Pith. sign in

REVIEW 5 major objections 6 minor 78 references

Tokenizing Electron Cloud in Protein-Ligand Interaction Learning

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read ECBind claims that tokenizing electron clouds alongside atoms improves protein-ligand affinity prediction, with per-structure Spearman up 15.58% over the best baseline.

desk verdict Electron-cloud pretraining gives a modest but consistent per-structure gain on MISATO, yet the hyperparameter ablation is too sloppy to fully rule out selection bias and the XTB filter needs a sensitivity check. read the letter →

arxiv 2505.19014 v2 pith:P3KQ3GLA submitted 2025-05-25 cs.LG physics.chem-phq-bio.QM

classification cs.LGphysics.chem-phq-bio.QM
keywords electroncloudbindingaffinitypredictiontokenizationvectorquantizationprotein-ligandinteractionspretrainingknowledgedistillationstructure-awaretransformer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that the electron cloud—the spatial distribution of electrons around atoms—carries binding-relevant information that atom-level models miss, and that this information can be captured by compressing raw electron densities into discrete tokens. It introduces ECBind, a pretrain-and-finetune model that resolves a binding site into patches centered on atoms, encodes each patch with a structure-aware transformer, and vector-quantizes the results into hierarchical codebooks covering both 3D electronic structure and 2D atomic identity. On relative binding affinity prediction, which asks how different ligands rank against the same protein, ECBind reports per-structure Pearson 0.431 and Spearman 0.445, exceeding the best baseline by 6.42% and 15.58%. A distilled student that never sees electron clouds at inference retains most of the teacher's accuracy. If true, the result means electron-cloud-aware pretraining is a practical way to get physically grounded interaction representations without paying the full quantum-chemistry cost at deployment.

What carries the argument

The central object is the electron-cloud tokenizer: a patchified, vector-quantized encoder-decoder that compresses tens of thousands of density samples into one token per atom, with each atom serving as the center of a patch that collects nearby density points. The encoder is a structure-aware transformer whose attention weights subtract a geometric pairwise-distance term from the classical attention matrix, making the tokens invariant to rigid motions. Each patch is quantized through two hierarchical codebooks, one for 3D electronic structure and one for 2D atomic attributes, and the tokenizer is trained by reconstructing the original electron density field plus a masked atom-type prediction task. An analogous full-atom tokenizer supplies structural context, learned attention weights merge the codebook tokens, and knowledge distillation transfers the merged representation to an electron-cloud-agnostic student model.

What would settle it

Re-run ECBind with the electron-density input replaced by matched random noise while keeping all other training details identical; if per-structure Pearson and Spearman remain near 0.431 and 0.445, the electron-cloud content is not the cause. Alternatively, restore the excluded non-converged complexes to the test set, using densities from a more robust method or a coarse surrogate, and check whether the advantage over the best baseline survives.

Watch

Extended reading notes

Core claim

The paper claims that replacing atom-only input with a tokenized representation of the joint electron-cloud and atomic structure improves the model's ability to predict how tightly a given ligand binds a protein, and in particular how different ligands rank against the same protein. The argument is that electronic features visible in the density—charge polarization, lone pairs, non-covalent interaction patterns—are not fully represented by ball-and-stick coordinates. The same architecture trained without electron-cloud pretraining (ECBind-nptn) keeps competitive overall correlations but performs clearly worse on per-structure metrics, while the full model (ECBind-ptn) achieves the best reported numbers on relative affinity, absolute affinity, and ligand specificity among the compared baselines.

Load-bearing premise

The load-bearing premise is that the fast semi-empirical electron densities computed on the prepared protein-ligand structures are accurate and informative enough to improve affinity ranking, and that discarding complexes where that computation fails to converge does not bias the benchmark.

Editorial extensions

If this is right

  • Relative affinity ranking within a fixed protein improves substantially: per-structure Pearson and Spearman rise by 6.42% and 15.58% over the best baseline.
  • The full electron-cloud-aware model clearly outperforms the non-pretrained variant on per-structure metrics, implying the density helps the model distinguish ligand-induced changes rather than just protein identity.
  • Absolute affinity and ligand specificity tasks also improve, showing the tokenizer transfers beyond pairwise ranking tasks.
  • The distilled electron-cloud-agnostic student loses only about 5.6% per-structure correlation on the relative task, so deployment without electron-cloud computation is plausible.
  • Ablations show that larger embedding dimensions and larger codebooks improve per-structure correlation up to a point, indicating the tokenizer's capacity matters for capturing electronic detail.

Reading between the lines

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

  • Not in the paper: ablating the electron-density input with matched random noise while keeping all other training details identical would settle whether the gains come from the density content or from the extra model capacity; the paper does not run this control.
  • Not in the paper: the tokenizer could transfer to pose ranking or binding-site prediction, where polarization and lone pairs should matter most; the paper evaluates only affinity and specificity.
  • Not in the paper: inspecting learned codebook entries for chemically interpretable patterns (e.g., lone-pair or polarization directions) could validate whether the tokens encode electronic effects; the paper shows representation distinctiveness but not codebook semantics.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces ECBind, a pretraining-finetuning framework that tokenizes GFN2-xTB electron densities of protein-ligand complexes into hierarchical vector-quantized codes and combines them with atom-level tokens for binding affinity and specificity prediction. The tokenizer is pretrained with electron-density reconstruction, masked atom-type prediction, and atom-level structure losses; downstream tasks are solved by an attention-based merge of the codebook tokens, and a knowledge-distilled electron-cloud-agnostic student model is also proposed. On MISATO relative-affinity prediction, ECBind reports per-structure Pearson 0.431 and Spearman 0.445, exceeding the best reported baseline GET (0.405/0.385), with additional results on LBA and LEP.

Significance. If the reported gains are robust, the paper makes a useful contribution by showing that quantum-chemical electron densities can be compressed into discrete tokens and integrated into structure-based affinity models. The work has clear strengths: a non-pretrained control (ECBind-nptn), a distilled student variant (ECBind-stdt), tokenizer-level ablations, loss curves, representation-similarity analysis, and efficiency measurements. The central empirical claim, however, is not yet established because the hyperparameter selection protocol in Appendix F is unspecified, the headline margins are small relative to the reported run-to-run variability, and the benchmark is built on a filtered dataset whose exclusion statistics are not reported.

major comments (5)
  1. [Appendix F, Table 4] Table 4 reports Pearson_PS/Spearman_PS for embedding dimension, codebook size, n_layer_enc, and n_layer_dec, but it does not state whether these values are validation or test metrics. The selected configuration (D=256, codebook_size=2048, n_layer_enc=3, n_layer_dec=3) is exactly the best row and matches the ECBind-ptn row in Table 1. Since the same per-structure metrics are the headline claim, selection on test metrics would directly inflate the reported 6.42% and 15.58% improvements; the sensitivity is large (Pearson_PS 0.314 at D=64 versus 0.431 at D=256). The paper must specify the selection protocol and report test metrics for all configurations, or provide a validation-based selection argument; otherwise the per-structure gain cannot be attributed to electron-cloud pretraining.
  2. [Appendix F, Table 4] The n_layer_enc block in Table 4 duplicates the D values (32, 64, 128, 256), which appears to be a copy error. As printed, the effect of encoder depth cannot be audited, and the claim that the final configuration is simply the best row cannot be verified for that axis. This needs to be corrected and the real encoder-depth sweep reported.
  3. [Table 1, Section 5.2] The reported headline margins are within the reported run-to-run variability. With three runs, ECBind-ptn has per-structure Pearson 0.431±0.022 versus GET 0.405±0.019, and per-structure Spearman 0.445±0.034 versus 0.385±0.056; treating the runs as independent, neither difference reaches conventional significance. The authors should report paired or bootstrap significance tests, or provide additional runs, before claiming state-of-the-art per-structure performance.
  4. [Section 5.1] The text states that 'Structures that fail to converge during computation are treated as low-quality data and excluded from the datasets,' but the paper does not report how many complexes were removed or whether removal correlates with binding-mode difficulty. Because the entire benchmark is built on this filtered set, the per-structure gains could reflect a test set that is systematically easier or different. The authors should report exclusion counts for each dataset and analyze the properties of excluded complexes, or otherwise demonstrate robustness to the filtering step.
  5. [General] No code, exact data splits, or preprocessing scripts are provided. Given the custom Reduce/MMFF/XTB preprocessing pipeline and the nonstandard filtered splits, the experiments are not reproducible from the text alone. The authors should release the code, split indices, and generated electron-density features, or at least a precise configuration file, so that the central comparison can be independently audited.
minor comments (6)
  1. [Section 4] There are several typos in the Related Work section, including 'andd hardly reveal' and 'ususlly'; these should be corrected.
  2. [Section 2] The sentence 'if we set the resolution as 0.5 Å, it will leads (20/0.5)^3 nodes' contains a grammatical error and an unstated grid extent; please rephrase and clarify the intended calculation.
  3. [Table 3] In the notation gallery, 'xui Eletron density’s position' should read 'Electron density’s position'.
  4. [Table 7] Table 7 lists hyperparameters for ECBind-nptn but not for the headline ECBind-ptn and ECBind-stdt models; the hyperparameters for all three variants should be given.
  5. [Figure 2] The 'hard' protein samples in Figure 2 are selected post hoc without a stated criterion; please explain how these examples were chosen so the visualization is not misleading.
  6. [Section 5.3] The heading 'Accurate reconstruction of electron cloud helps' is supported only indirectly by loss curves and tokenizer ablations; if the claim is causal, the paper should provide a direct link between reconstruction accuracy and downstream affinity metrics.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ECBind's electron-cloud tokenization and affinity prediction are empirically connected, not equivalent by construction.

full rationale

The paper's central claim—state-of-the-art per-structure Pearson and Spearman correlations on MISATO—is an empirical comparison against external baselines on an external benchmark. No equation in the paper defines the predicted affinity in terms of the pretraining reconstruction loss, the codebook entries, or any fitted constant; the pretraining objective (Eq. 4, MSE on electron densities; Eq. 5, CE on masked atom types) and the downstream objective (Sec. 3.3, MSE on affinity labels) are distinct. The method is therefore not self-definitional. Self-citations (GeoMHA adaptation, Surface-VQMAE, MAPE-PPI, ECloudGen, distillation references) are used as architectural inspirations or related work, not as a load-bearing 'uniqueness theorem' or as external proof of the headline result; there is no imported-uniqueness or ansatz-smuggling chain. The Appendix F hyperparameter ablation is a legitimate reporting concern: it reports per-structure metrics for different settings, the final configuration matches the best row, and the n_layer_enc block duplicates the D block, so a selection-on-test-set bias cannot be ruled out from the text. However, this is a statistical validity issue for the empirical claim, not circularity: the paper never derives the prediction from the hyperparameters, and the reported result is a measured test-set correlation rather than a fitted value renamed as a prediction. Overall, the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central result depends on the XTB electron density being a meaningful input signal, on the pretraining losses transferring to affinity, and on the random patch sampling preserving interaction-relevant information. These are empirical design assumptions rather than derived facts. No new physical entities are introduced, and no mathematical derivation is claimed.

free parameters (8)
  • codebook size |A_u(1)| = |A_u(2)| = 2048
    Chosen via ablation in Appendix F on per-structure Pearson/Spearman; controls quantization capacity.
  • embedding dimension D = 256
    Chosen via ablation in Appendix F on per-structure Pearson/Spearman.
  • encoder layers N_enc = 3
    Chosen via ablation in Appendix F on per-structure Pearson/Spearman.
  • decoder layers N_dec = 3
    Chosen via ablation in Appendix F on per-structure Pearson/Spearman.
  • atom masking ratio = 0.1
    Set in Section 3.1.3 for MAE-style atom type prediction; affects pretraining objective.
  • smoothing coefficient gamma_pt = 0.1
    Controls patch-element sampling probability p_jk in Section 3.1.1.
  • commitment loss weight alpha_u = 10.0
    Weights codebook commitment loss in Eq. 7.
  • patch neighborhood K / Kmax = not reported
    Patch size and sampling pool size are not specified in the paper; they determine how much electron cloud each token sees.
assumptions (5)
  • domain assumption GFN2-xTB electron densities approximate the true electronic structure of the binding complex.
    Section 5.1 computes electron cloud with XTB; the entire method depends on this signal being informative.
  • ad hoc to paper MSE reconstruction of electron densities plus masked atom type prediction is a suitable pretraining objective for binding affinity.
    No theoretical guarantee; validated only empirically in Sections 5.2 and 5.3.
  • ad hoc to paper Random patch-element selection with probability p_jk preserves interaction-relevant electron density information.
    Section 3.1.1 defines this sampling; if it discards chemically important regions, gains would not generalize.
  • domain assumption Experimental binding affinity labels in MISATO/LBA/LEP are reliable enough to benchmark.
    Standard benchmark assumption; the paper does not validate label quality beyond using the published datasets.
  • domain assumption Excluding structures where XTB fails to converge does not bias the benchmark.
    Section 5.1 excludes non-convergent structures; if these are systematically harder complexes, the reported gains could be inflated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tokenizing Electron Cloud in Protein-Ligand Interaction Learning." pith.science (2026). https://pith.science/paper/P3KQ3GLA

@misc{pith2026250519014,
  author       = {Pith},
  title        = {Pith review of: Tokenizing Electron Cloud in Protein-Ligand Interaction Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3KQ3GLA}},
  note         = {Machine review of arXiv:2505.19014}
}
read the original abstract

The affinity and specificity of protein-molecule binding directly impact functional outcomes, uncovering the mechanisms underlying biological regulation and signal transduction. Most deep-learning-based prediction approaches focus on structures of atoms or fragments. However, quantum chemical properties, such as electronic structures, are the key to unveiling interaction patterns but remain largely underexplored. To bridge this gap, we propose ECBind, a method for tokenizing electron cloud signals into quantized embeddings, enabling their integration into downstream tasks such as binding affinity prediction. By incorporating electron densities, ECBind helps uncover binding modes that cannot be fully represented by atom-level models. Specifically, to remove the redundancy inherent in electron cloud signals, a structure-aware transformer and hierarchical codebooks encode 3D binding sites enriched with electron structures into tokens. These tokenized codes are then used for specific tasks with labels. To extend its applicability to a wider range of scenarios, we utilize knowledge distillation to develop an electron-cloud-agnostic prediction model. Experimentally, ECBind demonstrates state-of-the-art performance across multiple tasks, achieving improvements of 6.42\% and 15.58\% in per-structure Pearson and Spearman correlation coefficients, respectively.

Figures

Figures reproduced from arXiv: 2505.19014 by the authors.

Figure 1
Figure 1. Pretraining electron-cloud-aware codebooks and finetuning on the specific tasks. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization: GET (blue) and ECBind (red) on per-structure metrics for ‘hard’ protein samples. the removal of data without labels and filtering based on convergence during electron cloud computation, the final training, validation, and test sets are split according to the benchmark1 , resulting in 9,964, 1,074, and 156 pairs. (ii) Absolute binding affinity with LBA [Liu et al., 2014]: Using a sequence identity spli… view at source ↗
Figure 3
Figure 3. Losses in tokenizers when training and testing, and metrics of variants of ECBind with/without different tokenizers. ligands is often of greater interest, so we evaluate average per-structure Pearson and per-structure Spearman correlation coefficient. To evaluate task (ii) absolute binding affinity, we consider Pearson and Spearman’s correlation and RMSE. To evaluate task (iii) ligand specificity, we use AUROC, AUPR… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Representation similarity of ‘key atoms’ from non-pretrained and pretrained ECBind. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 51 canonical work pages

  1. [1]

    Keir Adams, Kento Abeywardane, Jenna Fromer, and Connor W. Coley. Shepherd: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design, 2024. URL https://arxiv.org/abs/2411.04130

  2. [2]

    Gotennet: Rethinking efficient 3d equivariant graph neural networks

    Sarp Aykent and Tian Xia. Gotennet: Rethinking efficient 3d equivariant graph neural networks. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=5wxCQDtbMo

  3. [3]

    Gfn2-xtb-an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions

    Christoph Bannwarth, Sebastian Ehlert, and Stefan Grimme. Gfn2-xtb-an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions. Journal of chemical theory and computation, 15 3: 0 1652--1671, 2018. URL https://api.semanticscholar.org/CorpusID:73419235

  4. [4]

    Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. MACE : Higher order equivariant message passing neural networks for fast and accurate force fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum...

  5. [5]

    Estimating or propagating gradients through stochastic neurons for conditional computation, 2013

    Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation, 2013. URL https://arxiv.org/abs/1308.3432

  6. [6]

    A bird's-eye view of density-functional theory

    Klaus Capelle. A bird's-eye view of density-functional theory, 2006. URL https://arxiv.org/abs/cond-mat/0211443

  7. [7]

    Are high-degree representations really unnecessary in equivariant graph neural networks? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    Jiacheng Cen, Anyi Li, Ning Lin, Yuxiang Ren, Zihe Wang, and Wenbing Huang. Are high-degree representations really unnecessary in equivariant graph neural networks? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=M0ncNVuGYN

  8. [8]

    Self-supervised learning with random-projection quantizer for speech recognition

    Chung-Cheng Chiu, James Qin, Yu Zhang, Jiahui Yu, and Yonghui Wu. Self-supervised learning with random-projection quantizer for speech recognition. In International Conference on Machine Learning, 2022

Show all 78 references
  1. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

  2. [10]

    Insights into protein--ligand interactions: mechanisms, models, and methods

    Xing Du, Yi Li, Yuan-Ling Xia, Shi-Meng Ai, Jing Liang, Peng Sang, Xing-Lai Ji, and Shu-Qun Liu. Insights into protein--ligand interactions: mechanisms, models, and methods. International journal of molecular sciences, 17 0 (2): 0 144, 2016 a

  3. [11]

    Insights into protein–ligand interactions: Mechanisms, models, and methods

    Xing Du, Yi Li, Yuan-Ling Xia, Shi-Meng Ai, Jing Liang, Peng Sang, Xing lai Ji, and Shu-Qun Liu. Insights into protein–ligand interactions: Mechanisms, models, and methods. International Journal of Molecular Sciences, 17, 2016 b

  4. [12]

    Taming transformers for high-resolution image synthesis, 2021

    Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. URL https://arxiv.org/abs/2012.09841

  5. [13]

    Friesner, Jay L

    Richard A. Friesner, Jay L. Banks, Robert B. Murphy, Thomas A. Halgren, Jasna Klicic, Daniel T. Mainz, Matthew P. Repasky, Eric H. Knoll, Mee Shelley, Jason K. Perry, David E. Shaw, Perry Francis, and Peter S. Shenkin. Glide: a new approach for rapid, accurate docking and scor...

  6. [14]

    Fragment and geometry aware tokenization of molecules for structure-based drug design using language models

    Cong Fu, Xiner Li, Blake Olson, Heng Ji, and Shuiwang Ji. Fragment and geometry aware tokenization of molecules for structure-based drug design using language models. ArXiv, 2024. URL https://api.semanticscholar.org/CorpusID:271903635

  7. [15]

    Insights into the molecular mechanisms of protein-ligand interactions by molecular docking and molecular dynamics simulation: A case of oligopeptide binding protein

    Yi Fu, Ji Zhao, and Zhiguo Chen. Insights into the molecular mechanisms of protein-ligand interactions by molecular docking and molecular dynamics simulation: A case of oligopeptide binding protein. Computational and Mathematical Methods in Medicine, 2018, 2018

  8. [16]

    Directional message passing for molecular graphs, 2022

    Johannes Gasteiger, Janek Groß, and Stephan Günnemann. Directional message passing for molecular graphs, 2022. URL https://arxiv.org/abs/2003.03123

  9. [17]

    Gemnet: Universal directional graph neural networks for molecules, 2024

    Johannes Gasteiger, Florian Becker, and Stephan Günnemann. Gemnet: Universal directional graph neural networks for molecules, 2024. URL https://arxiv.org/abs/2106.08903

  10. [18]

    e3nn: Euclidean neural networks, 2022

    Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks, 2022. URL https://arxiv.org/abs/2207.09453

  11. [19]

    De novo molecular generation via connection-aware motif mining

    Zijie Geng, Shufang Xie, Yingce Xia, Lijun Wu, Tao Qin, Jie Wang, Yongdong Zhang, Feng Wu, and Tie-Yan Liu. De novo molecular generation via connection-aware motif mining. arXiv preprint arXiv:2302.01129, 2023

  12. [20]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In International Conference on Machine Learning, 2017

  13. [21]

    Grosse-Kunstleve, Nicholas K

    Ralf W. Grosse-Kunstleve, Nicholas K. Sauter, Nigel W. Moriarty, and Paul D. Adams. The computational crystallography toolbox: crystallographic algorithms in a reusable software framework. Journal of Applied Crystallography, 35: 0 126--136, 2002. URL https://api.semanticschola...

  14. [22]

    Adaptive fourier neural operators: Efficient token mixers for transformers, 2022

    John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Adaptive fourier neural operators: Efficient token mixers for transformers, 2022. URL https://arxiv.org/abs/2111.13587

  15. [23]

    Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J. Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q. Tran, Jonathan Deaton, Marius Wiggert, Rohil Badkundri, Irhum Shafkat, Jun Gong, Alexander Derry, Raul S. Molina, Neil Thomas, Yousuf Khan, Chetan Mishra, Carolyn K...

  16. [24]

    Masked autoencoders are scalable vision learners, 2021

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. URL https://arxiv.org/abs/2111.06377

  17. [25]

    Distilling the knowledge in a neural network, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. URL https://arxiv.org/abs/1503.02531

  18. [26]

    Equivariant diffusion for molecule generation in 3d

    Emiel Hoogeboom, V ctor Garcia Satorras, Cl \'e ment Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3d. In International Conference on Machine Learning, pages 8867--8887. PMLR, 2022

  19. [27]

    Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks

    Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks. In International Conference on Machine Learning. PMLR, 2023

  20. [28]

    Exploring protein--ligand binding affinity prediction with electron density-based geometric deep learning

    Clemens Isert, Kenneth Atz, Sereina Riniker, and Gisbert Schneider. Exploring protein--ligand binding affinity prediction with electron density-based geometric deep learning. RSC advances, 14 0 (7): 0 4492--4502, 2024

  21. [29]

    Junction tree variational autoencoder for molecular graph generation

    Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning, pages 2323--2332. PMLR, 2018

  22. [30]

    Unsupervised protein-ligand binding energy prediction via neural euler's rotation equation, 2023 a

    Wengong Jin, Siranush Sarkizova, Xun Chen, Nir Hacohen, and Caroline Uhler. Unsupervised protein-ligand binding energy prediction via neural euler's rotation equation, 2023 a . URL https://arxiv.org/abs/2301.10814

  23. [31]

    DSMB ind: an unsupervised generative modeling framework for binding energy prediction

    Wengong Jin, Caroline Uhler, and Nir Hacohen. DSMB ind: an unsupervised generative modeling framework for binding energy prediction. In NeurIPS 2023 Generative AI and Biology (GenBio) Workshop, 2023 b . URL https://openreview.net/forum?id=itp5RDMlNV

  24. [32]

    Generalist equivariant transformer towards 3d molecular interaction learning, 2024

    Xiangzhe Kong, Wenbing Huang, and Yang Liu. Generalist equivariant transformer towards 3d molecular interaction learning, 2024. URL https://openreview.net/forum?id=z3mPLBLfGY

  25. [33]

    Neural operator: learning maps between function spaces with applications to pdes

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: learning maps between function spaces with applications to pdes. J. Mach. Learn. Res., 24 0 (1), January 2023. ISSN 1532-4435

  26. [34]

    Intermolecular forces in biology

    Deborah Leckband and Jacob Israelachvili. Intermolecular forces in biology. Quarterly reviews of biophysics, 34 0 (2): 0 105--267, 2001

  27. [35]

    Autoregressive image generation using residual quantization

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523--11532, 2022

  28. [36]

    Geometry informed tokenization of molecules for language model generation, 2024

    Xiner Li, Limei Wang, Youzhi Luo, Carl Edwards, Shurui Gui, Yuchao Lin, Heng Ji, and Shuiwang Ji. Geometry informed tokenization of molecules for language model generation, 2024. URL https://arxiv.org/abs/2408.10120

  29. [37]

    Neural operator: Graph kernel network for partial differential equations, 2020 a

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations, 2020 a . URL https://arxiv.org/abs/2003.03485

  30. [38]

    Multipole graph neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Andrew Stuart, Kaushik Bhattacharya, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editor...

  31. [39]

    Fourier neural operator for parametric partial differential equations, 2021

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations, 2021. URL https://arxiv.org/abs/2010.08895

  32. [40]

    Equiformer: Equivariant graph attention transformer for 3d atomistic graphs

    Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=KwmPfARgOTD

  33. [41]

    Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations

    Yi-Lun Liao, Brandon M Wood, Abhishek Das, and Tess Smidt. Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=mCOBKZmrzD

  34. [42]

    Haitao Lin, Zhangyang Gao, Yongjie Xu, Lirong Wu, Ling Li, and Stan. Z. Li. Conditional local convolution for spatio-temporal meteorological forecasting, 2021

  35. [43]

    Haitao Lin, Yufei Huang, Meng Liu, Xuan Cindy Li, Shuiwang Ji, and Stan Z. Li. Diffbp: generative diffusion of 3d molecules for target protein binding. Chemical Science, 16: 0 1417 -- 1431, 2022

  36. [44]

    Haitao Lin, Yufei Huang, Odin Zhang, Yunfan Liu, Lirong Wu, Siyuan Li, Zhiyuan Chen, and Stan Z. Li. Functional-group-based diffusion for pocket-specific molecule generation and elaboration. In Thirty-seventh Conference on Neural Information Processing Systems, 2023 a . URL ht...

  37. [45]

    Haitao Lin, Lirong Wu, Yongjie Xu, Yufei Huang, Siyuan Li, Guojiang Zhao, and Stan Z. Li. Non-equispaced fourier neural solvers for PDE s. In ICLR 2023 Workshop on Physics for Machine Learning, 2023 b . URL https://openreview.net/forum?id=JDe8vCNfUO3

  38. [46]

    Pdb-wide collection of binding data: current status of the pdbbind database

    Zhihai Liu, Yan Li, Li Han, Jie Li, Jie Liu, Zhixiong Zhao, Wei Nie, Yuchen Liu, and Renxiao Wang. Pdb-wide collection of binding data: current status of the pdbbind database. Bioinformatics, 31 0 (3): 0 405--412, 10 2014. ISSN 1367-4803. doi:10.1093/bioinformatics/btu626. URL...

  39. [47]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3 0 (3): 0 218–229, March 2021. ISSN 2522-5839. doi:10.1038/s42256-0...

  40. [48]

    Rotamer density estimator is an unsupervised learner of the effect of mutations on protein-protein interaction

    Shitong Luo, Yufeng Su, Zuofan Wu, Chenpeng Su, Jian Peng, and Jianzhu Ma. Rotamer density estimator is an unsupervised learner of the effect of mutations on protein-protein interaction. bioRxiv, pages 2023--02, 2023

  41. [49]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators, 2022

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, Pedram Hassanzadeh, Karthik Kashinath, and Animashree Anandkumar. Fourcastnet: A global data-driven high-...

  42. [50]

    Prot2token: A multi-task framework for protein language processing using autoregressive language modeling

    Mahdi Pourmirzaei, Farzaneh Esmaili, Mohammadreza Pourmirzaei, Duolin Wang, and Dong Xu. Prot2token: A multi-task framework for protein language processing using autoregressive language modeling. bioRxiv, 2024. doi:10.1101/2024.05.31.596915

  43. [51]

    Revealing non-covalent interactions in molecular crystals through their experimental electron densities

    Gabriele Saleh, Carlo Gatti, Leonardo Lo Presti, and Julia Contreras-Garc \' a. Revealing non-covalent interactions in molecular crystals through their experimental electron densities. Chemistry--A European Journal, 18 0 (48): 0 15523--15536, 2012

  44. [52]

    E(n) equivariant graph neural networks, 2022

    Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks, 2022. URL https://arxiv.org/abs/2102.09844

  45. [53]

    Computational complexity of interacting electrons and fundamental limitations of density functional theory

    Norbert Schuch and Frank Verstraete. Computational complexity of interacting electrons and fundamental limitations of density functional theory. Nature physics, 5 0 (10): 0 732--735, 2009

  46. [54]

    u tt, P.-J. Kindermans, H. E. Sauceda, S. Chmiela, A. Tkatchenko, and K.-R. M\

    K. T. Sch\" u tt, P.-J. Kindermans, H. E. Sauceda, S. Chmiela, A. Tkatchenko, and K.-R. M\" u ller. Schnet: a continuous-filter convolutional neural network for modeling quantum interactions. In Proceedings of the 31st International Conference on Neural Information Processing ...

  47. [55]

    Theis, Michael Sattler, and Grzegorz Maria Popowicz

    Till Siebenmorgen, Filipe Menezes, Sabrina Benassou, Erinc Merdivan, Kieran Didi, Andr \'e Santos Dias Mour \ a o, Radosław Kitel, Pietro Li \`o , Stefan Kesselheim, Marie Piraud, Fabian J. Theis, Michael Sattler, and Grzegorz Maria Popowicz. Misato: machine learning dataset o...

  48. [56]

    Multi-scale representation learning on proteins

    Vignesh Ram Somnath, Charlotte Bunne, and Andreas Krause. Multi-scale representation learning on proteins. Advances in Neural Information Processing Systems, 34: 0 25244--25255, 2021

  49. [57]

    Surfpro: Functional protein design based on continuous surface, 2024

    Zhenqiao Song, Tinglin Huang, Lei Li, and Wengong Jin. Surfpro: Functional protein design based on continuous surface, 2024. URL https://arxiv.org/abs/2405.06693

  50. [58]

    a rk, Dominique Beaini, Gabriele Corso, Prudencio Tossou, Christian Dallago, Stephan G \

    Hannes St \"a rk, Dominique Beaini, Gabriele Corso, Prudencio Tossou, Christian Dallago, Stephan G \"u nnemann, and Pietro Li \'o . 3 D infomax improves GNN s for molecular property prediction. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and S...

  51. [59]

    Saprot: Protein language modeling with structure-aware vocabulary

    Jin Su, Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou, and Fajie Yuan. Saprot: Protein language modeling with structure-aware vocabulary. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=6MRm3G4NiU

  52. [60]

    Correia, and Michael M

    Freyr Sverrisson, Jean Feydy, Bruno E. Correia, and Michael M. Bronstein. Fast end-to-end learning on protein surfaces. bioRxiv, 2020. URL https://api.semanticscholar.org/CorpusID:230716669

  53. [61]

    Equivariant transformers for neural network based molecular potentials

    Philipp Th \"o lke and Gianni De Fabritiis. Equivariant transformers for neural network based molecular potentials. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=zNHzqZ9wrRB

  54. [62]

    Paolo Tosco, Nikolaus Stiefl, and Gregory A. Landrum. Bringing the mmff force field to the rdkit: implementation and validation. Journal of Cheminformatics, 6: 0 37 -- 37, 2014. URL https://api.semanticscholar.org/CorpusID:2302357

  55. [63]

    Raphael J. L. Townshend, Martin V \"o gele, Patricia Suriana, Alexander Derry, Alexander Powers, Yianni Laloudakis, Sidhika Balachandar, Brandon M. Anderson, Stephan Eismann, Risi Kondor, Russ B. Altman, and Ron O. Dror. Atom3d: Tasks on molecules in three dimensions. ArXiv, a...

  56. [64]

    Gilchrist, Johannes S \"o ding, and Martin Steinegger

    Michel van Kempen, Stephanie Kim, Charlotte Tumescheit, Milot Mirdita, Jeongjae Lee, Cameron L.M. Gilchrist, Johannes S \"o ding, and Martin Steinegger. Fast and accurate protein structure search with foldseek. Nature Biotechnology, 42: 0 243 -- 246, 2022

  57. [65]

    Learning protein representations via complete 3d graph networks

    Limei Wang, Haoran Liu, Yi Liu, Jerry Kurtin, and Shuiwang Ji. Learning protein representations via complete 3d graph networks. arXiv preprint arXiv:2207.12600, 2022

  58. [66]

    The pdbbind database: collection of binding affinities for protein-ligand complexes with known three-dimensional structures

    Renxiao Wang, Xueliang Fang, Yipin Lu, and Shaomeng Wang. The pdbbind database: collection of binding affinities for protein-ligand complexes with known three-dimensional structures. Journal of medicinal chemistry, 47 12: 0 2977--80, 2004. URL https://api.semanticscholar.org/C...

  59. [67]

    A new perspective on building efficient and expressive 3d equivariant graph neural networks

    weitao Du, Yuanqi Du, Limei Wang, Dieqiao Feng, Guifeng Wang, Shuiwang Ji, Carla P Gomes, and Zhi-Ming Ma. A new perspective on building efficient and expressive 3d equivariant graph neural networks. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. ...

  60. [68]

    Wright and Joseph E

    Matthew A. Wright and Joseph E. Gonzalez. Transformers are deep infinite-dimensional non-mercer binary kernel machines, 2021. URL https://arxiv.org/abs/2106.01506

  61. [69]

    Fang Wu and Stan Z. Li. Surface- VQMAE : Vector-quantized masked auto-encoders on molecular surfaces. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=szxtVHOh0C

  62. [70]

    Lirong Wu, Haitao Lin, Yufei Huang, and Stan Z. Li. Knowledge distillation improves graph structure augmentation for graph neural networks. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. UR...

  63. [71]

    Lirong Wu, Yijun Tian, Yufei Huang, Siyuan Li, Haitao Lin, Nitesh V Chawla, and Stan Z. Li. MAPE - PPI : Towards effective and efficient protein-protein interaction prediction via microenvironment-aware protein embedding. In The Twelfth International Conference on Learning Rep...

  64. [72]

    Lirong Wu, Yijun Tian, Haitao Lin, Yufei Huang, Siyuan Li, Nitesh V Chawla, and Stan Z. Li. Learning to predict mutation effects of protein-protein interactions by microenvironment-aware hierarchical prompt learning, 2024 b . URL https://arxiv.org/abs/2405.10348

  65. [73]

    Geodiff: A geometric diffusion model for molecular conformation generation

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923, 2022

  66. [74]

    Vector-quantized image modeling with improved vqgan, 2022

    Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan, 2022. URL https://arxiv.org/abs/2110.04627

  67. [75]

    Proteinmae: masked autoencoder for protein surface self-supervised learning

    Mingzhi Yuan, Ao Shen, Kexue Fu, Jiaming Guan, Yingfan Ma, Qin Qiao, and Manning Wang. Proteinmae: masked autoencoder for protein surface self-supervised learning. Bioinformatics, 39 0 (12): 0 btad724, 11 2023. ISSN 1367-4811. doi:10.1093/bioinformatics/btad724. URL https://do...

  68. [76]

    Ecloudgen: Access to broader chemical space for structure-based molecule generation

    Odin Zhang, Jieyu Jin, Haitao Lin, Jintu Zhang, Chenqing Hua, Yufei Huang, Huifeng Zhao, Chang-Yu Hsieh, and Tingjun Hou. Ecloudgen: Access to broader chemical space for structure-based molecule generation. bioRxiv, 2024. doi:10.1101/2024.06.03.597263. URL https://www.biorxiv....

  69. [77]

    Reliable data distillation on graph convolutional network

    Wentao Zhang, Xupeng Miao, Yingxia Shao, Jiawei Jiang, Lei Chen, Olivier Ruas, and Bin Cui. Reliable data distillation on graph convolutional network. In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data, SIGMOD '20, page 1399–1414, New York, NY...

  70. [78]

    Equipocket: an e(3)-equivariant geometric graph neural network for ligand binding site prediction

    Yang Zhang, Zhewei Wei, Ye Yuan, Chongxuan Li, and Wenbing Huang. Equipocket: an e(3)-equivariant geometric graph neural network for ligand binding site prediction. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2025

Pith tools

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