Use for identifying disease-associated non-coding variants.
• Explore for interpreting complex regulatory logic.

Topics

• DNA Language Models
• Genomic Sequence Analysis
• Cross-strand Interactions
• Enhancer Prediction
• Parameter Efficiency
• Non-coding Variants

Code references

• ML-Bioinfo-CEITEC/genomic_benchmarks
• HazyResearch/hyena-dna
• kuleshov-group/caduceus
• FunctionLab/sei-framework
• AIRI-Institute/GENA_LM

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Carbon, open source DNA model, 250x faster than Evo2-7B and runs on llama.cpp — Applying LLM architectures to DNA sequences enables highly efficient genomic analysis and structure prediction.
• Prediction and functional interpretation of inter-chromosomal genome architecture from DNA sequence with TwinC — TwinC predicts inter-chromosomal DNA contacts, revealing their mechanistic drivers and role in gene regulation.
• Deep learning of 777 K bulk transcriptomes reveals human–mouse gene conservation beyond DNA sequence similarity — Gene expression patterns, not just DNA, reveal critical human-mouse gene conservation differences.
• [R] Genomic Large Language Models — Genomic LLMs can identify functional similarities in DNA beyond sequence alignment, indicating learned regulatory patterns.
• Finding the molecular switches behind new infectious diseases — Co-Scientist accelerates identification of molecular disease targets by generating and refining hypotheses from diverse research data.
• 🔬ESM: The Bitter Lesson is Coming for Proteins - Alex Rives, BioHub — Scaling protein language models with vast, diverse data reveals emergent biological capabilities and enables programmable design.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.