Research map / Probabilistic and scientific modeling / Molecules and proteins
dnaHNet: A Scalable and Hierarchical Foundation Model for Genomic Sequence Learning
Arnav Shah, Junzhe Li, Parsa Idehpour, Adibvafa Fallahpour, Brandon Wang, Sukjun Hwang, BO WANG, Patrick D Hsu, Hani Goodarzi, Albert Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hierarchical dynamic chunking (dnaHNet) instead of StripedHyena2 | genomic autoregressive model · GTDB perplexity scaling | scaling exponent alpha: 0.04 -> 0.06 | Improved |
| use dnaHNet instead of StripedHyena2 at matched compute | genomic model · evaluation perplexity | perplexity gap: 0.078 -> 0.118 points as compute grows | Improved |
| use dnaHNet 2-stage (218M) instead of SH2 (166M) | inference · 1M nucleotides | FLOPs: 3.89x fewer FLOPs | Improved |
| use dnaHNet-XL instead of StripedHyena2-234M | inference · 262K nucleotides | latency: 2,352 ms -> 410 ms | Improved |
| use dnaHNet instead of StripedHyena2 compute-matched | zero-shot VEP · MaveDB E. coli | Spearman correlation | Improved |
| use dnaHNet instead of StripedHyena2 | zero-shot gene essentiality · DEG | AUROC | Improved |
| codon-aligned (3,2) compression instead of (2,2) (ablation) | dnaHNet · DEG | AUROC | Improved |
| increase compression to (3,3), 9x (ablation) | dnaHNet · MaveDB VEP | Spearman correlation | Hurt |
| use (2,2) instead of (3,2) compression (ablation) | dnaHNet · MaveDB VEP | Spearman correlation | No clear effect |
| go beyond two recursive stages (ablation) | dnaHNet · hierarchy depth | stability and returns | No clear effect |
| raise compression regularization coefficient alpha to >=0.03 (ablation) | dnaHNet · next nucleotide prediction | faithfulness to compression target | Hurt |
| allocate ~30% of parameters to encoder-decoder instead of 15% (ablation) | dnaHNet · genomic modeling | model quality | Improved |
Closest papers by meaning
- GeneZip: Region-Aware Compression for Long Context DNA Modeling (NeurIPS 2026 · similarity 0.89)
- BEND: Benchmarking DNA Language Models on Biologically Meaningful Tasks (ICLR 2024 · similarity 0.87)
- DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genomes (ICLR 2024 · similarity 0.87)
- Model Decides How to Tokenize: Adaptive DNA Sequence Tokenization with MxDNA (NeurIPS 2024 · similarity 0.87)
- Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis (NeurIPS 2025 · similarity 0.86)
- VQDNA: Unleashing the Power of Vector Quantization for Multi-Species Genomic Sequence Modeling (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
