Pith. sign in

REVIEW 4 major objections 6 minor 78 references

Hyperbolic Genome Embeddings

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Swapping Euclidean geometry for hyperbolic space improves genomic classification on 37 of 42 benchmark datasets.

desk verdict The 37/42 geometry claim is plausible but leans on an undertuned Euclidean baseline; the TEB benchmark and homology-splitting experiment carry the paper. read the letter →

arxiv 2507.21648 v1 pith:CKWA3YIY submitted 2025-07-29 cs.LG

Raiyan R. Khan , Philippe Chlenski , Itsik Pe'er This is my paper · ORCID
classification cs.LG
keywords hyperbolicneuralnetworksgenomesequenceclassificationLorentzmodelDNAlanguagemodelsGromovhyperbolicitytransposableelementsrepresentationlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Genome sequences inherit hierarchical structure from evolution, and this paper argues that genome classifiers should embed sequences in hyperbolic space, where tree-like data can be represented with low distortion, rather than in flat Euclidean space. To test that hypothesis, the authors build fully hyperbolic convolutional neural networks in the Lorentz model and match them layer-for-layer against equivalent Euclidean CNNs. Across 42 real-world genomic classification datasets, the hyperbolic models beat the Euclidean models on 37 tasks as measured by Matthews correlation coefficient, and the best hyperbolic runs outperform several much larger pretrained DNA language models on 7 of 28 Genome Understanding Evaluation datasets while using about 4.6 million parameters and no pretraining. The paper also introduces a transposable element benchmark and a distribution-level Gromov hyperbolicity analysis intended to show that genomic sequence embeddings are genuinely tree-like.

What carries the argument

The central object is the fully hyperbolic CNN built on the Lorentz model of hyperbolic space, the forward sheet of a two-sheeted hyperboloid in Minkowski space with constant negative curvature K. Convolution, batch normalization (via Fréchet mean and variance), residual connections, and multinomial logistic regression are all redefined so points remain on the hyperboloid, and curvature is a learnable parameter: HCNN-S uses one curvature across all layers, while HCNN-M uses a separate curvature per block. This machinery carries the argument because it is the only difference from the Euclidean control, so any performance gap is attributed to geometry. A secondary mechanism is the distribution-level Gromov δ-hyperbolicity analysis, which measures how close the embedding space is to a tree and calibrates against random samples of matched dimensionality.

What would settle it

Retrain the Euclidean CNN and both hyperbolic models on a subset of the TEB and GUE datasets with per-model learning-rate sweeps and matched width and capacity searches. If a tuned Euclidean model closes or reverses the MCC gap on most of these tasks, the hyperbolic advantage claimed in the paper would not survive fair comparison.

Watch

Extended reading notes

Core claim

The paper's central claim is that the inductive bias of negative curvature is a practical asset for DNA sequence classification: replacing the geometry of a compact CNN with Lorentz-hyperbolic geometry yields consistently higher MCC across most of a broad benchmark suite, with the largest gains on tasks where an evolutionary signal must be separated from background sequence. On seven GUE datasets, the best hyperbolic runs exceed the reported scores of DNA language models that have 1.7 to 543 times more parameters and were pretrained on whole genomes, indicating that geometry can substitute for scale on these tasks. The paper attributes this to the latent tree structure of genomic data, and supports the attribution with synthetic phylogeny simulations, a homology-splitting generalization experiment, and an empirical claim that final-layer embeddings of genomic datasets are more hyperbolic than dimensionality-matched Euclidean baselines.

Load-bearing premise

The load-bearing premise is that the Euclidean CNN is an equally well-tuned control, so the measured performance gap is caused by geometry rather than by differences in optimizer or hyperparameter quality.

Editorial extensions

If this is right

  • Geometry can substitute for scale: a 4.6M-parameter hyperbolic CNN with no pretraining matches or beats orders-of-magnitude-larger DNA language models on several GUE tasks.
  • The benefit is concentrated where an evolutionary signal sits in background noise, such as transcription factor binding sites, epigenetic marks, and transposable elements; promoter detection and Covid variant classification show little or no gain.
  • Hyperbolic models generalize to unseen homology branches better than Euclidean CNNs, which suggests the standard benchmarks may understate the geometric advantage.
  • Euclidean embeddings still improve when the classification head is hyperbolic, so the decision boundary, not only the representation, carries part of the benefit.
  • Genomic embedding spaces are more tree-like than equal-dimensional Gaussian baselines on 25 of 43 datasets, supporting the use of hyperbolic architectures when this distribution-level hyperbolicity is present.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A fairer geometry test would sweep learning rates and model widths separately for each optimizer; because the Euclidean control uses AdamW while the hyperbolic models use RiemannianAdam, some of the 37/42 gap could reflect optimizer fit rather than curvature, a possibility the paper does not exclude.
  • The distribution-based δ calibration could be reused as a cheap pre-training diagnostic: compute δ on a small sample of embeddings and choose hyperbolic versus Euclidean architecture only when the matched-dimensionality test shows tree structure.
  • If the geometric advantage is real, the natural next targets are tasks with explicit phylogenetic signal at scale, such as regulatory element annotation across species, variant-effect prediction, and metagenomic binning, none of which the paper tests.
  • The paper's 'emergent hyperbolicity' observation implies published claims that compare δworst across different dimensionalities need re-examination, since high dimension alone lowers δ.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes fully hyperbolic convolutional neural networks (HCNN-S and HCNN-M), built with Lorentz-model convolutions, batch normalization, and multinomial logistic regression, for DNA sequence classification. The authors evaluate these models on 42 distinct real-world genomic classification datasets spanning the Genome Understanding Evaluation (GUE), Genomic Benchmarks (GB), and a newly introduced Transposable Elements Benchmark (TEB), and they report that the hyperbolic models outperform an equivalent Euclidean CNN on 37 of 42 tasks as measured by MCC. They also compare the best HCNN runs against several DNA language models on GUE, report synthetic evolutionary simulations and a homology-splitting experiment, and develop a δ-hyperbolicity analysis of dataset embeddings. The main claims are that the hyperbolic inductive bias is responsible for the observed gains, that HCNNs can exceed DNA LMs on some tasks with far fewer parameters and no pretraining, and that genomic data exhibit measurable hyperbolic structure.

Significance. If the 37/42 result is robust, this is a significant empirical contribution: a compact, untrained CNN in hyperbolic geometry beating much larger pretrained DNA language models on seven GUE datasets is surprising and practically interesting. The paper also ships a new benchmark resource (TEB), reports code and dataset availability, uses five seeds per model, and honestly documents the failures on Covid variant classification and promoter detection. The δ-hyperbolicity analysis, with its dimensionality-matched calibration baselines, is a useful methodological addition even if exploratory. The main caveat is that the central causal claim about geometry is only as strong as the Euclidean control, and the control is not yet demonstrated to be equally well tuned.

major comments (4)
  1. [§5.1, Table 2; Appendix A.2.2, Table 3] The Euclidean control is not shown to be an equally tuned baseline. The CNN uses AdamW while both HCNNs use RiemannianAdam with an additional manifold learning rate of 2e-2, and the text states that hyperparameters are fixed 'when possible' without reporting learning-rate sweeps, per-dataset tuning, capacity matching, or convergence diagnostics. Because the headline 37/42 result, the synthetic-task results, and the homology-split results are all comparisons against this Euclidean control, the observed gains conflate geometry with optimizer choice and tuning quality. This is the load-bearing issue. The authors should add a control study that sweeps learning rates and optimizers for both geometries, reports validation curves, and checks parameter/capacity matching, or otherwise demonstrate that the advantage is not an artifact of an undertuned Euclidean baseline.
  2. [§5.1, Table 2] The statistical significance statement is based on five seeds per model and a Wilcoxon rank-sum test per dataset with no multiple-testing correction. With 42 uncorrected comparisons, the observation that the Euclidean CNN wins on two datasets is roughly what one would expect by chance even if the geometries were equivalent. Please report per-dataset effect sizes and confidence intervals, apply or justify a multiple-testing correction, and state explicitly whether the 37/42 count uses the better of HCNN-S and HCNN-M for each dataset, as the table appears to do.
  3. [§5.1, Table 5; Appendix A.4] The comparison with DNA language models mixes evaluation protocols: HCNN scores are from the authors' runs, and the text says they are 'best' over runs, while several DNA LM scores are taken from Zhou et al. (2024) and HCNN-S/M entries in Table 5 have no standard deviations. The 'seven of 28' claim should be accompanied by an apples-to-apples check on identical train/dev/test splits and metrics, and by reporting the variance of the HCNN scores. Without this, the strong statement about outperforming much larger pretrained models is not fully supported.
  4. [§4, Appendix A.6, Table 1] The synthetic experiments use a single simulation configuration: random ETE trees, GTR model with default parameters, and 10% label noise. There is no sensitivity analysis over tree depth, branch-length distribution, substitution-rate heterogeneity, or noise level, so the Scenario-specific conclusions (for example, that Scenario C gains are unrelated to Scenario A) are not robustly established. Please either add a sensitivity analysis or soften the mechanistic interpretation.
minor comments (6)
  1. [§2.2, Eq. (2)] The text says 'the remaining n−1 elements form the spacelike components,' but the Lorentz model as defined lives in R^(n+1), so the spacelike part has n components; please correct this.
  2. [Abstract and §5.1] The phrase 'orders of magnitude fewer parameters' is inaccurate for Caduceus-Ph (7.7M versus 4.6M, a factor of 1.7); it is only warranted for the largest Nucleotide Transformer models. Please rephrase to be precise.
  3. [Figure 11 caption] In the caption, 'HCCN-S' should be 'HCNN-S'.
  4. [§5.1, Table 2] The counting rule for 42 distinct datasets should be stated explicitly next to the table, including which row is treated as the overlap between GUE and GB and how wins are counted when HCNN-S and HCNN-M disagree.
  5. [Appendix A.4, Table 5] The HCNN parameter count is reported as an average over GUE datasets; please provide a model-card table with per-dataset sequence lengths and parameter counts, since the architecture depends on input length.
  6. [§5.2, Appendix A.9.1, Table 9] The δ-hyperbolicity values are computed from Euclidean CNN embeddings, but the main-text discussion sometimes reads as if it characterizes the datasets themselves; please clarify that these measurements are model-dependent and exploratory.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: benchmark results are externally grounded; self-citations are peripheral, and the Euclidean-control tuning gap is a fairness issue, not a circular step.

full rationale

The central claims—HCNNs outperform Euclidean CNNs on 37/42 datasets and beat large DNA language models on 7/28 GUE sets—are measured against independent external resources: GUE/GB/TEB data, GUE numbers reported by Zhou et al. (2024) for most DNA LMs, and external model families such as HyenaDNA and Caduceus-Ph that the authors benchmarked themselves. The hyperbolic machinery is adopted from Bdeir et al. (2024), an external method paper, and the Lorentz formulas are standard textbook mathematics. The self-citations (Chlenski et al. 2024 for a visualization setup, Chlenski et al. 2025 for the MANIFY sampling package, and Chen et al. 2025 for phylogenetic hyperbolic methods) are peripheral tools or related-work references; none of them supplies a premise that forces the paper's empirical outcome. The delta-hyperbolicity analysis measures embeddings extracted from the authors' trained Euclidean CNNs and compares their delta distributions to simulated manifolds of matched dimensionality; this is an empirical measurement of fitted embeddings, not a prediction that is equivalent to the definition of delta by construction. No equation in the paper is shown to reduce to its own input, and no fitted parameter is renamed as a prediction. The serious weakness in the paper is not circularity but control quality: Appendix A.2.2 states that 'we train the Euclidean CNN using the AdamW optimizer and the HCNNs using RiemannianAdam', with an additional manifold learning rate for the hyperbolic models and hyperparameters held fixed 'when possible'. This means the 37/42 headline may partly reflect optimizer or tuning differences rather than geometry alone, but that is a correctness/fairness concern, not a circular derivation. Accordingly, the circularity score is low.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on empirical benchmarking rather than a derivation, so the ledger is dominated by training choices (learning rates, optimizers, curvature initialization) and domain assumptions about biological signal. No result in the paper is equivalent by construction to its inputs; benchmark scores are measured against external datasets. The learnable curvature K stays near its -1 initialization and is not a fitted quantity that drives the headline results. No new physical or model entities are postulated; the TEB datasets are new resources, not invented entities.

free parameters (5)
  • Learnable manifold curvature K = near -1 (HCNN-S single K; HCNN-M per-block K1..K4)
    Each hyperbolic block learns a curvature (Section 3.2, A.5); fitted during training though it stays close to the -1 initialization.
  • Manifold learning rate = 2e-2
    Additional optimizer hyperparameter used only for the hyperbolic models (Table 3), chosen by hand with no sweep reported.
  • Per-benchmark learning rates = 1e-4 (TEB/GUE), 1e-5 (GB)
    Learning rate is fixed per benchmark (Table 3); no tuning search is reported for either geometry.
  • Synthetic label noise = 10% swapped labels
    A.6: noise is added by randomly swapping 10% of train/validation labels; the level is chosen by hand.
  • GTR model parameters in PYVOLVE = default parameters
    A.6: sequence evolution is simulated with PYVOLVE's GTR using default parameters, which shape the synthetic data but are not reported explicitly.
assumptions (5)
  • domain assumption DNA sequences carry latent hierarchical phylogenetic structure that a low-distortion tree-like geometry can exploit for classification
    Introduction (Section 1): this is the motivating premise for choosing hyperbolic space.
  • ad hoc to paper GTR simulation along random ETE trees with default parameters adequately reproduces the evolutionary signals relevant to genomic classification
    A.6: the three synthetic scenarios and the homology-splitting experiment depend on this simulation setup.
  • domain assumption Delta-hyperbolicity of trained Euclidean CNN embeddings approximates the intrinsic hyperbolicity of the datasets
    A.9.1: delta is computed on final-layer embeddings of Euclidean CNN test sets; disclosed, but the paper's claims are about datasets, not CNN embeddings.
  • domain assumption GERP scores capture evolutionary constraint at individual nucleotide positions
    A.10: the in silico mutagenesis experiment targets the highest-GERP positions as the most selectively constrained.
  • standard math The Lorentz model formulas and Gromov delta-hyperbolicity definitions are standard and correct
    Sections 2.2 and 3.3: the geometric machinery is taken as background without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyperbolic Genome Embeddings." pith.science (2026). https://pith.science/paper/CKWA3YIY

@misc{pith2026250721648,
  author       = {Pith},
  title        = {Pith review of: Hyperbolic Genome Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CKWA3YIY}},
  note         = {Machine review of arXiv:2507.21648}
}
read the original abstract

Current approaches to genomic sequence modeling often struggle to align the inductive biases of machine learning models with the evolutionarily-informed structure of biological systems. To this end, we formulate a novel application of hyperbolic CNNs that exploits this structure, enabling more expressive DNA sequence representations. Our strategy circumvents the need for explicit phylogenetic mapping while discerning key properties of sequences pertaining to core functional and regulatory behavior. Across 37 out of 42 genome interpretation benchmark datasets, our hyperbolic models outperform their Euclidean equivalents. Notably, our approach even surpasses state-of-the-art performance on seven GUE benchmark datasets, consistently outperforming many DNA language models while using orders of magnitude fewer parameters and avoiding pretraining. Our results include a novel set of benchmark datasets--the Transposable Elements Benchmark--which explores a major but understudied component of the genome with deep evolutionary significance. We further motivate our work by exploring how our hyperbolic models recognize genomic signal under various data-generating conditions and by constructing an empirical method for interpreting the hyperbolicity of dataset embeddings. Throughout these assessments, we find persistent evidence highlighting the potential of our hyperbolic framework as a robust paradigm for genome representation learning. Our code and benchmark datasets are available at https://github.com/rrkhan/HGE.

Figures

Figures reproduced from arXiv: 2507.21648 by the authors.

Figure 1
Figure 1. Overview of our HCNNs. Model inputs are sequences with latent phylogenetic structure [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The various plausible evolutionary scenarios informing genomic sequence learning. Leaf [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Decision boundaries learned by 2-dimensional HCNNs (circles) and CNNs (squares) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: The generalized block architecture for the CNNs/HCNNs. On the left, we delineate the [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: On the left, we show the average improvement in performance (MCC) on TEB datasets for [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Average values of K, the curvature parameter in the HCNNs, as they vary across each block of the model. Values are reported for models trained on each of the seven classification tasks in TEB. with default parameters. Four types of fixed-length sequences are generated …
Figure 7
Figure 7. Figure 7: Leaf node sequence classifications (with added noise) in Scenario A for the simulated [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Hamming distance matrix for all leaves in the simulated phylogenetic tree for Scenario A. [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Overview of the homology splitting experiment. A training and validation dataset (left) are [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: Distribution of scaled δ-hyperbolicity values across each genomic dataset. Colors delineate different task categories, while the bottom two rows provide reference distributions for δ values computed from a set of points sampled from the normal distribution on a Euclid…
Figure 11
Figure 11. Figure 11: Correlation between δworst and the performance differential between HCCN-S and CNN models. routliers includes outliers in the Pearson correlation coefficient calculation, while r excludes them (p < 0.05, except for HCNN-M r). dataset. Our methodology involves a struct…
Figure 12
Figure 12. Figure 12: Estimates for δworst (top) and δavg (bottom) using simulated data points from a wrapped normal distribution on manifolds with varying curvatures (K) and dimensionalities. When comparing representational shifts between perturbed processed pseudogene and background sequ…
Figure 13
Figure 13. Figure 13: Distribution of scaled δ-hyperbolicity values using embeddings from various DNA LMs. The distribution of each model is overlaid with the δ distribution of randomly sampled points from a Gaussian of equal dimensionality (red). An asterisk (*) denotes that the correspon…
Figure 14
Figure 14. Figure 14: HCNN embeddings for the processed pseudogene dataset, colored by model confidence on [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]
Figure 15
Figure 15. Figure 15: HCNN embeddings for a processed pseudogene sequence and a background sequence. [PITH_FULL_IMAGE:figures/full_fig_p030_15.png]
Figure 16
Figure 16. Figure 16: UMAP visualizations of the embeddings generated by the HCNN (left) and CNN (right) [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 65 canonical work pages

  1. [1]

    Topological implications of negative curvature for biological and social networks

    R \'e ka Albert, Bhaskar DasGupta, and Nasim Mobasheri. Topological implications of negative curvature for biological and social networks. Physical Review E, 89 0 (3): 0 032811, 2014

    work page 2014

  2. [2]

    Effective gene expression prediction from sequence by integrating long-range interactions

    Z iga Avsec, Vikram Agarwal, Daniel Visentin, Joseph R Ledsam, Agnieszka Grabska-Barwinska, Kyle R Taylor, Yannis Assael, John Jumper, Pushmeet Kohli, and David R Kelley. Effective gene expression prediction from sequence by integrating long-range interactions. Nature methods, 18 0 (10): 0 1196--1203, 2021

    2021

  3. [3]

    Fully hyperbolic convolutional neural networks for computer vision

    Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Fully hyperbolic convolutional neural networks for computer vision. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. URL https://openreview.net/forum?id=ekz1hN5QNh

    work page 2024

  4. [4]

    Riemannian adaptive optimization methods

    Gary B \' e cigneul and Octavian - Eugen Ganea. Riemannian adaptive optimization methods. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=r1eiqi09K7

    work page 2019

  5. [5]

    Hyperbolicity measures democracy in real-world networks

    Michele Borassi, Alessandro Chessa, and Guido Caldarelli. Hyperbolicity measures democracy in real-world networks. Physical Review E, 92 0 (3): 0 032812, 2015

    work page 2015

  6. [6]

    Transposable elements and the evolution of eukaryotic complexity

    Nathan J Bowen and I King Jordan. Transposable elements and the evolution of eukaryotic complexity. Current issues in molecular biology, 4 0 (3): 0 65--76, 2002

    work page 2002

  7. [7]

    High-coverage whole-genome sequencing of the expanded 1000 genomes project cohort including 602 trios

    Marta Byrska-Bishop, Uday S Evani, Xuefang Zhao, Anna O Basile, Haley J Abel, Allison A Regier, Andr \'e Corvelo, Wayne E Clarke, Rajeeva Musunuri, Kshithija Nagulapalli, et al. High-coverage whole-genome sequencing of the expanded 1000 genomes project cohort including 602 trios. Cell, 185 0 (18): 0 3426--3440, 2022

    work page 2022

  8. [8]

    Clough, and Marc Peter Deisenroth

    Benjamin Paul Chamberlain, James R. Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. CoRR, abs/1705.10359, 2017. URL http://arxiv.org/abs/1705.10359

    arXiv 2017

Show all 78 references
  1. [9]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher R \'e , and Jure Leskovec. Hyperbolic graph convolutional neural networks. Advances in neural information processing systems, 32, 2019

    2019

  2. [10]

    From trees to continuous embeddings and back: Hyperbolic hierarchical clustering

    Ines Chami, Albert Gu, Vaggos Chatziafratis, and Christopher R \'e . From trees to continuous embeddings and back: Hyperbolic hierarchical clustering. Advances in Neural Information Processing Systems, 33: 0 15065--15076, 2020 a

    2020

  3. [11]

    Low-dimensional hyperbolic knowledge graph embeddings

    Ines Chami, Adva Wolf, Da - Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher R \' e . Low-dimensional hyperbolic knowledge graph embeddings. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel R. Tetreault (eds.), Proceedings of the 58th Annual Meeting of the Associ...

    2020 doi

  4. [12]

    Average gromov hyperbolicity and the parisi ansatz

    Sourav Chatterjee and Leila Sloman. Average gromov hyperbolicity and the parisi ansatz. Advances in Mathematics, 376: 0 107417, 2021

    2021

  5. [13]

    Naesseth, and Itsik Pe'er

    Alex Chen, Philipe Chlenski, Kenneth Munyuza, Antonio Khalil Moretti, Christian A. Naesseth, and Itsik Pe'er. Variational combinatorial sequential monte carlo for bayesian phylogenetics in hyperbolic space, 2025. URL https://arxiv.org/abs/2501.17965

    2025 arXiv

  6. [14]

    A sequence-based global map of regulatory activity for deciphering human genetics

    Kathleen M Chen, Aaron K Wong, Olga G Troyanskaya, and Jian Zhou. A sequence-based global map of regulatory activity for deciphering human genetics. Nature genetics, 54 0 (7): 0 940--949, 2022 a

    2022

  7. [15]

    Fully hyperbolic neural networks

    Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguis...

    2022 doi

  8. [16]

    Hyperbolic pre-trained language model

    Weize Chen, Xu Han, Yankai Lin, Kaichen He, Ruobing Xie, Jie Zhou, Zhiyuan Liu, and Maosong Sun. Hyperbolic pre-trained language model. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  9. [17]

    Fast hyperboloid decision tree algorithms

    Philippe Chlenski, Ethan Turok, Antonio Khalil Moretti, and Itsik Pe'er. Fast hyperboloid decision tree algorithms. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=TTonmgTT9X

    2024

  10. [18]

    Manify: A python library for learning non-euclidean representations, 2025

    Philippe Chlenski, Kaizhu Du, Dylan Satow, and Itsik Pe'er. Manify: A python library for learning non-euclidean representations, 2025. URL https://arxiv.org/abs/2503.09576

    2025 arXiv

  11. [19]

    On computing the gromov hyperbolicity

    Nathann Cohen, David Coudert, and Aur \'e lien Lancin. On computing the gromov hyperbolicity. Journal of Experimental Algorithmics (JEA), 20: 0 1--18, 2015

    2015

  12. [20]

    To transformers and beyond: large language models for the genome

    Micaela E Consens, Cameron Dufault, Michael Wainberg, Duncan Forster, Mehran Karimzadeh, Hani Goodarzi, Fabian J Theis, Alan Moses, and Bo Wang. To transformers and beyond: large language models for the genome. arXiv preprint arXiv:2311.07621, 2023

    2023 arXiv

  13. [21]

    Distribution and intensity of constraint in mammalian genomic sequence

    Gregory M Cooper, Eric A Stone, George Asimenos, Eric D Green, Serafim Batzoglou, and Arend Sidow. Distribution and intensity of constraint in mammalian genomic sequence. Genome research, 15 0 (7): 0 901--913, 2005

    2005

  14. [22]

    Neural distance embeddings for biological sequences

    Gabriele Corso, Zhitao Ying, Michal P \'a ndy, Petar Veli c kovi \'c , Jure Leskovec, and Pietro Li \`o . Neural distance embeddings for biological sequences. Advances in Neural Information Processing Systems, 34: 0 18539--18551, 2021

    2021

  15. [23]

    Nucleotide transformer: building and evaluating robust foundation models for human genomics

    Hugo Dalla-Torre, Liam Gonzalez, Javier Mendoza-Revilla, Nicolas Lopez Carranza, Adam Henryk Grzywaczewski, Francesco Oteri, Christian Dallago, Evan Trop, Bernardo P de Almeida, Hassan Sirelkhatim, et al. Nucleotide transformer: building and evaluating robust foundation models...

    2024

  16. [24]

    Epd and epdnew, high-quality promoter resources in the next-generation sequencing era

    Ren \'e Dreos, Giovanna Ambrosini, Rouayda Cavin P \'e rier, and Philipp Bucher. Epd and epdnew, high-quality promoter resources in the next-generation sequencing era. Nucleic acids research, 41 0 (D1): 0 D157--D164, 2013

    2013

  17. [25]

    Roles of transposable elements in the different layers of gene expression regulation

    Denise Drongitis, Francesco Aniello, Laura Fucci, and Aldo Donizetti. Roles of transposable elements in the different layers of gene expression regulation. International Journal of Molecular Sciences, 20 0 (22): 0 5755, 2019

    2019

  18. [26]

    Sequence basis of transcription initiation in the human genome

    Kseniia Dudnyk, Donghong Cai, Chenlai Shi, Jian Xu, and Jian Zhou. Sequence basis of transcription initiation in the human genome. Science, 384 0 (6694): 0 eadj0116, 2024

    2024

  19. [27]

    The regulated retrotransposon transcriptome of mammalian cells

    Geoffrey J Faulkner, Yasumasa Kimura, Carsten O Daub, Shivangi Wani, Charles Plessy, Katharine M Irvine, Kate Schroder, Nicole Cloonan, Anita L Steptoe, Timo Lassmann, et al. The regulated retrotransposon transcriptome of mammalian cells. Nature Genetics, 41 0 (5): 0 563--571,...

    2009 doi

  20. [28]

    Computing the gromov hyperbolicity of a discrete metric space

    Herv \'e Fournier, Anas Ismail, and Antoine Vigneron. Computing the gromov hyperbolicity of a discrete metric space. Information Processing Letters, 115 0 (6-8): 0 576--579, 2015

    2015

  21. [29]

    Gencode reference annotation for the human and mouse genomes

    Adam Frankish, Mark Diekhans, Anne-Maud Ferreira, Rory Johnson, Irwin Jungreis, Jane Loveland, Jonathan M Mudge, Cristina Sisu, James Wright, Joel Armstrong, et al. Gencode reference annotation for the human and mouse genomes. Nucleic acids research, 47 0 (D1): 0 D766--D773, 2019

    2019

  22. [30]

    Hyperbolic neural networks

    Octavian Ganea, Gary B \'e cigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018

    2018

  23. [31]

    Genomic benchmarks: a collection of datasets for genomic sequence classification

    Katar \' na Gre s ov \'a , Vlastimil Martinek, David C ech \'a k, Petr S ime c ek, and Panagiotis Alexiou. Genomic benchmarks: a collection of datasets for genomic sequence classification. BMC Genomic Data, 24 0 (1): 0 25, 2023

    2023

  24. [32]

    Hyperbolic groups

    M Gromov. Hyperbolic groups. Essays in Group Theory, pages/Springer-Verlag, 1987

    1987

  25. [33]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=tEYskw1VY2

    2024

  26. [34]

    Hyperbolic attention networks

    Caglar Gulcehre, Misha Denil, Mateusz Malinowski, Ali Razavi, Razvan Pascanu, Karl Moritz Hermann, Peter Battaglia, Victor Bapst, David Raposo, Adam Santoro, et al. Hyperbolic attention networks. arXiv preprint arXiv:1805.09786, 2018

    2018 arXiv

  27. [35]

    Exploring network structure, dynamics, and function using networkx

    Aric Hagberg, Pieter J Swart, and Daniel A Schult. Exploring network structure, dynamics, and function using networkx. Technical report, Los Alamos National Laboratory (LANL), Los Alamos, NM (United States), 2008

    2008

  28. [36]

    Mobilization of transposable elements by environmental and endogenous factors

    Dustin C Hancks and Haig H Kazazian. Mobilization of transposable elements by environmental and endogenous factors. Human Molecular Genetics, 25 0 (R2): 0 R45--R50, 2016. doi:10.1093/hmg/ddw025

    2016 doi

  29. [37]

    Transposable elements

    Alexander Hayward and Cl \'e ment Gilbert. Transposable elements. Current Biology, 32 0 (17): 0 R904--R909, 2022

    2022

  30. [38]

    Capturing implicit hierarchical structure in 3d biomedical images with self-supervised hyperbolic representations

    Joy Hsu, Jeffrey Gu, Gong Wu, Wah Chiu, and Serena Yeung. Capturing implicit hierarchical structure in 3d biomedical images with self-supervised hyperbolic representations. Advances in neural information processing systems, 34: 0 5112--5123, 2021

    2021

  31. [39]

    Ete 3: reconstruction, analysis, and visualization of phylogenomic data

    Jaime Huerta-Cepas, Fran c ois Serra, and Peer Bork. Ete 3: reconstruction, analysis, and visualization of phylogenomic data. Molecular biology and evolution, 33 0 (6): 0 1635--1638, 2016

    2016

  32. [40]

    Visualising very large phylogenetic trees in three dimensional hyperbolic space

    Timothy Hughes, Young Hyun, and David A Liberles. Visualising very large phylogenetic trees in three dimensional hyperbolic space. BMC bioinformatics, 5: 0 1--6, 2004

    2004

  33. [41]

    Dnabert: pre-trained bidirectional encoder representations from transformers model for dna-language in genome

    Yanrong Ji, Zhihan Zhou, Han Liu, and Ramana V Davuluri. Dnabert: pre-trained bidirectional encoder representations from transformers model for dna-language in genome. Bioinformatics, 37 0 (15): 0 2112--2120, 2021

    2021

  34. [42]

    Learning hyperbolic embedding for phylogenetic tree placement and updates

    Yueyu Jiang, Puoya Tabaghi, and Siavash Mirarab. Learning hyperbolic embedding for phylogenetic tree placement and updates. Biology, 11 0 (9): 0 1256, 2022 a

    2022

  35. [43]

    Phylogenetic placement problem: A hyperbolic embedding approach

    Yueyu Jiang, Puoya Tabaghi, and Siavash Mirarab. Phylogenetic placement problem: A hyperbolic embedding approach. In RECOMB International Workshop on Comparative Genomics, pp.\ 68--85. Springer, 2022 b

    2022

  36. [44]

    Transposable elements: a common feature of neurodegenerative disorders

    Martin E J \"o nsson, Rebecca Garza, Per A Johansson, and Johan Jakobsson. Transposable elements: a common feature of neurodegenerative disorders. Mobile DNA, 11 0 (1): 0 1--15, 2020. doi:10.1186/s13100-020-00207-x

    2020 doi

  37. [45]

    Hyperbolic image embeddings

    Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 6418--6428, 2020

    2020

  38. [46]

    Controlling the fr \'e chet variance improves batch normalization on the symmetric positive definite manifold

    Reinmar J Kobler, Jun-ichiro Hirayama, and Motoaki Kawanabe. Controlling the fr \'e chet variance improves batch normalization on the symmetric positive definite manifold. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp...

    2022

  39. [47]

    Hyperplane margin classifiers on the multinomial manifold

    Guy Lebanon and John Lafferty. Hyperplane margin classifiers on the multinomial manifold. In Proceedings of the twenty-first international conference on Machine learning, pp.\ 66, 2004

    2004

  40. [48]

    Hyperbolic visual embedding learning for zero-shot recognition

    Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9273--9281, 2020

    2020

  41. [49]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  42. [50]

    Differentiating through the fr \'e chet mean

    Aaron Lou, Isay Katsman, Qingxuan Jiang, Serge Belongie, Ser-Nam Lim, and Christopher De Sa. Differentiating through the fr \'e chet mean. In International conference on machine learning, pp.\ 6393--6403. PMLR, 2020

    2020

  43. [51]

    Evolution is all you need: phylogenetic augmentation for contrastive learning

    Amy X Lu, Alex X Lu, and Alan Moses. Evolution is all you need: phylogenetic augmentation for contrastive learning. arXiv preprint arXiv:2012.13475, 2020

    2012 arXiv

  44. [52]

    Plantrep: a database of plant repetitive elements

    Xizhi Luo, Shiyu Chen, and Yu Zhang. Plantrep: a database of plant repetitive elements. Plant cell reports, pp.\ 1--4, 2022

    2022

  45. [53]

    Continuous hierarchical representations with poincar \'e variational auto-encoders

    Emile Mathieu, Charline Le Lan, Chris J Maddison, Ryota Tomioka, and Yee Whye Teh. Continuous hierarchical representations with poincar \'e variational auto-encoders. Advances in neural information processing systems, 32, 2019

    2019

  46. [54]

    A wrapped normal distribution on hyperbolic space for gradient-based learning

    Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, and Masanori Koyama. A wrapped normal distribution on hyperbolic space for gradient-based learning. In International Conference on Machine Learning, pp.\ 4693--4702. PMLR, 2019

    2019

  47. [55]

    Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution

    Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Michael Wornow, Callum Birch-Sykes, Stefano Massaroli, Aman Patel, Clayton Rabideau, Yoshua Bengio, et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. Advances in neural information proc...

    2024

  48. [56]

    Learning continuous hierarchies in the lorentz model of hyperbolic geometry

    Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, pp.\ 3779--3788. PMLR, 2018

    2018

  49. [57]

    Hyena hierarchy: Towards larger convolutional language models

    Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher R \'e . Hyena hierarchy: Towards larger convolutional language models. In International Conference on Machine Learning, pp.\ 28043--28078. PMLR, 2023

    2023

  50. [58]

    Autoencoding hyperbolic representation for adversarial generation

    Eric Qu and Dongmian Zou. Autoencoding hyperbolic representation for adversarial generation. arXiv preprint arXiv:2201.12825, 2022

    2022 arXiv

  51. [59]

    Caduceus: Bi-directional equivariant long-range DNA sequence modeling

    Yair Schiff, Chia Hsiang Kao, Aaron Gokaslan, Tri Dao, Albert Gu, and Volodymyr Kuleshov. Caduceus: Bi-directional equivariant long-range DNA sequence modeling. In First Workshop on Long-Context Foundation Models @ ICML 2024, 2024. URL https://openreview.net/forum?id=iYNSCJTlPO

    2024

  52. [60]

    The impact of transposable elements in adaptive evolution

    Lukas Schrader and J \"u rgen Schmitz. The impact of transposable elements in adaptive evolution. Molecular Ecology, 28 0 (6): 0 1537--1549, 2019

    2019

  53. [61]

    Widespread establishment and regulatory impact of alu exons in human genes

    Shihao Shen, Lan Lin, James J Cai, Peng Jiang, Emily J Kenkel, Miranda R Stroik, Shigeo Sato, Beverly L Davidson, and Yi Xing. Widespread establishment and regulatory impact of alu exons in human genes. Proceedings of the National Academy of Sciences, 108 0 (7): 0 2837--2842, ...

    2011 doi

  54. [62]

    Mixed-curvature variational autoencoders

    Ondrej Skopek, Octavian-Eugen Ganea, and Gary B \'e cigneul. Mixed-curvature variational autoencoders. In 8th international conference on learning representations (ICLR 2020)(virtual). International Conference on Learning Representations, 2020

    2020

  55. [63]

    Pyvolve: a flexible python module for simulating sequences along phylogenies

    Stephanie J Spielman and Claus O Wilke. Pyvolve: a flexible python module for simulating sequences along phylogenies. PloS one, 10 0 (9): 0 e0139047, 2015

    2015

  56. [64]

    Widespread contribution of transposable elements to the innovation of gene regulatory networks

    Vasavi Sundaram, Yong Cheng, Zhihai Ma, Daofeng Li, Xiaoyun Xing, Peter Edge, Michael P Snyder, and Ting Wang. Widespread contribution of transposable elements to the innovation of gene regulatory networks. Genome Research, 24 0 (12): 0 1963--1976, 2014. doi:10.1101/gr.168872.113

    1963 doi

  57. [65]

    Line-of-descent and genealogical processes, and their applications in population genetics models

    Simon Tavar \'e . Line-of-descent and genealogical processes, and their applications in population genetics models. Theoretical population biology, 26 0 (2): 0 119--164, 1984

    1984

  58. [66]

    Graphpart: homology partitioning for biological sequence analysis

    Felix Teufel, Magn \'u s Halld \'o r G \' slason, Jos \'e Juan Almagro Armenteros, Alexander Rosenberg Johansen, Ole Winther, and Henrik Nielsen. Graphpart: homology partitioning for biological sequence analysis. NAR genomics and bioinformatics, 5 0 (4): 0 lqad088, 2023

    2023

  59. [67]

    Complex hierarchical structures in single-cell genomics data unveiled by deep hyperbolic manifold learning

    Tian Tian, Cheng Zhong, Xiang Lin, Zhi Wei, and Hakon Hakonarson. Complex hierarchical structures in single-cell genomics data unveiled by deep hyperbolic manifold learning. Genome Research, 33 0 (2): 0 232--246, 2023

    2023

  60. [68]

    Poincare glove: Hyperbolic word embeddings

    Alexandru Tifrea, Gary B \' e cigneul, and Octavian - Eugen Ganea. Poincare glove: Hyperbolic word embeddings. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?i...

    2019

  61. [69]

    A field guide to eukaryotic transposable elements

    Jonathan N Wells and C \'e dric Feschotte. A field guide to eukaryotic transposable elements. Annual review of genetics, 54 0 (1): 0 539--561, 2020

    2020

  62. [70]

    Enhancing llm complex reasoning capability through hyperbolic geometry

    Menglin Yang, Aosong Feng, Bo Xiong, Jiahong Liu, Irwin King, and Rex Ying. Enhancing llm complex reasoning capability through hyperbolic geometry. In ICML 2024 Workshop on LLMs and Cognition, 2024

    2024

  63. [71]

    Deep learning for genomics: From early neural nets to modern large language models

    Tianwei Yue, Yuanxin Wang, Longxiang Zhang, Chunming Gu, Haoru Xue, Wenping Wang, Qi Lyu, and Yujie Dun. Deep learning for genomics: From early neural nets to modern large language models. International Journal of Molecular Sciences, 24 0 (21): 0 15858, 2023

    2023

  64. [72]

    Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale

    Jian Zhou. Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale. Nature genetics, 54 0 (5): 0 725--734, 2022

    2022

  65. [73]

    Hyperbolic geometry of gene expression

    Yuansheng Zhou and Tatyana O Sharpee. Hyperbolic geometry of gene expression. Iscience, 24 0 (3), 2021

    2021

  66. [74]

    Davuluri, and Han Liu

    Zhihan Zhou, Yanrong Ji, Weijian Li, Pratik Dutta, Ramana V. Davuluri, and Han Liu. DNABERT-2: efficient foundation model and benchmark for multi-species genomes. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . ...

    2024

  67. [75]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  68. [76]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  69. [77]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  70. [78]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.