Pith. sign in

REVIEW 3 major objections 6 minor 94 references

Latent features from frozen atomistic foundation models can be repurposed as fixed-length, chirality-aware 3D molecular descriptors that match or exceed 2D fingerprint baselines on drug-property benchmarks.

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 11:07 UTC pith:NHDMQ5YW

load-bearing objection Rem3Di is a solid, well-benchmarked methods paper whose pseudoscalar chirality encoding is a real contribution; the denoising-pretraining story is oversold. the 3 major comments →

arxiv 2607.19977 v1 pith:NHDMQ5YW submitted 2026-07-22 physics.chem-ph physics.atm-clusphysics.comp-ph

Rem3Di: Learning smooth, chiral 3D molecular descriptors from atomistic foundation models

Steffen Wedig , Felix Burton , Rokas Elijo\v{s}ius , Christoph Schran , Lars L. Schaaf This is my paper
classification physics.chem-ph physics.atm-clusphysics.comp-ph
keywords molecular descriptorsatomistic foundation modelsmachine-learned interatomic potentialschiralitypseudoscalar featuresdenoising pretrainingdrug-property predictiontransition-metal complexes
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.

Rem3Di tests a transfer claim: the internal per-atom features of machine-learned interatomic potentials, trained only to reproduce quantum-mechanical energies and forces, carry enough chemistry to serve as molecular descriptors for drug-property prediction. The framework aggregates those frozen features into a fixed-length, atom-order-invariant vector using attention pooling, adds pseudoscalar channels that flip sign under mirror reflection so enantiomers are distinguished, and pretrains the aggregation on unlabelled molecules by denoising corrupted atom features in an information-bottleneck setup. On public drug-property benchmarks the descriptor matches or exceeds published baselines without classical 2D fingerprints, with the largest gains on properties governed by 3D physics such as hydration free energy; on a scaffold-split optical-rotation benchmark it stays above chance where 2D fingerprints fall to near chance. The same descriptor also groups transition-metal complexes by metal centre, ligand chemistry, and geometry without predefined bonding rules. A sympathetic reader would care because the paper offers a route from simulation-trained representations to chemistry-aware machine learning, including stereochemistry that ordinary rotation-invariant features cannot express, without handcrafted features or task-specific labels at descriptor-building time.

Core claim

Core claim: latent per-atom features of frozen atomistic foundation models serve as transferable molecular descriptors once the right aggregation is learned. Rem3Di contracts them into a fixed-size embedding; pseudoscalar channels from two spherical-tensor products flip sign under reflection, separating enantiomers. Aggregation is learned without labels by denoising: the clean descriptor is the only clean input to a decoder reconstructing corrupted atom features. Evidence: the descriptor matches or exceeds published baselines on drug-property regression, predicts optical-rotation sign above chance on scaffold splits, and organises transition-metal complexes without bonding rules.

What carries the argument

The load-bearing object is the molecular descriptor: a fixed-length vector built from a frozen interatomic potential's per-atom features. Three mechanisms carry the argument. Pair-biased self-attention with distances encoded in a Bessel basis out to 32 Å contextualises potential features before attention-based set pooling contracts them; a chiral encoder forms pseudoscalars via two tensor products of equivariant features (minimal case: the scalar triple product c·(a×b)), giving rotation-invariant, reflection-odd channels; and a denoising objective forces the descriptor to be an information bottleneck by reconstructing noise-corrupted atom features with the clean descriptor as the only global

Load-bearing premise

The entire descriptor inherits its chemistry from the frozen backbone potential's latent space, and a single generated 3D conformer stands in for each molecule on the benchmarks; if either premise fails, Rem3Di cannot recover the missing signal.

What would settle it

Keep the Rem3Di aggregation fixed and swap the frozen backbone for one whose latent space provably lacks the relevant chemistry (or permute its feature channels as a control): the paper's own backbone ablation spans 0.63 to 0.16 on a normalised scale, so accuracy should collapse toward the parameter-free mean-aggregation baseline if the inheritance claim is right, and should not if the aggregation itself supplies the chemistry. Separately, benchmark-scale evaluation over several generated conformers per molecule would settle the conformer premise: if per-conformer descriptor variance is large

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

If this is right

  • Drug-property prediction can proceed without hand-crafted 2D fingerprints: the fine-tuned descriptor is the best matched-protocol model on all six regression endpoints tested (solubility, hydration free energy, Caco-2 permeability, acute toxicity, lipophilicity, plasma-protein binding), with its largest margin on hydration free energy, a property governed by 3D physics.
  • Stereochemistry becomes accessible to descriptor-based pipelines: under a scaffold split of an optical-rotation benchmark, the descriptor predicts stereocentre handedness at 77% and optical-rotation sign at 70% accuracy while 2D fingerprints fall near chance, so the pseudoscalar channels genuinely separate mirror-image molecules.
  • Transition-metal chemistry, where covalent bonding is ambiguous and 2D graphs are brittle, can be organised by the same descriptor: pretrained without labels, it clusters a large complex dataset by metal centre, geometry, and ligand chemistry.
  • Descriptor quality is dominated by the frozen backbone potential (normalised 0.63 for the strongest versus 0.16 for the weakest in the ablation) and requires both denoising pretraining and fine-tuning (0.96 combined versus 0.63 for the pretrained frozen descriptor, which only matches a parameter-free mean-aggregation baseline).
  • Pretraining needs no experimental labels: the denoising objective on unlabelled molecular corpora supplies the aggregation rules, and downstream regression accuracy grows with the amount of unlabelled pretraining data.

Where Pith is reading between the lines

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

  • A natural extension the authors leave implicit: because the descriptor is differentiable with respect to atomic coordinates — which the outlook itself notes — it could serve as a collective variable for enhanced-sampling simulations, letting the same embedding that predicts properties also drive exploration of conformational space.
  • The backbone-dependence result (0.63 vs 0.16 on the paper's normalised scale) implies an inheritance law the paper does not state: descriptor quality should track the latent-space quality of the underlying potential, so re-running the identical aggregation on newer backbones would directly measure whether Rem3Di's ceiling rises as foundation models improve.
  • The single-conformer protocol is the fragile link: all benchmarks use one generated top-one conformer per molecule, and only a short (200 fs) stability test on one small-molecule dataset supports conformer stability. In my reading, a benchmark-scale multi-conformer evaluation is needed before the smoothness claim is fully tested.
  • If the pseudoscalar channels genuinely encode handedness, they should transfer to endpoints the paper does not run — enantioselective binding, chiral-metabolite toxicity, or chromatographic retention — where the two mirror forms differ in outcome; predicting which enantiomer is the active or toxic one would be a sharp, falsifiable use of the same descriptor.

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

3 major / 6 minor

Summary. The paper introduces Rem3Di, a framework that converts per-atom latent features from frozen atomistic foundation MLIPs into a fixed-length, permutation-invariant, chirality-aware molecular descriptor. The descriptor is obtained by a transformer encoder with pair-biased attention and pooling-by-multi-head-attention; a chiral branch constructs pseudoscalar features via two successive spherical tensor products, so that the descriptor is O(3)-equivariant with parity-odd channels. The encoder is pretrained with a self-supervised denoising objective on unlabelled molecular corpora, then used frozen, with LoRA adapters, or fully fine-tuned for property prediction. The paper reports results on TDC/MoleculeNet drug-property benchmarks, a QM9-derived optical-rotation chiral benchmark under random and scaffold splits, and a qualitative UMAP analysis of transition-metal complexes.

Significance. If the central claims hold, Rem3Di provides a practical route from simulation-trained MLIP representations to transferable molecular descriptors, with a mathematically clean construction of pseudoscalar features that can distinguish enantiomers. The paper is unusually transparent about benchmark protocols: matched-protocol baseline bands, scaffold splits, coverage flags, and complete per-task results are reported, and the code is released. The pseudoscalar derivation (Section IIC) is a genuine, parameter-free consequence of O(3) representation theory and is a positive contribution. However, the evidence that the self-supervised pretraining produces a chemically meaningful global descriptor is weaker than the text claims: the information-bottleneck argument is not logically forced, and the paper's own ablation shows the pretrained frozen descriptor performs no better than a parameter-free mean-aggregation baseline.

major comments (3)
  1. [§II E 1, Eq. (20), Algorithm 1, Fig. 6] The central no-label pretraining claim rests on the statement that 'the only trainable global information path from the clean molecule to the decoder is through cross-attention to the descriptor M', making M an information bottleneck. This is not logically established. The decoder receives the corrupted per-atom features \tilde{S}^0 as its primary input; with noise scale σ=0.3 applied to standardized features, each atom's clean features can be approximately recovered from its own corrupted features with little or no global context. Nothing in the denoising loss (Eq. 20) forces the decoder to use M, and the VICReg variance and covariance regularizers are 'off' by default (Table II). The paper's own ablation (Fig. 6, left) is consistent with this concern: the pretrained frozen descriptor scores 0.63, identical to the parameter-free mean-aggregation baseline, while the large gain to 0.96 ap
  2. [§II C, §III B (QM9-OR)] The pseudoscalar construction is built atom-wise from local equivariant features (Eq. 11, Eq. 13). For a molecule with a single stereocentre, this local atom-wise pseudoscalar can capture handedness, and the QM9-OR results support that. However, the paper's broader claim that 'the descriptor distinguish enantiomers' and that it captures 'molecular handedness' is stronger than what is demonstrated: for multiple stereocentres, axial/planar chirality, or cases where the chiral information is delocalized, local atom-wise pseudoscalars may not suffice. The ChiralCat results in Appendix D do not resolve this, since the authors themselves state that the benchmark does not distinguish R/S configuration and that comparisons are of limited value. I suggest either narrowing the claims to local stereocentres or adding experiments on molecules with multiple stereocentres where a global pseudo-scalar
  3. [§III A, §III C] The descriptor is claimed to vary smoothly with 3D structure and to be conformation-aware, but all benchmark results use a single RDKit top-one conformer per molecule. The only conformer-robustness evidence is the 200 fs QM9 MD stability test in Fig. 4a, which is not a benchmark-level study. A load-bearing premise of the drug-property results is that single-conformer geometry is adequate for these tasks; this is plausible but unverified. Please provide a conformer-ensemble ablation on at least one regression and one classification endpoint (e.g., FreeSolv and BBB), or explicitly state and justify the single-conformer approximation as a known limitation. Without this, the smoothness/conformation-awareness claim is not quantitatively connected to the benchmark results.
minor comments (6)
  1. [Table I] The header contains a typo: 'toxisity' should be 'toxicity'.
  2. [§II E 1] Typo: 'direclty' should be 'directly'.
  3. [Appendix C, Algorithm 1, line 8] The stop-gradient notation sg(S0) is used, but it is not defined in the main text or the algorithm caption. Please define it explicitly, since the distinction between detached and non-detached features is important for the denoising objective.
  4. [Fig. 6] The annotation showing pretrained-frozen (0.63) equal to mean aggregation (0.63) is visually striking but the caption could be more explicit about its implication: it means the pretrained frozen descriptor does not outperform a parameter-free average of MLIP features on these tasks.
  5. [Appendix D, Table III] Rem3Di results are evaluated on a different split from the literature baselines. The text discloses this, but the table caption should state it directly so a casual reader does not make cross-split comparisons.
  6. [§III E / Fig. 7] The transition-metal analysis is purely qualitative (UMAP clustering). If the authors want to make a stronger claim about descriptor utility for TMCs, a quantitative evaluation (e.g., property prediction or similarity retrieval against RAC baselines) would be needed. As written, the qualitative claim is appropriate but should be labeled as such.

Circularity Check

0 steps flagged

No significant circularity: the pseudoscalar construction is a self-contained O(3) representation-theoretic derivation, and all downstream benchmarks use external labels.

full rationale

Walking the claimed derivation chain, no load-bearing step reduces to its own input. (i) Pseudoscalar construction (Sec. II C, Eqs. 1-15): the claim that two tensor products are required to reach a pseudoscalar follows directly from O(3) parity bookkeeping. A single product with proper inputs (pi_i = (-1)^{l_i}) forces l_a = l_b and even parity, so 'a single product therefore yields, at best, a pseudotensor of nonzero degree'; two products give parity pi_a*pi_b*pi_c = -1 iff l_a+l_b+l_c is odd (Eq. 13). The l=1 path is explicitly identified with the classical scalar triple product c.(a x b) (Eq. 15). This transparent identification is a derivation from standard Clebsch-Gordan algebra, not a renamed input, and no fitted label enters it. (ii) Aggregation: permutation invariance is constructed via identical per-atom/pair operations and PMA pooling (Sec. II D); geometry enters only as radial Bessel features. (iii) Denoising pretraining (Eq. 20, Algorithm 1): the loss is masked, variance-normalised MSE on corrupted per-atom features; pretraining corpora (PCQM4M, GEOM-drugs, tmQM) are unlabelled. The 'information bottleneck' claim (Sec. II E 1) is an architectural plausibility argument, not a derivation step; whether the decoder can shortcut via the noisy per-atom features is a robustness question, not a circular reduction. The paper's own ablation honestly tempers this claim: 'the pretrained frozen descriptor (0.63) matches both the randomly initialised, fully fine-tuned model (0.62) and the mean-aggregation baseline (0.63)' - a downward correction, not a circular inflation. (iv) Benchmarks: TDC/MoleculeNet and QM9-OR labels are external; the frozen-descriptor protocol trains only task heads on official/scaffold splits, so no fitted value is renamed as a prediction. (v) Self-citation: the frozen backbones (MACE-POLAR-1-M, MACE-OFF24) involve overlapping authors, but they are externally trained, quantum-mechanical potentials with separate validation; the base-model ablation (0.63 vs 0.16) honestly reports backbone influence. No uniqueness theorem or unverified prior ansatz is imported to force the paper's choices; stated limitations (ChiralCat 'direct comparisons between Rem3Di and the baselines are of limited value'; 'full fine-tuning degrades small-classification performance') corroborate that claims are not being manufactured. Verdict: self-contained derivation evaluated against external benchmarks; no circularity found.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The mathematical contribution is honest: the tensor-product construction is standard group theory, and the paper's new parameters are architecture hyperparameters, not physical constants. The real unstated dependencies are the frozen MLIP's latent-space quality and the representativeness of generated conformers; these are not free parameters but assumptions that transfer risk from the backbone into Rem3Di.

free parameters (7)
  • encoder width / descriptor dimension = 256
    Chosen by hand; sets capacity of the learned molecular descriptor.
  • encoder depth, heads, FF width = 4 blocks, 8 heads, 1024 FF
    Architecture hyperparameters chosen by hand; affect expressivity and results.
  • denoising noise scale sigma = 0.3
    Hand-chosen corruption scale; controls difficulty of the self-supervised task.
  • chiral channels K = 128
    Number of pseudoscalar channels; hand-chosen.
  • pair representation width and radial basis = d_pair=64, 16 Bessel, cutoff 32 A
    Hand-chosen geometric encoding hyperparameters.
  • LoRA rank/scale = r=16, alpha=32
    Hand-chosen adaptation capacity in fine-tuning.
  • VICReg regularization weights = lambda_var=25, lambda_cov=1, target std=1 (when on)
    Optional hand-chosen regularizers to prevent descriptor collapse.
axioms (6)
  • standard math O(3) representation theory: spherical tensor products, Clebsch-Gordan coefficients, and parity selection rules (Eqs. 1-13).
    Used to construct pseudoscalars; standard mathematical background.
  • domain assumption MACE latent features transform as O(3) irreps with well-defined parity.
    Required for the chiral encoder to produce parity-odd pseudoscalars; true for equivariant MACE but not for augmentation-based Orb (Fig 6 right).
  • domain assumption Frozen MLIP features contain transferable chemical information from QM energies and forces.
    The central premise of the paper; supported by ablations but not guaranteed.
  • domain assumption The denoising objective, with the clean descriptor as the only trainable global path, forces chemically meaningful aggregation.
    Plausible and supported by pretraining-size scaling, but not proven; the detached pair representation also provides geometric context to the decoder.
  • domain assumption Single RDKit top-one conformers are representative of molecular 3D structure for the benchmarks.
    Only weakly tested via short QM9 MD; conformer variability could affect 3D-sensitive endpoints.
  • domain assumption No substantial overlap between frozen MLIP training data and benchmark labels.
    Needed to rule out label leakage; no overlap analysis is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 24730 in / 14066 out tokens · 141213 ms · 2026-08-01T11:07:16.729675+00:00 · methodology

0 comments
read the original abstract

Foundation machine-learned interatomic potentials (MLIPs) are trained on large quantum-mechanical datasets and generalise across broad regions of chemical and configurational space. Beyond their usual role in accelerating sampling-based simulations, their internal representations encode chemically rich local atomic environments. Here, we introduce Rem3Di, a representation-learning framework that repurposes latent features from atomistic foundation models as transferable molecular descriptors for property prediction and virtual screening. Rem3Di combines a potential's per-atom features into a single fixed-length descriptor of the whole molecule that varies smoothly with three-dimensional structure and is invariant to the ordering of the atoms. The descriptor can be used directly or fine-tuned for specific prediction tasks. To capture molecular handedness, Rem3Di constructs pseudoscalar features, which are unchanged by rotation but reverse sign under mirror reflection. This lets the descriptor distinguish enantiomers, which can differ in activity and toxicity. The transformer is pretrained on large molecular datasets by reconstructing corrupted atom features, so no experimental labels are required. Across public drug-property benchmarks, Rem3Di matches or exceeds published baselines without relying on classical 2D fingerprints. Additionally, the same descriptor yields chemically meaningful differentiation of transition-metal complexes without predefined bonding rules or handcrafted representations. Rem3Di therefore provides a route from simulation-trained atomistic representations to transferable, chirality-aware molecular representations for chemical machine learning.

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

94 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    author Morgan, H. L. 10.1021/c160017a018 title The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service. journal Journal of Chemical Documentation volume 5 , pages 107--113 ( year 1965 )

    doi:10.1021/c160017a018 1965

  2. [2]

    & author Hahn, M

    author Rogers, D. & author Hahn, M. 10.1021/ci100050t title Extended-connectivity fingerprints . journal Journal of Chemical Information and Modeling volume 50 , pages 742--754 ( year 2010 )

    doi:10.1021/ci100050t 2010

  3. [3]

    10.1023/a:1010933404324 title Random forests

    author Breiman, L. 10.1023/a:1010933404324 title Random forests . journal Machine Learning volume 45 , pages 5--32 ( year 2001 )

    doi:10.1023/a:1010933404324 2001

  4. [4]

    author Jiang, D. et al. 10.1186/s13321-020-00479-8 title Could graph neural networks learn better molecular representation for drug discovery? a comparison study of descriptor-based and graph-based models . journal Journal of Cheminformatics volume 13 ( year 2021 )

    doi:10.1186/s13321-020-00479-8 2021

  5. [5]

    author Rasmussen, M. H. et al. 10.1186/s13321-025-01008-1 title Smiles all around: structure to smiles conversion for transition metal complexes . journal Journal of Cheminformatics volume 17 ( year 2025 )

    work page doi:10.1186/s13321-025-01008-1 2025

  6. [6]

    , author Tkatchenko, A

    author Rupp, M. , author Tkatchenko, A. , author Müller, K.-R. & author von Lilienfeld, O. A. 10.1103/physrevlett.108.058301 title Fast and accurate modeling of molecular atomization energies with machine learning . journal Physical Review Letters volume 108 , pages 058301 ( year 2012 )

    doi:10.1103/physrevlett.108.058301 2012

  7. [7]

    author Hansen, K. et al. 10.1021/acs.jpclett.5b00831 title Machine learning predictions of molecular properties: Accurate many-body potentials and nonlocality in chemical space . journal The Journal of Physical Chemistry Letters volume 6 , pages 2326--2331 ( year 2015 )

    doi:10.1021/acs.jpclett.5b00831 2015

  8. [8]

    & author Rupp, M

    author Huo, H. & author Rupp, M. 10.1088/2632-2153/aca005 title Unified representation of molecules and crystals for machine learning . journal Machine Learning: Science and Technology volume 3 , pages 045017 ( year 2022 )

    work page doi:10.1088/2632-2153/aca005 2022

  9. [9]

    & author Parrinello, M

    author Behler, J. & author Parrinello, M. 10.1103/physrevlett.98.146401 title Generalized neural-network representation of high-dimensional potential-energy surfaces . journal Physical Review Letters volume 98 , pages 146401 ( year 2007 )

    doi:10.1103/physrevlett.98.146401 2007

  10. [10]

    10.1063/1.3553717 title Atom-centered symmetry functions for constructing high-dimensional neural network potentials

    author Behler, J. 10.1063/1.3553717 title Atom-centered symmetry functions for constructing high-dimensional neural network potentials . journal The Journal of Chemical Physics volume 134 ( year 2011 )

    doi:10.1063/1.3553717 2011

  11. [11]

    author Bartók, A. P. , author Kondor, R. & author Csányi, G. 10.1103/physrevb.87.184115 title On representing chemical environments . journal Physical Review B volume 87 , pages 184115 ( year 2013 )

    doi:10.1103/physrevb.87.184115 2013

  12. [12]

    10.1103/physrevb.99.014104 title Atomic cluster expansion for accurate and transferable interatomic potentials

    author Drautz, R. 10.1103/physrevb.99.014104 title Atomic cluster expansion for accurate and transferable interatomic potentials . journal Physical Review B volume 99 , pages 014104 ( year 2019 )

    doi:10.1103/physrevb.99.014104 2019

  13. [13]

    & author von Lilienfeld, O

    author Huang, B. & author von Lilienfeld, O. A. 10.1038/s41557-020-0527-z title Quantum machine learning using atom-in-molecule-based fragments selected on the fly . journal Nature Chemistry volume 12 , pages 945--951 ( year 2020 )

    work page doi:10.1038/s41557-020-0527-z 2020

  14. [14]

    , author Bartók, A

    author De, S. , author Bartók, A. P. , author Csányi, G. & author Ceriotti, M. 10.1039/c6cp00415f title Comparing molecules and solids across structural and alchemical space . journal Physical Chemistry Chemical Physics volume 18 , pages 13754--13769 ( year 2016 )

    doi:10.1039/c6cp00415f 2016

  15. [15]

    author Schütt, K. T. et al. https://arxiv.org/abs/1706.08566 title Schnet: A continuous-filter convolutional neural network for modeling quantum interactions . In editor Guyon, I. et al. (eds.) booktitle Advances in Neural Information Processing Systems , vol. volume 30 , pages 991--1001 ( publisher Curran Associates, Inc. , year 2017 )

    Pith/arXiv arXiv 2017

  16. [16]

    , author Kov \'a cs, D

    author Batatia, I. , author Kov \'a cs, D. P. , author Simm, G. N. C. , author Ortner, C. & author Cs \'a nyi, G. https://arxiv.org/abs/2206.07697 title MACE : Higher order equivariant message passing neural networks for fast and accurate force fields . In editor Koyejo, S. et al. (eds.) booktitle Advances in Neural Information Processing Systems , vol. v...

    Pith/arXiv arXiv 2022

  17. [17]

    author Batzner, S. et al. 10.1038/s41467-022-29939-5 title E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials . journal Nature Communications volume 13 ( year 2022 )

    doi:10.1038/s41467-022-29939-5 2022

  18. [18]

    author Wang, Y. et al. 10.1038/s41467-023-43720-2 title Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing . journal Nature Communications volume 15 ( year 2024 )

    doi:10.1038/s41467-023-43720-2 2024

  19. [19]

    author Stärk, H. et al. https://proceedings.mlr.press/v162/stark22a.html title 3d infomax improves gnns for molecular property prediction . In editor Chaudhuri, K. et al. (eds.) booktitle Proceedings of the 39th International Conference on Machine Learning , vol. volume 162 of series Proceedings of Machine Learning Research , pages 20479--20502 ( publishe...

    2022

  20. [20]

    author Zhou, G. et al. https://openreview.net/forum?id=6K2RM6wVqKu title Uni-mol: A universal 3d molecular representation learning framework . In booktitle The Eleventh International Conference on Learning Representations (ICLR) ( publisher OpenReview.net , year 2023 )

    2023

  21. [21]

    author Pozdnyakov, S. N. et al. 10.1103/physrevlett.125.166001 title Incompleteness of atomic structure representations . journal Physical Review Letters volume 125 , pages 166001 ( year 2020 )

    work page doi:10.1103/physrevlett.125.166001 2020

  22. [22]

    author Bommasani, R. et al. 10.48550/ARXIV.2108.07258 title On the opportunities and risks of foundation models ( year 2021 )

    Pith/arXiv arXiv doi:10.48550/arxiv.2108.07258 2021

  23. [23]

    author Merchant, A. et al. 10.1038/s41586-023-06735-9 title Scaling deep learning for materials discovery . journal Nature volume 624 , pages 80--85 ( year 2023 )

    doi:10.1038/s41586-023-06735-9 2023

  24. [24]

    author Batatia, I. et al. 10.1063/5.0297006 title A foundation model for atomistic materials chemistry . journal The Journal of Chemical Physics volume 163 , pages 184110 ( year 2025 )

    doi:10.1063/5.0297006 2025

  25. [25]

    author Kovács, D. P. et al. 10.1021/jacs.4c07099 title Mace-off: Short-range transferable machine learning force fields for organic molecules . journal Journal of the American Chemical Society volume 147 , pages 17598--17611 ( year 2025 )

    doi:10.1021/jacs.4c07099 2025

  26. [26]

    author Wood, B. M. et al. https://openreview.net/forum?id=SvopaNxYWt title Uma: A family of universal models for atoms . In editor Belgrave, D. et al. (eds.) booktitle Advances in Neural Information Processing Systems , vol. volume 38 ( publisher Curran Associates, Inc. , year 2025 ). note Spotlight Poster

    2025

  27. [27]

    author Plé, T. et al. 10.26434/chemrxiv-2025-f1hgn-v3 title A foundation model for accurate atomistic simulations in drug design ( year 2025 )

    work page doi:10.26434/chemrxiv-2025-f1hgn-v3 2025

  28. [28]

    author Neumann, M. et al. 10.48550/ARXIV.2410.22570 title Orb: A fast, scalable neural network potential ( year 2024 )

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.22570 2024

  29. [29]

    author Rhodes, B. et al. 10.48550/ARXIV.2504.06231 title Orb-v3: atomistic simulation at scale ( year 2025 )

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.06231 2025

  30. [30]

    author Fu, X. et al. https://proceedings.mlr.press/v267/fu25h.html title Learning smooth and expressive interatomic potentials for physical property prediction . In editor Singh, A. et al. (eds.) booktitle Proceedings of the 42nd International Conference on Machine Learning , vol. volume 267 of series Proceedings of Machine Learning Research , pages 17875...

    2025

  31. [31]

    author Masters, M. R. , author Mahmoud, A. H. & author Lill, M. A. 10.1038/s41467-025-63947-5 title Investigating whether deep learning models for co-folding learn the physics of protein-ligand interactions . journal Nature Communications volume 16 ( year 2025 )

    work page doi:10.1038/s41467-025-63947-5 2025

  32. [32]

    author Abramson, J. et al. 10.1038/s41586-024-07487-w title Accurate structure prediction of biomolecular interactions with alphafold 3 . journal Nature volume 630 , pages 493--500 ( year 2024 )

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

  33. [33]

    author Krishna, R. et al. 10.1126/science.adl2528 title Generalized biomolecular modeling and design with rosettafold all-atom . journal Science volume 384 ( year 2024 )

    doi:10.1126/science.adl2528 2024

  34. [34]

    author Boitreaud, J. et al. 10.1101/2024.10.10.615955 title Chai-1: Decoding the molecular interactions of life ( year 2024 )

    doi:10.1101/2024.10.10.615955 2024

  35. [35]

    author Wohlwend, J. et al. 10.1101/2024.11.19.624167 title Boltz-1 democratizing biomolecular interaction modeling ( year 2024 )

    doi:10.1101/2024.11.19.624167 2024

  36. [36]

    , author Allen, B

    author Cournia, Z. , author Allen, B. & author Sherman, W. 10.1021/acs.jcim.7b00564 title Relative binding free energy calculations in drug discovery: Recent advances and practical considerations . journal Journal of Chemical Information and Modeling volume 57 , pages 2911--2937 ( year 2017 )

    doi:10.1021/acs.jcim.7b00564 2017

  37. [37]

    , author Cheung, B

    author Huh, M. , author Cheung, B. , author Wang, T. & author Isola, P. https://proceedings.mlr.press/v235/huh24a.html title Position: The platonic representation hypothesis . In editor Salakhutdinov, R. et al. (eds.) booktitle Proceedings of the 41st International Conference on Machine Learning , vol. volume 235 of series Proceedings of Machine Learning ...

    2024

  38. [38]

    & author Walsh, A

    author Li, Z. & author Walsh, A. 10.1038/s42256-026-01235-7 title Platonic representation of foundation machine learning interatomic potentials . journal Nature Machine Intelligence volume 8 , pages 830--840 ( year 2026 )

    work page doi:10.1038/s42256-026-01235-7 2026

  39. [39]

    , author Yang, S

    author Edamadaka, S. , author Yang, S. , author Li, J. & author Gómez-Bombarelli, R. 10.48550/ARXIV.2512.03750 title Universally converging representations of matter across scientific foundation models ( year 2025 )

    doi:10.48550/arxiv.2512.03750 2025

  40. [40]

    , author Nakajima, Y

    author Uchiyama, R. , author Nakajima, Y. , author Tanaka, Y. & author Seino, J. 10.48550/ARXIV.2602.03046 title Impact of local descriptors derived from machine learning potentials in graph neural networks for molecular property prediction ( year 2026 )

    work page doi:10.48550/arxiv.2602.03046 2026

  41. [41]

    author Kim, S. Y. , author Park, Y. J. & author Li, J. 10.1038/s41524-026-02167-x title Leveraging neural network interatomic potentials for a foundation model of chemistry . journal npj Computational Materials ( year 2026 )

    work page doi:10.1038/s41524-026-02167-x 2026

  42. [42]

    , author Ishihara, K

    author Shiota, T. , author Ishihara, K. & author Mizukami, W. 10.1039/D4DD00098F title Universal neural network potentials as descriptors: towards scalable chemical property prediction using quantum and classical computers . journal Digital Discovery volume 3 , pages 1714--1728 ( year 2024 )

    work page doi:10.1039/d4dd00098f 2024

  43. [43]

    author Bojan, M. et al. https://openreview.net/forum?id=9ZogcRkhoG title Representing local protein environments with machine learning force fields . In booktitle The Fourteenth International Conference on Learning Representations, ICLR 2026 ( publisher OpenReview.net , year 2026 )

    2026

  44. [44]

    author Zaheer, M. et al. https://arxiv.org/abs/1703.06114 title Deep sets . In editor Guyon, I. et al. (eds.) booktitle Advances in Neural Information Processing Systems , vol. volume 30 , pages 3391--3401 ( publisher Curran Associates, Inc. , year 2017 )

    Pith/arXiv arXiv 2017

  45. [45]

    author Lee, J. et al. https://arxiv.org/abs/1810.00825 title Set transformer: A framework for attention-based permutation-invariant neural networks . In editor Chaudhuri, K. & editor Salakhutdinov, R. (eds.) booktitle Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA , vol. volume ...

    Pith/arXiv arXiv 2019

  46. [46]

    author Jumper, J. et al. 10.1038/s41586-021-03819-2 title Highly accurate protein structure prediction with alphafold . journal Nature volume 596 , pages 583--589 ( year 2021 )

    doi:10.1038/s41586-021-03819-2 2021

  47. [47]

    author Dumitrescu, A. et al. https://openreview.net/forum?id=mXHTifc1Fn title E(3)-equivariant models cannot learn chirality: Field-based molecular generation . In booktitle The Thirteenth International Conference on Learning Representations, ICLR 2025 ( publisher OpenReview.net , year 2025 )

    2025

  48. [48]

    author Harris, A. B. , author Kamien, R. D. & author Lubensky, T. C. 10.1103/revmodphys.71.1745 title Molecular chirality and chiral parameters . journal Reviews of Modern Physics volume 71 , pages 1745--1757 ( year 1999 )

    work page doi:10.1103/revmodphys.71.1745 1999

  49. [49]

    , author Zhang, Z

    author Shi, R. , author Zhang, Z. , author Chen, L. , author Yu, G. & author Yang, Y. https://openreview.net/forum?id=E7mrnBuMFd title Learning molecular chirality via chiral determinant kernels . In booktitle International Conference on Learning Representations (ICLR) ( year 2026 )

    2026

  50. [50]

    A proper scalar is invariant under both

    note A pseudoscalar is invariant under proper rotations but changes sign under inversion. A proper scalar is invariant under both

  51. [51]

    & author Smidt, T

    author Geiger, M. & author Smidt, T. 10.48550/ARXIV.2207.09453 title e3nn: Euclidean neural networks ( year 2022 )

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

  52. [52]

    author Frank, J. T. , author Chmiela, S. , author Müller, K.-R. & author Unke, O. T. 10.1038/s42256-026-01195-y title Machine learning global atomic representations with euclidean fast attention . journal Nature Machine Intelligence volume 8 , pages 388--402 ( year 2026 )

    doi:10.1038/s42256-026-01195-y 2026

  53. [53]

    author Irwin, J. J. , author Sterling, T. , author Mysinger, M. M. , author Bolstad, E. S. & author Coleman, R. G. 10.1021/ci3001277 title Zinc: A free tool to discover chemistry for biology . journal Journal of Chemical Information and Modeling volume 52 , pages 1757--1768 ( year 2012 )

    doi:10.1021/ci3001277 2012

  54. [54]

    author Kim, S. et al. 10.1093/nar/gkae1059 title Pubchem 2025 update . journal Nucleic Acids Research volume 53 , pages D1516--D1525 ( year 2024 )

    doi:10.1093/nar/gkae1059 2025

  55. [55]

    author Mendez, D. et al. 10.1093/nar/gky1075 title Chembl: towards direct deposition of bioassay data . journal Nucleic Acids Research volume 47 , pages D930--D940 ( year 2018 )

    doi:10.1093/nar/gky1075 2018

  56. [56]

    author Cohen, O. et al. 10.1088/3050-287X/ae1799 title Torchsim: an efficient atomistic simulation engine in pytorch . journal AI for Science volume 1 , pages 025003 ( year 2025 )

    doi:10.1088/3050-287x/ae1799 2025

  57. [57]

    , author Jain, A

    author Ho, J. , author Jain, A. & author Abbeel, P. https://proceedings.neurips.cc/paper/2020/hash/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html title Denoising diffusion probabilistic models . In editor Larochelle, H. , editor Ranzato, M. , editor Hadsell, R. , editor Balcan, M.-F. & editor Lin, H.-T. (eds.) booktitle Advances in Neural Information Proc...

    2020

  58. [58]

    author Song, Y. et al. https://openreview.net/forum?id=PxTIG12RRHS title Score-based generative modeling through stochastic differential equations . In booktitle 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 ( publisher OpenReview.net , year 2021 )

    2021

  59. [59]

    , author Smidt, T

    author Liao, Y.-L. , author Smidt, T. E. , author Shuaibi, M. & author Das, A. https://openreview.net/forum?id=whGzYUbIWA title Generalizing denoising to non-equilibrium structures improves equivariant force fields . journal Transactions on Machine Learning Research ( year 2024 )

    2024

  60. [60]

    , author Guo, H

    author Liu, S. , author Guo, H. & author Tang, J. https://openreview.net/forum?id=CjTHVo1dvR title Molecular geometry pretraining with se(3)-invariant denoising distance matching . In booktitle The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 ( publisher OpenReview.net , year 2023 )

    2023

  61. [61]

    author Zaidi, S. et al. https://openreview.net/forum?id=tYIMtogyee title Pre-training via denoising for molecular property prediction . In booktitle The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 ( publisher OpenReview.net , year 2023 )

    2023

  62. [62]

    author Wijaya, K. T. et al. 10.48550/ARXIV.2411.03537 title Two-stage pretraining for molecular property prediction in the wild ( year 2024 )

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2411.03537 2024

  63. [63]

    , author Ni, Y

    author Feng, S. , author Ni, Y. , author Lan, Y. , author Ma, Z.-M. & author Ma, W.-Y. https://proceedings.mlr.press/v202/feng23c.html title Fractional denoising for 3d molecular pre-training . In editor Krause, A. et al. (eds.) booktitle Proceedings of the 40th International Conference on Machine Learning , vol. volume 202 of series Proceedings of Machin...

    2023

  64. [64]

    author Liu, Y. et al. https://openreview.net/forum?id=ym7pr83XQr title Denoise VAE : Learning molecule-adaptive noise distributions for denoising-based 3d molecular pre-training . In booktitle The Thirteenth International Conference on Learning Representations ( year 2025 )

    2025

  65. [65]

    author Joshi, C. K. et al. https://proceedings.mlr.press/v267/joshi25a.html title All-atom diffusion transformers: Unified generative modelling of molecules and materials . In editor Singh, A. et al. (eds.) booktitle Proceedings of the 42nd International Conference on Machine Learning , vol. volume 267 of series Proceedings of Machine Learning Research , ...

    2025

  66. [66]

    & author Walsh, A

    author Park, H. & author Walsh, A. 10.1038/s42256-026-01262-4 title Guiding generative models to uncover diverse and novel crystals via reinforcement learning . journal Nature Machine Intelligence ( year 2026 )

    doi:10.1038/s42256-026-01262-4 2026

  67. [67]

    author Geffner, T. et al. https://openreview.net/forum?id=RDerF20JYT title La-proteina: Atomistic protein generation via partially latent flow matching . In booktitle The Fourteenth International Conference on Learning Representations, ICLR 2026, Rio de Janeiro, Brazil, April 23-27, 2026 ( publisher OpenReview.net , year 2026 )

    2026

  68. [68]

    author Didi, K. et al. https://openreview.net/forum?id=qmCpJtFZra title Scaling atomistic protein binder design with generative pretraining and test-time compute . In booktitle International Conference on Learning Representations ( year 2026 ). note Oral Presentation

    2026

  69. [69]

    , author Ponce, J

    author Bardes, A. , author Ponce, J. & author LeCun, Y. https://openreview.net/forum?id=xm6YD62D1Ub title Vicreg: Variance-invariance-covariance regularization for self-supervised learning . In booktitle The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 ( publisher OpenReview.net , year 2022 )

    2022

  70. [70]

    , author Zhu, H

    author Zhou, Y. , author Zhu, H. , author Yuan, Y. , author Song, Z. & author Mort, B. C. 10.1021/acs.jcim.4c02374 title Machine learning classification of chirality and optical rotation using a simple one-hot encoded cartesian coordinate molecular representation . journal Journal of Chemical Information and Modeling volume 65 , pages 4281--4292 ( year 2025 )

    doi:10.1021/acs.jcim.4c02374 2025

  71. [71]

    author Wu, Z. et al. 10.1039/c7sc02664a title Moleculenet: a benchmark for molecular machine learning . journal Chemical Science volume 9 , pages 513--530 ( year 2018 )

    doi:10.1039/c7sc02664a 2018

  72. [72]

    author Sheridan, R. P. 10.1021/ci400084k title Time-split cross-validation as a method for estimating the goodness of prospective prediction . journal Journal of Chemical Information and Modeling volume 53 , pages 783--790 ( year 2013 )

    work page doi:10.1021/ci400084k 2013

  73. [73]

    author Ash, J. R. et al. 10.1021/acs.jcim.5c01609 title Practically significant method comparison protocols for machine learning in small molecule drug discovery . journal Journal of Chemical Information and Modeling volume 65 , pages 9398--9411 ( year 2025 )

    doi:10.1021/acs.jcim.5c01609 2025

  74. [74]

    author Hilt, D. E. & author Seegrist, D. W. 10.5962/bhl.title.68934 title Ridge, a computer program for calculating ridge regression estimates / ( publisher Dept. of Agriculture, Forest Service, Northeastern Forest Experiment Station, , year 1977 )

    work page doi:10.5962/bhl.title.68934 1977

  75. [75]

    author Hu, E. J. et al. https://openreview.net/forum?id=nZeVKeeFYf9 title Lora: Low-rank adaptation of large language models . In booktitle The Tenth International Conference on Learning Representations (ICLR 2022) ( publisher OpenReview.net , year 2022 )

    2022

  76. [76]

    & author Maeda, T

    author Nakata, M. & author Maeda, T. 10.1021/acs.jcim.3c00899 title Pubchemqc b3lyp/6-31g*//pm6 data set: The electronic structures of 86 million molecules using b3lyp/6-31g* calculations . journal Journal of Chemical Information and Modeling volume 63 , pages 5734--5754 ( year 2023 )

    doi:10.1021/acs.jcim.3c00899 2023

  77. [77]

    , author Dral, P

    author Ramakrishnan, R. , author Dral, P. O. , author Rupp, M. & author von Lilienfeld, O. A. 10.1038/sdata.2014.22 title Quantum chemistry structures and properties of 134 kilo molecules . journal Scientific Data volume 1 ( year 2014 )

    doi:10.1038/sdata.2014.22 2014

  78. [78]

    author Bemis, G. W. & author Murcko, M. A. 10.1021/jm9602928 title The properties of known drugs. 1. molecular frameworks . journal Journal of Medicinal Chemistry volume 39 , pages 2887--2893 ( year 1996 )

    doi:10.1021/jm9602928 1996

  79. [79]

    author Huang, K. et al. https://openreview.net/forum?id=8nvgnORnoWr title Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development . In editor Vanschoren, J. & editor Yeung, S.-K. (eds.) booktitle Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Bench...

    2021

  80. [80]

    author MacDermott-Opeskin, H. et al. 10.1021/acs.jcim.5c02106 title A computational community blind challenge on pan-coronavirus drug discovery data . journal Journal of Chemical Information and Modeling volume 66 , pages 3129--3149 ( year 2026 )

    work page doi:10.1021/acs.jcim.5c02106 2026

Showing first 80 references.