Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Rectified Flow For Structure Based Drug Design

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A rectified-flow generative model for structure-based drug design, FlowSBDD, reports state-of-the-art binding affinity (average AutoDock Vina dock score -8.50) and 75.0% diversity on CrossDocked2020 while sampling 100 molecules per pocket…

desk verdict Rectified-flow SBDD with fast sampling and an interesting refine-from-TargetDiff idea, but the bond-loss ablation rests on an equation constant in θ and the noise scale is tuned on the test set. read the letter →

arxiv 2412.01174 v1 pith:Q4YJ2QW4 submitted 2024-12-02 cs.LG cs.CE

Daiheng Zhang , Chengyue Gong , Qiang Liu This is my paper · ORCID
classification cs.LGcs.CE
keywords rectifiedflowstructure-baseddrugdesignligandgeneration3DmoleculebonddistancelossCrossDocked2020AutoDockVina
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

This paper tries to establish that rectified flow, a transport-based generative framework, is a strong alternative to diffusion models for structure-based drug design. It proposes FlowSBDD, which learns a velocity field that maps noisy initial atom states to ligand coordinates and types conditioned on a protein pocket, and it adds a bond-distance loss and flexible choice of initial distribution. On the CrossDocked2020 benchmark the model reports an average AutoDock Vina dock score of -8.50 and 75.0% diversity, surpassing the diffusion baselines DecompDiff (-8.39) and TargetDiff (-7.80) while sampling 100 molecules per pocket in 144 seconds. The importance is practical: if true, it means a simpler ODE-based sampler with a flexible objective can deliver comparable or better binding affinity and diversity at a fraction of the sampling cost of diffusion models.

What carries the argument

The central object is the rectified-flow ODE $dM_t/dt = v(M_t, t)$ with interpolated state $M_t = tM_1 + (1-t)M_0$, trained by matching the velocity field to $M_1 - M_0$. The paper wraps this in a SE(3)-equivariant graph neural network that updates atom embeddings and coordinates alternately, and it adds two extensions: a 'bond loss' that compares Euclidean distances between bonded atom pairs at the data endpoint and at the transported state, and a flexible initial distribution that replaces the standard Gaussian with other random distributions or with molecules generated by another model such as TargetDiff. The bond loss is the mechanism the paper credits for the largest performance jump in the ablation, and the flexible prior is the mechanism for refining existing samples.

What would settle it

Re-run the 'Best w/ B' setting with the printed bond loss exactly as written; since the loss is constant with respect to $\theta$, training with and without it must give identical results. If the reported gains vanish, the stated mechanism is falsified. Alternatively, inspect the training code to see whether the bond loss is applied to the network's output (e.g., $\phi(M_0 + v_\theta(M_t,t))$) rather than to $M_0$, in which case the printed equation would be a typo.

Watch

Extended reading notes

Core claim

The paper's central claim is that replacing the diffusion learning objective with a rectified-flow objective, where a network $v_\theta$ is trained to follow the straight-line interpolation between a noise distribution $M_0$ and the data $M_1$, yields a generative model of ligand molecules that binds to protein pockets at state-of-the-art affinity. The authors further claim that an auxiliary loss on bond Euclidean distances and the ability to start from a non-Gaussian or even a model-generated prior distribution improve the results, with the bond loss raising median QED by 17.6% and average Vina dock by 29.4% in their ablation. FlowSBDD is presented as a more flexible and faster alternative to diffusion-based SBDD, achieving -8.50 average Vina dock and 75.0% diversity on CrossDocked2020 while sampling 100 molecules per pocket in 144 seconds.

Load-bearing premise

The bond loss in Eq. (4) has no dependence on the velocity network's parameters, so the paper's central ablation claim that this loss boosts Vina dock by 29.4% and QED by 17.6% rests on the assumption that the implemented loss is different from the printed one and actually contributes a gradient.

Editorial extensions

If this is right

  • If FlowSBDD's numbers hold, rectified flow offers a faster sampling route than diffusion for SBDD: 144 seconds for 100 molecules per pocket versus 3428 seconds for TargetDiff and 6189 for DecompDiff.
  • The reported -8.50 average Vina dock and 75.0% diversity would put FlowSBDD ahead of current diffusion baselines on the binding-affinity metric while matching or exceeding their diversity.
  • The flexible-prior result suggests the same trained velocity field can serve as a refiner of molecules produced by other samplers, improving their dock scores (e.g., 100tar-Flow improves 100-step TargetDiff from -4.04 to -6.69 average Vina dock).
  • The bond Euclidean distance loss, if correctly implemented, provides a template for injecting task-specific geometric constraints into flow training beyond simple coordinate/type reconstruction.

Reading between the lines

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

  • The stated bond loss in Eq. (4) is independent of the network parameters, so the ablation gains attributed to it cannot come from that printed objective; a corrected loss that depends on $v_\theta$ would be needed to support the paper's interpretation of the ablation.
  • The speed of FlowSBDD could make it practical as the generative component in large-scale virtual screening pipelines, where diffusion samplers are often the bottleneck.
  • The flexible-prior idea points toward a general refinement paradigm: any existing docking or generative sampler's output can be fed as the initial distribution, turning the flow model into a post-processing step that potentially improves affinity without retraining the original model.
  • The strong dependence on noise scale ($N(0,0.001)$ far outperforming $N(0,1)$) suggests that the practical gains may come more from prior scale tuning than from the rectified-flow formulation itself; a fair comparison would require matching this tuning across baselines.
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

4 major / 6 minor

Summary. FlowSBDD applies rectified flow to structure-based drug design, learning an ODE that transports noisy atom coordinates and types to ligand coordinates conditioned on a protein pocket. The paper proposes adding a bond-distance loss and allows flexible prior distributions, including non-Gaussian noise and molecules generated by TargetDiff. On CrossDocked2020 the authors report an average Vina Dock score of -8.50 with 75.0% diversity, claim state-of-the-art performance over DecompDiff and TargetDiff, and report faster sampling (144 seconds per 100 molecules). The central contribution is the proposed bond-distance loss and the flexible-prior experiments, but the main claims rest on an ablation whose loss function, as written, is constant with respect to the model parameters.

Significance. If the method were correctly specified and the results reproducible, a rectified-flow formulation for SBDD with flexible priors and auxiliary losses would be a useful complement to diffusion-based methods, especially given the reported sampling speed advantage. The paper also draws on the marginal-preserving properties of rectified flow, which is a principled basis for generative modeling. However, as submitted, the key empirical claims are not supported: the bond loss of Eq. (4) has no dependence on the network parameters, the bond-loss ablation is confounded with a change in noise scale, and the headline setting is selected on the test set. The state-of-the-art claim therefore rests on an experimental setup that is internally inconsistent, and the contribution reduces to a plausible but unverified application of rectified flow.

major comments (4)
  1. [Section 2.2, Eq. (4)] The loss in Eq. (4), min_θ E_{M0,M1∼D} Σ_i ||b_i(M0) − b_i(M1)||, does not contain the network parameters θ or the velocity field v_θ; its gradient with respect to θ is identically zero. Table 2 therefore cannot support the claim that this bond loss improves Avg. Vina Dock by 29.4% and Med. QED by 17.6%. If the implemented loss is instead the one suggested by Eq. (3), evaluated using predicted coordinates, then Eq. (4) must be corrected and the ablation rerun. As printed, the central 'adaptive enhancement' is vacuous.
  2. [Section 3.3, Tables 2 and 3] The bond-loss ablation is confounded with a change in the noise scale. The caption of Table 2 states that 'Best w/ B' uses a Gaussian scaled by 0.01, but Table 3 reports identical numbers for N(0, 0.001). Either the scale is misreported, or the comparison changes both the noise scale and the loss at the same time; in neither case can the reported Vina Dock improvement be attributed specifically to the bond loss.
  3. [Section 3.3, Table 3] The prior noise scale is selected using the test set. The N(0, 0.001) row is chosen because it 'achieved the best QED and Vina Dock results' on the same 100 test complexes used for the headline benchmark numbers, and no validation split or model-selection procedure is described. The reported -8.50 Avg. Vina Dock is therefore an optimistic selection, and the comparison with baselines is biased in favor of FlowSBDD.
  4. [Section 3.1, Table 1] All benchmark numbers are point estimates with no error bars, standard deviations, or significance tests. The claimed improvement over DecompDiff in Avg. Vina Dock (-8.50 vs -8.39) is small, and without any measure of variability the state-of-the-art claim is not established. This concern is amplified by the test-set noise-scale selection in Table 3.
minor comments (6)
  1. [Section 2.2, Eq. (3)] Eq. (3) writes v(M_t, t) even though the text states that the velocity field is parameterized by θ; use v_θ(M_t, t) consistently.
  2. [Section 2.2, Theorem 1] The informal theorem uses dZ_t without defining Z_t; replace with dM_t or define the new process explicitly.
  3. [Abstract and Conclusion] The abstract claims 'state-of-the-art performance' but the conclusion says 'comparable performance to SOTA'; these statements should be reconciled.
  4. [Section 3.2, Diversity comparison] The statement that diversity 'exceeded Targetdiff by 4.2%' is ambiguous; report the absolute difference (0.03) as well as the relative percentage.
  5. [Section 3.3, Table 2 caption] The caption says 'Best w/ B' uses a Gaussian scaled by 0.01, but Table 3 shows the same numbers for N(0, 0.001); please correct the inconsistency and ensure all reported settings match their descriptions.
  6. [General reproducibility] No code, training hyperparameters, or details of the bond extraction function b_i are provided, which would be needed to reproduce the experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the SOTA benchmark numbers are externally evaluated with AutoDock Vina; the Eq. (4) bond-loss problem is a correctness/omitted-proof issue, not a circular reduction.

full rationale

No circular derivation was found. FlowSBDD's headline numbers (-8.50 Avg. Vina Dock, 75.0% Diversity, Table 1) are computed with AutoDock Vina on the CrossDocked2020 test split and compared with published baselines, so they are externally benchmarked rather than constructed from the model's own assumptions. The rectified-flow backbone is self-cited from the authors' prior works [13,14], but the benchmark comparison does not reduce to those citations. The informal Theorem 1 from [13] is invoked to justify adding an arbitrary bond cost; this theorem is self-cited and not reproduced in the paper, which is a citation-care concern, but it does not force the empirical Vina scores. The more serious problem is Eq. (4): as printed it contains no dependence on the network parameters θ, so the stated bond loss is effectively a no-op and cannot explain the Table 2 gains; this is an omitted or mis-stated proof and should be corrected, but it is an internal inconsistency rather than a circular equivalence. The central empirical claims remain self-contained against an external benchmark, so the appropriate circularity finding is low; the Eq. (4) issue should be handled as a correctness risk, not as circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central results depend on two free choices: the noise scale selected on the test set and the (mis-specified) bond loss. The background axioms are standard for the SBDD benchmark and the authors' own rectified flow theory; no new physical entities are introduced.

free parameters (1)
  • prior noise scale sigma = 0.001 (N(0,0.001))
    Chosen as the best among 6 priors on the test set metrics (Table 3); this is model selection on the benchmark, not a fixed training convention.
assumptions (4)
  • domain assumption Rectified flow can learn a transport map between any two distributions using the squared L2 loss in Eq. (1) (Liu et al. [13,14]).
    Invoked in Section 2.2 without proof; it is prior work by the same research group.
  • ad hoc to paper The bond-distance cost in Eq. (3) is a valid user-specified cost for the rectified flow optimal transport theorem (Theorem 1 in Section 2.2).
    The theorem is cited informally from [13], but the instantiation in Eq. (4) is constant w.r.t. θ, so the assumption that this loss can be optimized is not satisfied.
  • domain assumption AutoDock Vina scoring of generated molecules is a valid proxy for binding affinity in the benchmark.
    Used throughout Section 3 to rank models; the paper relies on the standard practice of prior SBDD papers but does not validate against experimental binding data.
  • domain assumption The CrossDocked2020 filtering (RMSD < 1, sequence identity < 30%) and the 100 test complexes are a representative benchmark.
    Adopted from prior works [6,7,15,17] without independent analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rectified Flow For Structure Based Drug Design." pith.science (2026). https://pith.science/paper/Q4YJ2QW4

@misc{pith2026241201174,
  author       = {Pith},
  title        = {Pith review of: Rectified Flow For Structure Based Drug Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q4YJ2QW4}},
  note         = {Machine review of arXiv:2412.01174}
}
read the original abstract

Deep generative models have achieved tremendous success in structure-based drug design in recent years, especially for generating 3D ligand molecules that bind to specific protein pocket. Notably, diffusion models have transformed ligand generation by providing exceptional quality and creativity. However, traditional diffusion models are restricted by their conventional learning objectives, which limit their broader applicability. In this work, we propose a new framework FlowSBDD, which is based on rectified flow model, allows us to flexibly incorporate additional loss to optimize specific target and introduce additional condition either as an extra input condition or replacing the initial Gaussian distribution. Extensive experiments on CrossDocked2020 show that our approach could achieve state-of-the-art performance on generating high-affinity molecules while maintaining proper molecular properties without specifically designing binding site, with up to -8.50 Avg. Vina Dock score and 75.0% Diversity.

Figures

Figures reproduced from arXiv: 2412.01174 by the authors.

Figure 1
Figure 1. Visualization of FlowSBDD in best setting(left two columns), TargetDiff(middle two [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Median Vina energy for different generated molecules (AR vs. Pocket2mol vs. Targetdiff [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

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

  1. MolFORM: Multi-modal Flow Matching for Structure-Based Drug Design

    cs.CE 2025-07 conditional novelty 5.0 of 10

    A flow-matching model with direct preference optimization fine-tuning generates protein-binding molecules faster than diffusion baselines, with improved docking scores on the CrossDocked2020 benchmark.

Reference graph

Works this paper leans on

21 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    Chemistry & biology 10(9), 787–797 (2003) 1

    Anderson, A.C.: The process of structure-based drug design. Chemistry & biology 10(9), 787–797 (2003) 1

    work page 2003

  2. [2]

    arXiv preprint arXiv:1705.04612 (2017) 1

    Bjerrum, E.J., Threlfall, R.: Molecular generation with recurrent neural networks (rnns). arXiv preprint arXiv:1705.04612 (2017) 1

    arXiv 2017

  3. [3]

    0: New docking methods, expanded force field, and python bindings

    Eberhardt, J., Santos-Martins, D., Tillack, A.F., Forli, S.: Autodock vina 1.2. 0: New docking methods, expanded force field, and python bindings. Journal of chemical information and modeling61(8), 3891–3898 (2021) 3, 4

    work page 2021

  4. [4]

    Journal of chemical information and modeling 60(9), 4200–4215 (2020) 1, 3

    Francoeur, P.G., Masuda, T., Sunseri, J., Jia, A., Iovanisci, R.B., Snyder, I., Koes, D.R.: Three-dimensional convolutional neural networks and a cross-docked data set for structure-based drug design. Journal of chemical information and modeling 60(9), 4200–4215 (2020) 1, 3

    work page 2020

  5. [5]

    Generating equilibrium molecules with deep neural networks

    Gebauer, N.W., Gastegger, M., Schütt, K.T.: Generating equilibrium molecules with deep neural networks. arXiv preprint arXiv:1810.11347 (2018) 1

    work page Pith review arXiv 2018

  6. [6]

    arXiv preprint arXiv:2303.03543 (2023) 1, 2, 3

    Guan, J., Qian, W.W., Peng, X., Su, Y ., Peng, J., Ma, J.: 3d equivariant diffusion for target-aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543 (2023) 1, 2, 3

    arXiv 2023

  7. [7]

    Guan, J., Zhou, X., Yang, Y ., Bao, Y ., Peng, J., Ma, J., Liu, Q., Wang, L., Gu, Q.: Decompdiff: Diffusion models with decomposed priors for structure-based drug design (2023) 1, 3

    work page 2023

  8. [8]

    Neuron 99(6), 1129–1143 (2018) 3

    Hollingsworth, S.A., Dror, R.O.: Molecular dynamics simulation for all. Neuron 99(6), 1129–1143 (2018) 3

    work page 2018

Show all 21 references
  1. [9]

    In: International conference on machine learning

    Hoogeboom, E., Satorras, V .G., Vignac, C., Welling, M.: Equivariant diffusion for molecule generation in 3d. In: International conference on machine learning. pp. 8867–8887. PMLR (2022) 2

    2022

  2. [10]

    In: International conference on machine learning

    Jin, W., Barzilay, R., Jaakkola, T.: Junction tree variational autoencoder for molecular graph generation. In: International conference on machine learning. pp. 2323–2332. PMLR (2018) 1

    2018

  3. [11]

    In: International conference on machine learning

    Kusner, M.J., Paige, B., Hernández-Lobato, J.M.: Grammar variational autoencoder. In: International conference on machine learning. pp. 1945–1954. PMLR (2017) 1

    2017

  4. [12]

    arXiv preprint arXiv:2204.09410 (2022) 1, 3

    Liu, M., Luo, Y ., Uchino, K., Maruhashi, K., Ji, S.: Generating 3d molecules for target protein binding. arXiv preprint arXiv:2204.09410 (2022) 1, 3

    2022 arXiv

  5. [13]

    arXiv preprint arXiv:2209.14577 (2022) 1, 2, 3

    Liu, Q.: Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577 (2022) 1, 2, 3

    2022 arXiv

  6. [14]

    arXiv preprint arXiv:2209.03003 (2022) 1, 2

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022) 1, 2

    2022 arXiv

  7. [15]

    Advances in Neural Information Processing Systems 34, 6229–6239 (2021) 1, 3

    Luo, S., Guan, J., Ma, J., Peng, J.: A 3d generative model for structure-based drug design. Advances in Neural Information Processing Systems 34, 6229–6239 (2021) 1, 3

    2021

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837–2845 (2021) 3

    2021

  9. [17]

    In: International Conference on Machine Learning

    Peng, X., Luo, S., Guan, J., Xie, Q., Peng, J., Ma, J.: Pocket2mol: Efficient molecular sampling based on 3d protein pockets. In: International Conference on Machine Learning. pp. 17644–17655. PMLR (2022) 1, 3

    2022

  10. [18]

    Frontiers in pharmacology 11, 565644 (2020) 1

    Polykovskiy, D., Zhebrak, A., Sanchez-Lengeling, B., Golovanov, S., Tatanov, O., Belyaev, S., Kurbanov, R., Artamonov, A., Aladinskiy, V ., Veselov, M., et al.: Molecular sets (moses): a benchmarking platform for molecular generation models. Frontiers in pharmacology 11, 56564...

    2020

  11. [19]

    Chemical science 13(9), 2701–2713 (2022) 3

    Ragoza, M., Masuda, T., Koes, D.R.: Generating 3d molecules conditional on receptor binding sites with deep generative models. Chemical science 13(9), 2701–2713 (2022) 3

    2022

  12. [20]

    In: International conference on machine learning

    Satorras, V .G., Hoogeboom, E., Welling, M.: E (n) equivariant graph neural networks. In: International conference on machine learning. pp. 9323–9332. PMLR (2021) 3

    2021

  13. [21]

    arXiv preprint arXiv:2011.12747 (2020) 1 6

    Simm, G.N., Pinsler, R., Csányi, G., Hernández-Lobato, J.M.: Symmetry-aware actor-critic for 3d molecular design. arXiv preprint arXiv:2011.12747 (2020) 1 6

    2020 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.