Multi-domain Distribution Learning for De Novo Drug Design

Adrian W. Dobbelstein; Arne Schneuing; Bruno Correia; Ilia Igashov; Michael Bronstein; Thomas Castiglione

REVIEW 3 major objections 4 minor 1 cited by

Multi-domain Distribution Learning for De Novo Drug Design

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper introduces DrugFlow, which combines flow matching with Markov bridges, and claims it reproduces the chemical, geometric, and physical distribution of protein–ligand complexes better than three established baselines.

desk verdict A genuine method paper with reusable components, but the SOTA distribution-learning claim is conditional on an unverified baseline-comparability assumption. read the letter →

arxiv 2508.17815 v1 pith:KT7AZWBM submitted 2025-08-25 cs.LG q-bio.BM

Arne Schneuing , Ilia Igashov , Adrian W. Dobbelstein , Thomas Castiglione , Michael Bronstein , Bruno Correia This is my paper · ORCID

classification cs.LGq-bio.BM

keywords denovodrugdesignstructure-basedflowmatchingMarkovbridgedistributionlearninguncertaintyestimationpreferencealignmentproteinflexibility

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 argues that generative models for structure-based drug design should be judged by how faithfully they reproduce the training distribution of protein–ligand complexes, not by absolute property scores. It introduces DrugFlow, a model that samples atom coordinates with conditional flow matching while generating atom and bond types with Markov bridges, and whose experiments show it tracks the filtered CrossDocked training set across geometry, drug-likeness, binding-efficiency, and interaction metrics more closely than Pocket2Mol, DiffSBDD, and TargetDiff. DrugFlow also adds an uncertainty head that flags out-of-distribution samples, a virtual atom type that lets molecule size be decided during sampling, and a flexible variant, FlexFlow, that simultaneously samples side-chain conformations. The paper further proposes a preference-alignment scheme, applied jointly to the continuous and discrete generative processes, that shifts sampling toward user-desired property values. If these results hold, the practical payoff is a generative model that can be retrained or fine-tuned on curated data to steer drug design while preserving physical plausibility.

What carries the argument

The load-bearing machinery is a heterogeneous graph neural network that, at each denoising step, operates on a 3D graph with ligand atoms, protein residues represented as C-alpha nodes carrying full residue vector features, and three edge types, using geometric vector perceptrons to keep coordinate predictions equivariant. It drives two coupled generative processes: Euclidean conditional flow matching for atom coordinates and side-chain torsion angles on a torus, and Markov bridge models for atom and bond types. A virtual node type acts as a remove-this-atom label, so molecule size can emerge during sampling; an uncertainty head outputs per-atom sigma, and the loss is modified to make sigma large where the predicted vector field is poor; preference alignment uses a DPO-style contrast between winning and losing samples, computed as loss differences in the coordinate flow and the discrete Markov bridges. A NERF step converts sampled side-chain angles back to atomic coordinates at every training and sampling step.

What would settle it

Retrain or fine-tune Pocket2Mol, DiffSBDD, and TargetDiff on the same filtered CrossDocked training set that DrugFlow used, then recompute the Wasserstein, Jensen-Shannon, and FCD distances to that filtered set; if any baseline then matches or beats DrugFlow's distances, the paper's central distribution-learning claim is falsified.

Watch

Extended reading notes

Core claim

DrugFlow is a multi-domain generative model for structure-based drug design that learns the joint distribution of ligand atom coordinates, atom types, and bond types conditioned on a protein pocket. The central discovery claimed is that this combination—continuous flow matching for Euclidean coordinates and discrete Markov bridges for graph types—yields a distribution learner whose samples lie closer to the training distribution than Pocket2Mol, DiffSBDD, and TargetDiff, across bond lengths and angles, atom and bond type frequencies, ring systems, QED, SA, logP, rotatable bonds, Vina and Gnina efficiency, and normalized protein-ligand interaction counts. The paper reports the largest gap in Fréchet ChemNet Distance (4.278 versus at least 11.637 for baselines) and the lowest Jensen-Shannon divergence for the joint distribution of QED, SA, logP, and Vina efficiency. It also claims the model learns the conditional distribution of molecule sizes, detects out-of-distribution samples through per-atom uncertainty, and, in its FlexFlow extension, recovers bound side-chain rotamer distributions; a preference-alignment update improves target properties beyond fine-tuning, at a moderate cost in molecular validity.

Load-bearing premise

The headline comparison assumes the three baseline models were trained and evaluated on the same filtered training set as DrugFlow; if the published baselines were trained on the original, unfiltered data, their distances to the filtered set would be inflated by data curation rather than by model deficiency, and the claimed margin would shrink or vanish.

Editorial extensions

If this is right

Distribution-centered evaluation becomes the primary benchmark for generative drug design: a model that wins on these distances can be retrained on curated data and will inherit the curated distribution's properties.
Molecule size no longer needs to be pre-specified: the virtual-node mechanism lets the model delete excess atoms during sampling, and clashes stay low within the trained deletion bandwidth.
Uncertainty scoring gives a usable out-of-distribution flag and correlates with ligand efficiency and steric clashes, so generated molecules can be ranked before expensive docking.
FlexFlow extends the same framework to side-chain conformational sampling, making target-aware generation possible for pockets that may be in unbound or flexible states.
The preference-alignment loss lets users push QED, SA, or Vina efficiency in a chosen direction while keeping the continuous and discrete domains coupled.

Reading between the lines

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

An implicit testable extension is to scale the virtual-node bandwidth: the paper shows the model removes about Nmax/2 atoms on average and fails beyond that bandwidth, so training with a larger Nmax or a heavy-tailed virtual-node distribution should enable fully adaptive sizing on large, crowded pockets.
Because the uncertainty score is trained end-to-end with the flow-matching loss, it could double as a training signal: weighting the flow-matching loss by the inverse predicted variance is a natural next step the paper does not explore.
The same multi-domain alignment scheme should transfer to other mixed continuous-discrete generative tasks, such as linker design or co-design of ligands with protein backbones, whenever a DPO-style win/lose comparison can be defined per modality.
The comparison also leaves a protocol question open: the reported margin assumes the baselines were trained on the same filtered distribution as DrugFlow, so retraining the baselines on that filtered set would be the cleanest way to confirm the margin is a property of the model and not of the evaluation setup.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Jan,

Here's my take on the DrugFlow paper. It is a genuinely strong method paper, and the new pieces are real: the per-atom uncertainty head with a clean MAP derivation, the virtual 'no atom' node type for adaptive molecule size, the side-chain angle flow on the torus with a polynomial scheduler, and the multi-domain DPO applied across flow and Markov bridge losses. The architecture is a sensible heterogeneous GNN with GVP layers. Appendix A's math is coherent, the empirical evaluation is broad with bootstrap and t-tests, and code is promised. This deserves a serious referee.

The soft spot is the baseline-comparability assumption. The paper removes training molecules that fail PoseBusters filters and filters DrugFlow samples by RDKit validity, but never says whether Pocket2Mol, DiffSBDD and TargetDiff were retrained on the filtered set or validity-filtered during evaluation. If they were not, every Wasserstein and JSD distance to the filtered training set is inflated for the baselines. That data-protocol gap is load-bearing for the SOTA distribution-learning claim; the bootstraps can't fix systematic bias. This is fixable, and I'd want it addressed before I believed the headline number. The OOD detection is suggestive but not a true OOD test—it's mostly tail-detection on the training distribution. The size adaptation experiment is honest about its 10-atom 'bandwidth.' The preference alignment results are interesting but come with a 10-20% validity drop.

Also, the 'Practical relevance' paragraph in the introduction is duplicated word-for-word, which is sloppy but not substantive.

Bottom line: the method is good, the math holds, and the components are reusable. The headline claim is conditional on an unverified baseline protocol. I'd accept it for peer review and push for a revised evaluation. I'd cite it for the components, but I'd be careful about citing the SOTA claim until the comparability question is resolved.

Best.

Referee Report

3 major / 4 minor

Summary. The paper introduces DrugFlow, a generative model for structure-based drug design that combines continuous flow matching for atom coordinates (and, in the FlexFlow variant, for protein side-chain torsion angles on the torus) with Markov bridge models for discrete atom and bond types. It adds three methodological contributions: an end-to-end trained per-atom uncertainty estimate intended to flag out-of-distribution samples, an adaptive size selection mechanism based on virtual 'no atom' nodes, and a multi-domain preference alignment (MDPA) loss adapted from DPO for both continuous and discrete domains. The evaluation is centered on distribution-matching metrics (Wasserstein distances and Jensen-Shannon divergences) against Pocket2Mol, DiffSBDD, and TargetDiff on CrossDocked, and the paper claims consistent state-of-the-art distribution learning performance.

Significance. If the claims hold, the paper makes a valuable contribution by shifting evaluation of structure-based drug design models toward distribution learning, which better reflects the training objective of likelihood-based generative models. The paper ships a reproducible codebase, transparent derivations in the appendices, and a set of falsifiable distribution-matching benchmarks. The uncertainty estimate, if validated quantitatively, could be practically useful for ranking generated molecules. The preference alignment scheme is a sensible extension of DPO to multi-domain generative models. The main risk is that the headline SOTA distribution-learning claim rests on a comparison protocol that may not be apples-to-apples.

major comments (3)

[Section 3.1, Tables 1-3] The paper states that ligands failing PoseBusters filters were removed from the training set and that DrugFlow samples were filtered by RDKit validity, but it does not state whether the baselines (Pocket2Mol, DiffSBDD, TargetDiff) were retrained on this filtered training set or whether their generated molecules were subjected to the same validity filter. Because every Wasserstein and Jensen-Shannon distance in Tables 1-3 is computed against the filtered training set, any mismatch in training distribution or in post-generation filtering would systematically inflate baseline distances. The bootstrap t-tests in Appendix B.10 resample from the same generated samples and cannot correct this systematic protocol difference. The authors should provide explicit details or re-run the baselines under identical conditions; without this, the central 'state-of-the-art distribution learner' claim is not fully supported.
[Section 5, Conclusion] The paper claims that DrugFlow 'consistently achieves state-of-the-art distribution learning performance across various orthogonal metrics,' but the comparison includes only three baselines, all from 2022-2023. More recent structure-based drug design models are not compared, and the paper itself notes a few metrics (e.g., logP and hydrophobic interactions in Tables 1 and 2) where another method ranks first. The authors should either broaden the baseline set or temper the 'state-of-the-art' wording to accurately reflect the scope of the comparison.
[Section 3.2] The out-of-distribution detection claim is supported only by qualitative histograms and correlation plots; no quantitative metric such as AUROC or precision-recall for OOD detection is reported. Given the abstract states that the uncertainty estimate is 'able to detect out-of-distribution samples,' the authors should provide a quantitative evaluation, especially because the uncertainty scores have a narrow range (0.85-0.92) as acknowledged in Section 3.2.

minor comments (4)

[Appendix A.5, Section 2.2] The empirical size prior p(N|M) is sampled from the training set; this is a practical heuristic rather than an end-to-end learned size selection. The phrase 'end-to-end size estimation' used in Section 2.2 and the Introduction should be qualified, as the model still requires a pre-specified maximum Nmax.
[Section 3.3] The authors acknowledge that with Nmax=10 the model 'fails to remove more' than about 10 atoms and that a fully adaptive regime would require scaling up Nmax. This limitation should be stated clearly in the main text rather than only in the discussion of Figure 3.
[Appendix B.10, Tables 14-16] Tables 14-16 contain typographical issues where standard deviations are concatenated with means without proper spacing (e.g., '0.082 (0.004)0.045(0.006)'); these should be fixed for readability.
[Section 2.4, Appendix A.4] The MDPA loss in Eq. 5 is presented with weights lambda_c, lambda_w, lambda_l, but the relationship between Eq. 5 and the derivation in Appendix A.4 (Eqs. 47-49) is not fully spelled out in the main text, making the loss harder to follow.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the distribution-learning and alignment claims are benchmarked externally, and the only noted comparability issue is a data-protocol threat, not a constructional circularity.

full rationale

DrugFlow's derivation chain is self-contained. The coordinate flow is standard conditional flow matching (Eq. 12), the torsional flow is Riemannian flow matching with an explicit geodesic scheduler (Eqs. 3-4, 13-17), the discrete graph process is a Markov bridge whose transition kernels and closed-form marginals are re-derived in the appendix (Eqs. 18-24), and the uncertainty head is a Gaussian NLL loss (Eq. 1, Appendix A.2). None of these definitions assumes the benchmark result. The central SOTA claim is tested against samples from Pocket2Mol, DiffSBDD, and TargetDiff using Wasserstein/JSD distances to the CrossDocked training set, FCD embeddings, and external oracles (QED, SA, Vina/Gnina, ProLIF, PoseBusters); no fitted parameter is renamed as a prediction. The Markov-bridge citation to Igashov et al. (2023) and the DiffSBDD citation to Schneuing et al. (2022) are self-citations by overlapping authors, but they are not load-bearing: the equations are re-derived and DiffSBDD is an external baseline. The PoseBusters filtering of the training set and RDKit-validity filtering of DrugFlow samples (Section 3.1) could make the baseline comparison unfair if baseline samples were not filtered or retrained, but that is an experimental-protocol validity threat, not a circularity: it does not make the Wasserstein/JSD numbers true by construction. The Appendix A.3 caveat that the uncertainty SDE is 'a purely hypothetical scenario' limits the interpretability of the uncertainty score but does not make the OOD evaluation circular, since the correlation with distribution tails is measured empirically.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

The central distribution-learning result rests mostly on standard generative-modeling theory plus a data-protocol assumption about baseline comparability. The auxiliary claims (uncertainty, size adaptation, preference alignment) depend on several hand-chosen hyperparameters (Nmax, scheduler k, alignment weights) and on the paper's own metric definitions. No fundamentally new physical entity or first-principles law is introduced; the virtual node and uncertainty score are computational devices.

free parameters (6)

Nmax (virtual nodes) = 10
Table 4. Sets the upper bandwidth of molecule-size adaptation. The paper explicitly notes that larger numbers are out-of-distribution and the model fails to remove them (Figure 3B).
Angle scheduler exponent k = 3
Eq. 3 and Table 4. Chosen to approximate an exponential scheduler with c=5 while satisfying boundary conditions. Affects FlexFlow sampling trajectories.
OOD regularization lambda = 10
Table 4. Stabilizes the uncertainty loss in Eq. 1. Chosen by hand; affects the uncertainty estimates and thus the OOD detection claim.
MDPA weights (beta, lambda_coord, lambda_atom, lambda_bond, lambda_w, lambda_l) = 100, 1, 0.5, 0.5, 1, 0.2
Table 4. Preference alignment results depend on these tuning choices. The paper does not report a sensitivity analysis.
Base loss weights (lambda_coord, lambda_chi, lambda_a, lambda_b) = unspecified
Eq. 25. The overall training loss is a weighted sum of flow matching and Markov bridge losses, but the weights are not given in the paper. If they are not 1, they are hidden free parameters.
Number of sampling steps = 500
Table 4. Used for all ODE sampling. Sample quality and the reported distributions depend on this choice.

assumptions (6)

standard math Flow matching and Markov bridge objectives trained by minimizing the stated losses produce samples distributed according to the training data distribution.
Section A.1. This is the theoretical foundation of the generative framework, taken from Lipman et al. and Igashov et al.
domain assumption The Gaussian error model on the flow-matching vector field, with the MAP regularizer, yields a calibrated per-atom uncertainty estimate.
Appendix A.2. The paper derives the loss but does not show calibration on held-out OOD data; the uncertainty is only examined on in-distribution tails.
domain assumption Baseline models (Pocket2Mol, DiffSBDD, TargetDiff) were trained on the same PoseBusters-filtered training set used to compute distribution distances.
Section 3.1, Dataset & Baselines. This is unstated and load-bearing for the SOTA distribution-learning comparison. If false, baseline distances to the filtered training set are inflated.
domain assumption Wasserstein/JSD/FCD distances between generated samples and the filtered CrossDocked training set are valid and sufficient measures of 'distribution learning' quality.
Section 3.1. The paper argues this evaluation philosophy is more meaningful than absolute metric values, but this remains a methodological choice.
ad hoc to paper The empirical prior p(N|M) for initial molecule size, plus Nmax=10 virtual nodes, is a sufficient starting distribution for learning conditional size distributions.
Section A.5, Number of nodes. The initial size is sampled from a histogram of training sizes given pocket residue count. The model only trims a small bandwidth, so the 'adaptive size' result is substantially constrained by this choice.
standard math NERF angle-to-coordinate conversion yields physically plausible structures during training and sampling.
Section 2.3. NERF is a standard protein-structure reconstruction tool; the paper relies on it for the FlexFlow extension.

invented entities (2)

Virtual 'no atom' node type
purpose: Allows the model to dynamically add or remove atoms during sampling, so molecule size can adapt to the pocket context.
This is a computational construct added to the discrete state space. Its usefulness is demonstrated only through in-paper experiments (Figure 3); there is no external falsifiable prediction attached.
Per-atom uncertainty score sigma_tot
purpose: Flags out-of-distribution or low-confidence generated atoms and ranks molecules at inference time.
Derived from the model's own flow-matching error variance. It is an internal model output with no external benchmark demonstrating calibrated OOD detection, so it is not independently evidenced.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-domain Distribution Learning for De Novo Drug Design." pith.science (2026). https://pith.science/paper/KT7AZWBM

@misc{pith2026250817815,
  author       = {Pith},
  title        = {Pith review of: Multi-domain Distribution Learning for De Novo Drug Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KT7AZWBM}},
  note         = {Machine review of arXiv:2508.17815}
}

read the original abstract

We introduce DrugFlow, a generative model for structure-based drug design that integrates continuous flow matching with discrete Markov bridges, demonstrating state-of-the-art performance in learning chemical, geometric, and physical aspects of three-dimensional protein-ligand data. We endow DrugFlow with an uncertainty estimate that is able to detect out-of-distribution samples. To further enhance the sampling process towards distribution regions with desirable metric values, we propose a joint preference alignment scheme applicable to both flow matching and Markov bridge frameworks. Furthermore, we extend our model to also explore the conformational landscape of the protein by jointly sampling side chain angles and molecules.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards
cs.LG 2026-07 reject novelty 6.0 of 10

A reinforcement-learning docking model with a cooperative stacking reward claims large gains on amyloid fibril poses, but its main evaluation set is model-generated rather than experimentally resolved.

Reference graph

Works this paper leans on

49 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

1x, y ∈ M and u, v ∈ TxM

w(α) = ((α +π) mod 2 π)−π wraps values within the range [−π,π ) and is applied element-wise. 1x, y ∈ M and u, v ∈ TxM. 17 Published as a conference paper at ICLR 2025 Here, we additionally use a schedulerκ(t), that satisfiesκ(0) = 1 andκ(1) = 0, to control the rate at which the geodesic distanced betweenx0 andx1 decreases (Chen & Lipman, 2023): d(xt,x1) =...

work page 2025
[2]

Riemannian flow matching on general geometries

Ricky TQ Chen and Yaron Lipman. Riemannian flow matching on general geometries. arXiv preprint arXiv:2302.03660,

arXiv
[3]

We compare distributions of the first two principal components in Figure

First, we apply PCA to the molecular embeddings computed by ChemNet, the neural network used to calculate the Fr´echet ChemNet Distance (FCD) (Preuer et al., 2018). We compare distributions of the first two principal components in Figure

work page 2018
[4]

Model Parameter D RUGFLOW DRUGFLOW-OOD F LEX FLOW DRUGFLOW-PA Num

Table 4: Model hyperparameters. Model Parameter D RUGFLOW DRUGFLOW-OOD F LEX FLOW DRUGFLOW-PA Num. weights (M) 12.1 12 .1 12 .6 12 .1 Model size (MB) 48.408 48.411 50.557 48.408 Training epochs 600 600 700 ≥50 Virtual nodesNmax 10 10 10 10 Sampling steps 500 500 500 500 Uncertainty head No Yes No No Flexible side chains No No Yes No OODλ – 10 – – Schedule...

work page 2025
[5]

In each plot, we compare training data (blue areas) and samples generated by different methods (black solid lines)

�� Figure 7: Distributions of the first two principle components of molecule embeddings computed by the FCD neural network. In each plot, we compare training data (blue areas) and samples generated by different methods (black solid lines). Figure 8: Distributions atom types (4 most popular types) and covalent bond types in the training dat...

work page 2025
[6]

PILOT: Equivariant diffusion for pocket conditioned de novo ligand generation with multi-objective guidance via importance sampling

11 Published as a conference paper at ICLR 2025 Julian Cremer, Tuan Le, Frank No ´e, Djork-Arn´e Clevert, and Kristof T Sch ¨utt. Pilot: Equivariant diffusion for pocket conditioned de novo ligand generation with multi-objective guidance via importance sampling. arXiv preprint arXiv:2405.14925,

work page Pith review arXiv 2025
[7]

Mixed continuous and categorical flow matching for 3d de novo molecule generation

Ian Dunn and David Ryan Koes. Mixed continuous and categorical flow matching for 3d de novo molecule generation. arXiv preprint arXiv:2404.19739,

arXiv
[9]

3d equiv- ariant diffusion for target-aware molecule generation and affinity prediction

Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equiv- ariant diffusion for target-aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543, 2023a. Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, and Quanquan Gu. Decompdiff: Diffusion models with dec...

arXiv

Show all 49 references

[10]

Equivariant 3d-conditional diffusion models for molecular linker design

12 Published as a conference paper at ICLR 2025 Ilia Igashov, Hannes St¨ark, Cl´ement Vignac, Victor Garcia Satorras, Pascal Frossard, Max Welling, Michael Bronstein, and Bruno Correia. Equivariant 3d-conditional diffusion models for molecular linker design. arXiv preprint arX...

2025 arXiv
[11]

Retrobridge: Modeling retrosynthesis with markov bridges

Ilia Igashov, Arne Schneuing, Marwin Segler, Michael Bronstein, and Bruno Correia. Retrobridge: Modeling retrosynthesis with markov bridges. arXiv preprint arXiv:2308.16212,

arXiv
[12]

Efficient 3d molecular generation with flow matching and scale optimal transport

Ross Irwin, Alessandro Tibo, Jon-Paul Janet, and Simon Olsson. Efficient 3d molecular generation with flow matching and scale optimal transport. arXiv preprint arXiv:2406.07266,

arXiv
[13]

Learning from protein structure with geometric vector perceptrons.arXiv preprint arXiv:2009.01411,

Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael JL Townshend, and Ron Dror. Learning from protein structure with geometric vector perceptrons.arXiv preprint arXiv:2009.01411,

2009 arXiv
[14]

Equivariant graph neural networks for 3d macromolecular structure

Bowen Jing, Stephan Eismann, Pratham N Soni, and Ron O Dror. Equivariant graph neural networks for 3d macromolecular structure. arXiv preprint arXiv:2106.03843,

arXiv
[16]

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

Haitao Lin, Yufei Huang, Meng Liu, Xuanjing Li, Shuiwang Ji, and Stan Z Li. Diffbp: Generative diffusion of 3d molecules for target protein binding. arXiv preprint arXiv:2211.11214,

arXiv
[17]

Ppflow: Target-aware peptide design with torsional flow matching

Haitao Lin, Odin Zhang, Huifeng Zhao, Dejun Jiang, Lirong Wu, Zicheng Liu, Yufei Huang, and Stan Z Li. Ppflow: Target-aware peptide design with torsional flow matching. bioRxiv, pp. 2024– 03,

2024
[18]

Flow matching for generative modeling

Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747,

arXiv
[19]

Generating 3d molecules for target protein binding

Meng Liu, Youzhi Luo, Kanji Uchino, Koji Maruhashi, and Shuiwang Ji. Generating 3d molecules for target protein binding. arXiv preprint arXiv:2204.09410,

arXiv
[22]

13 Published as a conference paper at ICLR 2025 Jerod Parsons, J Bradley Holmes, J Maurice Rojas, Jerry Tsai, and Charlie EM Strauss

URL https://openreview.net/forum?id=EKt4NQZ47U. 13 Published as a conference paper at ICLR 2025 Jerod Parsons, J Bradley Holmes, J Maurice Rojas, Jerry Tsai, and Charlie EM Strauss. Practi- cal conversion from torsion space to cartesian space for in silico protein synthesis. J...

2025
[24]

Structure-based drug design with equivariant diffusion models

Arne Schneuing, Yuanqi Du, Charles Harris, Arian Jamasb, Ilia Igashov, Weitao Du, Tom Blundell, Pietro Li´o, Carla Gomes, Max Welling, Michael Bronstein, and Bruno Correia. Structure-based drug design with equivariant diffusion models. arXiv preprint arXiv:2210.13695,

arXiv
[25]

Harmonic self-conditioned flow matching for multi-ligand docking and binding site design

Hannes St¨ark, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. Harmonic self-conditioned flow matching for multi-ligand docking and binding site design. arXiv preprint arXiv:2310.05764 ,

arXiv
[26]

URL https://doi.org/10.1038/nbt.3988

doi: 10.1038/nbt.3988. URL https://doi.org/10.1038/nbt.3988. Kyle Swanson, Gary Liu, Denise B Catacutan, Autumn Arnold, James Zou, and Jonathan M Stokes. Generative ai for designing and validating easily synthesizable and structurally novel antibiotics. Nature Machine Intellig...

doi
[27]

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. arXiv preprint arXiv:2209.14734,

arXiv
[28]

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

14 Published as a conference paper at ICLR 2025 Clement Vignac, Nagham Osman, Laura Toni, and Pascal Frossard. Midi: Mixed graph and 3d denoising diffusion for molecule generation. arXiv preprint arXiv:2302.09048,

2025 arXiv
[29]

W Patrick Walters, Matthew T Stahl, and Mark A Murcko

URL https://practicalcheminformatics.blogspot.com/2022/12/ identifying-ring-systems-in-molecules.html . W Patrick Walters, Matthew T Stahl, and Mark A Murcko. Virtual screening—an overview. Drug discovery today, 3(4):160–178,

2022
[30]

Fast protein backbone generation with se (3) flow matching

Jason Yim, Andrew Campbell, Andrew YK Foong, Michael Gastegger, Jos ´e Jim´enez-Luna, Sarah Lewis, Victor Garcia Satorras, Bastiaan S Veeling, Regina Barzilay, Tommi Jaakkola, et al. Fast protein backbone generation with se (3) flow matching. arXiv preprint arXiv:2310.05297, 2...

arXiv
[32]

based on a data point x1. Thus, the conditional flow matching loss amounts to LCFM(θ) = Et,q(x1),p(x0)∥vθ(xt,t )− ˙xt∥2, (7) where ˙xt = d/dtψt(x0|x1) is the time derivative of the conditional flowxt =ψt(x0|x1). For sampling, we obtain a samplex0 from the priorp and simulate t...

2022
[33]

(9) We use this setup with µt(x1) =tx1 + (1−t)x0, σ t(x1) =σ (10) to model the flow for ligand coordinates

and consider a Gaussian conditional probability path pt(x|x1) =N (x|µt(x1),σt(x1)2I) (8) with generating vector field ut(x|x1) = σ′ t(x1) σt(x1) (x−µt (x1)) +µ′ t(x1). (9) We use this setup with µt(x1) =tx1 + (1−t)x0, σ t(x1) =σ (10) to model the flow for ligand coordinates. T...

2023
[35]

out-of-distribution

(19) Each transition is given by p(zt+∆t|zt,z1 =y) = Cat(zt+∆t;Qtzt) (20) wherezt∈{ 0, 1}K is a one-hot representation of the current category andQt is a transition matrix parameterised as Qt :=Qt(y) =βtI + (1−βt)y1T K. (21) Any intermediate state of the Markov chain can be pr...

2025
[36]

(38) Applying this idea to the flow matching setup, we obtain the loss in Eq

19 Published as a conference paper at ICLR 2025 With priorp(ˆσ2) =N (1, 1/λ), we can then solve the maximum a posteriori (MAP) problem as ˆσ2 MAP = arg max ˆσ2 [logp(ˆσ2|y)] (35) = arg max ˆσ2 [logp(y|ˆσ2) + logp(ˆσ2)] (36) = arg min ˆσ2 [d 2 log ˆσ2 + 1 2ˆσ2∥y− ˆy∥2 + (ˆσ2− 1...

2025
[37]

A.3 M OTIVATION FOR THE FINAL UNCERTAINTY SCORE In the flow matching setting, we integrate an ordinary differential equation (ODE) along a time- dependent vector field. However, the true data generating vector field is not known and we only have access to the estimated neural ...

2023
[38]

(46) We propose to apply this framework to the Markov bridge models

logσ(−βT ( +DKL(p(xw t+∆t|xw 1,t )∥qφ(xw t+∆t|xw t )) −DKL(p(xw t+∆t|xw 1,t )∥qθ(xw t+∆t|xw t )) −DKL(p(xl t+∆t|xl 1,t )∥qφ(xl t+∆t|xl t)) +DKL(p(xl t+∆t|xl 1,t )∥qθ(xl t+∆t|xl t)))). (46) We propose to apply this framework to the Markov bridge models. Using Eq. 24, we can der...

2024
[39]

The overall loss function is defined as the weighted sum of ˜LMDPA and the regularization term: LMDPA(φ) =λMDPA ˜LMDPA(φ) +λwLw(φ) +λlLl(φ)

applied to both winning and losing samples. The overall loss function is defined as the weighted sum of ˜LMDPA and the regularization term: LMDPA(φ) =λMDPA ˜LMDPA(φ) +λwLw(φ) +λlLl(φ). (49) SettingλMDPA = λl = 0 corresponds to simple fine-tuning of the model on the winning sam...

2023
[40]

In all other cases, hydrogens are assumed to be implicit following normal valence assumptions

Featurization We consider the atom types{C, N, O, S, B, Br, Cl, P, I, F, NH, N+, O-} where +/- indicate charges and NH is a nitrogen atom with explicit hydrogen. In all other cases, hydrogens are assumed to be implicit following normal valence assumptions. Furthermore, DRUGFLO...

2022
[41]

All these functions are implemented with geometric vector perceptron (GVP) layers (Jing et al., 2020

It performs message passing operations using separate learnable message functions for each edge type, and separate update functions for each node type. All these functions are implemented with geometric vector perceptron (GVP) layers (Jing et al., 2020

2020
[42]

to ensure equivariance to global roto-translations. Number of nodes To choose a number of computational nodes during sampling, which represent an upper bound on the final number of atoms, we compute the categorical distribution p(N|M) 22 Published as a conference paper at ICLR...

2025
[46]

This model uses virtual nodes but not the uncertainty head; • D RUGFLOW (no virt. nodes), an identical model without virtual nodes; • D RUGFLOW-OOD, an identical model trained with additional uncertainty head to be able to detect out-of-distribution (OOD) samples. This model u...

2025
[47]

The authors of AliDiff provide sampled molecules and Vina scores, which we use for re-evaluation

are as reported by their authors (SA values were mapped to the original scale using SA= 10−9SAnorm). The authors of AliDiff provide sampled molecules and Vina scores, which we use for re-evaluation. 4 Bold values indicate the best performance, and un- derlined values indicate ...

2024
[48]

As shown in the last two rows of each matrix, the distri- bution differences between the preference aligned model and other methods are statistically signifi- cant everywhere

We perform pairwise comparisons of distributions for all possible pairs of methods and report the resulting p-values. As shown in the last two rows of each matrix, the distri- bution differences between the preference aligned model and other methods are statistically signifi- ...

2025
[49]

Wasserstein distances were computed using n = 20 bootstrap samples, each containing 500 datapoints

Binding efficiency Protein-ligand interactions Method Vina Gnina H-bond H-bond (acc.) H-bond (don.) π-stacking Hydrophobic POCKET2MOL 0.064 (0.005) 0.044 (0.005) 0.040(0.003) 0.026(0.002) 0.014 (0.001) 0.008(0.001)0.028(0.008) DIFFSBDD 0.086 (0.004) 0.043 (0.002) 0.047 (0.002)...

2025
[1994]

Pref- erence optimization for molecular language models

Ryan Park, Ryan Theisen, Rayees Rahman, Anna Cicho´nska, Marcel Patek, and Navriti Sahni. Pref- erence optimization for molecular language models. In NeurIPS 2023 Generative AI and Biology (GenBio) Workshop,

2023
[2014]

Diffdock: Dif- fusion steps, twists, and turns for molecular docking

Gabriele Corso, Hannes St ¨ark, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. Diffdock: Dif- fusion steps, twists, and turns for molecular docking. arXiv preprint arXiv:2210.01776,

arXiv
[2017]

Flowpacker: Protein side-chain packing with torsional flow match- ing

Jin Sub Lee and Philip M Kim. Flowpacker: Protein side-chain packing with torsional flow match- ing. bioRxiv, pp. 2024–07,

2024
[2018]

Fr´echet chemnet distance: A metric for generative models for molecules in drug discovery.arXiv preprint arXiv:1803.09518,

Kristina Preuer, Philipp Renz, Thomas Unterthiner, Sepp Hochreiter, and G ¨unter Klambauer. Fr´echet chemnet distance: A metric for generative models for molecules in drug discovery.arXiv preprint arXiv:1803.09518,

arXiv
[2019]

Multi-domain Distribution Learning for De Novo Drug Design

15 Published as a conference paper at ICLR 2025 Appendix for “Multi-domain Distribution Learning for De Novo Drug Design” TABLE OF CONTENTS A Extended Methods 17 A.1 Generative framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 A.1.1 Flow matching ....

2025
[2020]

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. arXiv preprint arXiv:2407.01648,

arXiv
[2021]

Estimating the mean and variance of the target probability distribution

David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 ieee international conference on neural networks (ICNN’94), volume 1, pp. 55–60. IEEE,

1994
[2022]

Decomposed direct prefer- ence optimization for structure-based drug design

Xiwei Cheng, Xiangxin Zhou, Yuwei Yang, Yu Bao, and Quanquan Gu. Decomposed direct prefer- ence optimization for structure-based drug design. arXiv preprint arXiv:2407.13981,
[2023]

Analog bits: Generating discrete data using diffusion models with self-conditioning

Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202,

arXiv
[2024]

Building normalizing flows with stochastic inter- polants

Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic inter- polants. arXiv preprint arXiv:2209.15571,

arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

1x, y ∈ M and u, v ∈ TxM

w(α) = ((α +π) mod 2 π)−π wraps values within the range [−π,π ) and is applied element-wise. 1x, y ∈ M and u, v ∈ TxM. 17 Published as a conference paper at ICLR 2025 Here, we additionally use a schedulerκ(t), that satisfiesκ(0) = 1 andκ(1) = 0, to control the rate at which the geodesic distanced betweenx0 andx1 decreases (Chen & Lipman, 2023): d(xt,x1) =...

work page 2025

[2] [2]

Riemannian flow matching on general geometries

Ricky TQ Chen and Yaron Lipman. Riemannian flow matching on general geometries. arXiv preprint arXiv:2302.03660,

arXiv

[3] [3]

We compare distributions of the first two principal components in Figure

First, we apply PCA to the molecular embeddings computed by ChemNet, the neural network used to calculate the Fr´echet ChemNet Distance (FCD) (Preuer et al., 2018). We compare distributions of the first two principal components in Figure

work page 2018

[4] [4]

Model Parameter D RUGFLOW DRUGFLOW-OOD F LEX FLOW DRUGFLOW-PA Num

Table 4: Model hyperparameters. Model Parameter D RUGFLOW DRUGFLOW-OOD F LEX FLOW DRUGFLOW-PA Num. weights (M) 12.1 12 .1 12 .6 12 .1 Model size (MB) 48.408 48.411 50.557 48.408 Training epochs 600 600 700 ≥50 Virtual nodesNmax 10 10 10 10 Sampling steps 500 500 500 500 Uncertainty head No Yes No No Flexible side chains No No Yes No OODλ – 10 – – Schedule...

work page 2025

[5] [5]

In each plot, we compare training data (blue areas) and samples generated by different methods (black solid lines)

�� Figure 7: Distributions of the first two principle components of molecule embeddings computed by the FCD neural network. In each plot, we compare training data (blue areas) and samples generated by different methods (black solid lines). Figure 8: Distributions atom types (4 most popular types) and covalent bond types in the training dat...

work page 2025

[6] [6]

PILOT: Equivariant diffusion for pocket conditioned de novo ligand generation with multi-objective guidance via importance sampling

11 Published as a conference paper at ICLR 2025 Julian Cremer, Tuan Le, Frank No ´e, Djork-Arn´e Clevert, and Kristof T Sch ¨utt. Pilot: Equivariant diffusion for pocket conditioned de novo ligand generation with multi-objective guidance via importance sampling. arXiv preprint arXiv:2405.14925,

work page Pith review arXiv 2025

[7] [7]

Mixed continuous and categorical flow matching for 3d de novo molecule generation

Ian Dunn and David Ryan Koes. Mixed continuous and categorical flow matching for 3d de novo molecule generation. arXiv preprint arXiv:2404.19739,

arXiv

[8] [9]

3d equiv- ariant diffusion for target-aware molecule generation and affinity prediction

Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equiv- ariant diffusion for target-aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543, 2023a. Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, and Quanquan Gu. Decompdiff: Diffusion models with dec...

arXiv