Pith. sign in

REVIEW 4 major objections 4 minor 101 references

A single configuration-driven equivariant stack can be retargeted across coarse-grained backmapping, NMR chemical-shift prediction, and mixed-field generation with competitive accuracy at the cost of a configuration change.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 13:29 UTC pith:6HU5VFHD

load-bearing objection A serious software-framework paper with a genuinely new generative capability, but the empirical case for 'competitive accuracy' isn't sealed yet. the 4 major comments →

arxiv 2607.19083 v2 pith:6HU5VFHD submitted 2026-07-21 cs.LG q-bio.BM

GEqTrain: A Configuration-Driven Framework for Retargeting Equivariant Graph Neural Networks Across 3D Scientific Tasks

Daniele Angioletti , Marco Nobile , Vittorio Limongelli This is my paper
classification cs.LG q-bio.BM
keywords equivariant graph neural networksconfiguration-driven frameworktyped fieldscoarse-grained backmappingNMR chemical shift predictionflow matchingequivariant generative modelingO(3) irreps
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the main bottleneck in equivariant graph learning is software organization, not architecture: one typed, configuration-driven stack can be reused across qualitatively different 3D scientific tasks by changing configuration alone. Raw data are mapped to typed node-, edge-, and graph-level fields, while model composition, losses, and training workflows are assembled declaratively. The same local equivariant backbone is then applied to coarse-grained-to-atomistic backmapping, NMR chemical-shift prediction in molecular solids, and a synthetic generative task. The paper does not claim to beat task-specialized systems; it claims competitive accuracy from one shared infrastructure. It also introduces a generative extension that treats higher-order equivariant fields, up to angular order l=3, as first-class flow-matching targets, jointly denoising coordinates, shape descriptors, and vector-valued attributes.

Core claim

The central claim is that a shared equivariant backbone, training loop, and typed-tensor interface can be retargeted to tasks as different as hierarchical backmapping, scalar and tensorial NMR shielding prediction, and mixed-field generation, with only dataset semantics, model composition, and objectives reconfigured. On the CSD-500 NMR benchmark, a single model trained from scratch is competitive with specialized methods; on alpha-carbon-trace backmapping of disordered proteins, a one-shot equivariant reconstruction outperforms published generative baselines in paired-reference RMSD; and on the synthetic LEGO benchmark, one flow-matching model simultaneously reconstructs Cartesian coordinat

What carries the argument

The machinery is a typed-field abstraction layered on a strictly local, edge-centric equivariant interaction stack. Raw dataset tensors are mapped by configuration to named node-, edge-, and graph-level fields with declared transformation behavior under O(3); geometry enters as an invariant radial embedding plus spherical-harmonic angular edge embeddings; and an interaction module updates latent edge states through equivariant layers, with explicit reduction stages and task-specific readouts. For generation, a flow-matching front end interpolates each declared field between data and noise along a linear path, stores the scheduler velocity as the target, and appends time conditioning, so fiel

Load-bearing premise

The load-bearing premise is that a strictly local neighborhood graph—cutoff radii of 7–10 angstroms for backmapping and 6 angstroms for NMR—contains enough information for every task; if longer-range electronic or structural effects are needed, the reported accuracy may not transfer beyond the tested systems.

What would settle it

Train the same NMR configuration on molecular solids with known long-range electronic effects, such as pi-stacked aromatics, charged frameworks, or hydrogen-bond networks extending beyond the 6-angstrom cutoff, and compare per-nucleus error against a nonlocal or enlarged-cutoff model; systematic degradation with long-range coupling would falsify the local-sufficiency premise. Alternatively, run the generative extension on a real mixed-field target—ligand coordinates plus pharmacophore or polarization fields—and check whether joint generation preserves physically valid structures.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, researchers can add or change a task—from scalar regression to tensorial output, or from prediction to generation—by editing configuration rather than rewriting the model, lowering the cost of experimentation.
  • A single local equivariant stack can serve as a common substrate for molecular reconstruction, physical-observable regression, and 3D generation, suggesting that task-specific architectures are not always necessary for competitive accuracy.
  • Equivariant fields of angular order up to l=3 can be treated as first-class generative targets, so future generative models need not be limited to coordinates plus scalars; orientation, local shape, and polarization-like fields can be generated jointly.
  • Strict locality plus chunked inference enables processing of much larger graphs under a fixed GPU memory budget (the synthetic scaling test reaches roughly 28,000 nodes), at the cost of longer serial runtime.
  • On paired-reference protein backmapping, a one-shot equivariant reconstruction can beat generative diffusion baselines in RMSD, suggesting that deterministic reconstruction followed by relaxation may be sufficient in some multiscale workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the local-cutoff assumption transfers to real systems, the configuration-driven interface could make domain-specific models—ligand pose generators, crystal-structure rankers—cheap to build from one codebase; this extension is untested beyond the three benchmarks.
  • The mixed-field generation capability suggests a direct test in realistic settings: generate a ligand pose jointly with pharmacophore or polarization fields and evaluate against physical-chemistry constraints; the paper leaves this to future work.
  • The LEGO benchmark is procedural and paired-reference, so its high-fidelity scores should not be read as chemical realism; a natural next experiment would port the same mixed-field flow to a real molecular ensemble with physical validity metrics.
  • The backmapping comparison uses paired-reference RMSD, which favors deterministic reconstructions; evaluating generative baselines by ensemble observables could change conclusions, an open question the paper acknowledges.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces GEqTrain, a configuration-driven framework for building equivariant graph neural networks from typed node-, edge-, and graph-level fields via Hydra configuration, with a shared local equivariant backbone and pluggable training objectives. The authors demonstrate the framework on three tasks: hierarchical coarse-grained-to-all-atom backmapping (HEroBM), NMR chemical-shift/shielding prediction in molecular solids, and a synthetic LEGO benchmark for mixed-field equivariant generation with a flow-matching extension (GEqDiff). The central claim is that a single configurable equivariant stack can be retargeted across qualitatively different tasks and achieve competitive accuracy at the cost of a configuration change, while GEqDiff extends generative modeling to jointly transport Cartesian positions and non-scalar equivariant fields up to l=3.

Significance. If the central claim were supported, GEqTrain would be a useful organizational contribution: it offers a typed, configurable interface that lowers the software overhead of moving between predictive and generative, scalar and tensorial tasks, and it ships code under an MIT license. The LEGO benchmark is a genuinely interesting stress test for mixed-field equivariant generation, and the design of field-wise corruptions and velocity heads is clean. However, the empirical support for the key phrase 'competitive accuracy' is not yet established: the comparisons in Tables 1-4 largely rely on published baseline numbers under unspecified or inconsistent evaluation protocols, and the generative capability is measured only with author-defined, hand-calibrated composite scores. The paper is honest in its Discussion about the limitations, but the abstract and contributions still assert 'competitive accuracy' and 'high fidelity' without the controlled comparisons needed for those claims.

major comments (4)
  1. [§5.1, Table 1] The headline PDB29k comparison is not controlled. HEroBM is trained on PDB3k, which is explicitly described as 'filtered from the PDB29k dataset,' and evaluated on the 'PDB29k test set,' but the paper never defines this test split or checks for overlap with the PDB3k training/validation structures. CG2AT and cg2all values are quoted from prior publications under unspecified splits, atom selections, and evaluation scripts. The reported margins are small (0.07 vs 0.08 Å backbone; 0.24 vs 0.31 Å all-heavy-atom), so protocol differences alone could invert the ranking. Because the abstract's 'competitive accuracy' claim rests on this table, this issue is load-bearing.
  2. [§5.1, Table 2] The Cα-trace comparison mixes evaluation conventions. HEroBM reports BB, SC, and ALL RMSD values computed with its own atom selection, while GenZProt, DiAMoNDBack, LDB, and LatCPB values are global RMSDs taken from Han et al. and Zhang et al. The ALL-row comparisons therefore compare different quantities, and the paper's wording that HEroBM 'can outperform current generative alternatives' in structural fidelity is not supported by an apples-to-apples measurement. The paper should either re-run at least one baseline under HEroBM's exact evaluation protocol or explicitly state that the ALL-row comparison is not controlled.
  3. [§5.2, Tables 3 and 4] NMR results are compared against published numbers without re-running baselines under GEqTrain's preprocessing, normalization, graph construction, or train/validation splits. For Table 3, the margins versus MR-3D-DenseNet and NMRNet are small (e.g., 0.41 vs 0.37 ppm for 1H; 3.41 vs 3.30 ppm for 13C), so protocol differences could invert the ranking. Table 4 reports ShiftML2/3 on 'the same CSD-test setting' but does not document whether the authors used identical data filtering, normalization, or atom types. Given the centrality of 'competitive accuracy,' this needs to be either replaced with a controlled comparison or explicitly downgraded to 'consistent with published values under different protocols.'
  4. [§5.3, Table 5 and SI §6.6] The generative capability claim for GEqDiff is evaluated entirely on the author-defined LEGO benchmark. The validity, shape, dipole, and pose scores are composite functions with hand-set constants (e.g., S_valid = 100 exp(−12.0V_eff − 0.9N_sev)/√K, the 0.60/0.20/0.20 weighting in S_μ, and the 0.50/1.20/1.2 coefficients in S_pose). Scores of 99–100 are reported as 'near-maximal' and 'high fidelity,' but without sensitivity analysis, comparison to an independent measure, or a clear interpretation of what the score values mean, these claims measure only the authors' chosen scoring. For a synthetic benchmark that is the only evidence supporting the mixed-field generation claim, the metric design should be justified or explicitly labeled as illustrative.
minor comments (4)
  1. [SI §5.2] The NMR model uses a 6.0 Å cutoff with a strictly local radius graph. The paper acknowledges this limitation in the Discussion, but Section 5.2 should state up front that the reported NMR accuracy is conditional on the local-neighborhood assumption, since chemical shifts can be sensitive to longer-range electronic effects.
  2. [References] Reference [54] is incomplete: it lacks author and title information, showing only 'Chemically transferable generative backmapping of coarse-grained proteins' and venue data.
  3. [Table 2] The LatCPB column shows values without standard deviations, and the table caption does not explain why. If the original source reports no uncertainty, this should be stated.
  4. [Table 3] The table caption labels the blocks as 'model evaluated directly on original CSD-500' versus 'modern-reference values on a comparable subset.' The 'comparable' subset is never precisely defined; please state how the subset was selected and how 'comparable' was checked.

Circularity Check

0 steps flagged

No circular derivation chain: central claims rest on external benchmarks; acknowledged synthetic benchmark is not an input-output loop.

full rationale

Walking the derivation chain, the paper's central claim—that a shared equivariant stack can be retargeted 'at the cost of a configuration change'—is supported by external benchmark evidence (PDB29k/PED for HEroBM, CSD-500 and ShiftML3 CSD-test for NMR). These benchmarks lie outside the fitted configuration, so they provide independent evidence rather than an input-output tautology. Protocol differences with quoted baselines are a comparison-quality or correctness concern, not a circular reduction. The generative LEGO case is explicitly framed as a 'controlled, synthetic benchmark' that 'isolates a geometric capability ... not to demonstrate chemical realism,' and its metrics are independently defined in SI §6.6 and evaluated on held-out assemblies; no equation in the paper makes a reported score equal to the training objective or to a fitted parameter by construction. Normalization statistics are fit only on training data and inverted for evaluation, so no fitted input is renamed as a prediction. The self-citation [48] records the origin of HEroBM, but the accuracy numbers are external RMSD comparisons; no uniqueness theorem, smuggled ansatz, or renaming of a known result is used to force the conclusions. Thus no significant circularity is present.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim relies on standard equivariant-GNN machinery and publicly available datasets. The main extra degrees of freedom are hand-calibrated LEGO metrics and un-reported loss weights; these do not affect the external-benchmark case studies but weaken the generative demonstration.

free parameters (3)
  • LEGO metric calibration constants = various (e.g., tau_ov=0.01, severe-overlap threshold 0.08, exponential rates 12.0, 0.9)
    The validity, pose, and dipole scores are defined with hand-set tolerances and exponential penalties in SI §6.6; these constants are not derived from any external benchmark and were presumably tuned to make scores interpretable. They directly affect the reported 'high fidelity' numbers.
  • Flow-matching loss weights λ_m = not reported
    The joint training objective L_FM = Σ λ_m ||...||² uses per-field weights that are not reported in the paper; the LEGO benchmark uses λ_r, λ_s, λ_μ (SI §6.4) with values untabulated. These weights affect the balance between coordinate and field fidelity.
  • NMR tensor loss weight = 10
    In SI §5.2, L = L_iso + 10 L_tensor; the factor 10 is a chosen hyperparameter not justified by any principle.
axioms (4)
  • standard math E(3)-equivariant message passing with spherical harmonics correctly encodes molecular symmetries.
    This is the foundation of the architecture (Section 3.4) and is accepted in the field.
  • domain assumption A strictly local radius graph is sufficient to capture the relevant features for each task.
    The reference backbone uses a cutoff graph (Section 3.3) with radius 7-10 Å in backmapping and 6.0 Å in NMR (SI §5.2); long-range couplings are assumed negligible or captured within the cutoff. The paper acknowledges this limitation in Discussion.
  • domain assumption The synthetic LEGO benchmark is representative of geometric ingredients in real molecular generation.
    Section 5.3 states the benchmark 'isolates geometric ingredients that recur in molecular generation'; transfer to physical systems is explicitly left open ('transfer to physically meaningful systems remains an open problem'). If this transfer fails, the GEqDiff contribution becomes a synthetic demo.
  • domain assumption The DFT-computed chemical shifts and molecular structures in the benchmark datasets are accurate enough to serve as ground truth.
    The paper relies on CSD-2k/CSD-500 and ShiftML3 datasets as labels; any systematic DFT error is inherited.

pith-pipeline@v1.3.0-alltime-deepseek · 33177 in / 15362 out tokens · 141598 ms · 2026-08-01T13:29:12.893269+00:00 · methodology

0 comments
read the original abstract

Equivariant graph neural networks provide a powerful modeling language for three-dimensional scientific data, but their reuse is often limited by implementations tied to specific tasks, outputs, and training regimes. We present GEqTrain, a configuration-driven framework that separates dataset semantics, model composition, and training objectives. Raw data are mapped to typed node-, edge-, and graph-level fields, while model stacks, losses, and training workflows are assembled declaratively through Hydra configurations. A shared equivariant backbone and training infrastructure can therefore be retargeted to a new task primarily through configuration. We demonstrate this flexibility on three different problems handled within one software stack: coarse-grained-to-atomistic backmapping of biomolecular systems, prediction of NMR chemical shifts in molecular solids, and equivariant generative modeling. Our aim is not to surpass individually optimized task-specific systems, but to show that a shared representation and training infrastructure can achieve competitive accuracy across qualitatively different tasks at the cost of a configuration change. We further introduce GEqDiff, a generative extension based on equivariant flow matching. GEqDiff treats user-defined equivariant fields as first-class generation targets, jointly transporting Cartesian positions and non-scalar node fields spanning representations up to l=3 within a single equivariant flow. We validate this capability on a controlled synthetic benchmark inspired by protein secondary-structure motifs, showing that fields with heterogeneous transformation properties can be reconstructed jointly and with high fidelity. By reducing the software overhead of moving between predictive and generative, scalar and tensorial settings, GEqTrain aims to make equivariant modeling more reproducible, extensible, and reusable.

Figures

Figures reproduced from arXiv: 2607.19083 by Daniele Angioletti, Marco Nobile, Vittorio Limongelli.

Figure 1
Figure 1. Figure 1: Overview of the GEqTrain workflow. Raw dataset tensors are mapped to typed [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Procedural LEGO benchmark. (a) Representative alpha-helix-like LEGO assembly. [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example mixed-field generation trajectory on the LEGO benchmark. (a) Brick-level [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: Memory and runtime scaling of full and chunked GEqTrain inference. A 28.96-million-parameter model was evaluated on synthetic directed graphs with fixed out-degree 32 on a single NVIDIA RTX A6000 GPU with 47.4 GB of memory. Top: peak CUDA memory reserved by the PyTorch allocator. Bottom: mean wall-clock inference time over three executions following one warm-up. Before the chunk budget is exceeded, chunked… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

101 extracted references · 18 canonical work pages

  1. [1]

    Stein and Farnaz Heidar-Zadeh and Meili Liu and Martin Head-Gordon and Luke Bertels and Hongxia Hao and Itai Leven and Teresa Head-Gordon , doi =

    Mojtaba Haghighatlari and Jie Li and Xingyi Guan and Oufan Zhang and Akshaya Das and Christopher J. Stein and Farnaz Heidar-Zadeh and Meili Liu and Martin Head-Gordon and Luke Bertels and Hongxia Hao and Itai Leven and Teresa Head-Gordon , doi =. NewtonNet: a Newtonian message passing network for deep learning of interatomic potentials and forces , volume...

  2. [2]

    and Chmiela, Stefan and Sauceda, Huziel E

    Unke, Oliver T. and Chmiela, Stefan and Sauceda, Huziel E. and Gastegger, Michael and Poltavsky, Igor and Sch. Machine Learning Force Fields , journal =. 2021 , doi =

    2021

  3. [3]

    International Conference on Learning Representations , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations , year=

  4. [4]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    E(n) Equivariant Graph Neural Networks , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  5. [5]

    Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties , author =. Phys. Rev. Lett. , volume =. 2018 , month =. doi:10.1103/PhysRevLett.120.145301 , url =

    doi:10.1103/physrevlett.120.145301 2018

  6. [6]

    and Riley, Patrick F

    Gilmer, Justin and Schoenholz, Samuel S. and Riley, Patrick F. and Vinyals, Oriol and Dahl, George E. , title =. Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =. 2017 , publisher =

    2017

  7. [7]

    Sauceda and Igor Poltavsky and Kristof T

    Stefan Chmiela and Alexandre Tkatchenko and Huziel E. Sauceda and Igor Poltavsky and Kristof T. Schütt and Klaus-Robert Müller , title =. Science Advances , volume =. 2017 , doi =. https://www.science.org/doi/pdf/10.1126/sciadv.1603015 , abstract =

    doi:10.1126/sciadv.1603015 2017

  8. [8]

    Stokes and Kevin Yang and Kyle Swanson and Wengong Jin and Andres Cubillos-Ruiz and Nina M

    Jonathan M. Stokes and Kevin Yang and Kyle Swanson and Wengong Jin and Andres Cubillos-Ruiz and Nina M. Donghia and Craig R. MacNair and Shawn French and Lindsey A. Carfrae and Zohar Bloom-Ackermann and Victoria M. Tran and Anush Chiappino-Pepe and Ahmed H. Badran and Ian W. Andrews and Emma J. Chory and George M. Church and Eric D. Brown and Tommi S. Jaa...

  9. [9]

    Utilizing graph machine learning within drug discovery and development , volume =

    Thomas Gaudelet and Ben Day and Arian R Jamasb and Jyothish Soman and Cristian Regep and Gertrude Liu and Jeremy B R Hayter and Richard Vickers and Charles Roberts and Jian Tang and David Roblin and Tom L Blundell and Michael M Bronstein and Jake P Taylor-King , doi =. Utilizing graph machine learning within drug discovery and development , volume =. Brie...

  10. [10]

    Graph neural networks for conditional de novo drug design , volume =

    Carlo Abate and Sergio Decherchi and Andrea Cavalli , doi =. Graph neural networks for conditional de novo drug design , volume =. WIREs Computational Molecular Science , month =

  11. [11]

    John Jumper and Richard Evans and Alexander Pritzel and Tim Green and Michael Figurnov and Olaf Ronneberger and Kathryn Tunyasuvunakool and Russ Bates and Augustin Žídek and Anna Potapenko and Alex Bridgland and Clemens Meyer and Simon A. A. Kohl and Andrew J. Ballard and Andrew Cowie and Bernardino Romera-Paredes and Stanislav Nikolov and Rishub Jain and...

  12. [12]

    Generative Models for Graph-Based Protein Design , url =

    Ingraham, John and Garg, Vikas and Barzilay, Regina and Jaakkola, Tommi , booktitle =. Generative Models for Graph-Based Protein Design , url =

  13. [13]

    Finkler and Stefan Goedecker and Jörg Behler , doi =

    Tsz Wai Ko and Jonas A. Finkler and Stefan Goedecker and Jörg Behler , doi =. A fourth-generation high-dimensional neural network potential with accurate electrostatics including non-local charge transfer , volume =. Nature Communications , month =

  14. [14]

    2019 , url =

    Omry Yadan , title =. 2019 , url =

    2019

  15. [15]

    Graph neural networks for materials science and chemistry , volume =

    Patrick Reiser and Marlen Neubert and André Eberhard and Luca Torresi and Chen Zhou and Chen Shao and Houssam Metni and Clint van Hoesel and Henrik Schopmans and Timo Sommer and Pascal Friederich , doi =. Graph neural networks for materials science and chemistry , volume =. Communications Materials , month =

  16. [16]

    Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces , author =. Phys. Rev. Lett. , volume =. 2007 , month =. doi:10.1103/PhysRevLett.98.146401 , url =

    doi:10.1103/physrevlett.98.146401 2007

  17. [17]

    The Journal of Chemical Physics , volume =

    Behler, Jörg , title =. The Journal of Chemical Physics , volume =. 2011 , month =. doi:10.1063/1.3553717 , url =

    doi:10.1063/1.3553717 2011

  18. [18]

    and Bruna, Joan and LeCun, Yann and Szlam, Arthur and Vandergheynst, Pierre , journal=

    Bronstein, Michael M. and Bruna, Joan and LeCun, Yann and Szlam, Arthur and Vandergheynst, Pierre , journal=. Geometric Deep Learning: Going beyond Euclidean data , year=

  19. [19]

    Bronstein and Joan Bruna and Taco Cohen and Petar Velickovic , title =

    Michael M. Bronstein and Joan Bruna and Taco Cohen and Petar Velickovic , title =. CoRR , volume =. 2021 , url =. 2104.13478 , timestamp =

    Pith/arXiv arXiv 2021

  20. [20]

    K. T. Schütt and H. E. Sauceda and P.-J. Kindermans and A. Tkatchenko and K.-R. Müller , doi =. SchNet – A deep learning architecture for molecules and materials , volume =. The Journal of Chemical Physics , month =

  21. [21]

    2022 , eprint=

    Directional Message Passing for Molecular Graphs , author=. 2022 , eprint=

    2022

  22. [22]

    Battaglia and Jessica B

    Peter W. Battaglia and Jessica B. Hamrick and Victor Bapst and Alvaro Sanchez. Relational inductive biases, deep learning, and graph networks , journal =. 2018 , url =. 1806.01261 , timestamp =

    Pith/arXiv arXiv 2018

  23. [23]

    and Ying, Rex and Leskovec, Jure , title =

    Hamilton, William L. and Ying, Rex and Leskovec, Jure , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

    2017

  24. [24]

    Structure-based protein and small molecule generation using EGNN and diffusion models: A comprehensive review , journal =

    Farzan Soleymani and Eric Paquet and Herna Lydia Viktor and Wojtek Michalowski , keywords =. Structure-based protein and small molecule generation using EGNN and diffusion models: A comprehensive review , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.csbj.2024.06.021 , url =

    work page doi:10.1016/j.csbj.2024.06.021 2024

  25. [25]

    Protein Interface Prediction using Graph Convolutional Networks , url =

    Fout, Alex and Byrd, Jonathon and Shariat, Basir and Ben-Hur, Asa , booktitle =. Protein Interface Prediction using Graph Convolutional Networks , url =

  26. [26]

    Improving cancer driver gene identification using multi-task learning on graph convolutional network , volume =

    Wei Peng and Qi Tang and Wei Dai and Tielin Chen , doi =. Improving cancer driver gene identification using multi-task learning on graph convolutional network , volume =. Briefings in Bioinformatics , month =

  27. [27]

    2021 , eprint=

    Learning from Protein Structure with Geometric Vector Perceptrons , author=. 2021 , eprint=

    2021

  28. [28]

    Tianjian Liang and Ze-Yu Sun and Rieko Ishima and Xiang-Qun Xie and Ying Xue and Wei Li and Zhiwei Feng , doi =. ProstaNet: A Novel Geometric Vector Perceptrons–Graph Neural Network Algorithm for Protein Stability Prediction in Single- and Multiple-Point Mutations with Experimental Validation , volume =. Research , month =

  29. [29]

    Mailoa and Mordechai Kornbluth and Nicola Molinari and Tess E

    Simon Batzner and Albert Musaelian and Lixin Sun and Mario Geiger and Jonathan P. Mailoa and Mordechai Kornbluth and Nicola Molinari and Tess E. Smidt and Boris Kozinsky , doi =. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials , volume =. Nature Communications , month =

  30. [30]

    Owen and Mordechai Kornbluth and Boris Kozinsky , doi =

    Albert Musaelian and Simon Batzner and Anders Johansson and Lixin Sun and Cameron J. Owen and Mordechai Kornbluth and Boris Kozinsky , doi =. Learning local equivariant representations for large-scale atomistic dynamics , volume =. Nature Communications , month =

  31. [31]

    Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , articleno =

    Kozinsky, Boris and Musaelian, Albert and Johansson, Anders and Batzner, Simon , title =. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , articleno =. 2023 , isbn =. doi:10.1145/3581784.3627041 , abstract =

    arXiv 2023

  32. [32]

    The Thirteenth International Conference on Learning Representations , year=

    GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks , author=. The Thirteenth International Conference on Learning Representations , year=

  33. [33]

    Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =

    Simeon, Guillem and De Fabritiis, Gianni , title =. Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =. 2023 , publisher =

    2023

  34. [34]

    Kinch and R

    Minkyung Baek and Frank DiMaio and Ivan Anishchenko and Justas Dauparas and Sergey Ovchinnikov and Gyu Rie Lee and Jue Wang and Qian Cong and Lisa N. Kinch and R. Dustin Schaeffer and Claudia Millán and Hahnbeom Park and Carson Adams and Caleb R. Glassman and Andy DeGiovanni and Jose H. Pereira and Andria V. Rodrigues and Alberdina A. van Dijk and Ana C. ...

  35. [35]

    Dral and Matthias Rupp and O

    Raghunathan Ramakrishnan and Pavlo O. Dral and Matthias Rupp and O. Anatole von Lilienfeld , doi =. Quantum chemistry structures and properties of 134 kilo molecules , volume =. Scientific Data , month =

  36. [36]

    2024 , eprint=

    EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations , author=. 2024 , eprint=

    2024

  37. [37]

    Pre-training with fractional denoising to enhance molecular property prediction , volume =

    Yuyan Ni and Shikun Feng and Xin Hong and Yuancheng Sun and Wei-Ying Ma and Zhi-Ming Ma and Qiwei Ye and Yanyan Lan , doi =. Pre-training with fractional denoising to enhance molecular property prediction , volume =. Nature Machine Intelligence , month =

  38. [38]

    CoRR , volume =

    Zhao Xu and Youzhi Luo and Xuan Zhang and Xinyi Xu and Yaochen Xie and Meng Liu and Kaleb Dickerson and Cheng Deng and Maho Nakata and Shuiwang Ji , title =. CoRR , volume =. 2021 , url =. 2110.01717 , timestamp =

    Pith/arXiv arXiv 2021

  39. [39]

    Pelaez and Guillem Simeon and Raimondas Galvelis and Antonio Mirarchi and Peter Eastman and Stefan Doerr and Philipp Thölke and Thomas E

    Raul P. Pelaez and Guillem Simeon and Raimondas Galvelis and Antonio Mirarchi and Peter Eastman and Stefan Doerr and Philipp Thölke and Thomas E. Markland and Gianni De Fabritiis , doi =. TorchMD-Net 2.0: Fast Neural Network Potentials for Molecular Simulations , volume =. Journal of Chemical Theory and Computation , month =

  40. [40]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Equivariant message passing for the prediction of tensorial properties and molecular spectra , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  41. [41]

    Scientific Data , year=

    Nandi, Surajit and Vegge, Tejs and Bhowmik, Arghya , title=. Scientific Data , year=. doi:10.1038/s41597-023-02690-2 , url=

    work page doi:10.1038/s41597-023-02690-2

  42. [42]

    and Khanna, Smayan and Ferguson, Andrew L

    Jones, Michael S. and Khanna, Smayan and Ferguson, Andrew L. , title=. Journal of Chemical Information and Modeling , year=. doi:10.1021/acs.jcim.4c02046 , url=

    work page doi:10.1021/acs.jcim.4c02046

  43. [43]

    2025 , eprint=

    FlowBack-Adjoint: Physics-Aware and Energy-Guided Conditional Flow-Matching for All-Atom Protein Backmapping , author=. 2025 , eprint=

    2025

  44. [44]

    Journal of Chemical Information and Modeling , year=

    Ugarte La Torre, Diego and Sugita, Yuji , title=. Journal of Chemical Information and Modeling , year=. doi:10.1021/acs.jcim.5c01281 , url=

    work page doi:10.1021/acs.jcim.5c01281

  45. [45]

    , title=

    Waltmann, Curt and Wang, Yihang and Yang, Chengxi and Kim, Siyoung and Voth, Gregory A. , title=. Journal of Chemical Theory and Computation , year=. doi:10.1021/acs.jctc.5c00459 , url=

    work page doi:10.1021/acs.jctc.5c00459

  46. [46]

    Journal of Cheminformatics , year=

    Chen, Haochen and Liang, Tao and Tan, Kai and Wu, Anan and Lu, Xin , title=. Journal of Cheminformatics , year=. doi:10.1186/s13321-024-00927-9 , url=

    work page doi:10.1186/s13321-024-00927-9

  47. [47]

    and Widmalm, Göran and Rönnols, Jerk

    Bånkestad, Maria and Dorst, Kevin M. and Widmalm, Göran and Rönnols, Jerk. Carbohydrate NMR chemical shift prediction by GeqShift employing E(3) equivariant graph neural networks. RSC Adv. 2024. doi:10.1039/D4RA03428G

    doi:10.1039/d4ra03428g 2024

  48. [48]

    Williamson and S

    D. Williamson and S. Ponte and I. Iglesias and N. Tonge and C. Cobas and E.K. Kemsley , abstract =. Chemical shift prediction in 13C NMR spectroscopy using ensembles of message passing neural networks (MPNNs) , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.jmr.2024.107795 , url =

    arXiv 2024

  49. [49]

    Communications Chemistry , year=

    Li, Yunrui and Xu, Hao and Kumar, Ambrish and Wang, Duo-Sheng and Heiss, Christian and Azadi, Parastoo and Hong, Pengyu , title=. Communications Chemistry , year=. doi:10.1038/s42004-025-01455-9 , url=

    work page doi:10.1038/s42004-025-01455-9

  50. [50]

    and Zheng, Shuangjia , title=

    Lu, Wei and Zhang, Jixian and Huang, Weifeng and Zhang, Ziqiao and Jia, Xiangyu and Wang, Zhenyu and Shi, Leilei and Li, Chengtao and Wolynes, Peter G. and Zheng, Shuangjia , title=. Nature Communications , year=. doi:10.1038/s41467-024-45461-2 , url=

    work page doi:10.1038/s41467-024-45461-2

  51. [51]

    Structure prediction of protein-ligand complexes from sequence information with Umol , journal=

    Bryant, Patrick and Kelkar, Atharva and Guljas, Andrea and Clementi, Cecilia and No. Structure prediction of protein-ligand complexes from sequence information with Umol , journal=. 2024 , month=. doi:10.1038/s41467-024-48837-6 , url=

    work page doi:10.1038/s41467-024-48837-6 2024

  52. [52]

    2025 , doi =

    Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy and Reveiz, Mateo and Thaler, Stephan and Somnath, Vignesh Ram and Getz, Noah and Portnoi, Tally and Roy, Julien and Stark, Hannes and Kwabi-Addo, David and Beaini, Dominique and Jaakkola, Tommi and Barzilay, Regina , title =. 2025 , doi =. https://www.biorxiv.org/content/early/2025/06/18/2025.06.14.6...

    2025

  53. [53]

    and Bambrick, Joshua and Bodenstein, Sebastian W

    Abramson, Josh and Adler, Jonas and Dunger, Jack and Evans, Richard and Green, Tim and Pritzel, Alexander and Ronneberger, Olaf and Willmore, Lindsay and Ballard, Andrew J. and Bambrick, Joshua and Bodenstein, Sebastian W. and Evans, David A. and Hung, Chia-Chun and O'Neill, Michael and Reiman, David and Tunyasuvunakool, Kathryn and Wu, Zachary and. Accur...

    doi:10.1038/s41586-024-07487-w 2024

  54. [54]

    and Anandkumar, Animashree , title=

    Qiao, Zhuoran and Nie, Weili and Vahdat, Arash and Miller, Thomas F. and Anandkumar, Animashree , title=. Nature Machine Intelligence , year=. doi:10.1038/s42256-024-00792-z , url=

    work page doi:10.1038/s42256-024-00792-z

  55. [55]

    and Lio, Pietro and Welling, Max and Bronstein, Michael and Correia, Bruno , title=

    Schneuing, Arne and Harris, Charles and Du, Yuanqi and Didi, Kieran and Jamasb, Arian and Igashov, Ilia and Du, Weitao and Gomes, Carla and Blundell, Tom L. and Lio, Pietro and Welling, Max and Bronstein, Michael and Correia, Bruno , title=. Nature Computational Science , year=. doi:10.1038/s43588-024-00737-x , url=

    doi:10.1038/s43588-024-00737-x

  56. [56]

    Geometric Transformer with Interatomic Positional Encoding , url =

    Wang, Yusong and Li, Shaoning and Wang, Tong and Shao, Bin and Zheng, Nanning and Liu, Tie-Yan , booktitle =. Geometric Transformer with Interatomic Positional Encoding , url =

  57. [57]

    Unke and Adil Kabylda and Huziel E

    Stefan Chmiela and Valentin Vassilev-Galindo and Oliver T. Unke and Adil Kabylda and Huziel E. Sauceda and Alexandre Tkatchenko and Klaus-Robert Müller , doi =. Accurate global machine learning force fields for molecules with hundreds of atoms , volume =. Science Advances , month =

  58. [58]

    Thorben Frank and Oliver T

    J. Thorben Frank and Oliver T. Unke and Klaus-Robert Müller and Stefan Chmiela , doi =. A Euclidean transformer for fast and stable machine learned force fields , volume =. Nature Communications , month =

  59. [59]

    and Lewis, Steven M

    Hintze, Bradley J. and Lewis, Steven M. and Richardson, Jane S. and Richardson, David C. , title =. Proteins: Structure, Function, and Bioinformatics , volume =. doi:https://doi.org/10.1002/prot.25039 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/prot.25039 , abstract =

    work page doi:10.1002/prot.25039

  60. [60]

    Bioinformatics , volume =

    Wang, Guoli and Dunbrack, Roland L., Jr , title = ". Bioinformatics , volume =. 2003 , month =. doi:10.1093/bioinformatics/btg224 , url =

    work page doi:10.1093/bioinformatics/btg224 2003

  61. [61]

    PED in 2021: a major update of the protein ensemble database for intrinsically disordered proteins , volume =

    Tamas Lazar and Elizabeth Martínez-Pérez and Federica Quaglia and András Hatos and Lucía B Chemes and Javier A Iserte and Nicolás A Méndez and Nicolás A Garrone and Tadeo E Saldaño and Julia Marchetti and Ana Julia Velez Rueda and Pau Bernadó and Martin Blackledge and Tiago N Cordeiro and Eric Fagerberg and Julie D Forman-Kay and Maria S Fornasari and Tob...

    2021

  62. [62]

    Chemically transferable generative backmapping of coarse-grained proteins , year =

    Yang, Soojung and G\'. Chemically transferable generative backmapping of coarse-grained proteins , year =. Proceedings of the 40th International Conference on Machine Learning , articleno =

  63. [63]

    Jones and Kirill Shmilovich and Andrew L

    Michael S. Jones and Kirill Shmilovich and Andrew L. Ferguson , doi =. DiAMoNDBack: Diffusion-Denoising Autoregressive Model for Non-Deterministic Backmapping of C-alpha Protein Traces , volume =. Journal of Chemical Theory and Computation , month =

  64. [64]

    Gumbart , doi =

    Yui Tik Pang and Lixinhao Yang and James C. Gumbart , doi =. From simple to complex: Reconstructing all-atom structures from coarse-grained models using cg2all , volume =. Structure , month =

  65. [65]

    Paulo C. T. Souza and Riccardo Alessandri and Jonathan Barnoud and Sebastian Thallmair and Ignacio Faustino and Fabian Grünewald and Ilias Patmanidis and Haleh Abdizadeh and Bart M. H. Bruininks and Tsjerk A. Wassenaar and Peter C. Kroon and Josef Melcr and Vincent Nieto and Valentina Corradi and Hanif M. Khan and Jan Domański and Matti Javanainen and Hec...

  66. [66]

    Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu , journal=

  67. [67]

    Experiment Tracking with Weights and Biases , year =

  68. [68]

    2024 , eprint=

    HEroBM: a deep equivariant graph neural network for universal backmapping from coarse-grained to all-atom representations , author=. 2024 , eprint=

    2024

  69. [69]

    Paruzzo and Albert Hofstetter and Félix Musil and Sandip De and Michele Ceriotti and Lyndon Emsley , doi =

    Federico M. Paruzzo and Albert Hofstetter and Félix Musil and Sandip De and Michele Ceriotti and Lyndon Emsley , doi =. Chemical shifts in molecular solids by machine learning , volume =. Nature Communications , month =

  70. [70]

    2024 , eprint=

    Towards a Unified Benchmark and Framework for Deep Learning-Based Prediction of Nuclear Magnetic Resonance Chemical Shifts , author=. 2024 , eprint=

    2024

  71. [71]

    The Journal of Chemical Physics , volume =

    Angioletti, Daniele and Raniolo, Stefano and Limongelli, Vittorio , title =. The Journal of Chemical Physics , volume =. 2025 , month =. doi:10.1063/5.0280330 , url =

    work page doi:10.1063/5.0280330 2025

  72. [72]

    , title =

    Yim, Jason and Stärk, Hannes and Corso, Gabriele and Jing, Bowen and Barzilay, Regina and Jaakkola, Tommi S. , title =. WIREs Computational Molecular Science , volume =. doi:https://doi.org/10.1002/wcms.1711 , url =. https://wires.onlinelibrary.wiley.com/doi/pdf/10.1002/wcms.1711 , abstract =

    work page doi:10.1002/wcms.1711

  73. [73]

    Proceedings of the 32nd International Conference on Machine Learning , pages =

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics , author =. Proceedings of the 32nd International Conference on Machine Learning , pages =. 2015 , editor =

    2015

  74. [74]

    Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , title =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =. 2020 , isbn =

    2020

  75. [75]

    2021 , eprint=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. 2021 , eprint=

    2021

  76. [76]

    Watson and David Juergens and Nathaniel R

    Joseph L. Watson and David Juergens and Nathaniel R. Bennett and Brian L. Trippe and Jason Yim and Helen E. Eisenach and Woody Ahern and Andrew J. Borst and Robert J. Ragotte and Lukas F. Milles and Basile I. M. Wicky and Nikita Hanikel and Samuel J. Pellock and Alexis Courbet and William Sheffler and Jue Wang and Preetham Venkatesh and Isaac Sappington a...

  77. [77]

    Structure-based drug design with geometric deep learning , journal =

    Clemens Isert and Kenneth Atz and Gisbert Schneider , abstract =. Structure-based drug design with geometric deep learning , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.sbi.2023.102548 , url =

    arXiv 2023

  78. [78]

    International Conference on Learning Representations , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations , year=

  79. [79]

    2022 , copyright =

    Geiger, Mario and Smidt, Tess , title =. 2022 , copyright =. doi:10.48550/ARXIV.2207.09453 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.09453 2022

  80. [80]

    2018 , eprint=

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds , author=. 2018 , eprint=

    2018

Showing first 80 references.