Pith. sign in

REVIEW 3 major objections 5 minor 78 references

DualEquiNet: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DualEquiNet claims that modeling biomolecules in both Euclidean and spherical-harmonics spaces, with bidirectional cross-space message passing and hierarchical pooling, beats single-space equivariant models on RNA and protein benchmarks.

desk verdict A genuinely new dual-space architecture with useful new RNA benchmarks, but the central E(3)-equivariance proof has a translation bug in the CSIP pooling equation, so the main theoretical claim does not hold as written. read the letter →

arxiv 2506.19862 v1 pith:OPDMET6K submitted 2025-06-10 q-bio.BM cs.AIcs.LG

Junjie Xu , Jiahao Zhang , Mangal Prakash , Xiang Zhang , Suhang Wang This is my paper
classification q-bio.BMcs.AIcs.LG
keywords geometricgraphneuralnetworksE(3)equivariancesphericalharmonicshierarchicalpoolingRNApropertypredictionproteinstructuremodelinglong-rangedependenciesbiomolecularbenchmarks
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

DualEquiNet sets out to solve a specific failure of geometric graph neural networks on large biomolecules: models that work in Euclidean coordinates see only local neighborhoods, while models that work in spherical-harmonics space capture rotational symmetries but at high cost. The paper's central claim is that keeping both representations and exchanging messages between them—through bidirectional cross-space message passing and a hierarchical pooling step—produces a scalable, E(3)-equivariant (rotation- and translation-consistent) network that captures atomic detail and long-range structural dependencies together. Reported error reductions over prior methods are 25.7 percent on CovidVaccine, 33.5 percent on Ribonanza, and 8.4 percent on Tc-ribo, with the lowest errors also on two newly introduced 3D benchmarks for solvent-accessible surface area and torsion angles. If the claim holds, dual-space hierarchical equivariant networks become a practical route to modeling RNA and proteins at full atomistic scale.

What carries the argument

The load-bearing object is the dual-space representation: each node carries Euclidean coordinates $x_i$, invariant scalar features $h_i$, and spherical-harmonics features $r_i$ up to degree $l_{max}$. Euclidean neighborhoods are defined by a distance cutoff, while SH neighborhoods are defined by cosine similarity of $r_i$ and $r_j$, creating edges that skip over spatial distance. Inside each DualEqui layer, within-space messages and cross-space messages are combined under multi-head attention, with the degree-wise inner product $r_i \odot r_j$ serving as the invariant that lets SH information pass into Euclidean updates and relative directions pass back through the spherical-harmonic basis. The Cross-Space Interaction Pooling aggregates atom features into residue-level features while projecting SH norms into coordinate updates and relative coordinates into SH updates. These components jointly carry the argument that local geometry and global structural similarity can be captured in one equivariant network without expensive Clebsch-Gordan tensor products.

What would settle it

Replace the SH-similarity neighborhood in Eq. (4) with a random graph that preserves each node's degree, and retrain on the N-chain and CovidVaccine tasks. If the error does not rise substantially, the long-range advantage attributed to SH similarity is not caused by the similarity criterion; if it rises sharply, the mechanism is confirmed. A second check: randomly permute a fraction of the SH feature vectors before computing cosine similarity; stable performance would show that the similarity measure itself is not load-bearing.

Watch

Extended reading notes

Core claim

The central discovery is that a geometric GNN does not have to choose between Euclidean and spherical-harmonics spaces: the paper constructs both, lets them exchange information, and shows this outperforms single-space baselines on RNA and protein tasks. The SH space is initialized from local Euclidean neighborhoods and then defines its own neighborhoods by cosine similarity of SH feature vectors, so residues that are far apart in space but locally similar can interact directly. A Cross-Space Interaction Pooling then aggregates atoms into residues while projecting each space into the other, preserving E(3) equivariance. On CovidVaccine, Ribonanza, and Tc-ribo the reported RMSE reductions are 25.7, 33.5, and 8.4 percent, and on the new SASA and TorsionAngle benchmarks the reductions are 3.1-28.8 percent and 2.5 percent. Ablation and neighborhood analyses support the claim that the two spaces play complementary roles: SH neighbors move closer in Euclidean distance across layers, while Euclidean neighbors diverge in SH similarity.

Load-bearing premise

The load-bearing assumption is that two residues with similar local-shape descriptors (computed with spherical harmonics) are structurally related in a way that matters for prediction, and that the input 3D structures are accurate enough to compute those descriptors.

Editorial extensions

If this is right

  • RNA reactivity and degradation prediction improves by 25.7-33.5 percent RMSE on standard benchmarks, indicating that dual-space message passing captures interactions that single-space geometric models miss.
  • The same architecture transfers to two newly introduced 3D structural benchmarks, SASA and torsion-angle prediction, with reported gains of 3.1-28.8 percent and 2.5 percent over previous best methods.
  • Because SH-similarity edges connect spatially distant residues directly, fewer layers are needed to propagate information, which the N-chain experiments show by reaching near-perfect accuracy even below the theoretical layer count.
  • E(3) equivariance is preserved through message passing and hierarchical pooling, so scalar predictions are invariant and coordinate-like outputs transform correctly under global rotations and translations.
  • Ablation results attribute clear gains to each component: replacing Cross-Space Interaction Pooling with mean pooling or removing cross-space messages increases RMSE on RNA property datasets.

Reading between the lines

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

  • An implicit corollary of the reported gains is that the dual-space construction could transfer to other large structured systems—protein complexes, chromatin, or materials—where spatially distant fragments share local geometry; the paper does not test this.
  • The sensitivity of N-chain accuracy to d_SH suggests a testable extension: replacing the fixed cosine threshold with a learned or adaptive similarity metric could improve robustness on real structures; this is an editorial suggestion, not a paper claim.
  • Because the neighborhood analysis shows SH neighbors converging in Euclidean distance over layers, one could probe whether intermediate coordinates correspond to physically plausible folding intermediates; the paper only reports the trend.
  • The reliance on RhoFold/AlphaFold structures points to an obvious next step the paper names in its conclusion: coupling the network with a differentiable structure predictor so it can run from sequence alone.
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

3 major / 5 minor

Summary. The paper introduces DualEquiNet, a hierarchical geometric GNN that constructs complementary representations in Euclidean (EU) and spherical-harmonics (SH) spaces. It defines EU neighborhoods by distance cutoff and SH neighborhoods by cosine similarity of SH features, performs bidirectional cross-space message passing, and pools atom-level features to residue level via a Cross-Space Interaction Pooling (CSIP). The authors claim E(3) equivariance, demonstrate expressivity on synthetic N-chain and L-fold tasks, and report state-of-the-art results on CovidVaccine, Ribonanza, Tc-ribo, and two new SASA and torsion-angle benchmarks.

Significance. The dual-space hierarchical design is a plausible and potentially scalable approach for large biomolecules, and the synthetic expressivity experiments are a useful addition. The paper would be a valuable contribution if the E(3)-equivariance proof were correct and the empirical claims reproducible. However, the CSIP coordinate pooling in Eq. (15) is not translation equivariant, which invalidates the central theoretical claim and undermines the interpretation of the benchmark gains as evidence for equivariant long-range modeling. The absence of released code/data and the small number of random splits further limit verification.

major comments (3)
  1. [§3.3, Eq. (15) and Appendix C] The CSIP Euclidean pooling x' = (1/|C|) Σ_{i∈C} α_i [x_i + γ Proj_SH(r_i)] is not E(3)-equivariant. Under a global translation t, x_i becomes x_i + t while r_i and Proj_SH(r_i) are unchanged; hence x' transforms to (1/|C|) Σ α_i (1+γ s_i)(x_i + t), with s_i = Σ_l w_l ||r_i^{(l)}||. Equivariance would require (1/|C|) Σ α_i (1+γ s_i) = 1, which is not enforced: α_i are sigmoid attention scores with no normalization, and γ and w_l are unconstrained. The proof in Appendix C omits the translation term, writing R x_i instead of R x_i + t in the first equality. Because x' is used as residue-level coordinates in Eq. (17), the downstream distance-based messages and final scalar predictions are not translation invariant. The central E(3)-equivariance claim is therefore unsupported.
  2. [§3.1, Eq. (4) and §4.6] The SH-similarity neighborhood is the paper's main mechanism for long-range dependencies, but the paper provides no theoretical or biological justification that cosine similarity of SH feature vectors is a meaningful proxy for functionally relevant structural similarity. The neighborhood analysis in Section 4.6 shows that SH neighbors become closer in Euclidean distance over layers, but this is an emergent property of the model and does not itself establish that the SH edges correspond to genuine long-range contacts. The paper should analyze the overlap of SH edges with known long-range contacts or at least report sensitivity of the real-data results to d_SH.
  3. [§4 and Appendix D.5] The empirical central claim of state-of-the-art performance is not currently verifiable: no code, trained models, or the two newly introduced datasets are released, and all results are averaged over only five random splits with hyperparameters selected by the authors via Optuna for all methods. Releasing the code and data, and ideally more seeds, would be necessary to support the reported margins of 25.7%, 33.5%, and 8.4%.
minor comments (5)
  1. [§3.2 and Appendix C.2] In Eq. (5) and in Appendix C.2, the message function for SH-space within-space messages is written as φ_EU but should be φ_SH; also, in C.2 the expression for m_EU→SH,ij contains an extra closing bracket.
  2. [Table 1 and Table 6] The captions refer to red/green highlighting that is not visible in the printed manuscript; the highlighting should be rendered or the captions should be updated.
  3. [Table 10] The abbreviation 'Lnt' should be 'L_res' (residue DualEqui layers) to be consistent with the notation in Eq. (17).
  4. [Appendix D.5] There is a typo: 'detialed' should be 'detailed'.
  5. [Appendix B.1] The composition formula for E(3) contains a stray semicolon and should read (R1,t1)·(R2,t2) = (R1R2, R1t2 + t1).

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the main benchmark claims rest on held-out test splits rather than on fitted quantities or self-citation chains.

full rationale

I walked the claimed derivation chain from the dual-space initialization (Eqs. 1-4), through the DualEqui layer updates (Eqs. 5-13), the CSIP pooling (Eqs. 14-16), and the hierarchy construction (Eq. 17). The benchmark predictions in Tables 2-4 are obtained by training on an 8:1:1 split, selecting hyperparameters with Optuna on the validation split, and evaluating on held-out test data; no target label or test statistic enters the network definition. The SH-similarity neighborhood (Eq. 4) is defined from SH features initialized by Eq. 2, but this is a modeling choice, not a derivation of the target from the target: the claimed long-range benefit is a falsifiable empirical hypothesis evaluated on the held-out benchmarks. The paper's self-citations ([21], [53], [74], [75]) are used for background or experimental protocol (e.g., 'We follow the experimental protocol in [53] to use eta-theta pseudotorsional backbone [68] as input'), not as the load-bearing justification for the equivariance or performance claims; [61], [68], and the benchmark sources are external. The Appendix C proof of E(3)-equivariance for Eq. (15) contains an algebraic gap in handling the translation term, as the proof writes x'(RX+t) with R x_i rather than (R x_i + t); this is a correctness and verification risk, but it is not a circularity because the claim does not reduce to its inputs by construction. Similarly, Optuna hyperparameter tuning is standard model selection and is not a fitted input renamed as a prediction. I therefore find no circular step that would warrant a score above 1.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The paper's central claim rests on standard equivariance math (sound), plus two paper-specific modeling assumptions: the SH-similarity neighborhood and the validity of predicted 3D structures. The main empirical claims depend on a large set of per-dataset hyperparameters tuned via Optuna, and on no shipped code or data.

free parameters (5)
  • d_SH (SH-space cosine similarity threshold) = 0.60 to 1.00, varies per dataset (Table 10)
    Threshold for SH-neighbor selection. The N-chain experiment shows model accuracy swings from 50% to 100% depending on this value, and per-dataset values are tuned via Optuna, so the central long-range-communication mechanism is highly sensitive to this fitted choice.
  • d_EU (Euclidean distance cutoff) = 3.03 to 10.44 (atom), 26.4 to 122.64 (residue), varies per dataset (Table 10)
    Spatial cutoff defining local EU neighborhoods; tuned per dataset via Optuna.
  • l_max (spherical harmonics degree) = 2 or 3, varies per dataset (Table 10)
    Maximum SH degree; a larger value increases expressivity but also cost, and the optimal value differs across datasets, indicating this choice is fitted rather than determined by the task.
  • Layer counts (L_atom, L_res), hidden size, number of attention heads = varies per dataset (Table 10)
    Architecture hyperparameters selected per dataset via Optuna on validation performance. This is standard hyperparameter tuning but means the final reported numbers are the result of validation-set selection.
  • Learnable weights w_l, gamma, epsilon, and all MLP parameters = trained
    Model parameters learned from data during training. Normal for an architecture paper, but worth listing because the pooling projections (Eqs. 15, 16) rely on learned coefficients to balance EU and SH contributions.
assumptions (4)
  • standard math E(3) equivariance of the message functions and pooling operations as proved in Appendix C.
    The proof relies on standard properties of Wigner D-matrices and the invariance of distances and SH inner products under rotations.
  • ad hoc to paper The SH-feature cosine similarity (Eq. 4) defines a chemically meaningful neighborhood for long-range structural dependencies.
    This is the key new modeling assumption. The paper provides no theoretical or empirical justification that cosine similarity of SH feature vectors corresponds to functional structural similarity; it is simply a design choice validated by the reported benchmarks.
  • domain assumption The input 3D structures are accurate enough for property prediction.
    RNA structures are generated with RhoFold and protein structures with AlphaFold2. The paper acknowledges in the conclusion that performance 'may depend on the quality of input 3D structures.' If the predicted structures are systematically inaccurate, all benchmarks inherit that error.
  • domain assumption The distance cutoff d_EU and pooling clusters (atoms grouped into residues) are given by the dataset and preprocessing.
    The hierarchical structure (atoms to residues) is assumed to be available and correctly assigned. The paper does not discuss how residues are identified beyond 'atoms in residue C'.
invented entities (2)
  • SH-space neighborhood defined by cosine similarity of SH features
    purpose: To connect distant atoms with similar spherical-harmonics-derived features, overcoming the locality of Euclidean message passing.
    This is the paper's central new construct. Its chemical validity rests entirely on the benchmark results in this paper; there is no external falsifiable prediction or independent test of the similarity measure itself.
  • Cross-Space Interaction Pooling (CSIP) with EU-to-SH and SH-to-EU projections
    purpose: To aggregate atom features into residue-level features while exchanging information between the two representation spaces.
    A new architectural component; its benefit is supported only by the paper's ablations. No independent verification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DualEquiNet: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules." pith.science (2026). https://pith.science/paper/OPDMET6K

@misc{pith2026250619862,
  author       = {Pith},
  title        = {Pith review of: DualEquiNet: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OPDMET6K}},
  note         = {Machine review of arXiv:2506.19862}
}
read the original abstract

Geometric graph neural networks (GNNs) that respect E(3) symmetries have achieved strong performance on small molecule modeling, but they face scalability and expressiveness challenges when applied to large biomolecules such as RNA and proteins. These systems require models that can simultaneously capture fine-grained atomic interactions, long-range dependencies across spatially distant components, and biologically relevant hierarchical structure, such as atoms forming residues, which in turn form higher-order domains. Existing geometric GNNs, which typically operate exclusively in either Euclidean or Spherical Harmonics space, are limited in their ability to capture both the fine-scale atomic details and the long-range, symmetry-aware dependencies required for modeling the multi-scale structure of large biomolecules. We introduce DualEquiNet, a Dual-Space Hierarchical Equivariant Network that constructs complementary representations in both Euclidean and Spherical Harmonics spaces to capture local geometry and global symmetry-aware features. DualEquiNet employs bidirectional cross-space message passing and a novel Cross-Space Interaction Pooling mechanism to hierarchically aggregate atomic features into biologically meaningful units, such as residues, enabling efficient and expressive multi-scale modeling for large biomolecular systems. DualEquiNet achieves state-of-the-art performance on multiple existing benchmarks for RNA property prediction and protein modeling, and outperforms prior methods on two newly introduced 3D structural benchmarks demonstrating its broad effectiveness across a range of large biomolecule modeling tasks.

Figures

Figures reproduced from arXiv: 2506.19862 by the authors.

Figure 1
Figure 1. Overview of DualEquiNet. (a) DualEquiNet: It begins with SH feature initialization, followed by multiple DualEqui layers applied at the atomic level. A CSIP module then aggregates atomic representations into residue-level features, which are further refined by additional DualEqui layers. (b) DualEqui Layer: Each layer first infers EU and SH neighborhoods, then performs both inner-space and cross-space message passin… view at source ↗
Figure 2
Figure 2. (a) Left: Average distance variation of EU and SH neighborhoods across layers. (b) Right: [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Visualization of real Spherical Harmonics table. [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 55 canonical work pages

  1. [1]

    Deeprna- twist: Language model guided rna torsion angle prediction with attention-inception network

    Abrar Rahman Abir, Md Toki Tahmid, Rafiqul Islam Rayan, and M Saifur Rahman. Deeprna- twist: Language model guided rna torsion angle prediction with attention-inception network. bioRxiv, pages 2024–10, 2024

    work page 2024

  2. [2]

    Rnasolo: a repository of cleaned pdb-derived rna 3d structures

    Bartosz Adamczyk, Maciej Antczak, and Marta Szachniuk. Rnasolo: a repository of cleaned pdb-derived rna 3d structures. Bioinformatics, 38(14):3668–3670, 2022

    work page 2022

  3. [3]

    Optuna: A next-generation hyperparameter optimization framework

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623–2631, 2019

    2019

  4. [4]

    On the bottleneck of graph neural networks and its practical implications

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. In International Conference on Learning Representations, 2021

    2021

  5. [5]

    Geometric deep learning on molecular representations

    Kenneth Atz, Francesca Grisoni, and Gisbert Schneider. Geometric deep learning on molecular representations. Nature Machine Intelligence, 3(12):1023–1032, 2021

    2021

  6. [6]

    Gotennet: Rethinking efficient 3d equivariant graph neural networks

    Sarp Aykent and Tian Xia. Gotennet: Rethinking efficient 3d equivariant graph neural networks. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  7. [7]

    Protein structure prediction and structural genomics

    David Baker and Andrej Sali. Protein structure prediction and structural genomics. Science, 294(5540):93–96, 2001

    work page 2001

  8. [8]

    Implicit graph neural networks: A monotone operator viewpoint

    Justin Baker, Qingsong Wang, Cory D Hauck, and Bao Wang. Implicit graph neural networks: A monotone operator viewpoint. In International Conference on Machine Learning , pages 1521–1548. PMLR, 2023

    work page 2023

Show all 78 references
  1. [9]

    Mace: Higher order equivariant message passing neural networks for fast and accurate force fields

    Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and Gábor Csányi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in Neural Information Processing Systems, 35:11423–11436, 2022

    2022

  2. [10]

    The protein data bank.Nucleic acids research, 28(1):235–242, 2000

    Helen M Berman, John Westbrook, Zukang Feng, Gary Gilliland, Talapady N Bhat, Helge Weissig, Ilya N Shindyalov, and Philip E Bourne. The protein data bank.Nucleic acids research, 28(1):235–242, 2000

    2000

  3. [11]

    Rna-torsionbert: lever- aging language models for rna 3d torsion angles prediction

    Clément Bernard, Guillaume Postic, Sahar Ghannay, and Fariza Tahi. Rna-torsionbert: lever- aging language models for rna 3d torsion angles prediction. Bioinformatics, 41(1):btaf004, 2025

    2025

  4. [13]

    Bekkers, and Max Welling

    Johannes Brandstetter, Rob Hesselink, Elise van der Pol, Erik J. Bekkers, and Max Welling. Geometric and Physical Quantities Improve E(3) Equivariant Message Passing, March 2022. arXiv:2110.02905 [cs, stat]

    2022 arXiv

  5. [14]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021

    2021 arXiv

  6. [15]

    Crystal structure of a group i ribozyme domain: principles of rna packing

    Jamie H Cate, Anne R Gooding, Elaine Podell, Kaihong Zhou, Barbara L Golden, Craig E Kundrot, Thomas R Cech, and Jennifer A Doudna. Crystal structure of a group i ribozyme domain: principles of rna packing. Science, 273(5282):1678–1685, 1996

    1996

  7. [16]

    Are high- degree representations really unnecessary in equivariant graph neural networks? In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    Jiacheng Cen, Anyi Li, Ning Lin, Yuxiang Ren, Zihe Wang, and Wenbing Huang. Are high- degree representations really unnecessary in equivariant graph neural networks? In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  8. [17]

    Are high- degree representations really unnecessary in equivariant graph neural networks? arXiv preprint arXiv:2410.11443, 2024

    Jiacheng Cen, Anyi Li, Ning Lin, Yuxiang Ren, Zihe Wang, and Wenbing Huang. Are high- degree representations really unnecessary in equivariant graph neural networks? arXiv preprint arXiv:2410.11443, 2024

    2024

  9. [18]

    Simple and deep graph convolutional networks

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In International conference on machine learning, pages 1725–1735. PMLR, 2020

    2020

  10. [19]

    Efficient and scalable implicit graph neural networks with virtual equilibrium

    Qi Chen, Yifei Wang, Yisen Wang, Jianlong Chang, Qi Tian, Jiansheng Yang, and Zhouchen Lin. Efficient and scalable implicit graph neural networks with virtual equilibrium. In 2022 IEEE International Conference on Big Data (Big Data), pages 864–873. IEEE, 2022. 12

    2022

  11. [20]

    Optimization-induced graph implicit nonlinear diffusion

    Qi Chen, Yifei Wang, Yisen Wang, Jiansheng Yang, and Zhouchen Lin. Optimization-induced graph implicit nonlinear diffusion. In International Conference on Machine Learning, pages 3648–3661. PMLR, 2022

    2022

  12. [21]

    A comprehensive survey on trustworthy graph neural networks: Privacy, robustness, fairness, and explainability

    Enyan Dai, Tianxiang Zhao, Huaisheng Zhu, Junjie Xu, Zhimeng Guo, Hui Liu, Jiliang Tang, and Suhang Wang. A comprehensive survey on trustworthy graph neural networks: Privacy, robustness, fairness, and explainability. arXiv preprint arXiv:2204.08570, 2022

    2022 arXiv

  13. [22]

    The limitless future of rna therapeutics

    Tulsi Ram Damase, Roman Sukhovershin, Christian Boada, Francesca Taraballi, Roderic I Pet- tigrew, and John P Cooke. The limitless future of rna therapeutics. Frontiers in bioengineering and biotechnology, 9:628137, 2021

    2021

  14. [23]

    Rna structure determination: From 2d to 3d

    Jie Deng, Xianyang Fang, Lin Huang, Shanshan Li, Lilei Xu, Keqiong Ye, Jinsong Zhang, Kaiming Zhang, and Qiangfeng Cliff Zhang. Rna structure determination: From 2d to 3d. Fundamental Research, 3(5):727–737, 2023

    2023

  15. [24]

    A hitchhiker’s guide to geometric gnns for 3d atomic systems.arXiv preprint arXiv:2312.07511, 2023

    Alexandre Duval, Simon V Mathis, Chaitanya K Joshi, Victor Schmidt, Santiago Miret, Fragkiskos D Malliaros, Taco Cohen, Pietro Liò, Yoshua Bengio, and Michael Bronstein. A hitchhiker’s guide to geometric gnns for 3d atomic systems.arXiv preprint arXiv:2312.07511, 2023

    2023 arXiv

  16. [25]

    Long range graph benchmark

    Vijay Prakash Dwivedi, Ladislav Rampášek, Michael Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long range graph benchmark. Advances in Neural Information Processing Systems, 35:22326–22340, 2022

    2022

  17. [26]

    Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. In ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

    2019

  18. [27]

    A euclidean trans- former for fast and stable machine learned force fields

    J Thorben Frank, Oliver T Unke, Klaus-Robert Müller, and Stefan Chmiela. A euclidean trans- former for fast and stable machine learned force fields. Nature Communications, 15(1):6539, 2024

    2024

  19. [28]

    Se (3)-transformers: 3d roto-translation equivariant attention networks

    Fabian Fuchs, Daniel Worrall, V olker Fischer, and Max Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. Advances in neural information processing systems, 33:1970–1981, 2020

    1970

  20. [29]

    Gemnet: Universal directional graph neural networks for molecules

    Johannes Gasteiger, Florian Becker, and Stephan Günnemann. Gemnet: Universal directional graph neural networks for molecules. Advances in Neural Information Processing Systems , 34:6790–6802, 2021

    2021

  21. [30]

    Fast and uncertainty-aware directional message passing for non-equilibrium molecules

    Johannes Gasteiger, Shankari Giri, Johannes T Margraf, and Stephan Günnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules. arXiv preprint arXiv:2011.14115, 2020

    2011 arXiv

  22. [31]

    Directional message passing for molecular graphs

    Johannes Gasteiger, Janek Groß, and Stephan Günnemann. Directional message passing for molecular graphs. arXiv preprint arXiv:2003.03123, 2020

    2003 arXiv

  23. [32]

    e3nn: Euclidean neural networks

    Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks. arXiv preprint arXiv:2207.09453, 2022

    2022 arXiv

  24. [33]

    Transformers meet directed graphs

    Simon Geisler, Yujia Li, Daniel J Mankowitz, Ali Taylan Cemgil, Stephan Günnemann, and Cosmin Paduraru. Transformers meet directed graphs. In International conference on machine learning, pages 11144–11172. PMLR, 2023

    2023

  25. [34]

    Tuning the performance of synthetic riboswitches using machine learning

    Ann-Christin Groher, Sven Jager, Christopher Schneider, Florian Groher, Kay Hamacher, and Beatrix Suess. Tuning the performance of synthetic riboswitches using machine learning. ACS synthetic biology, 8(1):34–44, 2018

    2018

  26. [35]

    Implicit graph neural networks

    Fangda Gu, Heng Chang, Wenwu Zhu, Somayeh Sojoudi, and Laurent El Ghaoui. Implicit graph neural networks. Advances in neural information processing systems, 33:11984–11995, 2020

    2020

  27. [36]

    A survey of geometric graph neural networks: Data structures, models and applications

    Jiaqi Han, Jiacheng Cen, Liming Wu, Zongzhao Li, Xiangzhe Kong, Rui Jiao, Ziyang Yu, Tingyang Xu, Fandi Wu, Zihe Wang, et al. A survey of geometric graph neural networks: Data structures, models and applications. arXiv preprint arXiv:2403.00485, 2024

    2024 arXiv

  28. [37]

    Geometrically equivariant graph neural networks: A survey

    Jiaqi Han, Yu Rong, Tingyang Xu, and Wenbing Huang. Geometrically equivariant graph neural networks: A survey. arXiv preprint arXiv:2202.07230, 2022

    2022 arXiv

  29. [38]

    Ribonanza: deep learning of rna structure through dual crowdsourcing

    Shujun He, Rui Huang, Jill Townley, Rachael C Kretsch, Thomas G Karagianes, David BT Cox, Hamish Blair, Dmitry Penzar, Valeriy Vyaltsev, Elizaveta Aristova, et al. Ribonanza: deep learning of rna structure through dual crowdsourcing. bioRxiv, 2024. 13

    2024

  30. [39]

    Rna-dependent rna polymerase: Structure, mechanism, and drug discovery for covid-19

    Yi Jiang, Wanchao Yin, and H Eric Xu. Rna-dependent rna polymerase: Structure, mechanism, and drug discovery for covid-19. Biochemical and biophysical research communications , 538:47–53, 2021

    2021

  31. [40]

    Learn- ing from protein structure with geometric vector perceptrons

    Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael JL Townshend, and Ron Dror. Learn- ing from protein structure with geometric vector perceptrons. arXiv preprint arXiv:2009.01411, 2020

    2009 arXiv

  32. [41]

    Joshi, Cristian Bodnar, Simon V

    Chaitanya K. Joshi, Cristian Bodnar, Simon V . Mathis, Taco Cohen, and Pietro Liò. On the expressive power of geometric graph neural networks. In International Conference on Machine Learning, 2023

    2023

  33. [42]

    Highly accurate protein structure prediction with alphafold

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ron- neberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583–589, 2021

    2021

  34. [43]

    Directional message passing for molecular graphs

    Johannes Klicpera, Janek Groß, Stephan Günnemann, et al. Directional message passing for molecular graphs. In ICLR, pages 1–13, 2020

    2020

  35. [44]

    On the generalization of equivariance and convolution in neural networks to the action of compact groups

    Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International conference on machine learning, pages 2747–2755. PMLR, 2018

    2018

  36. [45]

    Unbiased stochastic proximal solver for graph neural networks with equilibrium states

    Mingjie Li, Yifei Wang, Yisen Wang, and Zhouchen Lin. Unbiased stochastic proximal solver for graph neural networks with equilibrium states. In The Eleventh International Conference on Learning Representations, 2022

    2022

  37. [46]

    Scalable and effective implicit graph neural networks on large graphs

    Juncheng Liu, Bryan Hooi, Kenji Kawaguchi, Yiwei Wang, Chaosheng Dong, and Xiaokui Xiao. Scalable and effective implicit graph neural networks on large graphs. In The Twelfth International Conference on Learning Representations, 2024

    2024

  38. [47]

    Mgnni: Multiscale graph neural networks with implicit layers

    Juncheng Liu, Bryan Hooi, Kenji Kawaguchi, and Xiaokui Xiao. Mgnni: Multiscale graph neural networks with implicit layers. Advances in Neural Information Processing Systems , 35:21358–21370, 2022

    2022

  39. [48]

    Spherical message passing for 3d molecular graphs

    Yi Liu, Limei Wang, Meng Liu, Yuchao Lin, Xuan Zhang, Bora Oztekin, and Shuiwang Ji. Spherical message passing for 3d molecular graphs. In International Conference on Learning Representations, 2022

    2022

  40. [49]

    Dssr: an integrated software tool for dissecting the spatial structure of rna

    Xiang-Jun Lu, Harmen J Bussemaker, and Wilma K Olson. Dssr: an integrated software tool for dissecting the spatial structure of rna. Nucleic acids research, 43(21):e142–e142, 2015

    2015

  41. [50]

    Enhancing graph transformers with hierarchical distance structural encoding

    Yuankai Luo, Hongkang Li, Lei Shi, and Xiao-Ming Wu. Enhancing graph transformers with hierarchical distance structural encoding. Advances in Neural Information Processing Systems, 37:57150–57182, 2024

    2024

  42. [51]

    Transformers over directed acyclic graphs.Advances in Neural Information Processing Systems, 36:47764–47782, 2023

    Yuankai Luo, Veronika Thost, and Lei Shi. Transformers over directed acyclic graphs.Advances in Neural Information Processing Systems, 36:47764–47782, 2023

    2023

  43. [52]

    Role of solvent accessibility for aggregation-prone patches in protein folding

    Avinash Mishra, Shoba Ranganathan, B Jayaram, and Abdul Sattar. Role of solvent accessibility for aggregation-prone patches in protein folding. Scientific reports, 8(1):12896, 2018

    2018

  44. [53]

    SE(3)-hyena operator for scalable equivariant learning, 2024

    Artem Moskalev, Mangal Prakash, Junjie Xu, Rui Liao, and Tommaso Mansi. SE(3)-hyena operator for scalable equivariant learning, 2024

    2024

  45. [54]

    Revisiting over-smoothing and over-squashing using ollivier-ricci curvature

    Khang Nguyen, Nong Minh Hieu, Vinh Duc Nguyen, Nhat Ho, Stanley Osher, and Tan Minh Nguyen. Revisiting over-smoothing and over-squashing using ollivier-ricci curvature. In International Conference on Machine Learning, pages 25956–25979. PMLR, 2023

    2023

  46. [55]

    Bridging biomolecular modalities for knowledge transfer in bio-language models

    Mangal Prakash, Artem Moskalev, Peter DiMaggio Jr, Steven Combs, Tommaso Mansi, Justin Scheer, and Rui Liao. Bridging biomolecular modalities for knowledge transfer in bio-language models. In Neurips 2024 Workshop Foundation Models for Science: Progress, Opportunities, and Challenges

    2024

  47. [56]

    Recipe for a General, Powerful, Scalable Graph Transformer

    Ladislav Rampášek, Mikhail Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a General, Powerful, Scalable Graph Transformer. Advances in Neural Information Processing Systems, 35, 2022

    2022

  48. [57]

    mrna-based therapeutics—developing a new class of drugs

    Ugur Sahin, Katalin Karikó, and Özlem Türeci. mrna-based therapeutics—developing a new class of drugs. Nature reviews Drug discovery, 13(10):759–780, 2014. 14

    2014

  49. [58]

    E (n) equivariant graph neural networks

    Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323–9332. PMLR, 2021

    2021

  50. [59]

    Schnet: A continuous-filter convolutional neural network for modeling quantum interactions

    Kristof Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert Müller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in neural information processing systems, 30, 2017

    2017

  51. [60]

    Equivariant message passing for the prediction of tensorial properties and molecular spectra

    Kristof Schütt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pages 9377–9388. PMLR, 2021

    2021

  52. [61]

    E2efold-3d: End-to-end deep learning method for accurate de novo rna 3d structure prediction

    Tao Shen, Zhihang Hu, Zhangzhi Peng, Jiayang Chen, Peng Xiong, Liang Hong, Liangzhen Zheng, Yixuan Wang, Irwin King, Sheng Wang, et al. E2efold-3d: End-to-end deep learning method for accurate de novo rna 3d structure prediction. arXiv preprint arXiv:2207.01586, 2022

    2022 arXiv

  53. [62]

    Environment and exposure to solvent of protein atoms

    Andrew Shrake and John A Rupley. Environment and exposure to solvent of protein atoms. lysozyme and insulin. Journal of molecular biology, 79(2):351–371, 1973

    1973

  54. [63]

    Rna structural dynamics as captured by molecular simulations: a comprehensive overview

    Jiri Sponer, Giovanni Bussi, Miroslav Krepl, Pavel Banašš, Sandro Bottaro, Richard A Cunha, Alejandro Gil-Ley, Giovanni Pinamonti, Simón Poblete, Petr Jurecka, et al. Rna structural dynamics as captured by molecular simulations: a comprehensive overview. Chemical reviews, 118(...

    2018

  55. [64]

    Accelerating bayesian optimization for biological sequence design with denoising autoencoders

    Samuel Stanton, Wesley Maddox, Nate Gruver, Phillip Maffettone, Emily Delaney, Peyton Greenside, and Andrew Gordon Wilson. Accelerating bayesian optimization for biological sequence design with denoising autoencoders. In International conference on machine learning, pages 2045...

    2022

  56. [65]

    Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018

    2018 arXiv

  57. [66]

    Where did the gap go? reassessing the long-range graph benchmark

    Jan Tönshoff, Martin Ritzert, Eran Rosenbluth, and Martin Grohe. Where did the gap go? reassessing the long-range graph benchmark. Transactions on Machine Learning Research, 2024

    2024

  58. [67]

    Geometric deep learning of rna structure.Science, 373(6558):1047– 1051, 2021

    Raphael JL Townshend, Stephan Eismann, Andrew M Watkins, Ramya Rangan, Masha Karelina, Rhiju Das, and Ron O Dror. Geometric deep learning of rna structure.Science, 373(6558):1047– 1051, 2021

    2021

  59. [68]

    Evaluating and learning from rna pseudotorsional space: quantitative validation of a reduced representation for rna structure

    Leven M Wadley, Kevin S Keating, Carlos M Duarte, and Anna Marie Pyle. Evaluating and learning from rna pseudotorsional space: quantitative validation of a reduced representation for rna structure. Journal of molecular biology, 372(4):942–957, 2007

    2007

  60. [69]

    Deep learning models for predicting rna degradation via dual crowdsourcing

    Hannah K Wayment-Steele, Wipapat Kladwang, Andrew M Watkins, Do Soon Kim, Bojan Tunguz, Walter Reade, Maggie Demkin, Jonathan Romano, Roger Wellington-Oguri, John J Nicol, et al. Deep learning models for predicting rna degradation via dual crowdsourcing. Nature Machine Intelli...

    2022

  61. [70]

    Nodeformer: A scalable graph structure learning transformer for node classification

    Qitian Wu, Wentao Zhao, Zenan Li, David Wipf, and Junchi Yan. Nodeformer: A scalable graph structure learning transformer for node classification. In Advances in Neural Information Processing Systems, 2022

    2022

  62. [71]

    Sgformer: Simplifying and empowering transformers for large-graph representations

    Qitian Wu, Wentao Zhao, Chenxiao Yang, Hengrui Zhang, Fan Nie, Haitian Jiang, Yatao Bian, and Junchi Yan. Sgformer: Simplifying and empowering transformers for large-graph representations. Advances in Neural Information Processing Systems, 36:64753–64773, 2023

    2023

  63. [72]

    Representing long-range context for graph neural networks with global attention

    Zhanghao Wu, Paras Jain, Matthew Wright, Azalia Mirhoseini, Joseph E Gonzalez, and Ion Stoica. Representing long-range context for graph neural networks with global attention. Advances in neural information processing systems, 34:13266–13279, 2021

    2021

  64. [73]

    A comprehensive survey on graph neural networks

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    2020

  65. [74]

    Harmony: A multi-representation framework for rna property prediction

    Junjie Xu, Artem Moskalev, Tommaso Mansi, Mangal Prakash, and Rui Liao. Harmony: A multi-representation framework for rna property prediction. In ICLR 2025 Workshop on Machine Learning for Genomics Explorations. 15

    2025

  66. [75]

    Beyond sequence: Impact of geometric context for RNA property prediction

    Junjie Xu, Artem Moskalev, Tommaso Mansi, Mangal Prakash, and Rui Liao. Beyond sequence: Impact of geometric context for RNA property prediction. In NeurIPS 2024 Workshop on AI for New Drug Modalities, 2024

    2024

  67. [76]

    Helm: Hierarchical encoding for mrna language modeling

    Mehdi Yazdani-Jahromi, Mangal Prakash, Tommaso Mansi, Artem Moskalev, and Rui Liao. Helm: Hierarchical encoding for mrna language modeling. arXiv preprint arXiv:2410.12459, 2024

    2024 arXiv

  68. [77]

    Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34:28877–28888, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34:28877–28888, 2021

    2021

  69. [78]

    Improving equivariant graph neural networks on large geometric graphs via virtual nodes learning

    Yuelin Zhang, Jiacheng Cen, Jiaqi Han, Zhiqiang Zhang, Jun Zhou, and Wenbing Huang. Improving equivariant graph neural networks on large geometric graphs via virtual nodes learning. In Forty-first International Conference on Machine Learning

  70. [79]

    Long-Range

    Dongzhuoran Zhou, Evgeny Kharlamov, and Egor V . Kostylev. GLora: A benchmark to evaluate the ability to learn long-range dependencies in graphs. In The Thirteenth International Conference on Learning Representations, 2025. 16 A Related Work A.1 Invariant GNNs Graph Neural Net...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.