Pith. sign in

REVIEW 4 major objections 5 minor 19 references

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

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MolFORM claims that jointly aligning atom-type and 3D-coordinate preferences with multi-flow matching and Direct Preference Optimization yields larger gains in structure-based drug design than equivalently tuning a diffusion model.

desk verdict A plausible and fast flow-matching SBDD model with an interesting multi-flow DPO idea, but the SOTA claim is contradicted by its own table and the DPO loss derivation is missing. read the letter →

arxiv 2507.05503 v4 pith:7M3K6QDX submitted 2025-07-07 cs.CE

Jie Huang , Daiheng Zhang This is my paper
classification cs.CE
keywords structure-baseddrugdesignflowmatchingdiscretedirectpreferenceoptimizationmulti-modalgenerationequivariantgraphneuralnetworksbindingaffinityCrossDocked2020
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 proposes MolFORM, a non-autoregressive generative model that uses multi-flow matching to jointly generate atom types (discrete) and 3D coordinates (continuous) for molecules binding to a protein pocket. On the CrossDocked2020 benchmark it matches diffusion baselines such as TargetDiff before fine-tuning, and after a preference-guided fine-tuning step driven by the Vina docking score it reports state-of-the-art results across most metrics. The central claim is that the DPO step works best when applied to both flows at once, because the model can refine atom types and geometry separately. If correct, this shows flow-matching frameworks are not only competitive with diffusion for structure-based drug design, but better positioned for preference alignment.

What carries the argument

The central machinery is multi-flow matching with a Chamfer-augmented reparameterized loss, plus a multi-flow DPO loss. Flow matching learns a velocity field transporting a noise prior to the data distribution; here the continuous branch uses $L^2$ regression to the clean coordinates and the discrete branch uses cross-entropy to the clean atom type, with time sampled from a Beta-heavy schedule that concentrates on the final refinement phase. The DPO extension replaces the log-probability ratio with differences between the current model's and a frozen reference model's denoising errors, and the discrete loss under uniform noising reduces to a log-ratio of denoising distributions scaled by $\frac{1}{1-t}$ (Eq. 21). This two-branch structure is what allows preference signals to update atom types and coordinates separately.

What would settle it

Compute, on a held-out set of preference pairs, the true DPO objective $\log \frac{p_\theta(x^w_1|p)}{p_{\mathrm{ref}}(x^w_1|p)} - \log \frac{p_\theta(x^l_1|p)}{p_{\mathrm{ref}}(x^l_1|p)}$ and compare its sign with the denoising-error difference used in Eqs. 15-17; if the two disagree in sign for a substantial fraction of pairs, the DPO loss is not optimizing the preference it claims.

Watch

Extended reading notes

Core claim

MolFORM treats ligand generation as two coupled flow-matching processes: a discrete flow over atom types with uniform corruption and a continuous flow over 3D coordinates, trained with reparameterized losses that predict the clean target and a Chamfer distance that preserves geometric fidelity. The authors then derive a DPO objective for the multi-flow setting in which the standard log-likelihood ratio is replaced by differences in denoising errors: $L^2$ distance for coordinates, Chamfer distance for point clouds, and a rate-matrix expression for the discrete flow under uniform noising (Eq. 21). Fine-tuning with Vina score as the reward improves Vina Min from $-6.42$ to $-7.18$ and Vina Dock from $-7.50$ to $-8.13$ while keeping diversity at $0.77$, and the relative gains are larger than those achieved by applying DPO to TargetDiff (Table 4). The paper interprets this as evidence that factorizing discrete and continuous modalities lets DPO assign fine-grained preference signals and reduces gradient interference.

Load-bearing premise

The fine-tuning stage assumes that replacing the DPO log-likelihood ratio with differences of denoising errors ($L^2$ distance, Chamfer, and rate-matrix terms) yields a valid preference objective, and the discrete branch's final loss is stated without the intermediate algebra from Eqs. 18-20, so this equivalence carries the whole DPO claim.

Editorial extensions

If this is right

  • Flow-matching pipelines can match diffusion-based SBDD models on binding affinity and chemical property metrics while sampling 100 steps instead of 1000, taking about 69s for 100 molecules versus 3428s for TargetDiff.
  • Applying DPO to MolFORM yields larger relative improvements (QED +4%, SA +8%, Vina +14%, Vina Min +12%) than applying DPO to TargetDiff (2%, 2%, 2%, 3%), indicating that modality-factorized preference alignment extracts more from the same reward signal.
  • The Chamfer distance is load-bearing: ablating it from the DPO objective drops QED from 0.48 to 0.41 and Vina Min from -6.42 to -5.76 (Table 5).
  • DPO fine-tuning improves binding affinity while preserving diversity at 0.77, suggesting preference optimization here does not collapse the generated distribution.

Reading between the lines

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

  • The substitution of denoising errors for log-likelihood ratios is an assumption rather than a theorem; if it holds, the same recipe could extend DPO to other flow-matching models with structured outputs such as protein backbones or crystal structures, and if it fails, the reported gains may be coming from a different mechanism like implicit regularization toward the reference model.
  • Because the discrete DPO loss is derived under a specific uniform noising scheme, transferring the framework to masked or other continuous-time discrete flows would require re-deriving the rate-matrix ratio, so the approach may not carry over unchanged.
  • A testable extension is to run the same multi-flow DPO objective with a different reward, such as synthesizability or a multi-objective mix, to see whether the co-modeling benefit is tied to the Vina signal or generalizes across rewards.
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 / 5 minor

Summary. MolFORM is a flow-matching framework for structure-based drug design that jointly generates discrete atom types and continuous 3D coordinates by combining continuous conditional flow matching, discrete flow matching with uniform corruption, and a Chamfer-distance regularization term. A Direct Preference Optimization (DPO) fine-tuning stage is then applied, using Vina score as a reward, with a proposed multi-flow DPO co-modeling loss that aligns both modalities. On CrossDocked2020, the base model is competitive with TargetDiff, and DPO fine-tuning improves several metrics, leading the authors to claim state-of-the-art performance across nearly all metrics. The source code is publicly available.

Significance. If the multi-flow DPO approach is sound, this would be a useful and non-obvious contribution: it extends preference alignment beyond unimodal diffusion models to settings with discrete and continuous modalities, and it provides a flow-matching baseline for SBDD. The public code and the empirical comparison against TargetDiff-DPO are also potentially valuable. However, the central DPO derivation is incomplete, and the state-of-the-art claim is not supported by the paper's own tables, so the significance of the result is currently limited.

major comments (4)
  1. [§3.4, Eqs. (15)–(16)] The continuous DPO loss replaces the log-likelihood ratio log(pθ/pref) in Eq. (14) with a difference of instantaneous denoising errors (L2 and Chamfer distances) at a single time t, without any derivation. For flow matching, the marginal log-likelihood is given by the instantaneous change-of-variables formula log p0(x0) − ∫₀¹ div vθ(xt,t) dt, and the flow-matching loss is at best an upper bound on the negative log-likelihood only after integration over time with a specific weighting. A single-time squared error or a Chamfer distance has no established monotonic relationship to the log-density ratio. Because the central claim of the paper is that multi-flow DPO co-modeling yields superior alignment, this missing justification is load-bearing; the improvements in Table 4 could arise from additional training, the Chamfer regularization, or reward overfitting rather than from DPO itself.
  2. [§3.4, Eq. (21)] The discrete DPO loss in Eq. (21) does not follow from the preceding equations. Substituting the uniform-rate expression (Eq. (20)) into Eq. (18) yields, in addition to the log-ratio term, extra terms proportional to (p_ref − p_θ)/(1 − t) (from the R_ref − R_θ part of Eq. (18)). These terms are dropped in Eq. (21) without comment. Either provide the full algebra or state explicitly that these terms are neglected and justify why. Without this, the discrete DPO loss is unsubstantiated.
  3. [§4.2, Table 1] The claim in Section 4.2 that MolFORM-DPO achieves “state-of-the-art performance across nearly all metrics” is contradicted by the table's own entries: VoxBind has QED 0.54 vs. 0.50 and SA 0.54 vs. 0.51 for MolFORM-DPO, and MolCraft has substantially better geometric fidelity (JSDBL 0.2250 vs. 0.3215; JSDBA 0.2683 vs. 0.5584). MolFORM-DPO also has lower diversity (0.77) than the base MolFORM (0.78). The SOTA claim should be narrowed to the metrics where MolFORM-DPO is actually best (e.g., Vina Min and Vina Dock) or the presentation of Table 1 and the claim should be reconciled.
  4. [§4.2, Evaluation] All reported metrics are single point estimates without error bars, standard deviations, or multiple random seeds. Since the best and second-best values are often very close (e.g., Vina Score −6.16 for both MolFORM-DPO and VoxBind, and Vina Min −7.18 vs. −7.14 for LiGAN), it is unclear whether these differences are statistically meaningful. Report mean ± std over several sampling runs or multiple training seeds for the proposed method and, if feasible, for baselines.
minor comments (5)
  1. [§3.4] The text says the DPO loss is decomposed into “three parts 15, 16 and 17 as below, aligning with Equation (9),” but Eq. (15) and Eq. (16) are both continuous losses (L2 and Chamfer), while Eq. (17) is discrete. The alignment with L_pos, L_type, and L_Chamfer in Eq. (9) is not immediately clear and should be clarified.
  2. [§4.2, Table 5] The text introducing Table 5 says “We conducted ablation studies on the Chamfer DPO loss component,” but the table caption and rows describe “with and without Chamfer loss in Vanilla Multi-model flow matching.” Clarify whether this ablation is on the base model loss (Eq. (9)) or on the DPO loss (Eq. (16)).
  3. [§3.4, Eq. (14)] Eq. (14) writes the DPO loss as an expectation over t ∼ U[0,1] and the conditional denoising distribution pθ_{1|t}(x1|xt), but the standard DPO loss is defined over full samples x1. The connection between this conditional form and the full-sample likelihood ratio is not explained.
  4. [§4.1, DPO dataset] The DPO dataset description says “we sample two candidate ligands” but does not specify which model generates these candidates. State whether they are generated by the base MolFORM model, by a pretrained baseline, or by some other procedure, as this affects the interpretation of the preference data.
  5. [§2.2 and §6] The paper cites Zhang et al. (2024), a closely related flow-matching approach for SBDD, but does not compare against it empirically or discuss the relationship in the experiments. A brief comparison or explanation would help position the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the base flow-matching objectives and DPO fine-tuning losses are proposed training objectives, not fitted predictions, and the lone self-citation is not load-bearing.

full rationale

MolFORM's derivation chain is self-contained. The base model is trained with standard conditional flow matching losses (Eqs. 1-7) plus an explicit Chamfer term (Eq. 8), and the final training loss (Eq. 9) is a direct sum of those terms; no target evaluation metric is used to define these losses. The sampling procedures (Eqs. 10-13) follow from the learned vector field and discrete rate matrix. The DPO losses (Eqs. 15-21) are proposed objectives that use the model's own denoising predictions (and the reference model's predictions) as preference-alignment signals; they are not defined in terms of the reported metrics. Using the Vina score to construct preference pairs and then reporting Vina-based metrics is reward-guided fine-tuning, not a fitted parameter being renamed as a prediction, and therefore does not make the evaluation circular. The only self-citation (Zhang et al., 2024) appears in a survey list of prior diffusion models and is not load-bearing for any of the paper's claims. The main formal weakness is that Eqs. 15-16 replace the DPO log-likelihood ratio with single-time denoising errors and Eq. 21 omits a self-prediction term from the algebra of Eqs. 18-20; this is an unproven approximation and a correctness risk, but it is not a circular reduction because the proxy is not defined in terms of the target and the empirical gains could in principle fail. Under the stated criteria, no circular step can be exhibited.

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

The central claim depends on a handful of hyperparameters (Chamfer weight, DPO beta, time schedules) and on two structural assumptions: the MSE/Chamfer proxy for likelihood in DPO, and the uniform corruption scheme for discrete flow. No new physical entities are introduced. Most machinery is borrowed from cited prior work.

free parameters (4)
  • lambda (Chamfer loss weight) = not reported
    Balances the Chamfer loss in Eq. 9; chosen by the authors but value is omitted, so the final objective is not fully specified.
  • beta (DPO scaling) = 5
    Regularization coefficient in the DPO loss (Eq. 14 and training details section); controls deviation from the reference model.
  • time schedule weights = 0.02, 0.98, 1.9, 1.0
    Parameters of the training time-sampling distribution in Eq. 22, adopted from Geffner et al. (2025) and not tuned by this paper, but they shape the learned model.
  • sampling step allocation = 60 steps in [0,0.8], 40 in [0.8,1]
    Non-uniform discretization for Euler integration, chosen by hand to improve near-t=1 refinement.
assumptions (5)
  • standard math Conditional flow matching objective (Eq. 1) is a valid training target for generative modeling
    Invoked in Section 3.2; relies on Lipman et al. (2022) and Liu et al. (2022b).
  • standard math Uniform corruption process for discrete flow matching (Eq. 2) and its rate matrix formula (Eq. 20)
    Taken from Campbell et al. (2024); the DPO loss for discrete flow depends on this specific corruption and the resulting posterior in Eq. 13.
  • ad hoc to paper Denoising loss (MSE/Chamfer) is a monotonic proxy for the log-likelihood ratio in DPO
    Used without proof in Eqs. 15-16 to replace log(p_theta/p_ref) with differences of denoising errors; this is the load-bearing approximation for the DPO fine-tuning claim.
  • domain assumption Vina score is a valid reward and evaluation signal for binding affinity
    Vina is the only reward for DPO and a primary evaluation metric; the paper follows standard SBDD practice but this assumption is not interrogated.
  • domain assumption SE(3)-equivariant GNN architecture from Guan et al. (2023) provides sufficient expressivity
    Architecture is reused without modification or ablation in Section 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MolFORM: Multi-modal Flow Matching for Structure-Based Drug Design." pith.science (2026). https://pith.science/paper/7M3K6QDX

@misc{pith2026250705503,
  author       = {Pith},
  title        = {Pith review of: MolFORM: Multi-modal Flow Matching for Structure-Based Drug Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7M3K6QDX}},
  note         = {Machine review of arXiv:2507.05503}
}
read the original abstract

Structure-based drug design (SBDD) seeks to generate molecules that bind effectively to protein targets by leveraging their 3D structural information. While diffusion-based generative models have become the predominant approach for SBDD, alternative non-autoregressive frameworks remain relatively underexplored. In this work, we introduce MolFORM, a novel generative framework that jointly models discrete (atom types) and continuous (3D coordinates) molecular modalities using multi-flow matching. To further enhance generation quality, we incorporate a preference-guided fine-tuning stage based on Direct Preference Optimization (DPO), using Vina score as a reward signal. We propose a multi-modal flow DPO co-modeling strategy that simultaneously aligns discrete and continuous modalities, leading to consistent improvements across multiple evaluation metrics. The source code for MolFORM is publicly available at https://github.com/huang3170/MolForm.git.

Figures

Figures reproduced from arXiv: 2507.05503 by the authors.

Figure 1
Figure 1. Overview of MolFORM. This workflow can be summarized as two steps: 1) Employs multi-flow generation to construct the base model. 2) Applies DPO to fine-tune the dual modalities, using the Vina score as the reward. In our structure-based drug design (SBDD) task, we model each atom type as a categorical variable, where x ∈ K and K = {1, . . . , K} is the set of possible atom types with cardinality K. To model the gene… view at source ↗
Figure 2
Figure 2. Visualizations of reference molecules and generated ligands for protein pockets (4yhj) generated by Reference, Targetdiff, MolFORM and MolFORM-DPO. Vina score, QED, and SA are reported below. Chamfer loss To promote accurate geometric alignment between predicted and ground-truth molecular structures, we incorporate a Chamfer loss defined over atomic point clouds. Given two point sets xˆ1 = {xˆi} N i=1 and x1 = {xj}M… view at source ↗
Figure 3
Figure 3. Comparing the distribution for distances of allatom (top row) and carbon-carbon pairs (bottom row) for reference molecules in the test set (gray) and model generated molecules (color). Baselines. We select all baseline models from CBG￾bench (Lin et al., 2024) for comparison with our method. Early structure-based drug design (SBDD) methods are built on voxel grids with deep neural networks, such as LiGAN (Ragoza et a… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Median RMSD for rigid fragment before and after the force-field optimization. 5. Future Work Our model builds upon Guan et al. (2023) and achieves comparable or even superior performance across multiple evaluation metrics, indicating that the proposed MultiFlow￾based m…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 1 canonical work pages

  1. [3]

    Geffner, T., Didi, K., Zhang, Z., Reidenbach, D., Cao, Z., Yim, J., Geiger, M., Dallago, C., Kucukbenli, E., Vahdat, A., et al

    URL https:// diffusionflow.github.io/. Geffner, T., Didi, K., Zhang, Z., Reidenbach, D., Cao, Z., Yim, J., Geiger, M., Dallago, C., Kucukbenli, E., Vahdat, A., et al. Proteina: Scaling flow-based protein structure generative models.arXiv preprint arXiv:2503.00710,

    arXiv

  2. [4]

    W., Peng, X., Su, Y ., Peng, J., and Ma, J

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

    arXiv

  3. [6]

    Alphafold meets flow matching for generating protein ensembles.arXiv preprint arXiv:2402.04845,

    Jing, B., Berger, B., and Jaakkola, T. Alphafold meets flow matching for generating protein ensembles.arXiv preprint arXiv:2402.04845,

    arXiv

  4. [7]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    arXiv

  5. [8]

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

    arXiv

  6. [9]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y ., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    arXiv

  7. [11]

    L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow in- structions with human feedback, 2022.URL https://arxiv. org/abs/2203.02155, 13,

    arXiv 2022

  8. [12]

    O., Jamasb, A., Mahmood, O., Sresht, V ., and Saremi, S

    Pinheiro, P. O., Jamasb, A., Mahmood, O., Sresht, V ., and Saremi, S. Structure-based drug design by denoising voxel grids.arXiv preprint arXiv:2405.03961,

    arXiv

Show all 19 references
  1. [13]

    Molcraft: Structure-based drug design in continuous parameter space.arXiv preprint arXiv:2404.12141,

    Qu, Y ., Qiu, K., Song, Y ., Gong, J., Han, J., Zheng, M., Zhou, H., and Ma, W.-Y . Molcraft: Structure-based drug design in continuous parameter space.arXiv preprint arXiv:2404.12141,

    arXiv

  2. [14]

    D., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C. D., and Finn, C. Direct preference optimization: Your language model is secretly a reward model.arXiv preprint arXiv:2305.18290,

    arXiv

  3. [16]

    Diffusion model alignment using direct preference optimization.arXiv preprint arXiv:2311.12908,

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Pu- rushwalkam, S., Ermon, S., Xiong, C., Joty, S., and Naik, N. Diffusion model alignment using direct preference optimization.arXiv preprint arXiv:2311.12908,

    arXiv

  4. [17]

    Rectified flow for structure based drug design.arXiv preprint arXiv:2412.01174,

    Zhang, D., Gong, C., and Liu, Q. Rectified flow for structure based drug design.arXiv preprint arXiv:2412.01174,

    arXiv

  5. [18]

    Antigen-specific antibody design via direct energy-based preference optimization.arXiv preprint arXiv:2403.16576,

    Zhou, X., Xue, D., Chen, R., Zheng, Z., Wang, L., and Gu, Q. Antigen-specific antibody design via direct energy-based preference optimization.arXiv preprint arXiv:2403.16576,

    arXiv

  6. [19]

    M., Stiennon, N., Wu, J., Brown, T

    Ziegler, D. M., Stiennon, N., Wu, J., Brown, T. B., Radford, A., Amodei, D., Christiano, P., and Irving, G. Fine-tuning language models from human preferences.arXiv preprint arXiv:1909.08593,

    1909 arXiv

  7. [2003]

    Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design.arXiv preprint arXiv:2402.04997,

    Campbell, A., Yim, J., Barzilay, R., Rainforth, T., and Jaakkola, T. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design.arXiv preprint arXiv:2402.04997,

    arXiv

  8. [2019]

    L., Tseng, A

    Uehara, M., Zhao, Y ., Black, K., Hajiramezanali, E., Scalia, G., Diamant, N. L., Tseng, A. M., Levine, S., and Bian- calani, T. Feedback efficient online fine-tuning of diffu- sion models.arXiv preprint arXiv:2402.16359,

    arXiv

  9. [2022]

    Generating 3d molecules for target protein binding.arXiv preprint arXiv:2204.09410, 2022a

    Liu, M., Luo, Y ., Uchino, K., Maruhashi, K., and Ji, S. Generating 3d molecules for target protein binding.arXiv preprint arXiv:2204.09410, 2022a. Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint...

    arXiv

  10. [2023]

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

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

    arXiv

  11. [2024]

    Decom- posed direct preference optimization for structure-based drug design.arXiv preprint arXiv:2407.13981,

    Cheng, X., Zhou, X., Yang, Y ., Bao, Y ., and Gu, Q. Decom- posed direct preference optimization for structure-based drug design.arXiv preprint arXiv:2407.13981,

Pith tools

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