Pith. sign in

REVIEW 4 major objections 4 minor 68 references

ChemHyperMag claims that ADMET properties can be predicted more accurately by encoding molecular directionality—derived from electronegativity and partial charges—as a magnetic flow on a functional-group hypergraph, without needing 3D confo

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 18:34 UTC pith:3MERBHNN

load-bearing objection Novel architecture and clean math, but the empirical case is not yet credible: the ablation table has duplicated rows and baselines were not re-run in a common harness. the 4 major comments →

arxiv 2607.18332 v2 pith:3MERBHNN submitted 2026-07-19 cs.LG cs.AI

ChemHyperMag: Physics-informed magnetic hypergraph learning improves molecular ADMET prediction

Hexiao Ding , Hongzhao Chen , Jing Lan , Yufeng Jiang , Zihong Luo , Zehua Xiong , Tianlong Ruan , Yunlin Mao
show 7 more authors
Nga Chun Ng Gwing Kei Yip Gerald W.Y. Cheng Kate Inyoung Oh Jing Cai Liang-Ting Lin Jung Sun Yoo
This is my paper
classification cs.LG cs.AI
keywords ADMET predictionmagnetic Laplacianhypergraph neural networkscontrastive learningmolecular property predictionchemical potentialChebyshev spectral filteringdrug discovery
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.

The paper tries to establish that a standard weakness of molecular ADMET predictors—their reliance on undirected, pairwise atom-bond graphs—can be fixed by adding two ingredients: higher-order motifs and directional physics. ChemHyperMag lifts each molecule to a hypergraph whose hyperedges are rings, standard drug-like fragments, scaffolds, and bonds, and then defines a nonreversible flow biased by a chemical potential built from electronegativity and partial atomic charges. That flow is encoded in a Hermitian magnetic Laplacian, whose complex phases represent cyclic circulation while keeping the operator well-behaved for learning. On six ADMET endpoints the method reports the best or tied-best score on four, with notable gains on plasma-protein binding and hERG-related cardiotoxicity, and ablations attribute the gains to the hypergraph, magnetic operator, and contrastive phase perturbation. A sympathetic reader would take this as evidence that directional, motif-level encoding is a viable path to better ADMET prediction from SMILES alone.

Core claim

ChemHyperMag's central claim is that a potential-driven nonreversible Markov flow over a functional-group hypergraph captures ADMET-relevant physicochemistry that undirected 2D message passing misses. Starting from SMILES, atoms are assigned a chemical potential phi(u)=alpha chi(u)+beta Q(u), transitions are biased by exp(phi(v)-phi(u)), and the stationary flow is split into symmetric diffusion and skew-symmetric circulation. The circulation is encoded as complex phases exp(i 2πq tanh(R_uv)) in a magnetic adjacency, then processed by a magnetic Chebyshev encoder. Two stochastic views are generated by perturbing only the phases, and InfoNCE aligns them as a regularizer for missing labels. The

What carries the argument

The central object is the magnetic Laplacian Lmag = I - D^{-1/2}(Asym ⊙ Ξ(q))D^{-1/2} on the functional-group hypergraph. Its phase factor is Ξ(q)_uv = exp(i Θ(q)_uv) with phase Θ(q)_uv = 2πq tanh(R_uv), where R_uv is the ratio of skew-symmetric to symmetric stationary flow. The Hermitian, positive semidefinite Laplacian permits stable spectral propagation while preserving directional circulation. The phase is grounded in a chemical potential phi(u)=alpha chi(u)+beta Q(u) that biases the Markov chain, and the Chebyshev filter with magnitude-based ReLU keeps features real while using phase in message mixing.

Load-bearing premise

The load-bearing premise is that a simple linear combination of atomic electronegativity and computed partial charge is a trustworthy measure of the directionality that governs ADMET behaviour; if it does not reflect real physicochemistry, the magnetic phases encode noise rather than signal.

What would settle it

Train ChemHyperMag with the magnetic phases replaced by random or zero phases (q=0), keeping the hypergraph, Chebyshev encoder, and contrastive objective identical. If the randomized-phase model matches the full model on PPB and Cardiotoxicity-1, the claimed directional mechanism is not the cause of the gains. A second check: rerun under a scaffold split; the paper uses random 80/10/10 splits, and the authors note these may overestimate generalization, so if the gains vanish under scaffold splits the improvement is memorization of split similarity rather than general ADMET physics.

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

If this is right

  • ADMET prediction from 2D SMILES need not sacrifice directionality: asymmetric electronic effects can enter through operator-level phases without 3D conformers.
  • Encoding rings, fragments, and scaffolds as hyperedges gives the model a route to capture functional-group and motif effects that pairwise bond graphs approximate only by local aggregation.
  • Phase-perturbation contrastive learning provides a self-supervised regularizer that is cheap and operates under missing labels, improving sample efficiency.
  • The learned phases and potential weights (alpha, beta) supply an interpretable directional signal, potentially traceable to specific substructures or charge patterns.

Where Pith is reading between the lines

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

  • If the directional signal is real, the magnetic phases could be read as substructure-level maps of electron-poor and electron-rich character, giving medicinal chemists a diagnostic beyond attention weights—this interpretive step is not in the paper.
  • The sensitivity to charge and protonation protocols the paper acknowledges suggests a natural extension: making the chemical potential pH-dependent could turn a current weakness into a feature for endpoints like solubility and permeability.
  • The operator-level phase perturbation is a general augmentation idea; it could transfer to other directed or signed molecular and non-molecular graph tasks, with the hypergraph construction replaced by the appropriate incidence structure.
  • A strict test would compare random-phase or q=0 training against potential-derived phases; if the randomized-phase variant matches the full model, the physicochemistry is not doing the work and the gains come from the hypergraph alone.

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 ChemHyperMag, a 2D-based ADMET predictor that lifts SMILES-derived molecular graphs to functional-group hypergraphs, defines a chemical-potential-driven nonreversible Markov flow, encodes it with a Hermitian magnetic Laplacian, applies a magnetic Chebyshev encoder, and trains with masked multitask supervision plus phase-perturbation InfoNCE. Proposition 1 gives a standard proof that the magnetic Laplacian is Hermitian and positive semidefinite. The experiments report results on six ADMET endpoints, claiming best/tied-best on four endpoints and improvements of +0.061 on PPB and +0.055 on Cardiotoxicity-1 over baselines. The paper also claims improved sample efficiency, although no label-fraction experiment is presented.

Significance. If the empirical claims were verified, ChemHyperMag would be a useful conformer-free ADMET model: the combination of motif-level hyperedges and magnetic phase encoding is a plausible way to inject directionality, the operator-level contrastive view is principled, and the authors provide code. Proposition 1 is correctly proved, and the derivation of the phase map from the stationary flow is internally coherent. However, the significance is currently capped by two problems: Table 2's ablation contains duplicated rows that invalidate the module attribution, and Table 1's baseline numbers are not produced in a common experimental harness, so the claimed SOTA margins may be protocol artifacts. The theoretical contribution is also incremental relative to HyperMagNet and MagNet.

major comments (4)
  1. [Table 2 (§5.4)] The ablation table is corrupted. 'Only Magnetic' is numerically identical to 'Base Model' on all six endpoints; 'Magnetic + CL' is also identical to 'Base Model'; 'ChebNet + CL' and 'Magnetic + ChebNet' are identical to 'Only ChebNet'. Since Section 5.4 explicitly uses this table to conclude that the magnetic Laplacian, ChebNet, hypergraph, and CL 'contribute complementary gains', the central module-attribution claim is unsupported. The leave-one-out rows also report no standard deviation (only a .000 subscript), so the reader cannot assess variability. Please regenerate the ablation with correct toggling of modules, report per-seed statistics, and test whether the Full Model's margins over ablations are statistically meaningful.
  2. [Table 1 (§5.1, §5.4)] The headline comparison is not controlled. The table note says baseline numbers are 'taken from the original papers' while ChemHyperMag was evaluated under its own protocol, so differences may reflect split, feature, or tuning choices rather than the method. Section 5.1 says all methods used the same RDKit atom featurization but then states ChemHyperMag additionally received Pauling electronegativity and Gasteiger partial charges; if the baselines did not receive these descriptors, the +0.061 (PPB) and +0.055 (Cardiotoxicity-1) margins cannot be attributed to the magnetic hypergraph mechanism. In addition, Section 5.2 lists ADMET-AI, Chemprop v2, and ChemBERTa as included baselines, but no results for them appear in Table 1. Please re-run baselines in the same harness with identical splits and featurization, or report published numbers with explicit matching conditions, and include the m
  3. [Abstract, §5.4] The abstract claims improvements 'with fewer labeled samples', but no experiment varies the amount of labeled data. All reported runs use the full training split (80/10/10) with the original observation masks. A label-fraction or missing-label-rate sweep is needed to substantiate the sample-efficiency claim; if the claim is intended to mean that the model handles missing labels in multi-task settings, that should be stated and tested explicitly.
  4. [§6, §5.3] The paper's own Limitations section concedes that the directional signal relies on a heuristic chemical potential, that fixed-step stationary estimation and complex propagation introduce approximation risks, and that the magnitude projection removes phase. These are non-trivial caveats for the advertised physical mechanism. The manuscript reports no sensitivity analysis for the free parameters q, η, ρ, λ, α, β, or for the charge/protonation protocol that feeds Eq. (7). Given that the claimed improvement depends on these choices, please add a sensitivity study or otherwise bound the claim.
minor comments (4)
  1. [§3.2-§4.2] There is a notational mismatch: Section 3.3 defines the Markov state space as the atom set V, but Section 4.1/4.2 applies the magnetic operator to the incidence graph whose nodes include hyperedge nodes. Please state explicitly that the magnetic Laplacian in Eq. (14) is applied on the incidence node set, with hyperedge nodes treated as having zero chemical potential.
  2. [Table 1 note] The table note mentions MTGIB-UNet baselines [63], but no MTGIB-UNet column or result is reported in Table 1. Please either include those numbers or remove the citation from the note.
  3. [§4.2] The code rescales electronegativity by a constant factor 4.0, but Eq. (7) uses raw χ(u). Please reconcile the formula with the implementation or state the normalization explicitly.
  4. [Table 2 note] The note says 'Subscripts denote the absolute difference between test and validation results', but the rows display subscripts after the mean and no standard deviation for the leave-one-out and full-model rows. This is ambiguous and should be clarified, ideally by reporting mean ± std for every variant.

Circularity Check

0 steps flagged

No significant circularity: the model is trained end-to-end on ADMET labels; the physics-inspired potential, magnetic phase, and hypergraph are construction choices, and no reported prediction reduces by definition to a fitted input.

full rationale

ChemHyperMag is an end-to-end supervised multitask model. Its trainable parameters (α, β in Eq. 7, the magnetic charge q, and the encoder weights) are optimized against ADMET labels through the masked supervised loss in Eq. 2. This is standard supervised learning, not a claim to derive a physical constant or a target quantity from first principles. The magnetic phase, hypergraph incidence, and contrastive views are architectural and regularization choices; none of the reported endpoint predictions is defined as an algebraic function of the observed labels, nor is any fitted parameter renamed as an independent prediction. The reference list contains no self-citations, and the benchmark comparisons are against external MoleculeNet/MTGL-ADMET baselines, so there is no load-bearing self-citation chain. Section 6 candidly states that the directional signal 'relies on a heuristic chemical potential from electronegativity and Gasteiger charges, so outcomes can be sensitive to charge and protonation protocols and remain weakly mechanistically grounded for ADMET' — this weakens the physical-interpretability claim but is a modeling limitation, not a circular reduction. The abstract's 'fewer labeled samples' claim is not supported by a specific label-fraction experiment, and the duplicated rows in Table 2 (e.g., 'Only Magnetic' identical to 'Base Model') raise internal-consistency concerns about the ablation attributed to modules; however, these are empirical/correctness problems, not definitional circularity. No step of the derivation is equivalent by construction to its input, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The model introduces no new physical entities, but it relies on several domain assumptions: the chemical potential is a heuristic combination of known descriptors, and the stationary flow is computed via a fixed number of power iterations. The free parameters are the learned alpha/beta and hand-set hyperparameters (q, eta, rho, lambda), the last two of which are not reported.

free parameters (5)
  • alpha, beta = learned end-to-end
    Scaling weights in chemical potential phi(u)=alpha*chi(u)+beta*Q(u) (Eq. 7). These are optimized on the training data.
  • q = 0.1
    Magnetic charge controlling phase strength in Theta=2*pi*q*tanh(R) (Eq. 11). Chosen by hand, not learned.
  • eta = 0.15
    Teleportation rate for ergodicity (Eq. 9). Chosen by hand.
  • rho = not reported
    Perturbation magnitude for phase noise in contrastive views (Eq. 27). Not specified in Implementation Details.
  • lambda = not reported
    Weight of InfoNCE loss in L = L_sup + lambda*L_NCE. Not specified.
axioms (5)
  • domain assumption Electronegativity and Gasteiger partial charges capture relevant directionality for ADMET.
    The entire magnetic phase is derived from this potential (Eq. 7); the paper states it is heuristic.
  • domain assumption Nonreversible Markov flow on the functional-group hypergraph models molecular transport/partitioning relevant to ADMET.
    The model assumes this flow encodes actionable directional information.
  • domain assumption Random 80/10/10 splits are appropriate for evaluating generalization.
    The paper acknowledges random splits may overestimate scaffold generalization (Limitations).
  • standard math The teleported Markov chain has a unique stationary distribution (standard ergodic chain theory).
    Used to compute stationary flow eA in Section 4.2.
  • domain assumption Magnitude projection |z| retains sufficient information from the complex representations.
    The encoder discards phase after propagation, assuming phase only matters for mixing.

pith-pipeline@v1.3.0-alltime-deepseek · 15750 in / 12243 out tokens · 98912 ms · 2026-08-01T18:34:36.494485+00:00 · methodology

0 comments
read the original abstract

Accurate prediction of ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) is important for drug discovery. Most predictors use undirected molecular graphs and pairwise edges. This choice misses asymmetric interactions, nonreversible dynamics, and motif level effects from functional groups and ring systems. We propose ChemHyperMag for multitask ADMET prediction under missing labels. ChemHyperMag builds a functional group hypergraph from rings, BRICS fragments, Bemis-Murcko scaffolds, and bonds. It also defines a potential driven nonreversible flow guided by electronegativity and Gasteiger partial charges. The resulting circulation is encoded by a Hermitian magnetic Laplacian and processed with a magnetic Chebyshev encoder. We perturb magnetic phases to form stochastic views and train with an InfoNCE objective. Experiments on multiple ADMET benchmarks show improvements over recent methods with fewer labeled samples and no conformers. ChemHyperMag is scalable and provides interpretable directional signals through its magnetic phases.

Figures

Figures reproduced from arXiv: 2607.18332 by Gerald W.Y. Cheng, Gwing Kei Yip, Hexiao Ding, Hongzhao Chen, Jing Cai, Jing Lan, Jung Sun Yoo, Kate Inyoung Oh, Liang-ting Lin, Nga Chun Ng, Tianlong Ruan, Yufeng Jiang, Yunlin Mao, Zehua Xiong, Zihong Luo.

Figure 1
Figure 1. Figure 1: Workflow of ChemHyperMag. Stage 1 parses SMILES into a 2D molecular graph with RDKit and computes atom electronegativity and Gasteiger partial charges together with structural annotations. Stage 2 constructs a functional group hypergraph from rings, BRICS fragments, scaffolds, and bonds, then defines a chemical potential that induces directed flow under the hypergraph constraints. Stage 3 encodes this dire… view at source ↗
Figure 2
Figure 2. Figure 2: The model framework of ChemHyperMag. A. Internal structure of the Magnetic Chebyshev Encoder and subsequent gating mechanism; B. Brief overview of contrastive learning. and temperature τ > 0, define the batchwise scores s 12 ij = sim h (1) i , h(2) j  , s21 ij = sim h (2) i , h(1) j  . (17) The symmetric InfoNCE loss is LNCE = 1 2B X B i=1 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] 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

68 extracted references · 6 linked inside Pith

  1. [1]

    Moleculenet: a benchmark for molecular machine learning.Chemical science, 9(2):513–530, 2018

    Zhenqin Wu, Bharath Ramsundar, Evan N Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S Pappu, Karl Leswing, and Vijay Pande. Moleculenet: a benchmark for molecular machine learning.Chemical science, 9(2):513–530, 2018

    2018

  2. [2]

    Artificial intelligence foundation for therapeutic science.Nature chemical biology, 18(10):1033–1036, 2022

    Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Connor W Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik. Artificial intelligence foundation for therapeutic science.Nature chemical biology, 18(10):1033–1036, 2022

    2022

  3. [3]

    Smiles, a chemical language and information system

    David Weininger. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules.Journal of chemical information and computer sciences, 28(1):31–36, 1988. 11 Physics-informed magnetic hypergraph learning improves molecular ADMET prediction

    1988

  4. [4]

    An open source chemical structure curation pipeline using rdkit.Journal of Cheminformatics, 12(1):51, 2020

    A Patrícia Bento, Anne Hersey, Eloy Félix, Greg Landrum, Anna Gaulton, Francis Atkinson, Louisa J Bellis, Marleen De Veij, and Andrew R Leach. An open source chemical structure curation pipeline using rdkit.Journal of Cheminformatics, 12(1):51, 2020

    2020

  5. [5]

    Analyzing learned molecular representations for property prediction

    Kevin Yang, Kyle Swanson, Wengong Jin, Connor Coley, Philipp Eiden, Hua Gao, Angel Guzman-Perez, Timothy Hopper, Brian Kelley, Miriam Mathea, et al. Analyzing learned molecular representations for property prediction. Journal of chemical information and modeling, 59(8):3370–3388, 2019

    2019

  6. [6]

    Cornell university press, 1960

    Linus Pauling.The nature of the chemical bond and the structure of molecules and crystals: an introduction to modern structural chemistry, volume 18. Cornell university press, 1960

    1960

  7. [7]

    Iterative partial equalization of orbital electronegativity—a rapid access to atomic charges.Tetrahedron, 36(22):3219–3228, 1980

    Johann Gasteiger and Mario Marsili. Iterative partial equalization of orbital electronegativity—a rapid access to atomic charges.Tetrahedron, 36(22):3219–3228, 1980

    1980

  8. [8]

    Uni-mol: A universal 3d molecular representation learning framework

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecular representation learning framework. InThe eleventh international conference on learning representations, 2023

    2023

  9. [9]

    Data-driven quantum chemical property prediction leveraging 3d conformations with uni-mol+.Nature communications, 15(1):7104, 2024

    Shuqi Lu, Zhifeng Gao, Di He, Linfeng Zhang, and Guolin Ke. Data-driven quantum chemical property prediction leveraging 3d conformations with uni-mol+.Nature communications, 15(1):7104, 2024

    2024

  10. [10]

    Exploring molecular pretraining model at scale.Advances in Neural Information Processing Systems, 37:46956–46978, 2024

    Zhen Wang, Zhifeng Gao, Hang Zheng, Linfeng Zhang, Guolin Ke, et al. Exploring molecular pretraining model at scale.Advances in Neural Information Processing Systems, 37:46956–46978, 2024

    2024

  11. [11]

    Magnetic eigenmaps for community detection in directed networks.Physical Review E, 95(2):022302, 2017

    Michaël Fanuel, Carlos M Alaiz, and Johan AK Suykens. Magnetic eigenmaps for community detection in directed networks.Physical Review E, 95(2):022302, 2017

    2017

  12. [12]

    Magnet: A neural network for directed graphs.Advances in neural information processing systems, 34:27003–27015, 2021

    Xitong Zhang, Yixuan He, Nathan Brugnone, Michael Perlmutter, and Matthew Hirn. Magnet: A neural network for directed graphs.Advances in neural information processing systems, 34:27003–27015, 2021

    2021

  13. [13]

    Young, and Sinan Guven Aksoy

    Tatyana Benko, Martin Buck, Ilya Amburg, Stephen J. Young, and Sinan Guven Aksoy. Hypermagnet: A magnetic laplacian based hypergraph neural network.Transactions on Machine Learning Research, 2025

    2025

  14. [14]

    Directed graph contrastive learning.Advances in neural information processing systems, 34:19580–19593, 2021

    Zekun Tong, Yuxuan Liang, Henghui Ding, Yongxing Dai, Xinke Li, and Changhu Wang. Directed graph contrastive learning.Advances in neural information processing systems, 34:19580–19593, 2021

    2021

  15. [15]

    Extended-connectivity fingerprints.Journal of chemical information and modeling, 50(5):742–754, 2010

    David Rogers and Mathew Hahn. Extended-connectivity fingerprints.Journal of chemical information and modeling, 50(5):742–754, 2010

    2010

  16. [16]

    pkcsm: predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures.Journal of medicinal chemistry, 58(9):4066–4072, 2015

    Douglas EV Pires, Tom L Blundell, and David B Ascher. pkcsm: predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures.Journal of medicinal chemistry, 58(9):4066–4072, 2015

    2015

  17. [17]

    Swissadme: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules.Scientific reports, 7(1):42717, 2017

    Antoine Daina, Olivier Michielin, and Vincent Zoete. Swissadme: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules.Scientific reports, 7(1):42717, 2017

    2017

  18. [18]

    vnn web server for admet predictions

    Patric Schyman, Ruifeng Liu, Valmik Desai, and Anders Wallqvist. vnn web server for admet predictions. Frontiers in pharmacology, 8:889, 2017

    2017

  19. [19]

    Admetboost: a web server for accurate admet prediction.Journal of molecular modeling, 28(12):408, 2022

    Hao Tian, Rajas Ketkar, and Peng Tao. Admetboost: a web server for accurate admet prediction.Journal of molecular modeling, 28(12):408, 2022

    2022

  20. [20]

    Admetlab 2.0: an integrated online platform for accurate and comprehensive predictions of admet properties.Nucleic acids research, 49(W1):W5–W14, 2021

    Guoli Xiong, Zhenxing Wu, Jiacai Yi, Li Fu, Zhijiang Yang, Changyu Hsieh, Mingzhu Yin, Xiangxiang Zeng, Chengkun Wu, Aiping Lu, et al. Admetlab 2.0: an integrated online platform for accurate and comprehensive predictions of admet properties.Nucleic acids research, 49(W1):W5–W14, 2021

    2021

  21. [21]

    Li Fu, Shaohua Shi, Jiacai Yi, Ningning Wang, Yuanhang He, Zhenxing Wu, Jinfu Peng, Youchao Deng, Wenxuan Wang, Chengkun Wu, et al. Admetlab 3.0: an updated comprehensive online admet prediction platform enhanced with broader coverage, improved performance, api functionality and decision support.Nucleic acids research, 52(W1):W422–W431, 2024

    2024

  22. [22]

    Admet-ai: a machine learning admet platform for evaluation of large-scale chemical libraries

    Kyle Swanson, Parker Walther, Jeremy Leitz, Souhrid Mukherjee, Joseph C Wu, Rabindra V Shivnaraine, and James Zou. Admet-ai: a machine learning admet platform for evaluation of large-scale chemical libraries. Bioinformatics, 40(7):btae416, 2024

    2024

  23. [23]

    Potentialnet for molecular property prediction.ACS central science, 4(11):1520–1530, 2018

    Evan N Feinberg, Debnil Sur, Zhenqin Wu, Brooke E Husic, Huanghao Mai, Yang Li, Saisai Sun, Jianyi Yang, Bharath Ramsundar, and Vijay S Pande. Potentialnet for molecular property prediction.ACS central science, 4(11):1520–1530, 2018

    2018

  24. [24]

    Chemprop: a machine learning package for chemical property prediction.Journal of Chemical Information and Modeling, 64(1):9–17, 2023

    Esther Heid, Kevin P Greenman, Yunsie Chung, Shih-Cheng Li, David E Graff, Florence H Vermeire, Haoyang Wu, William H Green, and Charles J McGill. Chemprop: a machine learning package for chemical property prediction.Journal of Chemical Information and Modeling, 64(1):9–17, 2023. 12 Physics-informed magnetic hypergraph learning improves molecular ADMET prediction

    2023

  25. [25]

    Chemprop v2: An efficient, modular machine learning package for chemical property prediction.Journal of Chemical Information and Modeling, 2025

    David E Graff, Nathan K Morgan, Jackson W Burns, Anna C Doner, Brian Li, Shih-Cheng Li, Joel Manu, Angiras Menon, Hao-Wei Pang, Haoyang Wu, et al. Chemprop v2: An efficient, modular machine learning package for chemical property prediction.Journal of Chemical Information and Modeling, 2025

    2025

  26. [26]

    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

  27. [27]

    Admet property prediction via multi-task graph learning under adaptive auxiliary task selection.Iscience, 26(11), 2023

    Bing-Xue Du, Yi Xu, Siu-Ming Yiu, Hui Yu, and Jian-Yu Shi. Admet property prediction via multi-task graph learning under adaptive auxiliary task selection.Iscience, 26(11), 2023

    2023

  28. [28]

    Deepdelta: predicting admet improvements of molecular derivatives with deep learning.Journal of cheminformatics, 15(1):101, 2023

    Zachary Fralish, Ashley Chen, Paul Skaluba, and Daniel Reker. Deepdelta: predicting admet improvements of molecular derivatives with deep learning.Journal of cheminformatics, 15(1):101, 2023

    2023

  29. [29]

    Unified and explainable molecular representation learning for imperfectly annotated data from the hypergraph view.Nature Communications, 16(1):1–18, 2025

    Bowen Wang, Junyou Li, Donghao Zhou, Lanqing Li, Jinpeng Li, Ercheng Wang, Jianye Hao, Liang Shi, Chengqiang Lu, Jiezhong Qiu, et al. Unified and explainable molecular representation learning for imperfectly annotated data from the hypergraph view.Nature Communications, 16(1):1–18, 2025

    2025

  30. [30]

    Quantum-informed molecular representation learning enhancing admet property prediction.Journal of Chemical Information and Modeling, 64(13):5028–5040, 2024

    Jungwoo Kim, Woojae Chang, Hyunjun Ji, and InSuk Joung. Quantum-informed molecular representation learning enhancing admet property prediction.Journal of Chemical Information and Modeling, 64(13):5028–5040, 2024

    2024

  31. [31]

    Pretraining graph transformers with atom-in-a-molecule quantum properties for improved admet modeling.Journal of Cheminformatics, 17(1):25, 2025

    Alessio Fallani, Ramil Nugmanov, Jose Arjona-Medina, Jörg Kurt Wegner, Alexandre Tkatchenko, and Kostiantyn Chernichenko. Pretraining graph transformers with atom-in-a-molecule quantum properties for improved admet modeling.Journal of Cheminformatics, 17(1):25, 2025

    2025

  32. [32]

    Hypergraph neural networks

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 3558–3565, 2019

    2019

  33. [33]

    Recent advances in hypergraph neural networks: M.-r

    Mu-Rong Yang and Xin-Jian Xu. Recent advances in hypergraph neural networks: M.-r. yang, x.-j. xu.Journal of the Operations Research Society of China, pages 1–37, 2025

    2025

  34. [34]

    You are allset: A multiset function framework for hypergraph neural networks

    Eli Chien, Chao Pan, Jianhao Peng, and Olgica Milenkovic. You are allset: A multiset function framework for hypergraph neural networks. InInternational Conference on Learning Representations, 2022

    2022

  35. [35]

    Sheaf hypergraph networks.Advances in Neural Information Processing Systems, 36:12087–12099, 2023

    Iulia Duta, Giulia Cassarà, Fabrizio Silvestri, and Pietro Liò. Sheaf hypergraph networks.Advances in Neural Information Processing Systems, 36:12087–12099, 2023

    2023

  36. [36]

    Directed hypergraph representation learning for link prediction

    Zitong Ma, Wenbo Zhao, and Zhe Yang. Directed hypergraph representation learning for link prediction. In International conference on artificial intelligence and statistics, pages 3268–3276. PMLR, 2024

    2024

  37. [37]

    Directional sheaf hypergraph networks: Unifying learning on directed and undirected hypergraphs.arXiv preprint arXiv:2510.04727, 2025

    Emanuele Mule, Stefano Fiorini, Antonio Purificato, Federico Siciliano, Stefano Coniglio, and Fabrizio Silvestri. Directional sheaf hypergraph networks: Unifying learning on directed and undirected hypergraphs.arXiv preprint arXiv:2510.04727, 2025

    arXiv 2025

  38. [38]

    Molecular hypergraph grammar with its application to molecular optimization

    Hiroshi Kajino. Molecular hypergraph grammar with its application to molecular optimization. InInternational Conference on Machine Learning, pages 3183–3191. PMLR, 2019

    2019

  39. [39]

    Molecular hypergraph neural networks.The Journal of Chemical Physics, 160(14), 2024

    Junwu Chen and Philippe Schwaller. Molecular hypergraph neural networks.The Journal of Chemical Physics, 160(14), 2024

    2024

  40. [40]

    A hypergraph convolutional neural network for molecular properties prediction using functional group.arXiv, 2021

    Fangying Chen, Junyoung Park, and Jinkyoo Park. A hypergraph convolutional neural network for molecular properties prediction using functional group.arXiv, 2021

    2021

  41. [41]

    Msgnn: A spectral graph neural network based on a novel magnetic signed laplacian

    Yixuan He, Michael Perlmutter, Gesine Reinert, and Mihai Cucuringu. Msgnn: A spectral graph neural network based on a novel magnetic signed laplacian. InLearning on graphs conference, pages 40–1. PMLR, 2022

    2022

  42. [42]

    A spectral graph convolution for signed directed graphs via magnetic laplacian.Neural Networks, 164:562–574, 2023

    Taewook Ko, Yoonhyuk Choi, and Chong-Kwon Kim. A spectral graph convolution for signed directed graphs via magnetic laplacian.Neural Networks, 164:562–574, 2023

    2023

  43. [43]

    Exploitation of a latent mechanism in graph contrastive learning: Representation scattering.Advances in Neural Information Processing Systems, 37:115351–115376, 2024

    Dongxiao He, Lianze Shan, Jitao Zhao, Hengrui Zhang, Zhen Wang, and Weixiong Zhang. Exploitation of a latent mechanism in graph contrastive learning: Representation scattering.Advances in Neural Information Processing Systems, 37:115351–115376, 2024

    2024

  44. [44]

    Graph contrastive learning with augmentations.Advances in neural information processing systems, 33:5812–5823, 2020

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations.Advances in neural information processing systems, 33:5812–5823, 2020

    2020

  45. [45]

    Adversarial graph augmentation to improve graph contrastive learning.Advances in Neural Information Processing Systems, 34:15920–15933, 2021

    Susheel Suresh, Pan Li, Cong Hao, and Jennifer Neville. Adversarial graph augmentation to improve graph contrastive learning.Advances in Neural Information Processing Systems, 34:15920–15933, 2021

    2021

  46. [46]

    Prototypical graph contrastive learning.IEEE transactions on neural networks and learning systems, 35(2):2747–2758, 2022

    Shuai Lin, Chen Liu, Pan Zhou, Zi-Yuan Hu, Shuojia Wang, Ruihui Zhao, Yefeng Zheng, Liang Lin, Eric Xing, and Xiaodan Liang. Prototypical graph contrastive learning.IEEE transactions on neural networks and learning systems, 35(2):2747–2758, 2022. 13 Physics-informed magnetic hypergraph learning improves molecular ADMET prediction

    2022

  47. [47]

    An empirical study of graph contrastive learning

    Yanqiao Zhu, Yichen Xu, Qiang Liu, and Shu Wu. An empirical study of graph contrastive learning. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

    2021

  48. [48]

    Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence, 4(3):279–287, 2022

    Yuyang Wang, Jianren Wang, Zhonglin Cao, and Amir Barati Farimani. Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence, 4(3):279–287, 2022

    2022

  49. [49]

    Pre-training molecular graph representation with 3d geometry

    Shengchao Liu, Hanchen Wang, Weiyang Liu, Joan Lasenby, Hongyu Guo, and Jian Tang. Pre-training molecular graph representation with 3d geometry. InInternational Conference on Learning Representations, 2022

    2022

  50. [50]

    3d graph contrastive learning for molecular property prediction

    Kisung Moon, Hyeon-Jin Im, and Sunyoung Kwon. 3d graph contrastive learning for molecular property prediction. Bioinformatics, 39(6):btad371, 2023

    2023

  51. [51]

    Attention-wise masked graph contrastive learning for predicting molecular property.Briefings in bioinformatics, 23(5):bbac303, 2022

    Hui Liu, Yibiao Huang, Xuejun Liu, and Lei Deng. Attention-wise masked graph contrastive learning for predicting molecular property.Briefings in bioinformatics, 23(5):bbac303, 2022

    2022

  52. [52]

    Advancing drug discovery with enhanced chemical understanding via asymmetric contrastive multimodal learning.Journal of chemical information and modeling, 2025

    Yifei Wang, Yunrui Li, Lin Liu, Pengyu Hong, and Hao Xu. Advancing drug discovery with enhanced chemical understanding via asymmetric contrastive multimodal learning.Journal of chemical information and modeling, 2025

    2025

  53. [53]

    Universal graph contrastive learning with a novel laplacian perturbation

    Taewook Ko, Yoonhyuk Choi, and Chong-Kwon Kim. Universal graph contrastive learning with a novel laplacian perturbation. InUncertainty in Artificial Intelligence, pages 1098–1108. PMLR, 2023

    2023

  54. [54]

    Signed directed graph contrastive learning with laplacian augmentation.arXiv preprint arXiv:2301.05163, 2023

    Taewook Ko, Yoonhyuk Choi, and Chong-Kwon Kim. Signed directed graph contrastive learning with laplacian augmentation.arXiv preprint arXiv:2301.05163, 2023

    Pith/arXiv arXiv 2023

  55. [55]

    Toward general digraph contrastive learning: A dual spatial perspective.arXiv preprint arXiv:2510.16311, 2025

    Daohan Su, Yang Zhang, Xunkai Li, Rong-Hua Li, and Guoren Wang. Toward general digraph contrastive learning: A dual spatial perspective.arXiv preprint arXiv:2510.16311, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  57. [57]

    Ring perception using breadth-first search.Journal of chemical information and computer sciences, 36(5):986–991, 1996

    John Figueras. Ring perception using breadth-first search.Journal of chemical information and computer sciences, 36(5):986–991, 1996

    1996

  58. [58]

    On the art of compiling and using’drug-like’chemical fragment spaces.ChemMedChem, 3(10):1503, 2008

    Jorg Degen, Christof Wegscheid-Gerlach, Andrea Zaliani, and Matthias Rarey. On the art of compiling and using’drug-like’chemical fragment spaces.ChemMedChem, 3(10):1503, 2008

    2008

  59. [59]

    The properties of known drugs

    Guy W Bemis and Mark A Murcko. The properties of known drugs. 1. molecular frameworks.Journal of medicinal chemistry, 39(15):2887–2893, 1996

    1996

  60. [60]

    Convolutional neural networks on graphs with chebyshev approxi- mation, revisited

    Mingguo He, Zhewei Wei, and Ji-Rong Wen. Convolutional neural networks on graphs with chebyshev approxi- mation, revisited. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems, 2022

    2022

  61. [61]

    Convolutional neural networks on graphs with fast localized spectral filtering.Advances in neural information processing systems, 29, 2016

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering.Advances in neural information processing systems, 29, 2016

    2016

  62. [62]

    Mtgl-admet: a novel multi-task graph learning framework for admet prediction enhanced by status-theory and maximum flow

    Bing-Xue Du, Yi Xu, Siu-Ming Yiu, Hui Yu, and Jian-Yu Shi. Mtgl-admet: a novel multi-task graph learning framework for admet prediction enhanced by status-theory and maximum flow. InInternational Conference on Research in Computational Molecular Biology, pages 85–103. Springer, 2023

    2023

  63. [63]

    Mtgib-unet: A multi-task graph information bottleneck and uncertainty weighted network for admet prediction

    Xuqiang Li, Wenjie Du, Jun Xia, Jianmin Wang, Xiaoqi Wang, Yang Yang, and Yang Wang. Mtgib-unet: A multi-task graph information bottleneck and uncertainty weighted network for admet prediction. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 7518–7526, 2025

    2025

  64. [64]

    Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

    TN Kipf. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

    Pith/arXiv arXiv 2016

  65. [65]

    Graph attention networks.arXiv preprint arXiv:1710.10903, 2017

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.arXiv preprint arXiv:1710.10903, 2017

    Pith/arXiv arXiv 2017

  66. [66]

    Chemberta: large-scale self-supervised pretraining for molecular property prediction.arXiv preprint arXiv:2010.09885, 2020

    Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. Chemberta: large-scale self-supervised pretraining for molecular property prediction.arXiv preprint arXiv:2010.09885, 2020

    Pith/arXiv arXiv 2010

  67. [67]

    Modeling physico-chemical admet endpoints with multitask graph convolutional networks.Molecules, 25(1):44, 2019

    Floriane Montanari, Lara Kuhnke, Antonius Ter Laak, and Djork-Arné Clevert. Modeling physico-chemical admet endpoints with multitask graph convolutional networks.Molecules, 25(1):44, 2019

    2019

  68. [68]

    Bayesian flow network framework for chemistry tasks.Journal of Chemical Information and Modeling, 65(3):1178–1187, 2025

    Nianze Tao and Minori Abe. Bayesian flow network framework for chemistry tasks.Journal of Chemical Information and Modeling, 65(3):1178–1187, 2025. 14 Physics-informed magnetic hypergraph learning improves molecular ADMET prediction A Additional Details A.1 Notation Table 3: Summary of notation Symbol Description Problem Setup & Dataset GSet of molecular ...

    2025