Genetics

LDARNet: DNA Adaptive Representation Network with Learnable Tokenization for Genomic Modeling

Daria Ledneva, Denis Kuznetsov

Featured June 7, 2026

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Replacing rigid, fixed-length DNA segments with adaptive boundaries allows the model to automatically discover and focus on important biological patterns, leading to better performance with much less computing power.

In depth

The authors propose LDARNet, a hierarchical genomic foundation model that replaces fixed tokenization with dynamic chunking. By adapting H-Net's architecture to a bidirectional masked language modeling objective, the model learns to place sequence boundaries at biologically relevant landmarks, enabling a 120M-parameter model to outperform significantly larger baselines on epigenetic tasks.

Key Takeaways

1
Dynamic chunking allows the model to learn biologically meaningful sequence boundaries without supervision, aligning with promoter motifs and splice junctions.
2
The BiMamba-2 backbone enables efficient, long-range context modeling, allowing a compact 120M-parameter model to surpass 2.5B-parameter competitors on histone modification tasks.
3
A ratio-based regularizer stabilizes the compression process, ensuring the model maintains a consistent hierarchical structure during training.

Conceptual Flow

HIGH LEVEL

Methodology

The model learns to break DNA into smart, variable-sized pieces instead of using fixed-size blocks.

Raw DNA

Learn Boundaries

Compressed Representation

Results

The new method identifies important biological landmarks, leading to higher accuracy on complex tasks.

Fixed Grid

Learned Boundaries

Compare Accuracy

Higher Performance