Pith. sign in

REVIEW 3 major objections 6 minor 45 references

DBMol: Design of High-Affinity, Target-Specific Small Molecules through Structure Prediction Models

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read This paper claims that a differentiable biomolecular structure predictor can serve as the sole optimization signal for de novo small-molecule design, producing target-specific binders without reference ligands or task-specific retraining.

desk verdict A useful new method combination — differentiable Boltz-2 guidance for small-molecule design — but the held-out AF3 evaluation is too fragile (single seed, post-hoc threshold) to carry the 'competitive' claim as strongly as written. read the letter →

arxiv 2607.19237 v1 pith:UGXMTEAO submitted 2026-07-21 cs.LG

classification cs.LG
keywords denovodrugdesignstructurepredictionmodelsdifferentiableoptimizationmoleculargraphgenerationdiscreteflowmatchingbindingaffinitypocketcoveragetarget-specificligands
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

DBMol treats a structure prediction model (Boltz-2) as a differentiable scoring function for protein–ligand interaction. Starting from a generic linear molecule, it performs gradient descent on a relaxed molecular graph to improve predicted affinity, pocket contact, and complex confidence, then projects the noisy result back onto chemically valid molecules with a discrete flow-matching denoiser. The paper's central claim is that this Boltz-guided signal transfers: on a seven-target benchmark, DBMol reaches pocket coverage and AlphaFold-3-based success comparable to baselines that use reference ligands, pharmacophores, or 3D pockets, while using weaker supervision. The payoff, if true, is a route to designing ligands for new or poorly annotated protein targets with no known binder and no curated training data.

What carries the argument

The machinery is an alternating optimization–projection loop. Optimization operates in a relaxed molecular graph (atom-type and bond-type probability vectors, continuous charges) and minimizes a composite Boltz-2-derived loss: predicted affinity plus binder-probability logit, a distogram contact loss averaged over pocket residues plus a single anchor contact, and confidence terms such as ipTM, ipLDDT, and ipDE, negated as losses. Projection uses DeFoG, a discrete flow-matching denoiser trained on valid molecules, started at late denoising time t=0.8 so it repairs chemical invalidity without washing out the optimized signal. A final optional step moves the molecule to a synthesizable analog u

What would settle it

Evaluate the same generated molecules under multiple AF3 seeds and against a genuinely independent signal such as free-energy perturbation or experimental binding assays. If AF3 success varies sharply with the random seed, or if molecules that pass AF3 success show no measurable binding, the held-out transfer claim is undermined. A complementary check: optimize against AF3 instead of Boltz-2; if those molecules also score well on AF3 'held-out' metrics, the reported transfer may be a shared-family artifact rather than a real property.

Watch

Extended reading notes

Core claim

The paper's central claim is that modern structure-prediction models can be inverted for small-molecule design: instead of learning to generate ligands from complexes or docking scores, DBMol optimizes a continuous relaxation of a molecule graph against Boltz-2's predictions (affinity value, binder probability, distogram-based pocket contacts, and confidence terms), then uses a flow-matching model initialized at a late denoising time to map the optimized representation back to a valid discrete graph. The authors report that this optimization-projection pipeline raises Boltz-2 success from 0.13 to 0.47 over unconditional generation and, under held-out AlphaFold-3 evaluation not used during op

Load-bearing premise

The transfer claim rests on the assumption that the AF3-based evaluation protocol — a single fixed-seed AlphaFold-3 run per molecule with threshold-based success criteria and manually selected pocket residues — is a meaningful held-out proxy for real binding quality, even though AlphaFold-3 and Boltz-2 are related diffusion-based structure predictors.

Editorial extensions

If this is right

  • Structure-prediction models become a reusable, target-conditioned design signal: DBMol needs only a protein sequence and pocket positions, not curated protein–ligand complexes or known ligands.
  • New or poorly annotated targets become feasible; on the de novo targets tested here, DBMol matches a pocket-conditioned baseline without a predefined 3D pocket or reference ligand.
  • Improvements in structure-prediction accuracy or affinity prediction should flow directly into better generated molecules, because the optimization signal is read off the model rather than retrained per target.
  • Synthesis-aware post-processing can raise the fraction of molecules with a solved retrosynthetic route from 0.06 to 0.33–0.39 while retaining most of the AF3 success, provided the mapped molecule stays close in size.
  • The weak correlations between structure-prediction objectives and Vina docking scores imply that docking-based reference scores and structure-prediction metrics capture different aspects of a complex, so rankings can change depending on which held-out oracle is used.

Reading between the lines

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

  • Swapping the optimizer is the decisive control: optimize against a different or corrupted structure predictor and see whether AF3 metrics still improve; if they do, the guidance is generic, if not, the effect may be Boltz-specific.
  • Because AF3 and Boltz-2 are from the same diffusion-based structure-prediction family, 'held-out' is not a fully independent oracle; experimental binding measurements or free-energy perturbation on a sample of generated molecules would settle the transfer claim.
  • Making the starting molecule a diverse scaffold rather than a fixed-size generic chain could expand the chemical space DBMol explores, since the paper fixes size from the reference ligand and begins from a simple linear molecule.
  • The authors' similarity analysis (low overlap with the reference ligand but enriched known actives) suggests structure-predictor guidance may find alternative chemotypes; that hypothesis is testable by synthesizing and assaying DBMol hits around the discovered scaffolds.
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 / 6 minor

Summary. The paper proposes DBMol, a framework for de novo small-molecule design that uses a differentiable structure prediction model (Boltz-2) as an optimization signal. Starting from a simple discrete molecule, it relaxes atom/bond/charge variables and performs constrained gradient descent on a composite objective combining affinity, contact, and confidence terms. A discrete flow-matching denoiser (DeFoG) then projects the optimized relaxed representation back to valid molecular graphs at a late denoising time; an optional SynFormer step maps molecules toward synthesizable space. The method is evaluated on seven LIT-PCBA targets with fixed hyperparameters and no use of AF3 or Vina during optimization. The main results (Table 1) report held-out AF3-based metrics (BCov, AF3 Success, iPAE, Dist) plus Vina and diversity, showing DBMol at average rank 3.83, between CGFlow-ZS (3.33) and DeFoG (5.50), with Boltz2 Success 0.47. The central claim is that optimizing a Boltz-2 objective transfers to held-out AF3 structural metrics, and that DBMol is competitive with reference-ligand-, pharmacophore-, and pocket-conditioned baselines despite weaker supervision. Additional de novo targets (LGR4, CD47) and ablations are reported.

Significance. If the transfer claim holds, the work is significant: it offers a way to use structure prediction models as general-purpose differentiable scoring functions for ligand design without curated pocket-ligand pairs or task-specific retraining. The paper has real strengths: fixed hyperparameters across targets; use of AF3 as a held-out evaluator not touched by optimization; explicit runtime disclosure; random sampling of 100 valid molecules without metric-based selection; and a synthesis-aware robustness check. The main weakness is that the held-out evaluation protocol itself is not shown to be stable enough to carry the small performance gaps on which the competitive claim rests.

major comments (3)
  1. [App. C.7.1; Table 1] The central transfer claim rests on an evaluation protocol that uses one AF3 diffusion sample per molecule at a fixed seed. AF3 is stochastic, so every number in Table 1 is a point estimate with no error bar. The gaps supporting 'competitive' are small: AF3 Success 0.92 vs 0.89 for DeFoG, BCov 0.51 vs 0.50 for DiffSBDD, average rank 3.83 vs 3.67 for DiffSBDD/SynCoGen and 3.33 for CGFlow-ZS. Without seed sensitivity or confidence intervals, it is not established that these differences exceed run-to-run noise. Moreover, App. C.7.1 states that the ipTM>=0.6 threshold was selected because it separated unconditional generation from stronger methods; this is a post hoc threshold on the same data used to report success. Please report multiple AF3 samples/seeds and a threshold sweep, or pre-specify the threshold, and do not rely on AF3 Success as a headline metric in its current form.
  2. [Table 1; App. C.6] The comparison with CGFlow-ZS is not on equal footing: App. C.6 states that the CGFlow-ZS row is averaged over 6 of the 7 LIT-PCBA targets (without 3a2i), while DBMol and the other baselines are averaged over all 7. The average-rank claim in Sec. 4.2 directly uses this number. Recompute CGFlow-ZS on the same 7 targets, or restrict the rank comparison to the common subset and state this explicitly.
  3. [Sec. 4.3; Table 3] The de novo experiments are too weak to support the paper's intended-use claim. On CD47, DBMol achieves BCov 0.03, AF3 Success 0.08, Dist 28.5, and only ties DiffSBDD in average rank; the text acknowledges low absolute pocket engagement but still concludes that DBMol 'remains competitive' in this setting. With two targets, one of which has near-zero signal, this is an inconclusive pilot rather than evidence for the claimed flexibility. Either add more targets with meaningful signal or explicitly present this section as a feasibility demonstration without a competitiveness claim.
minor comments (6)
  1. [App. C.1] The sentence 'The exact weighted composition of L_conf is provided in App. C' is self-referential; the weights (1.0, 0.3, 0.3, 0.1, 0.3, 0.1) appear later in the same section. Please clarify the pointer.
  2. [Table 2] The DBMol row renders the 'Avg. Rank / Syn.' columns as '1.830.06'; add a separator and define Syn. in the caption.
  3. [Figure 3] Define whether 'Success' in the figure is Boltz2 Success, and state the metric in the caption. The left panel shows validity dropping as denoising time increases; ensure both panels have fully legible axis labels.
  4. [App. C.3] 'polydectime' appears to be a typo for 'polynomial decay' or the intended time-distortion name; please correct.
  5. [Sec. 4.1] State clearly how many valid molecules are used for each baseline and whether baselines also use random subsampling. DBMol's oversampling/validity filtering is described, but the analogous details for baselines are missing.
  6. [Title and Abstract] 'High-affinity' in the title and abstract is stronger than what is actually evaluated; all evidence is predicted structure/confidence metrics, not experimental affinity. Consider rewording to 'predicted high-affinity' or similar.

Circularity Check

1 steps flagged · score 5.0 of 10

De novo AF3 evaluation is partially self-referential: the target pocket is defined from DBMol's own Boltz-2 output; the main LIT-PCBA comparison is independent.

  1. self definitional [Sec. 4.3 (de novo targets, Tab. 3); App. C.6 (DiffSBDD setup)]
    "Since these proteins do not have a pre-defined 3D pocket, we define it using the pocket region induced by a high-scoring DBMol-generated molecule, selected by the average of Boltz-2 predicted BCov and ipTM."

    The de novo AF3 metrics are presented as held-out structural evaluation, but the pocket residues that define BCov, AF3 Success, iPAE, and Dist are taken from a high-scoring DBMol molecule's Boltz-2 prediction—the same differentiable signal DBMol optimizes. DBMol is therefore scored against a target that DBMol itself proposed, and DiffSBDD is evaluated in a pocket chosen by DBMol. The reported LGR4 improvement (AF3 Success 0.71 vs 0.51) is not an independent test; it partially reduces to agreement with the method's own pocket proposal. The LIT-PCBA results use reference-derived pockets, so this is partial rather than total circularity.

full rationale

The core LIT-PCBA claim is not circular: DBMol optimizes Boltz-2, while the primary comparison uses AF3 metrics, Vina, and diversity that are not part of the optimization; the AF3 threshold was fixed before seeing DBMol results, and the conclusions rely on several metrics rather than AF3 Success alone. The paper explicitly labels Boltz2 Success as a diagnostic proxy-aligned reference metric rather than as primary evidence, so the 0.13->0.47 proxy improvement is a disclosed sanity check, not a load-bearing prediction. The main circularity is confined to the de novo section, where the evaluation pocket is defined from a DBMol-generated molecule's Boltz-2 prediction, making the Tab. 3 comparison self-referential at the level of target definition. The post-hoc AF3 threshold and single-seed AF3 sampling are statistical fragility concerns, not definitional circularity. Self-citations to DeFoG and SynCoGen supply implemented components or benchmark protocols, not an unverified premise, so they do not further raise the score. Overall, the paper has one partial circular evaluation step, while the central held-out comparison retains independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are mild (weights and thresholds are disclosed), but the most critical implicit input is the domain assumption that predictor confidence/affinity outputs are useful design signals, which the paper cannot validate experimentally. The AF3 Success threshold was selected using baseline separation and therefore carries a residual fitting burden.

free parameters (6)
  • loss weights w1 (contact), w2 (confidence) = w1 = w2 = 2.5 across LIT-PCBA
    Chosen by limited hyperparameter search; the paper states only a modest tuning budget; these weights directly trade off proxy optimization against held-out transfer.
  • denoising initialization time t0 = 0.8
    Set to balance validity and signal preservation; Figure 3 shows strong dependence of both success and validity on this value; effectively tuned.
  • contact distance cutoff d_cutoff = 8.0 A
    Defines the pocket-contact loss; the pocket itself is defined by residues within 3.5 A of the reference ligand.
  • AF3 Success thresholds (ipTM >= 0.6, residue contact < 3.5 A) = ipTM 0.6 threshold; 3.5 A contact
    The authors state the threshold was selected to separate unconditional generation from stronger baselines on AF3 metrics — this is a form of criteria fitting on the baselines, though it is disclosed.
  • L_conf component weights = ipTM 1.0, ipLDDT 0.3, ipDE 0.3, PDE 0.1, confidence 0.3, pTM 0.1
    Hand-set weighting of Boltz-2 confidence terms; no sensitivity analysis.
  • optimization steps K = 50 and learning rate eta = 2.0 = 50 steps, eta=2.0
    Fixed across targets but effectively chosen because the right panel of Fig. 3 shows saturation and validity drop; not fully justified as a principled choice.
assumptions (5)
  • domain assumption Boltz-2 predicted affinity, binder probability, contact distogram, and ipTM are differentiable and informative for small-molecule ligand design.
    The entire optimization stage relies on Boltz-2 gradients being meaningful design signals; no experimental validation is provided.
  • domain assumption AlphaFold-3 ipTM >= 0.6 and pocket-residue contact within 3.5 A constitute a meaningful 'success' criterion for pocket-specific binding.
    AF3 Success is the headline held-out metric, but it is a thresholded predictor-based criterion, not a binding measurement.
  • domain assumption The denoiser DeFoG can act as a projection operator that preserves the optimized signal while restoring chemical validity.
    DeFoG is trained on valid graphs only; the paper assumes the gradient signal survives the late-time denoising, supported only by the reported proxy numbers.
  • ad hoc to paper The relaxation of atom/bond categorical variables to probability simplices and charges to continuous scalars preserves the Boltz-2 scoring landscape.
    The relaxed representation is engineered to make Boltz-2 differentiable; the paper itself states it has 'no strict probabilistic interpretation'.
  • domain assumption AF3 evaluation with one fixed seed per ligand is sufficient to rank methods.
    The evaluation protocol runs AF3 once per molecule with a fixed seed; no variance is reported, although AF3 is a stochastic sampler.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DBMol: Design of High-Affinity, Target-Specific Small Molecules through Structure Prediction Models." pith.science (2026). https://pith.science/paper/UGXMTEAO

@misc{pith2026260719237,
  author       = {Pith},
  title        = {Pith review of: DBMol: Design of High-Affinity, Target-Specific Small Molecules through Structure Prediction Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGXMTEAO}},
  note         = {Machine review of arXiv:2607.19237}
}
read the original abstract

Designing small molecule ligands that bind with high affinity to specific protein pockets is a fundamental goal in drug discovery, as small molecules constitute a major fraction of approved therapeutics. Recent breakthroughs in structure prediction, such as AlphaFold-3 and Boltz-2, enable accurate biomolecular interaction prediction and show promise as foundation models for downstream tasks, including binding affinity prediction. We propose to leverage these models and introduce DBMol, a new structure predictor-guided framework for de novo small molecule design. DBMol formulates an alternating optimization and projection process. In the optimization stage, DBMol starts from an initial molecule and uses gradient-based optimization to improve pocket-specific interactions and predicted binding affinity using a structure prediction model. In the projection stage, a flow-matching model maps the optimized molecular graph to discrete and chemically valid molecules. Experiments show that DBMol effectively optimizes the Boltz-2 affinity proxy and generates molecules with strong predicted affinity and specificity under Boltz-2 evaluation. To reduce self-confirmation bias, we further evaluate generated molecules using held-out metrics, including AF3-based evaluation. DBMol substantially improves pocket coverage while maintaining molecular diversity over unconditional generation, and is competitive under held-out metrics despite the absence of reference-ligand supervision. These results support the promise of structure prediction models as effective optimization signals for de novo molecular design.

Figures

Figures reproduced from arXiv: 2607.19237 by the authors.

Figure 1
Figure 1. Framework of DBMol. (I) Optimization, where Boltz-2-conditioned gradient updates start from G0, an uninformative initial molecule graph visualized here as a simple linear carbon chain, and refine relaxed molecule representation to improve predicted binding affinity and pocket specificity for a given protein pocket; and (II). Projection, where a flow-matching-based molecule denoising model maps a noisy molecule Gt (i… view at source ↗
Figure 2
Figure 2. Protein–ligand interactions predicted by Boltz-2. For each target, the reference ligand and its interaction pattern are shown in beige in the upper left, while three molecules found via DBMol after SynFormer post-processing are shown in blue. The protein backbone and molecular surface are shown in gray, interacting protein side chains are highlighted in purple, and hydrogen bonds are shown as dashed light-blue lines… view at source ↗
Figure 3
Figure 3. Success rate based on Boltz2 and molecular validity as functions of denoising time (left) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Ligand-centric anchoring loss helps to stabilize the optimization. [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Effect of ipTM on Estrogen Receptor (ESR; 2p15) [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 3 linked inside Pith

  1. [1]

    Pharmacogenomics of gpcr drug targets.Cell, 172(1):41–54, 2018

    Alexander S Hauser, Sreenivas Chavali, Ikuo Masuho, Leonie J Jahn, Kirill A Martemyanov, David E Gloriam, and M Madan Babu. Pharmacogenomics of gpcr drug targets.Cell, 172(1):41–54, 2018

  2. [2]

    Structure-based drug design with equiv- ariant diffusion models.Nature Computational Science, 4(12):899–909, 2024

    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 equiv- ariant diffusion models.Nature Computational Science, 4(12):899–909, 2024

  3. [3]

    Flowr: Flow matching for structure-aware de novo, interaction- and fragment-based ligand generation.ArXiv, abs/2504.10564, 2025

    Julian Cremer, Ross Irwin, Alessandro Tibo, Jon Paul Janet, Simon Olsson, and Djork-Arn ´e Clevert. Flowr: Flow matching for structure-aware de novo, interaction- and fragment-based ligand generation.ArXiv, abs/2504.10564, 2025

  4. [4]

    Pocket2mol: Efficient molecular sampling based on 3d protein pockets.ArXiv, abs/2205.07249, 2022

    Xingang Peng, Shitong Luo, Jiaqi Guan, Qi Xie, Jian Peng, and Jianzhu Ma. Pocket2mol: Efficient molecular sampling based on 3d protein pockets.ArXiv, abs/2205.07249, 2022

  5. [5]

    Yael Ziv, Fergus Imrie, Brian Marsden, and Charlotte M. Deane. Molsnapper: Conditioning Diffusion for Structure-Based Drug Design.Journal of Chemical Information and Modeling, 65(9):4263–4273, 2025. PMID: 40248896

  6. [6]

    Keir Adams, Kento Abeywardane, Jenna Fromer, and Connor W. Coley. ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design, 2024

  7. [7]

    Generative artificial intelligence for navi- gating synthesizable chemical space.arXiv preprint arXiv:2410.03494, 2024

    Wenhao Gao, Shitong Luo, and Connor W Coley. Generative artificial intelligence for navi- gating synthesizable chemical space.arXiv preprint arXiv:2410.03494, 2024

  8. [8]

    Compositional flows for 3d molecule and synthesis pathway co-design

    Tony Shen, Seonghwan Seo, Ross Irwin, Kieran Didi, Simon Olsson, Woo Youn Kim, and Martin Ester. Compositional flows for 3d molecule and synthesis pathway co-design. In Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025

Show all 45 references
  1. [9]

    Syncogen: Synthe- sizable 3d molecule generation via joint reaction and coordinate modeling.arXiv preprint arXiv:2507.11818, 2025

    Andrei Rekesh, Miruna Cretu, Dmytro Shevchuk, Vignesh Ram Somnath, Pietro Li `o, Robert A Batey, Mike Tyers, Michał Koziarski, and Cheng-Hao Liu. Syncogen: Synthe- sizable 3d molecule generation via joint reaction and coordinate modeling.arXiv preprint arXiv:2507.11818, 2025

  2. [10]

    Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, pages 1–3, 2024

    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, pages 1–3, 2024

  3. [11]

    Boltz-2: Towards accurate and efficient binding affinity prediction.BioRxiv, pages 2025–06, 2025

    Saro Passaro, Gabriele Corso, Jeremy Wohlwend, Mateo Reveiz, Stephan Thaler, Vignesh Ram Somnath, Noah Getz, Tally Portnoi, Julien Roy, Hannes Stark, et al. Boltz-2: Towards accurate and efficient binding affinity prediction.BioRxiv, pages 2025–06, 2025

  4. [12]

    Defog: Discrete flow matching for graph generation

    Yiming Qin, Manuel Madeira, Dorina Thanou, and Pascal Frossard. Defog: Discrete flow matching for graph generation. InProceedings of the 42nd International Conference on Ma- chine Learning (ICML), 2025. 10

  5. [13]

    Lit-pcba: An unbiased data set for machine learning and virtual screening.Journal of Chemical Information and Modeling, 60(9):4263–4273, 2020

    Viet-Khoa Tran-Nguyen, C ´elien Jacquemard, and Didier Rognan. Lit-pcba: An unbiased data set for machine learning and virtual screening.Journal of Chemical Information and Modeling, 60(9):4263–4273, 2020. PMID: 32282202

  6. [14]

    Hirzel, Ryan P

    Rafael G ´omez-Bombarelli, David Kristjanson Duvenaud, Jos ´e Miguel Hern ´andez-Lobato, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and Al´an Aspuru-Guzik. Au- tomatic chemical design using a data-driven continuous representation of molecules.ACS Central Sc...

  7. [15]

    Viraj Bagal, Rishal Aggarwal, P. K. Vinod, and U. Deva Priyakumar. Molgpt: Molecular generation using a transformer-decoder model.J. Chem. Inf. Model., 62(9):2064–2076, 2022

  8. [16]

    Kusner, Brooks Paige, and Jos ´e Miguel Hern ´andez-Lobato

    Matt J. Kusner, Brooks Paige, and Jos ´e Miguel Hern ´andez-Lobato. Grammar variational au- toencoder. In Doina Precup and Yee Whye Teh, editors,Proceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, pages 19...

  9. [17]

    Constrained graph variational autoencoders for molecule design

    Qi Liu, Miltiadis Allamanis, Marc Brockschmidt, and Alexander Gaunt. Constrained graph variational autoencoders for molecule design. InAdvances in Neural Information Processing Systems (NeurIPS), 2018

  10. [18]

    Junction tree variational autoencoder for molecular graph generation, 2019

    Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation, 2019

  11. [19]

    Moflow: an invertible flow model for generating molecular graphs

    Chengxi Zang and Fei Wang. Moflow: an invertible flow model for generating molecular graphs. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 617–626, 2020

  12. [20]

    Digress: Discrete denoising diffusion for graph generation

    Clement Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, V olkan Cevher, and Pas- cal Frossard. Digress: Discrete denoising diffusion for graph generation. InInternational Conference on Learning Representations (ICLR), 2023

  13. [21]

    Equivariant diffusion for molecule generation in 3d

    Emiel Hoogeboom, Vıctor Garcia Satorras, Cl ´ement Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3d. InInternational conference on machine learning, pages 8867–8887. PMLR, 2022

  14. [22]

    Midi: Mixed graph and 3d denoising diffusion for molecule generation

    Cl ´ement Vignac, Nagham Osman, Laura Toni, and Pascal Frossard. Midi: Mixed graph and 3d denoising diffusion for molecule generation. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML/PKDD), 2023

  15. [23]

    Semlaflow–efficient 3d molecular generation with latent attention and equivariant flow matching

    Ross Irwin, Alessandro Tibo, Jon Paul Janet, and Simon Olsson. Semlaflow–efficient 3d molecular generation with latent attention and equivariant flow matching. InThe 28th In- ternational Conference on Artificial Intelligence and Statistics, 2025

  16. [24]

    Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary W

    Chaitanya K. Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary W. Ulissi. All-atom diffusion transformers: Unified genera- tive modelling of molecules and materials. InInternational Conference on Machine Learning, 2025

  17. [25]

    Tabasco: A fast, simplified model for molecular generation with improved physical quality, 2025

    Carlos V onessen, Charles Harris, Miruna Cretu, and Pietro Li `o. Tabasco: A fast, simplified model for molecular generation with improved physical quality, 2025

  18. [26]

    Synflownet: Design of diverse and novel molecules with synthesis constraints

    Miruna Cretu, Charles Harris, Ilia Igashov, Arne Schneuing, Marwin Segler, Bruno Correia, Julien Roy, Emmanuel Bengio, and Pietro Lio. Synflownet: Design of diverse and novel molecules with synthesis constraints. InThe Thirteenth International Conference on Learning Representa...

  19. [27]

    RGFN: Synthesizable molec- ular generation using GFlowNets.Advances in Neural Information Processing Systems, 37:46908–46955, 2024

    Michał Koziarski, Andrei Rekesh, Dmytro Shevchuk, Almer van der Sloot, Piotr Gai ´nski, Yoshua Bengio, Chenghao Liu, Mike Tyers, and Robert Batey. RGFN: Synthesizable molec- ular generation using GFlowNets.Advances in Neural Information Processing Systems, 37:46908–46955, 2024

  20. [28]

    All-atom inverse protein folding through discrete flow matching

    Kai Yi, Kiarash Jamali, and Sjors HW Scheres. All-atom inverse protein folding through discrete flow matching. InForty-second International Conference on Machine Learning, 2025. 11

  21. [29]

    Protein hunter: ex- ploiting structure hallucination within diffusion for protein design.bioRxiv, pages 2025–10, 2025

    Yehlin Cho, Griffin Rangel, Gaurav Bhardwaj, and Sergey Ovchinnikov. Protein hunter: ex- ploiting structure hallucination within diffusion for protein design.bioRxiv, pages 2025–10, 2025

  22. [30]

    One-shot design of functional protein binders with bindcraft.Nature, 646(8084):483–492, 2025

    Martin Pacesa, Lennart Nickel, Christian Schellhaas, Joseph Schmidt, Ekaterina Pyatova, Lu- cas Kissling, Patrick Barendse, Jagrity Choudhury, Srajan Kapoor, Ana Alcaraz-Serna, et al. One-shot design of functional protein binders with bindcraft.Nature, 646(8084):483–492, 2025

  23. [31]

    Pxdesign: Fast, modular, and accurate de novo design of protein binders.bioRxiv, pages 2025–08, 2025

    Protenix Team, Milong Ren, Jinyuan Sun, Jiaqi Guan, Cong Liu, Chengyue Gong, Yuzhe Wang, Lan Wang, Qixu Cai, Wenzhi Ma, et al. Pxdesign: Fast, modular, and accurate de novo design of protein binders.bioRxiv, pages 2025–08, 2025

  24. [32]

    Boltzdesign1: Inverting all-atom structure prediction model for generalized biomolecular binder design.bioRxiv, pages 2025–04, 2025

    Yehlin Cho, Martin Pacesa, Zhidian Zhang, Bruno E Correia, and Sergey Ovchinnikov. Boltzdesign1: Inverting all-atom structure prediction model for generalized biomolecular binder design.bioRxiv, pages 2025–04, 2025

  25. [33]

    Aizynthfinder: a fast, robust and flexible open-source software for ret- rosynthetic planning.Journal of Cheminformatics, 2020

    Samuel Genheden, Amol Thakkar, Veronika Chadimov´a, Jean-Louis Reymond, Ola Reymond, and Esben Bjerrum. Aizynthfinder: a fast, robust and flexible open-source software for ret- rosynthetic planning.Journal of Cheminformatics, 2020. Published 17 November 2020

  26. [34]

    L. Wang, F. Hu, Q. Cui, et al. Structural insights into the lgr4–rspo2–znrf3 complexes regulat- ing wnt/β-catenin signaling.Nature Communications, 16:362, 2025

  27. [35]

    Cd47 as a promising therapeutic target in oncology.Frontiers in Immunol- ogy, 2022

    Hai Zhao, Shuangshuang Song, Junwei Ma, Zhiyong Yan, Hongwei Xie, Ying Feng, and Shusheng Che. Cd47 as a promising therapeutic target in oncology.Frontiers in Immunol- ogy, 2022. Published 2022 Aug 22

  28. [36]

    Have protein-ligand co-folding methods moved beyond memorisation?BioRxiv, pages 2025–02, 2025

    Peter ˇSkrinjar, J´erˆome Eberhardt, Janani Durairaj, and Torsten Schwede. Have protein-ligand co-folding methods moved beyond memorisation?BioRxiv, pages 2025–02, 2025

  29. [37]

    Openbind: Open protein–ligand binding data

    OpenBind. Openbind: Open protein–ligand binding data. https://zenodo.org/records/20026661, 2026

  30. [38]

    Gen- erative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design

    Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. Gen- erative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. InInternational Conference on Machine Learning (ICML), 2024

  31. [39]

    Sparse training of discrete diffusion models for graph generation

    Yiming Qin, Clement Vignac, and Pascal Frossard. Sparse training of discrete diffusion models for graph generation. InArXiv, 2023

  32. [40]

    How powerful are graph neural networks? InInternational Conference on Learning Representations (ICLR), 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations (ICLR), 2019

  33. [41]

    On structural expressive power of graph transformers

    Wenhao Zhu, Tianyu Wen, Guojie Song, Liang Wang, and Bo Zheng. On structural expressive power of graph transformers. InACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2023

  34. [42]

    Graph inductive biases in transformers without message pass- ing

    Liheng Ma, Chen Lin, Derek Lim, Adriana Romero-Soriano, Puneet K Dokania, Mark Coates, Philip Torr, and Ser-Nam Lim. Graph inductive biases in transformers without message pass- ing. InInternational Conference on Machine Learning (ICML), 2023

  35. [43]

    Cometh: A continuous- time discrete-state graph diffusion model

    Antoine Siraudin, Fragkiskos D Malliaros, and Christopher Morris. Cometh: A continuous- time discrete-state graph diffusion model. InArXiv, 2024

  36. [44]

    Zinc 15–ligand discovery for everyone

    Teague Sterling and John J Irwin. Zinc 15–ligand discovery for everyone. InJournal of Chemical Information and Modeling, 2015

  37. [45]

    ref element

    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, 2024. 12 Appendix Table of Contents A Discussion...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.