Pith. sign in

REVIEW 3 major objections 5 minor 59 references

IBEX: Information-Bottleneck-EXplored Coarse-to-Fine Molecular Generation under Limited Data

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read IBEX claims that training a diffusion generator on scaffold hopping plus a rigid-body physics refinement lifts zero-shot docking success from 53% to 64% on CBGBench.

desk verdict A useful, plausibly real result — SH training transfers zero-shot to de novo docking — but the evaluation protocol is under-specified enough that the headline gain could be an artifact, and the theory is thinner than advertised. read the letter →

arxiv 2508.10775 v1 pith:J66HEQT3 submitted 2025-08-14 cs.LG q-bio.BM

Dong Xu , Zhangfan Yang , Jenna Xinyi Yao , Shuangbao Song , Zexuan Zhu , Junkai Ji This is my paper
classification cs.LGq-bio.BM
keywords structure-baseddrugdesigndiffusionmodelsscaffoldhoppinginformationbottleneckPAC-Bayesboundrigid-bodyrefinementzero-shotgeneralizationlimiteddata
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 is trying to establish that the information bottleneck in structure-based drug design is not the model architecture but the information content of the training task. It shows that a scaffold-hopping task—keep the side chains fixed, regenerate the molecular core—gives each sample higher information density than de novo generation, tightening a PAC-Bayes bound on test error. On top of that, it adds a cheap rigid-body refinement of the generated pose, minimizing a five-term physics energy with L-BFGS. With the same TargetDiff network, these two changes push zero-shot docking success from 53% to 64%, improve mean Vina from -7.41 to -8.07 kcal/mol, and beat the baseline in 57 of 100 pockets.

What carries the argument

The load-bearing mechanism is the pairing of a task-level information bottleneck with rigid-body physics refinement. The task indicator selects a masking pattern after Bemis–Murcko decomposition of the ligand into scaffold and side chains; the scaffold-hopping (SH) task fixes side chains and regenerates the core, giving the highest information density $\rho \propto I(Z;X)/4$, which tightens the PAC-Bayes test-risk bound $R \propto \rho \times |D|$. The second stage treats the generated ligand as a rigid body and minimizes the five-term energy $E_{\mathrm{phys}}(P,M) = W^T (E_{g1}+E_{g2}+E_{rep}+E_{hyd}+E_{hd})$ over 3 translations and 3 rotations via finite-difference L-BFGS. The generative

What would settle it

Re-calibrate $W^T$ on a panel of pockets from a different residue-class distribution than the test set (e.g., metal-rich or highly charged pockets) and run the pipeline; if the Vina-score improvement over TargetDiff shrinks or reverses, the fixed-weight surrogate does not transfer. Similarly, train the same SH masking with a different backbone (e.g., MolCRAFT) and check whether the 53% to 64% success improvement still holds.

Watch

Extended reading notes

Core claim

The central discovery is that the effective information density of a training example, not the model capacity, is what limits generalization when protein–ligand complexes are scarce. Comparing three masking tasks derived from the same Bemis–Murcko decomposition, IBEX finds an ordering $\rho_{\mathrm{SH}} > \rho_{\mathrm{DN}} > \rho_{\mathrm{SC}}$ and shows that the scaffold-hopping task, despite its higher difficulty, produces earlier capacity compression, lower gradient-SNR variance, and a smaller train–test gap. Retaining the TargetDiff architecture unchanged, IBEX then refines the coarse pose by minimizing $E_{\mathrm{phys}}(P,M) = W^T (E_{g1}+E_{g2}+E_{rep}+E_{hyd}+E_{hd})$ over the six

Load-bearing premise

The five-term physics energy with weights set once on a small calibration panel is a transferable surrogate for binding affinity, so refining with those fixed weights improves held-out docking scores.

Editorial extensions

If this is right

  • A model trained only on scaffold-hopping pairs generalizes zero-shot to de novo generation on held-out pockets, raising docking success from 53% to 64% without fine-tuning.
  • A post-hoc rigid-body refinement that takes under one second per molecule can improve docking energies without modifying the generative model.
  • Task construction—what is masked during training—matters as much as architecture under data scarcity.
  • The PAC-Bayes information-density ordering gives a principled basis for choosing among generative tasks when data are limited.
  • The same pipeline raises QED by 25% while maintaining state-of-the-art validity (0.96) and uniqueness (0.9507) on 100-molecule sets.

Reading between the lines

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

  • The information-density measure $Z=(\bar{n},\bar{d},\bar{t},\bar{k})$ could serve as a dataset-difficulty diagnostic for other protein–ligand benchmarks, with the prediction that higher $\rho$ correlates with tighter generalization gaps.
  • Because the refinement weights $W^T$ were calibrated on a small panel, the docking gains may be pocket-type-dependent; an editor's testable extension is to re-calibrate the surrogate on diverse pockets and check whether the reported improvements persist.
  • Since the architecture is frozen, IBEX decouples representation learning from pose optimization; one could swap in different backbones (e.g., MolCRAFT, DiffBP) with the same SH task to test whether the gains transfer.
  • The ordering $\rho_{\mathrm{SH}} > \rho_{\mathrm{DN}} > \rho_{\mathrm{SC}}$ suggests a general design rule for scarce-data molecular generation: maximize per-sample mutual information even at the cost of higher task failure rate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes IBEX, a two-stage pipeline for structure-based molecular generation under data scarcity. The first stage trains an SE(3)-equivariant diffusion generator (stated to keep the TargetDiff architecture and hyperparameters) on the scaffold-hopping (SH) task, in which side-chain context is fixed and the model proposes alternative scaffolds; the second stage refines sampled poses by L-BFGS over six rigid-body degrees of freedom, minimizing a weighted sum of five short-range physics terms. On CBGBench (CrossDocked2020), the paper reports raising zero-shot docking success from 53% to 64%, improving mean Vina from -7.41 to -8.07 kcal/mol, and achieving the best median Vina energy in 57 of 100 pockets. A PAC-Bayes information-bottleneck analysis is invoked to argue that SH has higher information density and a tighter generalization bound. The paper includes ablations (SH/SC/DN, with and without physics refinement) and comparisons against several generative models, but several load-bearing details, including the inference-time context protocol and the PAC-Bayes derivation, are missing or deferred to appendices that are not provided.

Significance. If the empirical claims are substantiated, the result is significant: it would show that constrained scaffold-hopping training transfers zero-shot to de novo generation and that a simple rigid-body physics refinement improves docking scores, all without changing the generative architecture. The design of keeping the TargetDiff backbone and isolating the training-task change is a strength, as is the use of a 100-pocket held-out benchmark with per-pocket medians. However, the central empirical comparison depends on an unspecified inference protocol for the SH-trained model, and the information-theoretic support is asserted rather than demonstrated, with key appendices missing. In its current form the paper is a promising empirical report but does not yet establish the theoretical or experimental claims as stated.

major comments (3)
  1. [Methods, 3D Pocket-aware Diffusion; Results, Table 1] The inference-time context for the SH-trained model is never specified. Methods defines SH as 'receives C and proposes alternative scaffolds' (Figure 2), while at inference the text only says 'ancestral sampling yields a coarse pose M0 = Gθ(P)' without stating whether the side-chain set C is provided. Table 1 compares IBEX-SH against de novo baselines (TargetDiff, DiffSBDD, MolCRAFT, etc.). If CBGBench's SH evaluation supplies the reference ligand's side chains, then the comparison is not de novo and the 53%→64% gain may be an artifact of a reduced search space. If the SH model is evaluated with empty context, the paper must explain how a model trained with fixed side chains handles an empty context. Please state explicitly the task and context mask used for each row of Table 1 and report a true de novo evaluation of the SH-trained model.
  2. [Results, Task-dependent Information-Bottleneck Analysis] The claim that SH 'tightens the PAC-Bayes bound on test risk by 38% and 47% relative to DN and SC' is unsupported in the manuscript. No PAC-Bayes risk bound is stated, no prior is defined, no mutual-information estimator is described, and no numerical computation of I(Z; X) is shown; the text only says 'For details, please refer to Appendix B', which is not provided. Moreover, ρ = I(Z; X)/4 relies on a 4-D summary Z=(n̄,d̄,t̄,k̄) that appears to be chosen post hoc to match observed outcomes. This makes the ordering ρ_SH > ρ_DN > ρ_SC circular as an explanation of generalization. Please provide the full derivation, estimator, and validation (e.g., a pre-specified Z or a held-out complexity measure).
  3. [Methods, Physics-guided Position Refinement, Eq. (8)] The physics energy Ephys(P,M) = W^T (Eg1+Eg2+Erep+Ehyd+Ehd) is central to the coarse-to-fine claim, but the five terms, the functional forms, the calibration panel, and the weight values are not described; Appendix D is referenced but missing. Saying that weights W^T 'were set once on a small calibration panel' is insufficient without panel composition and sensitivity analysis. The ablation shows only a small gain from PR (Dock 63.50→63.69, Vina Score -5.06→-5.23), so this issue does not invalidate the SH result, but it prevents assessment of the refinement contribution and of its transferability across pockets.
minor comments (5)
  1. [Table 1] The row 'IBEX-DN ✗ ✗' exactly reproduces the TargetDiff row (-5.71, 38.21, -6.43, 47.09, -7.41, 51.99, etc.). If this is a re-run, state so; if it is the published TargetDiff baseline, clarify why the identical architecture and hyperparameters yield exactly identical numbers, and explain what 'IBEX-DN' adds as a control.
  2. [Results, Task Difficulty under Geometric Constraints] The sentence 'SH task exposes the network to explicit side chain–pocket interactions during training but leaves these atoms un-denoised; at test time, the model must denoise them from scratch' is confusing and appears to contradict the Methods definition of SH (fixed side chain context). Please clarify whether 'these atoms' are side-chain atoms and whether they are provided or masked at test time.
  3. [Tables and references] There are typographical inconsistencies: 'MolCARFT' should be 'MolCRAFT' in Table 1; the Boopathy et al. 2023a and 2023b references are identical; the author name '˘Skrinjar' is malformed. Please correct these.
  4. [Methods, Physics-guided Position Refinement] The finite-difference step size ε=10^-3 and L-BFGS learning rate 0.1 are stated as fixed, but no convergence criterion, maximum iteration behavior, or sensitivity analysis is reported. The statement 'in under one second' should be supported by a runtime measurement.
  5. [Results, Batch Generative Performance Evaluation] For the out-of-distribution pocket 9F7W, the comparison with Pocket2Mol is anecdotal (one pocket, one model). If this is intended as evidence of diversity, consider reporting over a small set of held-out pockets with standard errors.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; central SH-transfer and physics-refinement claims are empirically supported and not forced by construction.

full rationale

Walked the claimed derivation chain. (1) Information-bottleneck analysis: rho = I(Z;X)/4 is computed from empirical distributions of the hand-crafted summary Z = (nbar,dbar,tbar,kbar). The claim that SH has the largest rho and hence a tighter PAC-Bayes bound is a data-dependent quantity logically independent of the test-risk comparison; the empirical generalization ordering could in principle have failed to match it. Even if Z is a post-hoc interpretive choice, no equation in the paper defines Z in terms of the observed docking outcomes, so the confirmation is not by construction. (2) Physics refinement: W in Eq. 8 is calibrated once and held fixed, the test metric is AutoDock Vina rather than Ephys, and the refinement is only a 6-DoF rigid-body search, so improved Vina is an empirical outcome, not a fitted prediction. (3) The IBEX-DN row exactly reproduces TargetDiff, which is a proper control showing the architecture is held fixed. (4) The only overlapping-author citation is [Yang et al. 2024] for the weak pocket perturbation in Eq. 5, a non-load-bearing detail. Caveat: the paper never explicitly specifies the inference-time context for the SH-trained model (Methods says 'SH receives C' but inference is written as M0 = Gtheta(P) with no C). If the SH model is evaluated with side-chain context while baselines are de novo, the headline comparison would be a task-mismatch artifact. This is a serious evaluation-protocol gap, but it is a validity concern, not a demonstrated circular reduction, so it does not raise the circularity score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central performance claim relies on the physics refinement weights W, which are calibrated rather than derived. The theoretical motivation relies on a chosen 4D encoding and an externally sourced PAC-Bayes bound, both assumed to transfer to this setting. No new physical entities are introduced, but the information density rho is a newly coined summary of the model task.

free parameters (3)
  • Physics energy weights W
    Weights of the five-term Ephys, set once on a small calibration panel; values not reported in main text. They determine the refinement objective and directly affect the docking improvement.
  • Step size epsilon for finite differences = 10^-3
    Used for gradient approximation in PR; fixed after scale normalization.
  • L-BFGS learning rate = 0.1
    Fixed learning rate for refinement updates.
assumptions (4)
  • domain assumption PAC-Bayes information-bottleneck bound (Wang et al. 2022) applies to 3D molecular generation and controls test risk via mutual information
    Invoked in the task-dependent information-bottleneck analysis to derive rho and claim bound tightening; not re-derived or adapted to the diffusion setting in the visible text.
  • ad hoc to paper The 4D summary Z=(n_bar, d_bar, t_bar, k_bar) accurately captures the information content of a protein-ligand complex relevant to generation
    Defined in the analysis section; this encoding is constructed by the authors and is not an established or externally validated representation for PAC-Bayes bounds.
  • domain assumption Bemis-Murcko decomposition cleanly separates scaffold and side-chain, supporting the SH/SC/DN task definitions
    Used in Figure 2 to define tasks; standard in medicinal chemistry but sensitive to ring/linker edge cases.
  • domain assumption AutoDock Vina scores are a valid and sufficient evaluation of generated molecule binding quality
    Used as the primary success metric; no comparison to alternative scoring or experimental binding data is provided.
invented entities (1)
  • Information density rho = I(Z; X)/4
    purpose: Quantify per-sample information content of each task and justify SH as the high-information regime
    A new composite quantity defined for this paper; its predictive value for generalization is asserted but not validated against external benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IBEX: Information-Bottleneck-EXplored Coarse-to-Fine Molecular Generation under Limited Data." pith.science (2026). https://pith.science/paper/J66HEQT3

@misc{pith2026250810775,
  author       = {Pith},
  title        = {Pith review of: IBEX: Information-Bottleneck-EXplored Coarse-to-Fine Molecular Generation under Limited Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J66HEQT3}},
  note         = {Machine review of arXiv:2508.10775}
}
abstract

Three-dimensional generative models increasingly drive structure-based drug discovery, yet it remains constrained by the scarce publicly available protein-ligand complexes. Under such data scarcity, almost all existing pipelines struggle to learn transferable geometric priors and consequently overfit to training-set biases. As such, we present IBEX, an Information-Bottleneck-EXplored coarse-to-fine pipeline to tackle the chronic shortage of protein-ligand complex data in structure-based drug design. Specifically, we use PAC-Bayesian information-bottleneck theory to quantify the information density of each sample. This analysis reveals how different masking strategies affect generalization and indicates that, compared with conventional de novo generation, the constrained Scaffold Hopping task endows the model with greater effective capacity and improved transfer performance. IBEX retains the original TargetDiff architecture and hyperparameters for training to generate molecules compatible with the binding pocket; it then applies an L-BFGS optimization step to finely refine each conformation by optimizing five physics-based terms and adjusting six translational and rotational degrees of freedom in under one second. With only these modifications, IBEX raises the zero-shot docking success rate on CBGBench CrossDocked2020-based from 53% to 64%, improves the mean Vina score from $-7.41 kcal mol^{-1}$ to $-8.07 kcal mol^{-1}$, and achieves the best median Vina energy in 57 of 100 pockets versus 3 for the original TargetDiff. IBEX also increases the QED by 25%, achieves state-of-the-art validity and diversity, and markedly reduces extrapolation error.

Figures

Figures reproduced from arXiv: 2508.10775 by the authors.

Figure 1
Figure 1. Conceptual overview of the IBEX pipeline. Start [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Workflow of the Bemis–Murcko decomposition. Starting from the full ligand structure, the algorithm (i) iden￾tifies all ring systems and the linkers that connect them, (ii) removes peripheral side chains that are not part of a ring or linker, and (iii) collapses any fused ring junctions to gener￾ate the unique Bemis–Murcko scaffold. its placement in the pocket using a lightweight, gradient￾based search over rigid-bod… view at source ↗
Figure 3
Figure 3. Hexagon–bin density maps for SC, DN, and SH. Top panels: mean edge length ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Harder geometry implies higher Im, supporting the ordering above. IBEX exhibits the highest failure rate, followed by DecompDiff, TargetDiff, and MolCRAFT. (Pocket2Mol (Peng et al. 2022), GraphBP (Liu et al. 2022a), FLAG (Zhang and Liu 2023)). (Refer to Appendix E.) Ta…
Figure 5
Figure 5. Figure 5: Training and validation dynamics of IBEX models on three generation tasks. Panels show atom-type classification [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Train–test comparison on 18 normalised metrics for three IBEX settings. For each IBEX variant we track [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Per-target docking performance of five generative pipelines on 100 held-out CBGBench (same as CrossDocked2020) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Docking poses and drug-likeness of median-score candidates on four CBGBench test pockets. Rows list the crystal [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 45 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Abramson, J.; Adler, J.; Dunger, J.; Evans, R.; Green, T.; Pritzel, A.; and et al. 2024. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630(8016): 493--500

    work page 2024

  4. [4]

    L.; Long, P

    Bartlett, P. L.; Long, P. M.; Lugosi, G.; and Tsigler, A. 2020. Benign overfitting in linear regression. Proceedings of the National Academy of Sciences, 117(48): 30063--30070

    work page 2020

  5. [5]

    P.; Hersey, A.; F \'e lix, E.; Landrum, G.; Gaulton, A.; Atkinson, F.; Bellis, L

    Bento, A. P.; Hersey, A.; F \'e lix, E.; Landrum, G.; Gaulton, A.; Atkinson, F.; Bellis, L. J.; De Veij, M.; and Leach, A. R. 2020. An open source chemical structure curation pipeline using RDKit. Journal of Cheminformatics, 12(1): 51

    work page 2020

  6. [6]

    Betker, J.; Goh, G.; Jing, L.; Brooks, T.; Wang, J.; Li, L.; Ouyang, L.; Zhuang, J.; Lee, J.; Guo, Y.; et al. 2023. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2(3): 8

    2023

  7. [7]

    Biroli, G.; Bonnaire, T.; De Bortoli, V.; and M \'e zard, M. 2024. Dynamical regimes of diffusion models. Nature Communications, 15(1): 9957

    work page 2024

  8. [8]

    Boopathy, A.; Liu, K.; Hwang, J.; Ge, S.; Mohammedsaleh, A.; and Fiete, I. R. 2023 a . Model-agnostic measure of generalization difficulty. In International Conference on Machine Learning, 2857--2884. PMLR

    work page 2023

Show all 59 references
  1. [9]

    Boopathy, A.; Liu, K.; Hwang, J.; Ge, S.; Mohammedsaleh, A.; and Fiete, I. R. 2023 b . Model-agnostic measure of generalization difficulty. In International Conference on Machine Learning, 2857--2884. PMLR

    2023

  2. [10]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  3. [11]

    Cai, H.; Shen, C.; Jian, T.; Zhang, X.; Chen, T.; Han, X.; Yang, Z.; Dang, W.; Hsieh, C.-Y.; Kang, Y.; et al. 2024. CarsiDock: a deep learning paradigm for accurate protein--ligand docking and screening based on large-scale pre-training. Chemical Science, 15(4): 1449--1471

    2024

  4. [12]

    Corso, G.; St \"a rk, H.; Jing, B.; Barzilay, R.; and Jaakkola, T. 2022. Diffdock: Diffusion steps, twists, and turns for molecular docking. arXiv preprint arXiv:2210.01776

    2022 arXiv

  5. [13]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language tec...

    2019

  6. [14]

    B.; Snyder, I.; and Koes, D

    Francoeur, P.; Masuda, T.; Sunseri, J.; Jia, A.; Iovanisci, R. B.; Snyder, I.; and Koes, D. R. 2020. 3D Convolutional neural networks and a crossdocked dataset for structure-based drug design. Journal of chemical information and modeling

    2020

  7. [15]

    E.; and Coley, C

    Graff, D. E.; and Coley, C. W. 2022. MolPAL: Software for Sample Efficient High-Throughput Virtual Screening. In AI for Accelerated Materials Design NeurIPS 2022 Workshop

    2022

  8. [16]

    W.; Peng, X.; Su, Y.; Peng, J.; and Ma, J

    Guan, J.; Qian, W. W.; Peng, X.; Su, Y.; Peng, J.; and Ma, J. 2023. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543

    2023 arXiv

  9. [17]

    Guan, J.; Zhou, X.; Yang, Y.; Bao, Y.; Peng, J.; Ma, J.; Liu, Q.; Wang, L.; and Gu, Q. 2024. DecompDiff: diffusion models with decomposed priors for structure-based drug design. arXiv preprint arXiv:2403.07902

    2024 arXiv

  10. [18]

    G.; Vignac, C.; and Welling, M

    Hoogeboom, E.; Satorras, V. G.; Vignac, C.; and Welling, M. 2022. Equivariant diffusion for molecule generation in 3d. In International conference on machine learning, 8867--8887. PMLR

    2022

  11. [19]

    Huang, Z.; Yang, L.; Zhou, X.; Zhang, Z.; Zhang, W.; Zheng, X.; Chen, J.; Wang, Y.; Cui, B.; and Yang, W. 2024. Protein-ligand interaction prior for binding-aware 3d molecule diffusion models. In The Twelfth International Conference on Learning Representations

    2024

  12. [20]

    Huang, Z.; Zhou, P.; Yan, S.; and Lin, L. 2023. Scalelong: Towards more stable training of diffusion model via scaling network long skip connection. Advances in Neural Information Processing Systems, 36: 70376--70401

    2023

  13. [21]

    R.; van der Schaar, M.; and Deane, C

    Imrie, F.; Bradley, A. R.; van der Schaar, M.; and Deane, C. M. 2020. Deep generative models for 3D linker design. Journal of chemical information and modeling, 60(4): 1983--1995

    2020

  14. [22]

    Jiang, Y.; Zhang, G.; You, J.; Zhang, H.; Yao, R.; Xie, H.; Zhang, L.; Xia, Z.; Dai, M.; Wu, Y.; et al. 2024. Pocketflow is a data-and-knowledge-driven structure-based molecular generative model. Nature Machine Intelligence, 6(3): 326--337

    2024

  15. [23]

    Jin, W.; Barzilay, R.; and Jaakkola, T. 2018. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning, 2323--2332. PMLR

    2018

  16. [24]

    Liao, Z.; Xie, L.; Mamitsuka, H.; and Zhu, S. 2023. Sc2Mol: a scaffold-based two-step molecule generator with variational autoencoder and transformer. Bioinformatics, 39(1): btac814

    2023

  17. [25]

    Lim, J.; Hwang, S.-Y.; Moon, S.; Kim, S.; and Kim, W. Y. 2020. Scaffold-based molecular design with a graph generative model. Chemical science, 11(4): 1153--1164

    2020

  18. [26]

    Lin, H.; Huang, Y.; Zhang, O.; Liu, Y.; Wu, L.; Li, S.; Chen, Z.; and Li, S. Z. 2023. Functional-group-based diffusion for pocket-specific molecule generation and elaboration. Advances in Neural Information Processing Systems, 36: 34603--34626

    2023

  19. [27]

    Lin, H.; Huang, Y.; Zhang, O.; Ma, S.; Liu, M.; Li, X.; Wu, L.; Wang, J.; Hou, T.; and Li, S. Z. 2025. Diffbp: Generative diffusion of 3d molecules for target protein binding. Chemical Science, 16(3): 1417--1431

    2025

  20. [28]

    Lin, H.; Zhao, G.; Zhang, O.; Huang, Y.; Wu, L.; Liu, Z.; Li, S.; Tan, C.; Gao, Z.; and Li, S. Z. 2024. CBGBench: fill in the blank of protein-molecule complex binding graph. arXiv preprint arXiv:2406.10840

    2024 arXiv

  21. [29]

    Liu, M.; Luo, Y.; Uchino, K.; Maruhashi, K.; and Ji, S. 2022 a . Generating 3d molecules for target protein binding. arXiv preprint arXiv:2204.09410

    2022 arXiv

  22. [30]

    S.; Michaud, E.; Tegmark, M.; and Williams, M

    Liu, Z.; Kitouni, O.; Nolte, N. S.; Michaud, E.; Tegmark, M.; and Williams, M. 2022 b . Towards understanding grokking: An effective theory of representation learning. Advances in Neural Information Processing Systems, 35: 34651--34663

    2022

  23. [31]

    Luo, S.; Guan, J.; Ma, J.; and Peng, J. 2021. A 3D generative model for structure-based drug design. Advances in Neural Information Processing Systems, 34: 6229--6239

    2021

  24. [32]

    Luo, Y.; Yan, K.; and Ji, S. 2021. Graphdf: A discrete flow model for molecular graph generation. In International conference on machine learning, 7192--7203. PMLR

    2021

  25. [33]

    Lyu, Y.; Liu, X.; Song, M.; Wang, X.; Peng, Y.; Zeng, T.; and Jing, L. 2023. Recognizable information bottleneck. arXiv preprint arXiv:2304.14618

    2023 arXiv

  26. [34]

    R.; Mahmoud, A

    Masters, M. R.; Mahmoud, A. H.; and Lill, M. A. 2024. Do Deep Learning Models for Co-Folding Learn the Physics of Protein-Ligand Interactions? bioRxiv

    2024

  27. [35]

    Morehead, A.; and Cheng, J. 2024. Geometry-complete diffusion for 3D molecule generation and optimization. Communications Chemistry, 7(1): 150

    2024

  28. [36]

    Nittinger, E.; Yoluk, O.; Tibo, A.; Olanders, G.; and Tyrchan, C. 2025. Co-folding, the future of docking – prediction of allosteric and orthosteric ligands. Artificial Intelligence in the Life Sciences, 8: 100136

    2025

  29. [37]

    Peng, X.; Luo, S.; Guan, J.; Xie, Q.; Peng, J.; and Ma, J. 2022. Pocket2mol: Efficient molecular sampling based on 3d protein pockets. In International conference on machine learning, 17644--17655. PMLR

    2022

  30. [38]

    O.; Jamasb, A.; Mahmood, O.; Sresht, V.; and Saremi, S

    Pinheiro, P. O.; Jamasb, A.; Mahmood, O.; Sresht, V.; and Saremi, S. 2024. Structure-based drug design by denoising voxel grids. arXiv preprint arXiv:2405.03961

    2024 arXiv

  31. [39]

    Power, A.; Burda, Y.; Edwards, H.; Babuschkin, I.; and Misra, V. 2022. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177

    2022 arXiv

  32. [40]

    Qu, Y.; Qiu, K.; Song, Y.; Gong, J.; Han, J.; Zheng, M.; Zhou, H.; and Ma, W.-Y. 2024. MolCRAFT: structure-based drug design in continuous parameter space. arXiv preprint arXiv:2404.12141

    2024 arXiv

  33. [41]

    Ragoza, M.; Masuda, T.; and Koes, D. R. 2022. Generating 3D molecules conditional on receptor binding sites with deep generative models . Chem Sci, 13: 2701--2713

    2022

  34. [42]

    Rohlfs, C. 2025. Generalization in neural networks: A broad survey. Neurocomputing, 611: 128701

    2025

  35. [43]

    V.; and Katritch, V

    Sadybekov, A. V.; and Katritch, V. 2023. Computational approaches streamlining drug discovery. Nature, 616(7958): 673--685

    2023

  36. [44]

    J.; Adasme, M

    Salentin, S.; Schreiber, S.; Haupt, V. J.; Adasme, M. F.; and Schroeder, M. 2015. PLIP: fully automated protein--ligand interaction profiler. Nucleic acids research, 43(W1): W443--W447

    2015

  37. [45]

    L.; Lio, P.; et al

    Schneuing, A.; Harris, C.; Du, Y.; Didi, K.; Jamasb, A.; Igashov, I.; Du, W.; Gomes, C.; Blundell, T. L.; Lio, P.; et al. 2024. Structure-based drug design with equivariant diffusion models. Nature Computational Science, 4(12): 899--909

    2024

  38. [46]

    Shi, C.; Xu, M.; Zhu, Z.; Zhang, W.; Zhang, M.; and Tang, J. 2020. GraphAF : a Flow-based Autoregressive Model for Molecular Graph Generation. In International Conference on Learning Representations

    2020

  39. [47]

    Trott, O.; and Olson, A. J. 2010. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of computational chemistry, 31(2): 455--461

    2010

  40. [48]

    u nze, G.; and Str \

    Useini, A.; Schwerin, I. K.; K \"u nze, G.; and Str \"a ter, N. 2024. Structural Studies on the Binding Mode of Bisphenols to PPAR . Biomolecules, 14(6): 640

    2024

  41. [49]

    S krinjar, P.; Eberhardt, J.; Durairaj, J.; and Schwede, T. 2025. Have protein-ligand co-folding methods moved beyond memorisation? bioRxiv

    2025

  42. [50]

    Wang, R.; Fang, X.; Lu, Y.; Yang, C.-Y.; and Wang, S. 2005. The PDBbind database: methodologies and updates. Journal of medicinal chemistry, 48(12): 4111--4119

    2005

  43. [51]

    E.; Sun, J.; Chen, X.; and Zheng, Y

    Wang, Z.; Huang, S.-L.; Kuruoglu, E. E.; Sun, J.; Chen, X.; and Zheng, Y. 2022. PAC-Bayes Information Bottleneck. In International Conference on Learning Representations

    2022

  44. [52]

    Wu, K.; Xia, Y.; Deng, P.; Liu, R.; Zhang, Y.; Guo, H.; Cui, Y.; Pei, Q.; Wu, L.; Xie, S.; et al. 2024. TamGen: drug design with target-aware molecule generation through a chemical language model. Nature Communications, 15(1): 9360

    2024

  45. [53]

    Xu, M.; Yu, L.; Song, Y.; Shi, C.; Ermon, S.; and Tang, J. 2022. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923

    2022 arXiv

  46. [54]

    Yang, Z.; Ji, J.; He, S.; Li, J.; He, T.; Bai, R.; Zhu, Z.; and Ong, Y. S. 2024. Dockformer: A transformer-based molecular docking paradigm for large-scale virtual screening. arXiv preprint arXiv:2411.06740

    2024 arXiv

  47. [55]

    Zhang, O.; Zhang, J.; Jin, J.; Zhang, X.; Hu, R.; Shen, C.; Cao, H.; Du, H.; Kang, Y.; Deng, Y.; et al. 2023 a . ResGen is a pocket-aware 3D molecular generation model based on parallel multiscale modelling. Nature Machine Intelligence, 5(9): 1020--1030

    2023

  48. [56]

    Zhang, X.; Zhang, O.; Shen, C.; Qu, W.; Chen, S.; Cao, H.; Kang, Y.; Wang, Z.; Wang, E.; Zhang, J.; et al. 2023 b . Efficient and accurate large library ligand docking with KarmaDock. Nature Computational Science, 3(9): 789--804

    2023

  49. [57]

    Zhang, Y.; Zhang, Z.; Ke, D.; Pan, X.; Wang, X.; Xiao, X.; and Ji, C. 2024. FragGrow: A web server for structure-based drug design by fragment growing within constraints. Journal of Chemical Information and Modeling, 64(10): 3970--3976

    2024

  50. [58]

    Zhang, Z.; and Liu, Q. 2023. Learning subpocket prototypes for generalizable structure-based drug design. In International Conference on Machine Learning, 41382--41398. PMLR

    2023

  51. [59]

    Zhou, G.; Gao, Z.; Ding, Q.; Zheng, H.; Xu, H.; Wei, Z.; Zhang, L.; and Ke, G. 2023. Uni-mol: A universal 3d molecular representation learning framework. In International Conference on Learning Representations

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.