Pith. sign in

REVIEW 5 major objections 7 minor 1 cited by

Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration

T0 review · 5 major / 7 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 Gradient-guided Bayesian Flow Networks, the paper claims, beat diffusion and optimization baselines on binding affinity, synthetic feasibility, and selectivity by steering 3D drug generation in parameter space, not sample space.

desk verdict Useful evaluation scheme and a plausible guidance idea, but the Tweedie derivation is wrong and the empirical gains are partly circular. read the letter →

arxiv 2508.21468 v1 pith:GPOL5IEH submitted 2025-08-29 cs.LG cs.AI

Seungyeon Choi , Hwanhee Kim , Chihyun Park , Dahyeon Lee , Seungyong Lee , Yoonju Kim , Hyoungjoon Park , Sein Kwon
show 2 more authors
Youngwan Jo Sanghyun Park
This is my paper
classification cs.LGcs.AI
keywords structure-baseddrugdesign3DmoleculargenerationBayesianFlowNetworkgradientguidancebindingaffinitysyntheticfeasibilityselectivityconditional
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 sets out to fix a practical gap in structure-based drug design: generative models are usually judged only by docking-score binding affinity, while real drug candidates also need to be synthesizable and selective. Its central claim is that the standard way of steering generative models toward desired properties — gradient guidance in a diffusion model's sample space — is fundamentally ill-suited to 3D molecules, because atom coordinates and atom types live in different mathematical spaces and noisy intermediate molecules are chemically meaningless. The proposed alternative, CByG, extends Bayesian Flow Networks so that property guidance is injected directly into the parameter-space Bayesian updates, which the paper shows can be rewritten as score-gradient updates via Tweedie's formula for both continuous and categorical variables. With an uncertainty-aware predictor supplying the property gradient, CByG generates target-pocket-specific molecules with reported state-of-the-art binding affinity, retrosynthetic feasibility, and selectivity — all at sampling time, with no retraining. If these results hold, controllable 3D molecular generation becomes a matter of swapping in a property predictor rather than rebuilding the generative model for each new objective.

What carries the argument

The load-bearing mechanism is the gradient-form Bayesian update: the BFN receiver update θ_i = h(θ_{i-1}, y, α) is rewritten as a score-gradient step (Eqs. 7 and 10) via Tweedie's formula, so conditioning on a property l reduces to adding ∇ log p(l | x) — the gradient of an external predictor — to the parameter update. Three supporting pieces carry the empirical argument: (1) the predictor is a Bayesian neural network outputting mean and variance, so the guidance signal is scaled by predicted uncertainty (confidence-weighted steering); (2) the predictor is trained on a composite label Score = DS/(−20 × SA), which fuses docking score and synthetic accessibility into one guidance signal; and (

What would settle it

Test CBYG-generated molecules with a scoring method that was never used in training or evaluation (for example, a different docking engine or a physics-based estimator such as free-energy perturbation), and compare against baseline-generated molecules; if the large affinity and selectivity gains do not survive under that independent scorer, the reported gains are an artifact of aligning the predictor with the evaluator. A second check targets the theory: run the guided sampling in a case where the conditional distribution p(m | l) is known exactly and verify that the gradient-updated parameter

Watch

Extended reading notes

Core claim

The core discovery is that the Bayesian update of a BFN is secretly a score-based update: for continuous variables (atom coordinates), the Gaussian sender–receiver structure makes the update equivalent to a step along the score ∇ log p(x) via Tweedie's formula; for categorical variables (atom types), the same identity holds after reparameterizing one-hot vectors through a Gaussian sender, so the update becomes a softmax of previous parameters exponentiated by the score. Conditioning on a property l replaces the unconditional score with the conditional score ∇ log p(l|x), splitting each update into an unconditional generation term plus a guidance term. Because updates live in a continuous par

Load-bearing premise

The empirical case rests on a single load-bearing premise: that the Bayesian neural network predictor trained on CrossDocked2020 with the composite label Score = DS/(−20 × SA) gives trustworthy, target-specific property estimates, and that adding its gradient to the BFN update steers sampled molecules toward the intended properties — while the same families of docking and scoring tools are then used to measure the improvement.

Editorial extensions

If this is right

  • Property control moves to sampling time: adding a new objective (a different docking score, a toxicity filter, a selectivity target) means swapping or composing external predictors, not retraining the generative model.
  • Joint guidance over coordinates and atom types becomes well-defined and stable, because both modalities are updated in the same continuous parameter space rather than through discrete argmax sampling, which should yield chemically coherent molecules.
  • Multi-property optimization is expressible as a single composite label (as with Score = DS/(−20 × SA)), giving drug-discovery pipelines a direct handle on the affinity–synthesizability trade-off.
  • Pre-docking affinity scores that already match post-docking scores of strong baselines imply the generated poses are natively favorable, reducing reliance on a separate docking stage.
  • If the guidance stability claim generalizes, the same parameter-space steering should apply to other hybrid continuous/categorical generation tasks, not just molecules.

Reading between the lines

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

  • Because the guidance signal is a swap-in predictor, an obvious next step the paper leaves implicit is to use the AiZynthFinder 'Solved' rate itself as a guidance label; the paper concedes that roughly half of generated molecules remain synthetically infeasible, so directly steering on retrosynthetic success would close that gap.
  • The paper demonstrates sampling-time control for pairs of objectives (affinity + synthesizability, on-target affinity + off-target avoidance); composing all three into one guided run, and testing how independent gradients interact, is a direct extension.
  • The gradient-form Bayesian update is not molecule-specific: any hybrid continuous/categorical generation task — protein sequence–structure co-design, for instance — faces the same modality-coupling problem, so the parameter-space guidance recipe has a natural neighboring application.
  • The selectivity dataset construction (kinase panels, on/off-target pocket pairs) could serve as training signal rather than just evaluation; an off-target-aware predictor would let the framework steer explicitly to avoid side effects.
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 / 7 minor

Summary. The paper proposes CByG, a Bayesian Flow Network (BFN) extended with gradient-based property guidance for structure-based drug design. The authors claim a theoretical result: the Bayesian update for both continuous coordinates and categorical atom types can be reformulated as a score-based (Tweedie) update, enabling stable conditional generation in parameter space. They introduce a new evaluation scheme using multiple docking tools, AiZynthFinder retrosynthetic feasibility, and a kinase selectivity dataset, and report state-of-the-art results on CrossDocked2020 (Table 1). The central theoretical claim is in Section 4; the method is applied in Section 5; experiments span Sections 6 and appendices.

Significance. If the theoretical derivation and empirical claims held, this would be a significant contribution: it would provide a principled way to inject property guidance into BFN without retraining, handling continuous and discrete modalities uniformly, and it would broaden SBDD evaluation beyond single-tool docking scores. The proposed multi-docking evaluation and the retrosynthetic-feasibility benchmark are useful and address real limitations of current practice. However, the theoretical derivation is not sound: Tweedie's formula is applied to the wrong variable, and the conditional update is asserted rather than proved. The empirical evaluation also selects guidance scales on the test metrics and reports no error bars or significance tests. These issues directly undermine the paper's main claims, although the evaluation framework and the heuristic guidance idea could be valuable in a revised form.

major comments (5)
  1. [§4.2, Eq. (7)] Tweedie's formula is applied to the wrong variable. From Eq. (5), the noisy quantity is w = (ρ_i θ_i^x − ρ_{i−1}θ_{i−1}^x)/α, with x | w ~ N(w, α^{-1}I). Tweedie gives E[x|w] = w + α^{-1}∇_w log p(w), not w = x + α^{-1}∇_x log p(x) as used in Eq. (7). The score on the right is evaluated at the clean variable, while Tweedie's score is with respect to the noisy variable; the two coincide only in special limits. For a concrete counterexample, with p(x)=N(0,1), α=1, and ρ_i=ρ_{i−1}+1, Eq. (7) reduces θ_i to ρ_{i−1}θ_{i−1}/ρ_i, dropping the sender noise realization ε that is present in the reparameterized update two lines above. The appended Justification on p.29 states Tweedie for y and then substitutes ∇_x log p(x), so it does not repair the step.
  2. [§4.3, Eq. (10)] The discrete analogue repeats the continuous error and additionally mishandles the noise scale. With sender y = α(K e_x − 1) + sqrt(αK)ε, the reparameterized variable q = y/(αK) + 1/K has conditional mean e_x and variance (1/(αK))I. Tweedie would give E[e_x|q] = q + (1/(αK))∇_q log p(q). Instead, Eq. (10) writes y = α(K e_x − 1) + ∇_{e_x} log p(e_x), dropping the 1/(αK) factor and equating a scalar-scaled noise vector with a score vector. The gradient form of the discrete Bayesian update is therefore not derived.
  3. [§5.2, Eqs. (14)–(15) and Algorithm 1] Even if the Tweedie reformulation were correct, the conditional update is asserted by analogy to classifier guidance. Replacing ∇ log p(x) with ∇ log p(x|l) and then adding ∇ log p(l|x) is a heuristic; no theorem shows that iterating Eqs. (14)–(15) samples from the conditional BFN posterior p_ϕ(m|p,l) defined in Eq. (12). In addition, the uncertainty weighting in Algorithm 1 (σ^2_ϑ · λ · ∇ ...) is not derived from the stated theory and appears ad hoc. The manuscript's claim in Section 2 of 'rigorously establishes its theoretical foundations' is therefore unsupported.
  4. [§6.2, Table 4] The guidance scales λ_x=40, λ_v=40 used for the headline results in Table 1 are selected by comparing directly on the same evaluation metrics and test proteins in Table 4. No validation split or model-selection procedure is described. Tables 1–5 report no error bars, standard deviations, or significance tests. Under this protocol, the reported improvements may reflect selection on the test set rather than a genuine advantage of the method.
  5. [Appendix D.2] The property predictor is trained with label Score = DS/(-20*SA), where DS is a docking score and SA is the synthetic accessibility score. The evaluation in Section 6 measures docking scores (Vina, SMINA, GNINA) and SA, so the guidance is optimizing a fitted surrogate of the evaluator. This overlap does not by itself invalidate the experiments, but it substantially weakens the claim of general controllable generation. A held-out predictor, an independent evaluation oracle, or a prospective docking study would be needed to support the state-of-the-art claim.
minor comments (7)
  1. [General] The paper contains several typos and formatting inconsistencies, e.g., 'Yonei University' (should be Yonsei), 'CB YG' vs 'CBYG' vs 'CByG' vs 'CB Y G', and 'CbyG' in Table 2.
  2. [Definition 4.1, Eq. (6)] The statement of Tweedie's formula is imprecise: it uses E[μ_x|x] = x + Σ_x ∇_x log p(x), but the left-hand side should be a conditional expectation of the clean variable given the noisy observation, and the score should be with respect to the noisy variable. The notation also switches between x, z, and μ without clear definitions. The label 'Definition 5.1' should be 4.1.
  3. [§4.2] The sender distribution is written ambiguously as pS(y | X; αI) = N(X, α^{-1}I); it is not clear whether the variance is α^{-1}I or whether the mean/variance roles are as stated. This ambiguity propagates to Eqs. (5) and (7).
  4. [Eq. (15)] The Softmax expression is ambiguous: Softmax(e^{α(K·e_x−1)+∇_{e_x} log p(e_x|l)} · θ_{i−1}) mixes exponentiated vectors with a parameter vector in a way that is not clearly elementwise; the notation should be defined explicitly.
  5. [Appendix D.3, Algorithm 1] Line 17 uses h = σ^2_ϑ · λ_v · ∇_{e_v} log p_l(...) with e_v = GumbelSoftmax(v̂), but GumbelSoftmax is not described or justified in the main text or appendix. The relationship between e_v and the categorical sender variable y_v is unclear.
  6. [Appendix D.1, Eq. (34)] The predictive variance formula appears to have an index error: the last term should be the squared mean of the averaged means, not μ^2_{ϑ,i}(x). The current expression is dimensionally inconsistent.
  7. [Table 1] The header 'Score.' and 'Dock.' are visually ambiguous (periods are easily missed). The table would benefit from clearer column labels such as 'Pre-docking' and 'Post-docking'.

Circularity Check

3 steps flagged · score 6.0 of 10

Core theoretical 'gradient-based BFN' is produced by renaming sender noise as a score, and the headline affinity/SA results optimize a surrogate trained on the same evaluation scores; central claims partially reduce to their inputs.

  1. self definitional [Section 4.2, Eq. (7)]
    "= αx + θx i−1 · ρi−1 ρi + √αi ρi · ϵ = α ρi · x + ρi−1 ρi · θx i−1 + 1 ρi ∇x log p(x) (7)"

    The equality chain rewrites the sender-noise term √α_i/ρ_i · ε as (1/ρ_i)∇_x log p(x). But ε is an independent Gaussian noise realization in the sender pS(y|x;αI)=N(x,α^{-1}I), while ∇_x log p(x) is a deterministic function of the clean variable x. No Tweedie identity equates the two; Tweedie's formula gives E[x|y] = y + α^{-1}∇_y log p(y), not y + α^{-1}∇_x log p(x). Thus the 'gradient-based Bayesian update' is not derived: the score term is introduced by renaming the noise term. The later conditional update (Eq. 14) inherits this construction, so the claimed theoretical grounding of guidance is an ansatz presented as a derivation.

  2. self definitional [Section 4.3, Eq. (10)]
    ", where y = α (K · ex − 1) + √ αK · ϵ = α (K · ex − 1) + ∇ex log p(ex)"

    The same reduction is performed for categorical variables: the sender noise √(αK)·ε is replaced by ∇_{ex} log p(ex). The resulting softmax update is exactly the original BFN categorical Bayesian update with the noise relabeled as a 'score'. Proposition 4.2 therefore does not establish a gradient-based reformulation; it asserts, by construction, that the noise term equals the score. The discrete guidance equation (15) is built on this asserted equality, so its theoretical status is no stronger than the renaming.

1 more flagged steps
  1. fitted input called prediction [Appendix D.2 and Section 6.2, Table 1]
    "Finally, the property predictor’s scoring function is defined as Score = ( DS −20 × SA ), where DS means docking score and SA means synthetic accessibility score."

    The Bayesian property predictor that provides the guidance signal is trained to predict a score constructed from docking score (DS) and synthetic accessibility (SA). The headline evaluation (Table 1) then measures generated molecules with SMINA/GNINA/Vina docking scores and SA. Therefore the reported 'High Affinity' and SA improvements are not independent predictions: they are the result of gradient-guided optimization against a fitted surrogate of the evaluation scoring function. The claim of substantial superiority over baselines is partly built into the setup, because the baselines are not given access to this fitted surrogate. Some independent content remains (selectivity benchmark, PoseBusters, comparisons that do not use the surrogate), so the circularity is partial rather than total

full rationale

The paper does not rely on a load-bearing self-citation chain: citations to Graves et al. and other prior work are external, and no uniqueness theorem is imported from the authors' own papers. The main circularity is internal. In §4.2–4.3, the supposedly derived score-based form of the BFN update is obtained by identifying the sender noise with ∇_x log p(x) (Eqs. 7 and 10). This is a definitional renaming, not a consequence of Tweedie's formula; the appended 'Justification' boxes state Tweedie for ∇_y log p(y), but the main derivation uses ∇_x log p(x), and the two are not equal in general. Consequently, the conditional guidance in Eqs. (14)–(15) is an asserted analogy to classifier guidance rather than a proved sampling rule. On the empirical side, the property predictor is trained on Score = DS/(-20·SA), and the evaluation table reports DS and SA, so the affinity/synthesizability gains are partly forced by optimizing a fitted surrogate of the evaluator. Baselines lack this surrogate, making the comparison favorable by construction to a degree. However, the paper also evaluates on selectivity, PoseCheck, and AiZynthFinder, which are not the training targets of the surrogate, so the empirical contribution is not entirely circular. Overall: partial circularity in both the theoretical and empirical central claims, warranting a score of 6.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical or model entities are postulated; the BNN property predictor and GumbelSoftmax guidance are engineering components, not invented entities with independent evidence requirements.

free parameters (4)
  • Guidance scales lambda_x, lambda_v = lambda_x=40, lambda_v=40
    Chosen by ablation on the test benchmark (Appendix H.1, Table 4); the paper reports the best-performing combination rather than a held-out selection.
  • Composite score normalization -20 in Score = DS/(-20*SA) = -20
    Hand-chosen divisor for the predictor training label (Appendix D.2); no sensitivity analysis is provided.
  • beta-NLL hyperparameter beta = not specified
    Used in the predictor loss (Eq. 36) with no stated value or selection procedure.
  • Selectivity dataset thresholds = TM-score > 0.4, RMSD < 1 A, pocket radius 5 A
    Hand-chosen cutoffs for building on/off-target pairs (Appendix I); they affect the selectivity benchmark.
assumptions (6)
  • ad hoc to paper Tweedie's formula can be applied to the rearranged BFN update with the score taken with respect to the clean predicted sample x.
    Central to Sections 4.2-4.3 and Eqs. (5)-(10); the observed noisy variable is y, so the variables in the score are conflated. This is a nonstandard application, not a standard theorem invocation.
  • domain assumption The property predictor p(l|m,p) trained on CrossDocked-derived Score = DS/(-20*SA) is a reliable conditional density for affinity, SA, and selectivity across test proteins.
    Algorithm 1 and Section 5.2 rely on this for guidance; no independent calibration or held-out predictor evaluation is given.
  • domain assumption The pretrained MolCRAFT BFN output network remains a valid generative model under externally injected gradients.
    Section 5.2 and Algorithm 1 perturb parameter updates without retraining; no proof shows that guided trajectories stay on the learned data manifold.
  • domain assumption The conditional distribution factorizes into independent coordinate and atom-type integrals (Eq. 12).
    Eq. (12) separates x and v terms, yet the paper criticizes diffusion methods for ignoring cross-modal interactions; guidance in Eqs. (14)-(15) is applied separately to theta_x and theta_v.
  • domain assumption AlphaFold3-predicted structures of kinase pockets are accurate enough for selectivity evaluation.
    Appendix I constructs the selectivity benchmark from AlphaFold3 structures without experimental validation.
  • domain assumption Docking scores from Vina, SMINA, and GNINA are valid proxies for binding affinity and selectivity.
    Standard in the field, but both the evaluation and the guidance labels assume this; no experimental binding data are used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration." pith.science (2026). https://pith.science/paper/GPOL5IEH

@misc{pith2026250821468,
  author       = {Pith},
  title        = {Pith review of: Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GPOL5IEH}},
  note         = {Machine review of arXiv:2508.21468}
}
read the original abstract

Recent advances in Structure-based Drug Design (SBDD) have leveraged generative models for 3D molecular generation, predominantly evaluating model performance by binding affinity to target proteins. However, practical drug discovery necessitates high binding affinity along with synthetic feasibility and selectivity, critical properties that were largely neglected in previous evaluations. To address this gap, we identify fundamental limitations of conventional diffusion-based generative models in effectively guiding molecule generation toward these diverse pharmacological properties. We propose CByG, a novel framework extending Bayesian Flow Network into a gradient-based conditional generative model that robustly integrates property-specific guidance. Additionally, we introduce a comprehensive evaluation scheme incorporating practical benchmarks for binding affinity, synthetic feasibility, and selectivity, overcoming the limitations of conventional evaluation methods. Extensive experiments demonstrate that our proposed CByG framework significantly outperforms baseline models across multiple essential evaluation criteria, highlighting its effectiveness and practicality for real-world drug discovery applications.

Figures

Figures reproduced from arXiv: 2508.21468 by the authors.

Figure 1
Figure 1. Gradient trajectory for target properties [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. A comparative illustration of at￾tribute conditioning using p(l | xt) versus p(l | x0) in the image and SBDD domains. Necessity of Posterior Sampling in Guidance. In gradient-based generative frameworks such as diffusion models, conditional generation typically leverages a poste￾rior conditioned on labels (attributes) l, known as the con￾ditional score function ∇xt p(l | xt), which is learned via a dedicated neural … view at source ↗
Figure 3
Figure 3. Schematic illustration comparing the guidance propagation mechanisms of diffusion-based [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Cumulative distribu￾tion function of strain energy To achieve a more precise assessment of the realistic synthesizabil￾ity of generated molecules beyond conventional Synthetic Acces￾sibility (SA) scores (as discussed in Section 2), we introduced the AiZynthFinder bench…
Figure 4
Figure 4. Figure 4: Visualizations of reference molecules and generated ligands for protein pockets (PDB ID: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Guidance score dynamics of three model types throughout the generation pro￾cess evaluations of synthetic accessibility in future SBDD research. Notably, models with high SA scores exhibited practical retrosynthesis success rates of less than 50%, emphasizing the necess…
Figure 7
Figure 7. Figure 7: Schematic illustration of Bayesian updates for each variable type [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Visualization of generated ligands for protein pockets, with a reference molecule (left) and [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Uncertainty-Calibrated Diffusion for Reliable 3D Molecular Graph Generation

    cs.LG 2026-06 unverdicted novelty 4.0 of 10

    UCD adjusts diffusion-based 3D molecular graph generation to handle epistemic uncertainty, improving sample quality and reaching new benchmark performance.

Reference graph

Works this paper leans on

57 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Accurate structure prediction of biomolecular interactions with alphafold 3

    Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, 630(8016):493–500, 2024

    2024

  2. [2]

    Protein sequence modelling with bayesian flow networks

    Timothy Atkinson, Thomas D Barrett, Scott Cameron, Bora Guloglu, Matthew Greenig, Char- lie B Tan, Louis Robinson, Alex Graves, Liviu Copoiu, and Alexandre Laterre. Protein sequence modelling with bayesian flow networks. Nature Communications, 16(1):3197, 2025

    work page 2025

  3. [3]

    Geometric deep learning methods and applications in 3d structure-based drug design

    Qifeng Bai, Tingyang Xu, Junzhou Huang, and Horacio Pérez-Sánchez. Geometric deep learning methods and applications in 3d structure-based drug design. Drug Discovery Today, page 104024, 2024

    work page 2024

  4. [4]

    Equivariant energy-guided sde for inverse molecular design

    Fan Bao, Min Zhao, Zhongkai Hao, Peiyao Li, Chongxuan Li, and Jun Zhu. Equivariant energy-guided sde for inverse molecular design. arXiv preprint arXiv:2209.15408, 2022

    arXiv 2022

  5. [5]

    Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences

    Martin Buttenschoen, Garrett M Morris, and Charlotte M Deane. Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chemical Science, 15(9):3130–3139, 2024

    work page 2024

  6. [6]

    Pid- iff: Physics informed diffusion model for protein pocket-specific 3d molecular generation

    Seungyeon Choi, Sangmin Seo, Byung Ju Kim, Chihyun Park, and Sanghyun Park. Pid- iff: Physics informed diffusion model for protein pocket-specific 3d molecular generation. Computers in Biology and Medicine, 180:108865, 2024

    work page 2024

  7. [7]

    Diffusion posterior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh Interna- tional Conference on Learning Representations

    work page

  8. [8]

    Uniprot: a worldwide hub of protein knowledge

    UniProt Consortium. Uniprot: a worldwide hub of protein knowledge. Nucleic acids research, 47(D1):D506–D515, 2019

    work page 2019

Show all 57 references
  1. [9]

    Multi-parameter optimization: identifying high quality compounds with a balance of properties

    Matthew D Segall. Multi-parameter optimization: identifying high quality compounds with a balance of properties. Current pharmaceutical design, 18(9):1292–1310, 2012

    2012

  2. [10]

    Comprehensive analysis of kinase inhibitor selectivity

    Mindy I Davis, Jeremy P Hunt, Sanna Herrgard, Pietro Ciceri, Lisa M Wodicka, Gabriel Pallares, Michael Hocker, Daniel K Treiber, and Patrick P Zarrinkar. Comprehensive analysis of kinase inhibitor selectivity. Nature biotechnology, 29(11):1046–1051, 2011

    2011

  3. [11]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  4. [12]

    Autodock vina 1.2

    Jerome Eberhardt, Diogo Santos-Martins, Andreas F Tillack, and Stefano Forli. Autodock vina 1.2. 0: New docking methods, expanded force field, and python bindings. Journal of chemical information and modeling, 61(8):3891–3898, 2021

    2021

  5. [13]

    Tweedie’s formula and selection bias

    Bradley Efron. Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106(496):1602–1614, 2011

    2011

  6. [14]

    Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions

    Peter Ertl and Ansgar Schuffenhauer. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of cheminfor- matics, 1:1–11, 2009

    2009

  7. [15]

    Three-dimensional convolutional neural networks and a cross- docked data set for structure-based drug design

    Paul G Francoeur, Tomohide Masuda, Jocelyn Sunseri, Andrew Jia, Richard B Iovanisci, Ian Snyder, and David R Koes. Three-dimensional convolutional neural networks and a cross- docked data set for structure-based drug design. Journal of chemical information and modeling, 60(9):...

    2020

  8. [16]

    Reinforced genetic algorithm for structure-based drug design

    Tianfan Fu, Wenhao Gao, Connor Coley, and Jimeng Sun. Reinforced genetic algorithm for structure-based drug design. Advances in Neural Information Processing Systems , 35: 12325–12338, 2022. 10

    2022

  9. [17]

    Rethinking specificity in sbdd: Leveraging delta score and energy-guided diffusion

    Bowen Gao, Minsi Ren, Yuyan Ni, Yanwen Huang, Bo Qiang, Zhi-Ming Ma, Wei-Ying Ma, and Yanyan Lan. Rethinking specificity in sbdd: Leveraging delta score and energy-guided diffusion. arXiv preprint arXiv:2403.12987, 2024

    2024 arXiv

  10. [18]

    Bayesian flow networks

    Alex Graves, Rupesh Kumar Srivastava, Timothy Atkinson, and Faustino Gomez. Bayesian flow networks. arXiv preprint arXiv:2308.07037, 2023

    2023 arXiv

  11. [19]

    Aligning target-aware molecule diffusion models with exact energy optimization

    Siyi Gu, Minkai Xu, Alexander Powers, Weili Nie, Tomas Geffner, Karsten Kreis, Jure Leskovec, Arash Vahdat, and Stefano Ermon. Aligning target-aware molecule diffusion models with exact energy optimization. Advances in Neural Information Processing Systems, 37:44040–44063, 2024

    2024

  12. [20]

    3d equivariant diffusion for target-aware molecule generation and affinity prediction.arXiv preprint arXiv:2303.03543, 2023

    Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equivariant diffusion for target-aware molecule generation and affinity prediction.arXiv preprint arXiv:2303.03543, 2023

    2023 arXiv

  13. [21]

    Decompdiff: diffusion models with decomposed priors for structure- based drug design

    Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, and Quanquan Gu. Decompdiff: diffusion models with decomposed priors for structure- based drug design. arXiv preprint arXiv:2403.07902, 2024

    2024 arXiv

  14. [22]

    Gradient guidance for diffusion models: An optimization perspective

    Yingqing Guo, Hui Yuan, Yukang Yang, Minshuo Chen, and Mengdi Wang. Gradient guidance for diffusion models: An optimization perspective. arXiv preprint arXiv:2404.14743, 2024

    2024 arXiv

  15. [23]

    Training- free multi-objective diffusion model for 3d molecule generation

    Xu Han, Caihua Shan, Yifei Shen, Can Xu, Han Yang, Xiang Li, and Dongsheng Li. Training- free multi-objective diffusion model for 3d molecule generation. In The Twelfth International Conference on Learning Representations, 2023

    2023

  16. [24]

    Posecheck: Generative models for 3d structure-based drug design produce unrealistic poses

    Charles Harris, Kieran Didi, Arian Jamasb, Chaitanya Joshi, Simon Mathis, Pietro Lio, and Tom Blundell. Posecheck: Generative models for 3d structure-based drug design produce unrealistic poses. In NeurIPS 2023 Generative AI and Biology (GenBio) Workshop

    2023

  17. [25]

    Protein-ligand interaction prior for binding- aware 3d molecule diffusion models

    Zhilin Huang, Ling Yang, Xiangxin Zhou, Zhilong Zhang, Wentao Zhang, Xiawu Zheng, Jie Chen, Yu Wang, CUI Bin, and Wenming Yang. Protein-ligand interaction prior for binding- aware 3d molecule diffusion models. In The Twelfth International Conference on Learning Representations, 2023

    2023

  18. [26]

    Rational approaches to improving selectivity in drug design

    David J Huggins, Woody Sherman, and Bruce Tidor. Rational approaches to improving selectivity in drug design. Journal of medicinal chemistry, 55(4):1424–1444, 2012

    2012

  19. [27]

    A quantitative analysis of kinase inhibitor selectivity

    Mazen W Karaman, Sanna Herrgard, Daniel K Treiber, Paul Gallant, Corey E Atteridge, Brian T Campbell, Katrina W Chan, Pietro Ciceri, Mindy I Davis, Philip T Edeen, et al. A quantitative analysis of kinase inhibitor selectivity. Nature biotechnology, 26(1):127–132, 2008

    2008

  20. [28]

    Noise2score: tweedie’s approach to self-supervised image denoising without clean images

    Kwanyoung Kim and Jong Chul Ye. Noise2score: tweedie’s approach to self-supervised image denoising without clean images. Advances in Neural Information Processing Systems , 34: 864–874, 2021

    2021

  21. [29]

    Recent developments in structure-based drug design

    Gerhard Klebe. Recent developments in structure-based drug design. Journal of molecular medicine, 78:269–281, 2000

    2000

  22. [30]

    Lessons learned in empirical scoring with smina from the csar 2011 benchmarking exercise

    David Ryan Koes, Matthew P Baumgartner, and Carlos J Camacho. Lessons learned in empirical scoring with smina from the csar 2011 benchmarking exercise. Journal of chemical information and modeling, 53(8):1893–1904, 2013

    2011

  23. [31]

    A 3d generative model for structure-based drug design

    Shitong Luo, Jiaqi Guan, Jianzhu Ma, and Jian Peng. A 3d generative model for structure-based drug design. Advances in Neural Information Processing Systems, 34:6229–6239, 2021

    2021

  24. [32]

    Gnina 1.0: molecular docking with deep learning

    Andrew T McNutt, Paul Francoeur, Rishal Aggarwal, Tomohide Masuda, Rocco Meli, Matthew Ragoza, Jocelyn Sunseri, and David Ryan Koes. Gnina 1.0: molecular docking with deep learning. Journal of cheminformatics, 13(1):43, 2021

    2021

  25. [33]

    3d molecule generation by denoising voxel grids

    Pedro O O Pinheiro, Joshua Rackers, Joseph Kleinhenz, Michael Maser, Omar Mahmood, Andrew Watkins, Stephen Ra, Vishnu Sresht, and Saeed Saremi. 3d molecule generation by denoising voxel grids. Advances in Neural Information Processing Systems, 36:69077–69097, 2023. 11

    2023

  26. [34]

    Pocket2mol: Efficient molecular sampling based on 3d protein pockets

    Xingang Peng, Shitong Luo, Jiaqi Guan, Qi Xie, Jian Peng, and Jianzhu Ma. Pocket2mol: Efficient molecular sampling based on 3d protein pockets. In International Conference on Machine Learning, pages 17644–17655. PMLR, 2022

    2022

  27. [35]

    Moldiff: Addressing the atom-bond inconsistency problem in 3d molecule diffusion generation

    Xingang Peng, Jiaqi Guan, Qiang Liu, and Jianzhu Ma. Moldiff: Addressing the atom-bond inconsistency problem in 3d molecule diffusion generation. arXiv preprint arXiv:2305.07508, 2023

    2023 arXiv

  28. [36]

    Geometric deep learning for structure-based ligand design

    Alexander Powers, Helen Yu, Patricia Suriana, Rohan Koodli, Tianyu Lu, Joseph Paggi, and Ron Dror. Geometric deep learning for structure-based ligand design. ACS Central Science, 2023

    2023

  29. [37]

    Molcraft: Structure-based drug design in continuous parameter space

    Yanru Qu, Keyue Qiu, Yuxuan Song, Jingjing Gong, Jiawei Han, Mingyue Zheng, Hao Zhou, and Wei-Ying Ma. Molcraft: Structure-based drug design in continuous parameter space. In International Conference on Machine Learning, pages 41749–41768. PMLR, 2024

    2024

  30. [38]

    Generating 3d molecules con- ditional on receptor binding sites with deep generative models

    Matthew Ragoza, Tomohide Masuda, and David Ryan Koes. Generating 3d molecules con- ditional on receptor binding sites with deep generative models. Chemical science, 13(9): 2701–2713, 2022

    2022

  31. [39]

    Structure-based drug design with equivariant diffusion models, 2023

    Arne Schneuing, Yuanqi Du, Charles Harris, Arian Jamasb, Ilia Igashov, Weitao Du, Tom Blundell, Pietro Lió, Carla Gomes, Max Welling, Michael Bronstein, and Bruno Correia. Structure-based drug design with equivariant diffusion models, 2023

    2023

  32. [40]

    Structure-based drug design with equivariant diffusion models

    Arne Schneuing, Charles Harris, Yuanqi Du, Kieran Didi, Arian Jamasb, Ilia Igashov, Weitao Du, Carla Gomes, Tom L Blundell, Pietro Lio, et al. Structure-based drug design with equivariant diffusion models. Nature Computational Science, 4(12):899–909, 2024

    2024

  33. [41]

    On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks

    Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. arXiv preprint arXiv:2203.09168, 2022

    2022 arXiv

  34. [42]

    Tacogfn: Target-conditioned gflownet for structure-based drug design

    Tony Shen, Seonghwan Seo, Grayson Lee, Mohit Pandey, Jason R Smith, Artem Cherkasov, Woo Youn Kim, and Martin Ester. Tacogfn: Target-conditioned gflownet for structure-based drug design. Transactions on Machine Learning Research

  35. [43]

    Tacogfn: Target-conditioned gflownet for structure-based drug design

    Tony Shen, Seonghwan Seo, Grayson Lee, Mohit Pandey, Jason R Smith, Artem Cherkasov, Woo Youn Kim, and Martin Ester. Tacogfn: Target-conditioned gflownet for structure-based drug design. arXiv preprint arXiv:2310.03223, 2023

    2023 arXiv

  36. [44]

    From target to drug: generative modeling for the multimodal structure-based ligand design.Molecular pharmaceutics, 16(10):4282–4291, 2019

    Miha Skalic, Davide Sabbadin, Boris Sattarov, Simone Sciabola, and Gianni De Fabritiis. From target to drug: generative modeling for the multimodal structure-based ligand design.Molecular pharmaceutics, 16(10):4282–4291, 2019

    2019

  37. [45]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  38. [46]

    Unified generative modeling of 3d molecules with bayesian flow networks

    Yuxuan Song, Jingjing Gong, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules with bayesian flow networks. In The Twelfth International Conference on Learning Representations, 2023

    2023

  39. [47]

    Selective optimization of side activities: another way for drug discovery

    Camille G Wermuth. Selective optimization of side activities: another way for drug discovery. Journal of medicinal chemistry, 47(6):1303–1314, 2004

    2004

  40. [48]

    Learning subpocket prototypes for generalizable structure-based drug design

    Zaixi Zhang and Qi Liu. Learning subpocket prototypes for generalizable structure-based drug design. In International Conference on Machine Learning, pages 41382–41398. PMLR, 2023

    2023

  41. [49]

    Molecule generation for target pro- tein binding with structural motifs

    Zaixi Zhang, Yaosen Min, Shuxin Zheng, and Qi Liu. Molecule generation for target pro- tein binding with structural motifs. In The Eleventh International Conference on Learning Representations, 2023

    2023

  42. [50]

    A systematic survey in geometric deep learning for structure-based drug design

    Zaixi Zhang, Jiaxian Yan, Qi Liu, Enhong Chen, and Marinka Zitnik. A systematic survey in geometric deep learning for structure-based drug design. arXiv preprint arXiv:2306.11768, 2023. 12

    2023 arXiv

  43. [51]

    known unknown

    Xiangxin Zhou, Xiwei Cheng, Yuwei Yang, Yu Bao, Liang Wang, and Quanquan Gu. Decom- popt: Controllable and decomposed diffusion models for structure-based molecular optimization. In The Twelfth International Conference on Learning Representations. 13 A Detail description of Ba...

  44. [52]

    Sample from output distribution: An intermediate sample m′ i is drawn from the output distribution m′ i ∼ pO(· |θi−1, ti) given the current parameter θi−1 and scheduled noise level ti

  45. [53]

    Sample from sender distribution: A noisy observation yi is then drawn from the sender distribution, conditional on m′ i, via y′ i ∼ pS(· |m′ i, αi), where αi is the accuracy (inverse noise variance) prescribed for step i

  46. [54]

    Here h(θi−1, yi, αi) computes the posterior parameter after observing yi with precision αi, given the prior θi−1

    Bayesian parameter update: The distribution parameter is updated by incorporating the observation yi through the Bayesian update function: θi = h(θi−1, yi, αi). Here h(θi−1, yi, αi) computes the posterior parameter after observing yi with precision αi, given the prior θi−1. Re...

  47. [55]

    In contrast, diffusion models explicitly add random noise to samples at each step to maintain stochasticity

    Implicit versus Explicit Noise Injection: BFN deterministically update parameters given noisy observations, implicitly capturing uncertainty. In contrast, diffusion models explicitly add random noise to samples at each step to maintain stochasticity

  48. [56]

    Parameter-space versus Sample-space Updates: Diffusion models perform both training and inference entirely within the sample space. Conversely, BFN operate fundamentally within the parameter space, integrating information from the sample space through occa- sional noisy observ...

  49. [57]

    Only Generation Type

    and DrugGPS [48], integrated chemically meaningful fragments, thereby improving the structural realism of generated ligands. In parallel, diffusion-based generative methods have emerged, achieving remarkable success across various generative tasks such as image and text synthe...

Pith tools

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