Pith. sign in

REVIEW 5 major objections 6 minor 34 references

MODA: A Unified 3D Diffusion Framework for Multi-Task Target-Aware Molecular Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MODA claims that one mask-and-denoise diffusion pass can replace separate models for four molecular-editing tasks.

desk verdict MODA's multi-task masked diffusion is a plausible extension of existing unified-generation ideas, but the zero-shot de novo claim and the 'Bayesian' scheduler do not survive close reading. read the letter →

arxiv 2507.07201 v1 pith:2OAPFQAS submitted 2025-07-09 q-bio.BM cs.AIcs.LG

Dong Xu , Zhangfan Yang , Sisi Yuan , Jenna Xinyi Yao , Jiangqiang Li , Junkai Ji This is my paper
classification q-bio.BMcs.AIcs.LG
keywords 3Dmoleculargenerationdiffusionmodelsmulti-tasklearningBayesianmaskschedulingfragmentgrowinglinkerdesignscaffoldhoppingside-chaindecoration
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

MODA proposes that structure-based molecular design can be reduced to one masked-denosing task. Instead of training separate models for fragment growing, linker design, scaffold hopping, and side-chain decoration, it masks a contiguous spatial fragment of a molecule and reconstructs it in a single diffusion pass conditioned on the protein pocket. The claim is that a single multi-task backbone learns shared geometric and chemical priors, matches or beats task-specific models, and transfers zero-shot to de novo design and lead optimisation without pretrain-finetune or force-field refinement. If true, one trained model could serve many editing tasks and cut data and engineering costs.

What carries the argument

The central object is the Bayesian mask scheduler and the family of deterministic masking operators $F_t$. For each task, $F_t$ flags which atoms to remove and regenerate: up to two atoms on the shortest path between anchors for linker design, an exit atom and its two-bond neighbourhood for fragment growing, non-ring terminal groups for side-chain decoration, and the Bemis-Murcko core for scaffold hopping. The mechanism couples these masks with a joint continuous-discrete denoising diffusion process over coordinates and atom types, with the protein pocket perturbed by small Gaussian noise as regularisation. The load-bearing identity is $p_{\theta}(m \mid t) = p_{\theta}(t \mid m)\,p_{\theta}(m)/p_{\theta}(t)$: the paper claims that training on mask-conditioned reconstruction alone, without task labels, induces task-conditional priors through Bayes' rule and gives the model self-routing capability at inference.

What would settle it

Observe whether the model's output distribution changes for a fixed molecule and fixed mask when the underlying intended task in the training data differs; if the output is unchanged, the Bayesian task-conditional prior is not exerting any identifiable influence.

Watch

Extended reading notes

Core claim

The central claim is that one-stage multi-task training can replace the two-stage pretrain-finetune pipeline for target-aware 3D molecular generation. On a training set built from a cross-docked protein-ligand dataset, each molecule is split into context atoms and target atoms by a deterministic task mask, and the model denoises the noisy target in one pass. The paper argues that because the same molecule appears with different masks and task labels, accurate reconstruction forces the model to learn $p_{\theta}(t \mid M)$ internally, and by Bayes' rule induces a universal task-conditional prior $p_{\theta}(m \mid t)$ even though $t$ never appears in the loss. The reported multi-task model achieves the lowest Vina docking energies, higher improvement rates, and better substructure and geometric fidelity than six diffusion baselines and the single-task variant, while the multi-task-trained single-task evaluation shows that multi-task exposure alone transfers zero-shot to de novo tasks. All gains are reported without force-field refinement.

Load-bearing premise

The load-bearing premise is that a deterministic mask pattern alone can tell the model which task is being asked, even though task labels never appear in the loss and the same molecule can appear with the same mask under different tasks.

Editorial extensions

If this is right

  • The same backbone is claimed to generate molecules for all four editing tasks plus de novo design without task switches or auxiliary classifiers.
  • The pretrain-finetune gap disappears: the fully multi-task model is reported to match the pretrained variant on most metrics, so two-stage pipelines would no longer be necessary for these tasks.
  • Zero-shot transfer follows from the shared priors: multi-task variants keep negative Vina scores and high improvement rates on unseen de novo and lead-optimisation benchmarks without force-field refinement.
  • Multi-task training improves geometric fidelity, reported as lower bond-length divergence and reduced ligand-protein clashes relative to single-task training.
  • Ranking robustness improves: drug-likeness and synthesizability stay high for the multi-task model as the candidate list grows, while single-task models degrade beyond the top few candidates.

Reading between the lines

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

  • If the self-routing claim holds, the same framework could absorb additional editing operations, such as R-group enumeration, scaffold extension, or bioisosteric replacement, simply by adding new deterministic mask operators.
  • A direct test of the Bayesian assumption would use a fixed molecule and fixed mask that are compatible with two tasks and check whether the model's output distribution changes when the task semantics differ; if the outputs are identical, the claimed prior is not identifiable.
  • Because the paper states that the shared backbone can be distilled into compact student models, a practical extension would be to measure how much of the cross-task transfer survives distillation, which the paper does not demonstrate.
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 MODA, a 3D diffusion framework for target-aware molecular generation that aims to unify four structure-based editing tasks—fragment growing, linker design, scaffold hopping, and side-chain decoration—through a single masked-reconstruction objective. The authors compare three training paradigms (single-task Model-A, multi-task-then-single-task Model-B, full multi-task Model-C) against six diffusion baselines on docking, chemical property, substructure, and geometry metrics. They also report zero-shot de novo design and lead-optimization results. The core methodological claim is that a 'Bayesian mask scheduler' enables the model to learn shared priors across tasks and transfer zero-shot, so that one trained backbone can replace task-specific pipelines.

Significance. If the claims were fully supported, MODA would be a notable step toward a unified 3D molecular generator, and the empirical breadth—three training paradigms, four editing tasks, six baselines, and multiple evaluation metrics—is commendable. The paper also correctly identifies a real limitation in current structure-based design pipelines, namely the one-task-one-model mindset. However, the two load-bearing methodological contributions are not established: the Bayesian mask scheduler is not actually implemented (task labels never appear in the loss), and the zero-shot de novo results are not reproducible because the paper does not specify how the number of generated atoms is set or whether all-atom masks ever appear in training. These gaps undermine the central 'single-stage multi-task routine can replace two-stage workflows' claim as stated.

major comments (5)
  1. [Section III-B, Eqs. (7)-(8)] The derivation of the 'Bayesian mask scheduler' is not implemented. The training loss in Eq. (7) conditions on the mask M(i) only, and the task label t(i) never appears in the conditioning or in any auxiliary loss; thus pθ(t|m) in Eq. (8) is not a quantity the model is trained to represent. Since the masking operator F_t is deterministic but the same molecule can appear with different masks and task labels in the dataset, the input mask does not uniquely determine the task, and no mechanism is given for the network to infer the intended task from the mask pattern. The claimed 'self-routing capability' in Section III-B(b) is therefore unsupported, and the 'Bayesian' framing in the abstract and contribution (O3) is misleading. I recommend removing the Bayesian terminology or adding an explicit task-prediction head and a proper posterior over tasks.
  2. [Table VII and Sections III-B/III-E] The task space is inconsistent. Section III-B defines T = {LINKER, FRAGMENT, SIDE_CHAIN, SCAFFOLD}, while Table VII is titled 'MASK CONSTRUCTION RULES FOR THE FIVE TASKS' and includes DENOVO with M_i = 1. Section III-E states that each molecule is assigned to one or more of the four editing tasks, never to DENOVO. If all-atom masks are never seen during training, the zero-shot de novo results in Tables III and V are not reproducible from the described method, because the model has not learned to denoise from an all-ones mask. The paper must clarify whether DENOVO masks appear in training or provide a principled way to condition on an unseen mask.
  3. [Section III-D and Table III] The inference procedure for de novo generation lacks a rule for the number of output atoms. In Eq. (15), initial positions are sampled around the pocket center of mass, but the number of atoms N_M is not specified. Without a mechanism to set N_M, the stable negative Vina scores and high improvement rates reported for Model-C in Table III cannot be reproduced. Please specify how N_M is selected (e.g., from the reference-ligand size distribution, a learned prior, or a separate length predictor) and report the resulting molecular-size distribution.
  4. [Tables I-IV and Section II] The empirical claims are presented without error bars or statistical significance tests, and no code is provided. Several conclusions rest on small differences, such as Table IV Linker rows where Model-C (-8.53) and DiffBP (-8.74) are close on Vina Dock, or Table I Scaffold rows where Model-A-FG exceeds Model-C on Evina and IMP. Without multiple seeds and standard deviations, the statement that multi-task training 'surpasses' the baselines is not statistically supported. I recommend reporting variance across at least three seeds and releasing code and data to enable verification.
  5. [Section II.B and Table I] The abstract and Section II.B state that Model-B 'trails in novelty,' but no novelty metric is defined anywhere in the paper or tables. Table I reports a 'Unique' column (likely chemical validity/uniqueness) but not novelty relative to the training set or reference molecules. The comparison between Model-B and Model-C on novelty is therefore unverifiable. Please define the novelty metric or remove this claim from the abstract and results.
minor comments (6)
  1. [Section III-B] There is a typo in the training-objective paragraph: 'every ask–specific mask' should read 'every task-specific mask'.
  2. [Abstract and Table VII] The abstract says 'four editing scenarios' while Table VII lists five tasks including DENOVO. This count should be reconciled, and the role of DENOVO should be stated consistently throughout.
  3. [Section III-D] The initialization variance σ^2 in Eqs. (14) and (15) is never specified. Please give the value used in experiments, and also state the noise schedule parameters for Eq. (10)-(11).
  4. [Table III and Section II] The abbreviation 'PT' in Model-C-PT is not defined when first used in Table III. If it means pretrained, please define it in Section II.A or the table caption.
  5. [Section II.B] 'UniSeq correctness' is presumably 'unique correctness'; if a different metric is intended, please define it explicitly.
  6. [Section II and Related Work] The paper claims data and engineering cost savings from a unified model, but no runtime or parameter-count comparison is provided. A simple table of training/inference time and model size would strengthen the practical claim that a single-stage routine can replace two-stage pipelines.

Circularity Check

2 steps flagged · score 6.0 of 10

Central 'Bayesian task-conditional prior' step is a tautology, and the zero-shot de novo claim is either trained-in or under-specified.

  1. self definitional [Section III-B, Training objective and Inference, Eqs. (7)-(8)]
    "Because different tasks impose mutually incompatible constraints, accurate reconstruction forces the model to learn a shared latent mapping pθ(t | M) internally. Via Bayes’ rule this induces a universal task-conditional prior pθ(m | t) = pθ(t | m) pθ(m) pθ(t), even though t never appears in the loss."

    The training loss in Eq. (7) is L(θ)=Σ log pθ(m(i) | M(i)); the task label t, the posterior pθ(t|M), the marginal pθ(t), and pθ(m) are never parameterized, trained, or inferred. Bayes' rule is an algebraic identity for any joint distribution, so writing pθ(m|t) as pθ(t|m)pθ(m)/pθ(t) is vacuous unless the network actually computes pθ(t|M). The paper asserts that 'accurate reconstruction forces' this posterior to emerge, but the objective contains no task term and the mask M is a deterministic function of m and t, so nothing forces a unique posterior over t. The 'universal task-conditional prior' and the resulting 'self-routing' are therefore the conclusion being assumed, not consequences of the stated objective.

  2. other [Section III-B task space vs. Table VII and Section III-E dataset construction; abstract zero-shot claim]
    "Let T = {LINKER , FRAGMENT , SIDE CHAIN , SCAFFOLD } denote the unified task space (see Table VII for detailed definitions)... TABLE VII MASK CONSTRUCTION RULES FOR THE FIVE TASKS . Task t Rule M = Ft(m) DENOVO Mi ≡ 1 (all atoms)."

    Table VII presents DENOVO with the all-atoms mask rule and calls it one of five tasks, while Section III-B defines T with only four tasks and Section III-E assigns molecules to FRAGMENT, LINKER, SCAFFOLD, or SIDE_CHAIN, never specifying how DENOVO masks enter training. If all-ones masks were included in the dynamic task sampling described in Section III-E, then the 'zero-shot' de novo evaluation is not zero-shot: the all-ones mask format was a training input by construction. If all-ones masks were excluded, the inference procedure supplies no mechanism for choosing the number of output atoms N_M or an empty context; Eq. (15) only initializes positions around CoM(P).

full rationale

The paper's empirical comparisons against external diffusion baselines are not themselves circular and may be valid as reported. The circularity is concentrated in the claimed theoretical mechanism: the Bayesian mask scheduler is justified by Eq. (8), but Eq. (8) is a Bayes-rule tautology over quantities the model never learns, so the key claim that the network acquires a task-conditional prior and self-routing capability reduces to an assertion rather than a derivation. The zero-shot de novo claim is also internally inconsistent: DENOVO appears as a mask-construction task in Table VII but is absent from the formal task space and dataset assignment, so the reported 'zero-shot' gains are either an in-distribution training result mislabeled as zero-shot or an undefined generation setting. No load-bearing self-citation chain was found, and the comparison tables stand as independent empirical material, which keeps the score at 6 rather than higher.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on hand-coded mask rules, an unspecified noise schedule and initialization variance, and a formal Bayes identity that is not implemented as inference over tasks. No new physical or chemical entities are introduced.

free parameters (5)
  • Diffusion noise schedule (alpha_t) = not specified
    The beta schedule for the forward process in Eqs. (10)-(11) is not reported; required to reproduce the denoising steps.
  • Protein perturbation noise standard deviation = 0.12
    Chosen in Eq. (12) as a regularizer; no sensitivity analysis is given.
  • Initialization variance sigma^2 = not specified
    Used in Eqs. (14)-(15) to initialize ligand atom positions; value not reported.
  • Mask size thresholds = LINKER <=2 atoms; FRAGMENT 2-bond neighborhood; SIDE CHAIN terminal groups; SCAFFOLD Bemis-Murcko core
    Hand-coded rules in Table VII define the task distribution; different thresholds would change results.
  • Task sampling distribution = not specified
    Section III-E says a task is sampled each iteration but the distribution is not stated; assumed to be uniform.
assumptions (4)
  • ad hoc to paper Bayes' rule applies to p_theta(m|t) even though t never appears in the loss (Eqs. (7)-(8)).
    Section III-B claims p_theta(t|m) is learned internally from masks, but no posterior over tasks is defined or trained; this is the load-bearing justification for the 'Bayesian' scheduler.
  • domain assumption CrossDocked2020 covers the chemical space needed for all four tasks.
    Training and evaluation rely on CrossDocked2020 (Section III-E); no evidence is provided that it is sufficient for side-chain and scaffold tasks.
  • domain assumption Vina docking scores and the custom metrics MPBG, LBE, Pass/Good/Excellent are valid proxies for binding and drug-likeness.
    All conclusions about potency rest on these metrics; no experimental validation is provided.
  • domain assumption The SE(3)-equivariant backbone can represent the joint distribution over atom types and coordinates.
    The architecture is referenced but not specified; universal approximation capacity and training stability are assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MODA: A Unified 3D Diffusion Framework for Multi-Task Target-Aware Molecular Generation." pith.science (2026). https://pith.science/paper/2OAPFQAS

@misc{pith2026250707201,
  author       = {Pith},
  title        = {Pith review of: MODA: A Unified 3D Diffusion Framework for Multi-Task Target-Aware Molecular Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2OAPFQAS}},
  note         = {Machine review of arXiv:2507.07201}
}
read the original abstract

Three-dimensional molecular generators based on diffusion models can now reach near-crystallographic accuracy, yet they remain fragmented across tasks. SMILES-only inputs, two-stage pretrain-finetune pipelines, and one-task-one-model practices hinder stereochemical fidelity, task alignment, and zero-shot transfer. We introduce MODA, a diffusion framework that unifies fragment growing, linker design, scaffold hopping, and side-chain decoration with a Bayesian mask scheduler. During training, a contiguous spatial fragment is masked and then denoised in one pass, enabling the model to learn shared geometric and chemical priors across tasks. Multi-task training yields a universal backbone that surpasses six diffusion baselines and three training paradigms on substructure, chemical property, interaction, and geometry. Model-C reduces ligand-protein clashes and substructure divergences while maintaining Lipinski compliance, whereas Model-B preserves similarity but trails in novelty and binding affinity. Zero-shot de novo design and lead-optimisation tests confirm stable negative Vina scores and high improvement rates without force-field refinement. These results demonstrate that a single-stage multi-task diffusion routine can replace two-stage workflows for structure-based molecular design.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 19 canonical work pages

  1. [1]

    Highly accurate protein structure prediction with alphafold,

    J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. ˇZ´ıdek, A. Potapenko et al., “Highly accurate protein structure prediction with alphafold,” nature, vol. 596, no. 7873, pp. 583–589, 2021

    work page 2021

  2. [2]

    Accurate structure prediction of biomolecular interactions with alphafold 3,

    J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, O. Ronneberger, L. Willmore, A. J. Ballard, J. Bambricket al., “Accurate structure prediction of biomolecular interactions with alphafold 3,” Nature, vol. 630, no. 8016, pp. 493–500, 2024

    work page 2024

  3. [3]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , 2019, pp. 4171–4186

    2019

  4. [4]

    Improving language understanding by generative pre-training,

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever et al., “Improving language understanding by generative pre-training,” 2018

    2018

  5. [5]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  6. [6]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019

    2019

  7. [7]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 000–16 009

    2022

  8. [8]

    Generation of 3d molecules in pockets via a language model,

    W. Feng, L. Wang, Z. Lin, Y . Zhu, H. Wang, J. Dong, R. Bai, H. Wang, J. Zhou, W. Peng et al., “Generation of 3d molecules in pockets via a language model,” Nature Machine Intelligence, vol. 6, no. 1, pp. 62–73, 2024

    work page 2024

Show all 34 references
  1. [9]

    Genmol: A drug discovery generalist with discrete diffusion,

    S. Lee, K. Kreis, S. P. Veccham, M. Liu, D. Reidenbach, Y . Peng, S. Paliwal, W. Nie, and A. Vahdat, “Genmol: A drug discovery generalist with discrete diffusion,” arXiv preprint arXiv:2501.06158 , 2025

    2025 arXiv

  2. [10]

    Molgpt: molecular generation using a transformer-decoder model,

    V . Bagal, R. Aggarwal, P. Vinod, and U. D. Priyakumar, “Molgpt: molecular generation using a transformer-decoder model,” Journal of chemical information and modeling, vol. 62, no. 9, pp. 2064–2076, 2021

    2021

  3. [11]

    Gotta be safe: a new framework for molecular design,

    E. Noutahi, C. Gabellini, M. Craig, J. S. Lim, and P. Tossou, “Gotta be safe: a new framework for molecular design,” Digital Discovery, vol. 3, no. 4, pp. 796–804, 2024

    2024

  4. [12]

    Functional-group-based diffusion for pocket-specific molecule generation and elaboration,

    H. Lin, Y . Huang, O. Zhang, Y . Liu, L. Wu, S. Li, Z. Chen, and S. Z. Li, “Functional-group-based diffusion for pocket-specific molecule generation and elaboration,” Advances in Neural Information Processing Systems, vol. 36, pp. 34 603–34 626, 2023

    2023

  5. [13]

    Fraggen: towards 3d geometry reli- able fragment-based molecular generation,

    O. Zhang, Y . Huang, S. Cheng, M. Yu, X. Zhang, H. Lin, Y . Zeng, M. Wang, Z. Wu, H. Zhao et al., “Fraggen: towards 3d geometry reli- able fragment-based molecular generation,” Chemical Science, vol. 15, no. 46, pp. 19 452–19 465, 2024

    2024

  6. [14]

    Equivariant 3d-conditional diffusion model for molecular linker design,

    I. Igashov, H. St ¨ark, C. Vignac, A. Schneuing, V . G. Satorras, P. Frossard, M. Welling, M. Bronstein, and B. Correia, “Equivariant 3d-conditional diffusion model for molecular linker design,” Nature Machine Intelli- gence, vol. 6, no. 4, pp. 417–427, 2024

    2024

  7. [15]

    3d equivariant diffusion for target-aware molecule generation and affinity prediction,

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

    2023 arXiv

  8. [16]

    A 3d generative model for structure- based drug design,

    S. Luo, J. Guan, J. Ma, and J. Peng, “A 3d generative model for structure- based drug design,” Advances in Neural Information Processing Systems, vol. 34, pp. 6229–6239, 2021

    2021

  9. [17]

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

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

    2024 arXiv

  10. [18]

    Molecule generation for target protein binding with structural motifs,

    Z. Zhang, Y . Min, S. Zheng, and Q. Liu, “Molecule generation for target protein binding with structural motifs,” in The eleventh international conference on learning representations , 2023

    2023

  11. [19]

    Cbgbench: fill in the blank of protein-molecule complex binding graph,

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

    2024 arXiv

  12. [20]

    Structure-based drug design with equivariant diffusion models,

    A. Schneuing, C. Harris, Y . Du, K. Didi, A. Jamasb, I. Igashov, W. Du, C. Gomes, T. L. Blundell, P. Lioet al., “Structure-based drug design with equivariant diffusion models,” Nature Computational Science , vol. 4, no. 12, pp. 899–909, 2024

    2024

  13. [21]

    Diffbp: Generative diffusion of 3d molecules for target protein binding,

    H. Lin, Y . Huang, O. Zhang, S. Ma, M. Liu, X. Li, L. Wu, J. Wang, T. Hou, and S. Z. Li, “Diffbp: Generative diffusion of 3d molecules for target protein binding,” Chemical Science, vol. 16, no. 3, pp. 1417–1431, 2025

    2025

  14. [22]

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

    P. G. Francoeur, T. Masuda, J. Sunseri, A. Jia, R. B. Iovanisci, I. Snyder, and D. R. Koes, “Three-dimensional convolutional neural networks and a cross-docked data set for structure-based drug design,” Journal of chemical information and modeling, vol. 60, no. 9, pp. 4200–4215, 2020

    2020

  15. [23]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” Advances in neural infor- mation processing systems , vol. 35, pp. 26 565–26 577, 2022

    2022

  16. [24]

    Geodiff: A geometric diffusion model for molecular conformation generation,

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

    2022 arXiv

  17. [25]

    Deep generative models for 3d linker design,

    F. Imrie, A. R. Bradley, M. van der Schaar, and C. M. Deane, “Deep generative models for 3d linker design,” Journal of chemical information and modeling, vol. 60, no. 4, pp. 1983–1995, 2020

    1983

  18. [26]

    Molgrow: A graph normalizing flow for hierarchical molecular generation,

    M. Kuznetsov and D. Polykovskiy, “Molgrow: A graph normalizing flow for hierarchical molecular generation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 35, no. 9, 2021, pp. 8226– 8234

    2021

  19. [27]

    Scaffold-based molecular design with a graph generative model,

    J. Lim, S.-Y . Hwang, S. Moon, S. Kim, and W. Y . Kim, “Scaffold-based molecular design with a graph generative model,” Chemical science , vol. 11, no. 4, pp. 1153–1164, 2020

    2020

  20. [28]

    Accelerated rational protac design via deep learning and molecular simulations,

    S. Zheng, Y . Tan, Z. Wang, C. Li, Z. Zhang, X. Sang, H. Chen, and Y . Yang, “Accelerated rational protac design via deep learning and molecular simulations,” Nature Machine Intelligence , vol. 4, no. 9, pp. 739–748, 2022

    2022

  21. [29]

    Chemberta: large-scale self-supervised pretraining for molecular property prediction,

    S. Chithrananda, G. Grand, and B. Ramsundar, “Chemberta: large-scale self-supervised pretraining for molecular property prediction,” arXiv preprint arXiv:2010.09885, 2020

    2010 arXiv

  22. [30]

    Mol-bert: An effective molecular representation with bert for molecular property prediction,

    J. Li and X. Jiang, “Mol-bert: An effective molecular representation with bert for molecular property prediction,” Wireless Communications and Mobile Computing, vol. 2021, no. 1, p. 7181815, 2021

    2021

  23. [31]

    Translation between molecules and natural language,

    C. Edwards, T. Lai, K. Ros, G. Honke, K. Cho, and H. Ji, “Translation between molecules and natural language,” arXiv preprint arXiv:2204.11817, 2022

    2022 arXiv

  24. [32]

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

    G. Zhou, Z. Gao, Q. Ding, H. Zheng, H. Xu, Z. Wei, L. Zhang, and G. Ke, “Uni-mol: A universal 3d molecular representation learning framework,” 2023

    2023

  25. [33]

    Unimomo: Unified generative modeling of 3d molecules for de novo binder design,

    X. Kong, Z. Zhang, Z. Zhang, R. Jiao, J. Ma, K. Liu, W. Huang, and Y . Liu, “Unimomo: Unified generative modeling of 3d molecules for de novo binder design,” arXiv preprint arXiv:2503.19300 , 2025

    2025 arXiv

  26. [34]

    Mol-instructions: A large-scale biomolecular instruction dataset for large language models,

    Y . Fang, X. Liang, N. Zhang, K. Liu, R. Huang, Z. Chen, X. Fan, and H. Chen, “Mol-instructions: A large-scale biomolecular instruction dataset for large language models,” arXiv preprint arXiv:2306.08018 , 2023

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.