MolPIF: A Parameter Interpolation Flow Model for Molecule Generation

Chuanlong Zeng; Dan Teng; Duanhua Cao; Duo An; Jiacheng Xiong; Junjie Wang; Mingyue Zheng; Qian Shi; Shuangjia Zheng; Wenkai Xiang

REVIEW 4 major objections 6 minor 2 cited by

MolPIF: A Parameter Interpolation Flow Model for Molecule Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MolPIF claims that a generative flow can unify continuous coordinates and discrete atom types by interpolating distribution parameters instead of samples, and reports stronger structure-based drug design results on CrossDocked2020 than…

desk verdict Useful incremental SBDD model with released code and a solid empirical section, but the theoretical claims as written are contradicted by the deployed schedule and the discrete loss looks mis-derived. read the letter →

arxiv 2507.13762 v4 pith:LDSARH7D submitted 2025-07-18 cs.LG q-bio.BM

Yaowei Jin , Junjie Wang , Yufan Tang , Wenkai Xiang , Duanhua Cao , Dan Teng , Zhehuan Fan , Jiacheng Xiong

show 6 more authors

Xia Sheng Chuanlong Zeng Duo An Mingyue Zheng Shuangjia Zheng Qian Shi

This is my paper

classification cs.LGq-bio.BM

keywords structure-baseddrugdesignmoleculargenerationparameterinterpolationflowmatchingDirichletdistributionFisher-Raometricleadoptimization

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

MolPIF proposes that the right place to interpolate between noise and data in a generative flow is not sample space but the parameter space of the probability distributions describing molecular features. The paper argues that this shift unifies continuous atomic coordinates, modeled as Gaussian distributions, and discrete atom types, modeled as Dirichlet distributions, in a single iterative refinement loop: the continuous branch theoretically recovers Wasserstein-2 optimal transport, and the discrete branch follows Fisher-Rao geodesics. On the CrossDocked2020 benchmark the resulting model reports better binding affinity, drug-likeness, geometric fidelity, and chemical space coverage than prior structure-based drug design generators, and it extends to lead optimization by fixing atoms while generating the rest. A reader should care because the parameter-space objective is meant to fix a known mismatch: treating categorical atom types with Euclidean metrics distorts the probability simplex, and MolPIF claims a principled replacement.

What carries the argument

The central object is the Parameter Interpolation Flow (PIF), a generative framework that replaces sample-space interpolation with interpolation of the parameters of the distributions that describe the data. For a data point represented as a Dirac distribution, PIF writes the data parameters $\theta_{\mathrm{data}}$ (a Gaussian mean and zero variance for coordinates, a one-hot Dirichlet concentration for atom type) and a prior $\theta_{\mathrm{prior}}$, then defines $\theta_t = f(t)\theta_{\mathrm{data}} + (1-f(t))\theta_{\mathrm{prior}}$ with $f(t)=1-\gamma t$, trains a network to predict the data parameters from the blended parameters at nearby time steps using KL divergence, and samples by iteratively refining the parameters from prior to data. This object does the work of unifying the two modalities: the same interpolation rule applies to Gaussian and Dirichlet parameters, the continuous branch inherits a $W_2$-optimal transport interpretation, and the discrete branch lives on the statistical manifold where Fisher-Rao geodesics are the natural straight lines. A geometry-enhanced training strategy that randomly masks ligand atoms and reconstructs them is the second mechanism; it forces the network to act as a structural imputer and is reported to improve fidelity, with effects that depend on the prior.

What would settle it

Evaluate Eq. (11) numerically against the textbook KL divergence of two Dirichlet distributions for fixed parameter vectors, e.g. predicted parameters (2,2,1) and interpolated parameters (1,2,3); if the two numbers differ, the discrete branch trains on a loss that is not the claimed KL, and the Fisher-Rao-geodesic interpretation of the stated objective collapses.

Watch

Extended reading notes

Core claim

MolPIF's central claim is that a generative model for molecules should learn a time-dependent path through the parameters of the data distributions, not through the samples themselves. For each ligand atom, the coordinates are represented by a Gaussian with mean at the atom position and a vanishing standard deviation, and the atom type by a Dirichlet concentration vector set to a one-hot vector; at training time these parameters are linearly blended with a prior according to a schedule $f(t)=1-\gamma t$, and the network predicts the data-side parameters from the blended state, with the loss measured as the KL divergence between the predicted and the true interpolated distribution. The paper argues that this parameter-space interpolation recovers the Wasserstein-2 optimal transport map for continuous coordinates and, because the Dirichlet is an exponential-family distribution, produces probability paths that follow Fisher-Rao (exponential) geodesics for discrete atom types. On the CrossDocked2020 test set, with 100 molecules sampled per protein, MolPIF reports the best mean Vina Score (-6.64 kcal/mol), Vina Min (-7.41), and Vina Dock (-8.09) among compared models, the highest QED (0.59), low strain energy percentiles, and close-to-reference distributions of ring sizes, bond lengths, and bond angles; on an out-of-distribution PoseBusters subset it reports generated molecules that beat the reference molecules in binding affinity, drug-likeness, and synthetic accessibility. The same conditioning mechanism used to fix known atoms also supports lead optimization, with 25.58% of scaffold-hopping outputs and 63.33% of fragment-growth outputs improving on the reference Vina score.

Load-bearing premise

The load-bearing premise is that Eq. (11) really equals the KL divergence between the predicted and interpolated Dirichlet distributions; as printed it contains only one of the two special-function terms the true divergence needs and no derivation is given, so if it is not exactly that divergence, the discrete branch is not optimizing the Fisher-Rao-consistent objective the paper claims for atom types.

Editorial extensions

If this is right

A single flow mechanism, with one interpolation rule, covers both continuous coordinates and discrete atom types; separate denoising branches with mismatched metrics are not required.
The continuous branch inherits a $W_2$-optimal transport interpretation, so MolPIF's coordinate paths are the shortest Wasserstein geodesics between prior and data for the Gaussian case, and the discrete branch follows Fisher-Rao geodesics rather than Euclidean lines across the simplex.
On the CrossDocked2020 test set, MolPIF reports the best mean Vina Score, Vina Min, and Vina Dock among compared models, the highest QED, and close-to-reference distributions of ring sizes, bond lengths, and bond angles.
On a PoseBusters out-of-distribution subset, MolPIF-generated molecules are reported to surpass the reference molecules in binding affinity, drug-likeness, and synthetic accessibility, which the paper says no baseline achieved in that setting.
Conditional generation with fixed atoms enables lead optimization across scaffold hopping, fragment growth, and side-chain modification, with reported improvements over reference Vina scores in 25.58% of scaffold-hopping outputs and 63.33% of fragment-growth outputs.

Reading between the lines

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

If the parameter-space interpolation recipe works for Gaussians and Dirichlets, the same construction should transfer to other exponential-family parameters, such as Poisson rates or categorical logits, giving a single training objective for mixed-type tabular or graph generation without modality-specific losses; the paper does not test this transfer.
The reported interaction between the Laplace prior and the masking strategy suggests a testable prediction not pursued in the paper: heavy-tailed coordinate priors should systematically improve reconstruction of masked substructures, while light-tailed priors should over-smooth them.
Because inference is iterative parameter refinement with conditioning implemented by concatenating fixed-atom parameters, MolPIF's lead-optimization mechanism may extend naturally to other partial-observation tasks, such as completing a ligand from a docked fragment with the protein context held fixed, where the same interpolation path applies.

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.

Desk Editor's Note

read the letter

MolPIF is a plausible incremental step in structure-based drug design, with released code, weights, and a solid empirical section, but the theoretical claims in the abstract and Section 2.2 are not supported by the equations as written. The schedule f(t)=1−γt gives f(0)=1 and f(1)=0.991, so the interpolation starts at the data and never approaches the prior, directly contradicting the stated boundary conditions θ0=θprior, θ1=θdata. That means the W2-optimal transport and Fisher–Rao geodesic statements are properties of a path the paper does not actually use, and the same f(t) enters the losses, so the training objective doesn't match the claimed geometry. This is a load-bearing flaw in the theory, not a cosmetic typo.

What's genuinely new: the masking strategy for atomic imputation, the use of a convex schedule to emphasize fine-grained structure, the Laplace prior comparison, and the PoseBusters OOD evaluation. The ablation against standard flow matching under the same architecture is a nice control, and the authors are transparent about limitations like rigid pockets and lack of attribute control.

The soft spots beyond the boundary issue: Eq (11), the discrete KL loss, appears mis-derived. The expression uses ψ(θ̂v,t)−ψ(1) where a KL between two Dirichlets would require digamma differences at both parameter vectors, and no derivation is given. The empirical gains over MolCRAFT are small (Vina Score −6.64 vs −6.55) with no error bars or significance tests, and hyperparameters like γ, ε0, and mask probabilities were tuned on the benchmark. That said, the results are not faked; they're just presented without uncertainty.

Who is this for? Someone building SBDD models will find the implementation details useful, and the PoseBusters generalization check is a good practice to copy. The theoretical framing should be ignored until corrected. A serious referee should engage because the released code and benchmark results deserve scrutiny, but the authors need to fix the schedule, derive the discrete loss properly, and restate the optimality claims for the actual path.

Yes, send it to peer review.

Referee Report

4 major / 6 minor

Summary. The paper proposes MolPIF, a structure-based drug design generative model built on a Parameter Interpolation Flow (PIF) framework. Instead of interpolating between samples, MolPIF interpolates the parameters of the distributions that describe atomic coordinates (Gaussian) and atom types (Dirichlet), and trains a conditional model to predict the data parameters from a time-dependent interpolation via KL-divergence losses. The authors claim that PIF theoretically recovers W2-optimal transport for continuous coordinates and Fisher-Rao geodesics for discrete atom types, and they report favorable empirical results on CrossDocked2020 and PoseBusters, including binding affinity, chemical validity, geometric fidelity, chemical space coverage, lead optimization, and flexible prior selection.

Significance. If the theoretical and empirical claims are correct, the paper offers a genuinely unified treatment of continuous and discrete molecular variables in one flow-based framework, which would be a useful contribution to the SBDD and generative-modeling communities. The authors provide code, model weights, and extensive evaluations, including an out-of-distribution PoseBusters benchmark and lead-optimization case studies. However, two load-bearing technical issues—the boundary inconsistency of the deployed schedule and the incorrect-looking Dirichlet KL loss—must be resolved, and the empirical claims need uncertainty quantification before the results can be taken at face value.

major comments (4)

[§2.2, Eq. (9)] The definition θ_t = f(t)θ(xdata) + (1−f(t))θprior is stated to have boundary conditions θ0 = θprior and θ1 = θ(xdata), but the deployed schedule f(t) = 1 − γt with γ = 0.009 gives f(0) = 1 and f(1) = 0.991. Thus θ0 = θ(xdata) and θ1 ≈ 0.991·θ(xdata) + 0.009·θprior, so the trajectory never starts from the prior. This contradicts the sampling process described in §2.2, and it means the W2 and Fisher-Rao theoretical claims, which are derived for a path connecting prior to data, are not realized by the equations as written. For the Gaussian case, the W2 geodesic corresponds to f(t) = t. The ablation in Table 3 explicitly selects the endpoint-violating schedule over f(t) = t, so the paper must either correct the schedule or restate the theoretical results for the actual path.
[§2.3, Eq. (11)] The proposed discrete loss is not the KL divergence between Dirichlet distributions. The KL between Dir(α) and Dir(β) is ln[Γ(α0)/Γ(β0)] + Σ(α_i−β_i)(ψ(α_i)−ψ(α0)) with α0 = Σα_i. Equation (11) contains (θ̂_v,t−θ_v,t)^T (ψ(θ̂_v,t)−ψ(1)), which has the wrong sign for a KL, uses ψ(1) instead of ψ of the sum of concentration parameters, and is given without derivation. If the model is trained with this expression, the discrete branch is not optimizing the claimed Fisher-Rao-consistent objective. The authors must derive the loss from the stated model or explicitly identify the actual objective being optimized.
[§3.2.1, Table 1] The reported improvements over MolCRAFT are small (e.g., Vina Score −6.64 vs −6.55, a difference of 0.09 kcal/mol) and no error bars, confidence intervals, or multiple-seed results are reported. Since the abstract makes a strong claim of outperforming baselines across binding affinity and other metrics, the analysis needs uncertainty quantification, such as means and standard deviations over at least three training runs or bootstrap confidence intervals on the test set.
[§3.2.6, Table 3] The schedule decay γ is selected on the same benchmark test set used for the final comparison (the text states that γ = 0.009 is the 'optimal setting' based on this ablation). This makes the reported performance a fitted optimum rather than a predictive result and biases the comparison against baselines that did not receive the same test-set tuning. Please clarify whether a validation split was used and, if not, provide a separate model-selection procedure or acknowledge the tuning in the interpretation.

minor comments (6)

[§2.3, Eq. (10)] The notation θ̂(1)_x and θ(1)_x is not defined in the text; please clarify that these denote the predicted and true data parameters, and explain how the scalar θ(1)_x is derived from the mean and standard deviation.
[§2.2] The sentence 'PIF characterizes the W2-optimal transport structure between any two one-dimensional distributions of the same type' is too broad and unsupported; either provide a proof or limit the claim to the Gaussian case for which the construction is standard.
[Fig. 1] Figure 1 is difficult to read in the manuscript; a higher-resolution version with legible labels would improve the presentation.
[§2.3, Eq. (10)] The loss in Eq. (10) has a denominator 2γtε0^2 that is singular at t = 0; please discuss how this is handled in practice, for example by clipping t or excluding t = 0.
[Throughout] The spelling 'Fisher–Rao' and 'Fisher-Rao' is inconsistent; please unify the orthography.
[§3.2.4] The explanation for why the Laplace prior interacts well with the masking mechanism is framed as an established phenomenon, but the evidence presented is indirect; consider presenting it as a hypothesis that motivated the experiments rather than as a conclusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the PIF objective and the main empirical claims are self-contained; the noted schedule inconsistency and test-set hyperparameter selection are correctness/selection concerns, not reductions of predictions to fitted inputs.

full rationale

MolPIF's central derivation is the parameter-space interpolation θ_t = f(t)θ_data + (1−f(t))θ_prior with Gaussian and Dirichlet likelihoods and closed-form KL losses. Equations (10) and (11) follow from that model definition; no fitted parameter is renamed as a prediction. The W2 and Fisher–Rao statements are mathematical properties claimed for the interpolation, not empirical predictions derived from the fitted model. The empirical comparison is anchored by a separate PoseBusters OOD evaluation (Section 3.2.5) with sequence-identity filtering, so the main outperformance claim has an external check that does not reduce to the training objective. The self-citations (UniTransformer, MolCRAFT, PILOT filtering) are architectural or baseline references and are not used to justify the theoretical claims. Two concerns are worth noting without being circularity: Eq (9) with γ=0.009 violates the stated boundary conditions θ0=θprior, θ1=θdata, so the theoretical W2/Fisher–Rao path is not realized by the equations as written; and Table 3 selects the schedule and γ on the same CrossDocked test set, which is test-set selection bias rather than a statistical forcing of the prediction. Score 2 reflects minor self-citation and these non-circular caveats.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the Gaussian/Dirichlet parameterization, on the assumption that parameter interpolation yields a valid and theoretically meaningful path, and on the correctness of the KL losses. The exponential schedule and masking probabilities are hand-chosen or tuned, and the W2/Fisher-Rao statements are asserted without proof. No new physical entities are introduced.

free parameters (7)

gamma (schedule decay) = 0.009
Controls f(t)=1-gamma^t; chosen by ablation over {0.005,0.009,0.02} in Table 3.
epsilon_0 (prior standard deviation) = 1
Hyperparameter for the Gaussian prior on coordinates; set in Section 2.5.
Pm (mask activation probability) = 0.3
Probability of enabling the masking mechanism during training; set in Section 2.5.
Pam (per-atom mask probability) = 0.3
Probability each atom is masked in geometry-enhanced training; set in Section 2.5.
beta0 = 1
Defined in Section 2.5 but never used in the equations or results; possibly a leftover or prior parameter.
number of sampling steps = 100
Inference steps; set in Section 2.5, not ablated.
lambda_x, lambda_v (loss weights) = not reported
Mentioned as hyperparameters after Eq (11) but no values or ablation are given.

assumptions (6)

domain assumption Atomic coordinates are Gaussian with isotropic variance; atom types are Dirichlet.
Eq (1)-(2) in Section 2.3; this parameterization fixes the geometry of the interpolation and is not derived.
domain assumption Linear interpolation of distribution parameters in Eq (5)-(6) defines a meaningful probability path.
The central construction of PIF in Section 2.2; no proof that the induced path has the claimed optimality for general distributions.
ad hoc to paper The loss in Eq (10)-(11) is the KL divergence between predicted and interpolated distributions.
Stated after Eq (10)-(11) but the Dirichlet KL expression does not match the standard form; no derivation is supplied.
ad hoc to paper The exponential schedule f(t)=1-gamma^t is a valid and beneficial schedule.
Eq (9) is justified only empirically via ablation in Table 3; it breaks the W2 optimality that holds for f(t)=t.
domain assumption The protein pocket is rigid.
Stated as a limitation in Section 4; the model does not account for induced fit.
ad hoc to paper Random masking forces the model to learn local geometric constraints.
Geometry-enhanced learning strategy in Section 2.4; no evidence isolates the masking effect from other changes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MolPIF: A Parameter Interpolation Flow Model for Molecule Generation." pith.science (2026). https://pith.science/paper/LDSARH7D

@misc{pith2026250713762,
  author       = {Pith},
  title        = {Pith review of: MolPIF: A Parameter Interpolation Flow Model for Molecule Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LDSARH7D}},
  note         = {Machine review of arXiv:2507.13762}
}

read the original abstract

Motivation: Structure-based drug design (SBDD) has advanced with deep generative models, but bridging the gap between continuous atomic coordinates and discrete atom types remains a challenge. Current approaches, such as diffusion and flow matching models, often fail to unify these heterogeneous modalities, relying on separate strategies or ill-fitting Euclidean metrics for discrete variables. This lack of a consistent framework limits generative models' ability to capture the geometric and chemical structure of protein-ligand complexes. Results: We present MolPIF, a parameter interpolation flow mechanism designed to unify the generation of continuous and discrete molecular variables. Unlike traditional flow models that operate in sample space, MolPIF interpolates between distributions in the parameter space, theoretically recovering Wasserstein-2 optimal transport for continuous coordinates and establishing Fisher-Rao geodesics for discrete atom types. We further incorporate a geometry-enhanced learning strategy to improve the capture of atomic contexts. Extensive evaluations on the CrossDocked2020 dataset demonstrate that MolPIF outperforms baselines in binding affinity, chemical validity, geometric fidelity and chemical space coverage. Additionally, MolPIF exhibits versatility in lead optimization and offers flexible prior distribution selection (such as Laplace), establishing a robust paradigm for SBDD. Availability: Source code is freely available at https://github.com/BLEACH366/MolPIF. Supplementary information: Supplementary data are available at Bioinformatics.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

ParetoPilot: Zero-Surrogate Offline Multi-Objective Optimization via Infer-Perturb-Guide Diffusion
cs.LG 2026-06 unverdicted novelty 6.0 of 10

ParetoPilot is a zero-surrogate diffusion framework for offline MOO that uses an IPG engine to steer generation via inferred objective directions and orthogonal perturbations, outperforming 14 baselines on 51 tasks.
ParetoPilot: Zero-Surrogate Offline Multi-Objective Optimization via Infer-Perturb-Guide Diffusion
cs.LG 2026-06 conditional novelty 6.0 of 10

ParetoPilot uses Infer-Perturb-Guide inside reverse diffusion to push samples to the Pareto front without surrogates, ranking best among 16 methods on 51 offline MOO tasks.

Reference graph

Works this paper leans on

14 extracted references · 4 canonical work pages · cited by 1 Pith paper

[6]

URL https://arxiv.org/abs/2412.04431. C. Harris, K. Didi, A. R. Jamasb, C. K. Joshi, S. V. Mathis, P. Lio, and T. Blundell. Benchmarking generated poses: How rational is structure-based drug design with generativ e models? arXiv preprint arXiv:2308.07413 ,

arXiv
[7]

URL https://arxiv.org/abs/2503.00975. H. Lin, G. Zhao, O. Zhang, Y. Huang, L. Wu, Z. Liu, S. Li, C. Tan, Z. Gao, and S. Z. Li. Cbgbench: ﬁll in the blank of protein-molecule complex binding graph. arXiv preprint arXiv:2406.10840,

work page Pith review arXiv
[10]

URL https://arxiv.org/abs/2205.07249. K. Qiu, Y. Song, Z. Fan, P. Liu, Z. Zhang, M. Zheng, H. Zhou, and W.-Y. Ma. Piloting structure-based drug design via modality-speciﬁc optimal schedule. arXiv preprint arXiv:2505.07286,

arXiv
[11]

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

arXiv
[12]

Stark, B

H. Stark, B. Jing, C. Wang, G. Corso, B. Berger, R. Barzilay, and T. Jaakkola. Dirichlet ﬂow matching with applications t o dna sequence design. arXiv preprint arXiv:2402.05841 ,

arXiv
[1997]

Lipman, R

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

arXiv
[1999]

Graves, R

A. Graves, R. K. Srivastava, T. Atkinson, and F. Gomez. Bayesian ﬂow networks. arXiv preprint arXiv:2308.07037 ,

arXiv
[2012]

B. Boll, D. Gonzalez-Alvarado, and C. Schn¨ orr. Generative modeling of discrete joint distributions by e-geodesic ﬂow matching on assignment manifolds. arXiv preprint arXiv:2402.07846,

arXiv

Show all 14 references

[2017]

Z. Tian, Y. Jin, Z. Liu, R. Yuan, X. Tan, Q. Chen, W. Xue, and Y. Guo. Audiox: Diﬀusion transformer for anything-to-audi o generation. arXiv preprint arXiv:2503.10522 ,

arXiv
[2020]

URL https://api.semanticscholar.org/CorpusID:219792684. L. Zhao, X. Ding, L. Yu, and L. Akoglu. Improving and unifying discrete&continuous-time discrete denoising diﬀusion. CoRR, 2024

2024
[2021]

X. Peng, F. Guo, R. Guo, J. Sun, J. Guan, Y. Jia, Y. Xu, Y. Huang, M. Zhang, J. Peng, et al. Atom-level generative foundation model for molecular interaction with pockets. bioRxiv, pages 2024–10,

2024
[2023]

J. Guan, X. Zhou, Y. Yang, Y. Bao, J. Peng, J. Ma, Q. Liu, L. Wang, and Q. Gu. Decompdiﬀ: diﬀusion models with Bioinformatics, YEAR, Volume XX, Issue x 9 decomposed priors for structure-based drug design. arXiv preprint arXiv:2403.07902 ,

arXiv
[2024]

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

arXiv
[2025]

URL https://arxiv.org/abs/2504.10564. T. Cui, C. Tang, M. Su, S. Zhang, Y. Li, L. Bai, Y. Dong, X. Gong, and W. Ouyang. Geometry-enhanced pretraining on interatomic potentials. Nature Machine Intelligence , 6(4): 428–436,

arXiv

Pith tools

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

[1] [6]

URL https://arxiv.org/abs/2412.04431. C. Harris, K. Didi, A. R. Jamasb, C. K. Joshi, S. V. Mathis, P. Lio, and T. Blundell. Benchmarking generated poses: How rational is structure-based drug design with generativ e models? arXiv preprint arXiv:2308.07413 ,

arXiv

[2] [7]

URL https://arxiv.org/abs/2503.00975. H. Lin, G. Zhao, O. Zhang, Y. Huang, L. Wu, Z. Liu, S. Li, C. Tan, Z. Gao, and S. Z. Li. Cbgbench: ﬁll in the blank of protein-molecule complex binding graph. arXiv preprint arXiv:2406.10840,

work page Pith review arXiv

[3] [10]

URL https://arxiv.org/abs/2205.07249. K. Qiu, Y. Song, Z. Fan, P. Liu, Z. Zhang, M. Zheng, H. Zhou, and W.-Y. Ma. Piloting structure-based drug design via modality-speciﬁc optimal schedule. arXiv preprint arXiv:2505.07286,

arXiv

[4] [11]

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

arXiv

[5] [12]

Stark, B

H. Stark, B. Jing, C. Wang, G. Corso, B. Berger, R. Barzilay, and T. Jaakkola. Dirichlet ﬂow matching with applications t o dna sequence design. arXiv preprint arXiv:2402.05841 ,

arXiv

[6] [1997]

Lipman, R

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

arXiv

[7] [1999]

Graves, R

A. Graves, R. K. Srivastava, T. Atkinson, and F. Gomez. Bayesian ﬂow networks. arXiv preprint arXiv:2308.07037 ,

arXiv

[8] [2012]

B. Boll, D. Gonzalez-Alvarado, and C. Schn¨ orr. Generative modeling of discrete joint distributions by e-geodesic ﬂow matching on assignment manifolds. arXiv preprint arXiv:2402.07846,

arXiv