Research map / Label-efficient and robust learning / Recognition and long-tail
Hyperbolic Genome Embeddings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies fully hyperbolic CNNs (Lorentz model) to DNA sequence classification, introduces the Transposable Elements Benchmark, and analyzes dataset hyperbolicity via delta-hyperbolicity distributions.
Problem
Genomic sequence models use Euclidean inductive biases that poorly match the hierarchical, evolutionarily structured nature of DNA, and existing hyperbolic genome methods require explicit phylogenetic information.
Approach
Train small fully hyperbolic CNNs (Lorentz model, learnable curvature, HCNN-S and HCNN-M variants) end-to-end on one-hot DNA sequences without pretraining or explicit phylogenetic mapping, paired with a distributional delta-hyperbolicity analysis of dataset embeddings.
Key results
- HCNNs outperform the Euclidean CNN on 37 of 42 genome interpretation datasets by MCC, with 29 statistically significant improvements.
- A HCNN on GUE beats DNA language models on seven of 28 datasets while using 4.6M parameters versus 87M to 2.5B for the LMs.
- Scaled HCNNs (about 43M parameters) reach SOTA on 12 GUE datasets, versus 11 for DNABERT-2 and 5 for NT-2500M-multi.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hyperbolic sequence embeddings with a single curvature instead of Euclidean embeddings | CNN for genomic classification · 42 genome interpretation datasets | MCC | Helped |
| Use hyperbolic sequence embeddings with a single curvature instead of Euclidean embeddings | CNN · TEB — LTR Copia | MCC: 54.73 ±1.45 -> 64.58 ±3.07 | Helped |
| Use hyperbolic sequence embeddings with multiple curvatures instead of Euclidean embeddings | CNN · TEB — LINEs | MCC: 70.63 ±1.24 -> 77.10 ±2.92 | Helped |
| Use hyperbolic sequence embeddings with multiple curvatures instead of Euclidean embeddings | CNN · TEB — SINEs | MCC: 85.15 ±1.64 -> 81.85 ±2.95 | No clear effect |
| Use hyperbolic sequence embeddings with a single curvature instead of Euclidean embeddings | CNN · TEB — CMC-EnSpm | MCC: 72.18 ±0.32 -> 80.98 ±1.48 | Helped |
| Use hyperbolic sequence embeddings with multiple curvatures instead of Euclidean embeddings | CNN · TEB — hAT-Ac | MCC: 87.45 ±0.90 -> 91.04 ±1.58 | Helped |
| Use hyperbolic sequence embeddings with a single curvature instead of Euclidean embeddings | CNN · TEB — processed pseudogenes | MCC: 60.66 ±0.82 -> 68.30 ±0.93 | Helped |
| Use hyperbolic sequence embeddings with multiple curvatures instead of Euclidean embeddings | CNN · Synthetic Scenario C — real sequences | MCC: 66.22 ±0.44 -> 69.30 ±2.34 | Helped |
| Use hyperbolic sequence embeddings with a single curvature instead of Euclidean embeddings | CNN · Synthetic Scenario C — real sequences | MCC: 66.22 ±0.44 -> 73.62 ±0.62 | Helped |
| Use single curvature rather than multiple curvatures (ablation) | HCNN · TEB — SINEs | MCC: 81.85 ±2.95 -> 85.45 ±1.16 | Helped |
| Use multiple curvatures rather than a single curvature (ablation) | HCNN · TEB — LTR Copia | MCC: 64.58 ±3.07 -> 68.05 ±2.80 | Helped |
| Replace a Euclidean classifier head with Lorentzian MLR (ablation) | Euclidean CNN · TEB | MCC | Helped |
Surprising findings
- Hyperbolic models show no benefit on promoter detection tasks and are much worse on the Covid variant task (HCNN-M MCC/F1 14.81 vs CNN 66.43).
- Weak anticorrelation between delta_worst and relative performance gains (r = -0.35 for HCNN-S, -0.21 for HCNN-M).
Limitations
- Hyperbolic advantage is absent or reversed on some tasks such as promoter detection and Covid variant classification.
- Delta-hyperbolicity correlates only weakly with performance gains, so it is an imperfect predictor.
- Comparisons with large DNA LMs use reported or partially re-run results, and HCNNs are trained from scratch without pretraining.
Experimental setup
Datasets: Transposable Elements Benchmark GUE Genomic Benchmarks synthetic phylogenetic sequences (pyvolve GTR)
Benchmarks: Transposable Elements Benchmark GUE Genomic Benchmarks
Metrics: MCC F1-score delta-hyperbolicity
Baselines: Euclidean CNN HyenaDNA Caduceus-Ph DNABERT DNABERT-2 Nucleotide Transformer
Builds on: fully hyperbolic neural networks (HCNN) Riemannian Adam GUE benchmark Genomic Benchmarks Gromov delta-hyperbolicity
Closest papers by meaning
- What DNA Foundation Models Learn Beyond Sequence Composition (NeurIPS 2026 · similarity 0.87)
- DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genomes (ICLR 2024 · similarity 0.87)
- Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.86)
- VQDNA: Unleashing the Power of Vector Quantization for Multi-Species Genomic Sequence Modeling (ICML 2024 · similarity 0.86)
- BEND: Benchmarking DNA Language Models on Biologically Meaningful Tasks (ICLR 2024 · similarity 0.85)
- DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA (NeurIPS 2024 · similarity 0.85)
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.