acceptodds
Under review as a conference paper at ICLR 2027

Self-Improving Bayesian Flow Networks via Group Relative Policy Optimization for Structure-Based Drug Design

Abstract

Structure-based drug design (SBDD) requires molecules that bind favorably to a protein pocket while maintaining plausible conformations and suitable chemical properties. Generative models learn from protein–ligand complexes, but fitting the training distribution alone does not directly optimize these properties. On-policy reinforcement learning offers a way for generators to learn from evaluations of their own candidates. Applying it to Bayesian flow networks (BFNs) requires identifying the stochastic decisions whose likelihoods support policy updates. In the molecular BFN sampler considered here, the network predicts clean coordinates and atom-type probabilities, while the sampler draws new coordinate-belief means and atom-type logits. We introduce GRAB (Group-Relative Alignment of Bayesian flow networks), a self-improving framework that defines these sampled quantities as a joint policy action. For this Gaussian sampler, we derive a closed-form conditional action density that is differentiable through the network predictions, with covariance fixed by the sampling schedules. Terminal rewards from docking affinity, strain energy, drug-likeness (QED), and synthetic accessibility (SA) support Group Relative Policy Optimization (GRPO), without a learned value model or gradients through reconstruction or scoring. Generation retains the original sampler without additional property-gradient guidance. On 100 CrossDocked2020 test pockets, one GRAB configuration achieves the lowest mean Vina Dock and median strain energy among the generators in the main comparison, while maintaining favorable drug-likeness, synthetic accessibility, and ligand diversity.

Primary area:
generative models
Keywords:
Bayesian flow networks, group relative policy optimization, structure-based drug design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_f8HROpenReview ↗
Rating
2
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GRAB, a framework for fine-tuning molecular Bayesian flow networks using group-relative policy optimization. It formulates MolCRAFT’s sampled coordinate belief means and atom-type logits as joint actions and optimizes their likelihoods using terminal molecular rewards.

Critical strengths

The policy formulation explicitly accounts for the stochastic variables sampled by the generator. The framework accommodates non-differentiable molecular rewards without requiring property-predictor gradients during inference. Reported improvements in docking and molecular quality demonstrate practical potential.

Critical weaknesses

Limited methodological novelty. CByG already investigates property-driven control of molecular BFN belief updates, BFN-DPO establishes multi-reward BFN alignment, and FTDiff applies trajectory-level GRPO to diffusion-based SBDD. Against these precedents, GRAB’s novelty largely reduces to defining actions and Gaussian likelihoods for a particular BFN sampler. The manuscript does not sufficiently establish why this adaptation constitutes a substantial methodological advance.

Insufficient isolation of the proposed contributions. The experiments do not establish the individual benefits of joint-action optimization, log-density scaling, and grouping choices over a straightforward policy-gradient implementation.

Limited comparative evidence. The BFN-DPO comparison is not matched in candidate-generation budget, complicating attribution of improvements to the optimization method. Single-run results and evaluation on one backbone further limit conclusions about robustness and generality.

Questions for authors

What nontrivial methodological obstacle does GRAB resolve beyond applying standard GRPO to the known Gaussian sampling distributions? Please clarify its distinction from CByG, BFN-DPO, and FTDiff.

Which existing results establish the necessity of joint-action optimization, log-density scaling, and the proposed grouping strategy?

Please clarify the total generation and oracle-evaluation budgets for GRAB and BFN-DPO, and whether the reported advantage remains supported under comparable budgets.

7 comments
AuthorsOpenReview ↗

Response to Reviewer f8HR (W1/Q1)

W1/Q1: BFN-specific Contribution and Distinction from Prior Work

Thank you for asking us to clarify the methodological contribution. GRAB’s methodological contribution is an on-policy formulation that makes the joint belief-update process of a molecular BFN trainable from black-box molecular evaluations. This enables the generator to improve through repeated generation–evaluation–update cycles, while retaining its original sampling recursion and requiring no additional property-gradient guidance at inference.

1. BFN-specific formulation

Knowing the Gaussian sampling distributions does not by itself specify the appropriate policy likelihood. In a molecular BFN, the network predicts clean coordinates and atom-type probabilities, but the sampler draws coordinate-belief means and type logits. The terminal reward evaluates a reconstructed molecule downstream of these decisions. GRAB connects these levels by defining the sampled pair (ai=(μi,yi))(a_i=(\mu_i,y_i)) as the joint action and placing softmax in the deterministic state transition. We derive the gradients of the joint action log-density for both coordinate and atom-type predictions, allowing terminal molecular scores to guide updates to both branches through the recorded sampling trajectories (Sec. 4.1; App. A.1–A.2). This is the methodological role of the formulation: it specifies how the existing BFN sampling dynamics can support reward-driven policy learning.

2. Distinction from prior work

CByG [1] uses property-predictor gradients to control sampling with a fixed generator. GRAB instead updates generator parameters from on-policy terminal evaluations and subsequently generates without additional property-gradient guidance.

BFN-DPO [2] learns from reward-derived preference pairs through a reference-relative objective based on BFN sender/receiver distributions. GRAB uses the actions actually sampled in fresh rollouts and their group-relative advantages for policy updates.

FTDiff [3] already applies trajectory-level GRPO and explicitly formulates a joint coordinate/type likelihood ratio for diffusion. GRAB defines its actions over BFN belief variables. Its practical objective also clips a scaled joint action ratio, whereas FTDiff separately clips coordinate/type terms. This difference alone does not establish empirical superiority.

Our contribution is the explicit BFN policy formulation and the resulting training procedure. The score-function derivation concerns an objective with a frozen terminal output rule, and the implemented clipped objective is a surrogate (App. A.2–A.4). The comparisons with other methods evaluate the complete procedure; the additional D=1 comparison in W2/Q2 examines log-density scaling. We have clarified the distinction from prior work in Sec. 2 and the role of the joint-action formulation in Sec. 4.1.

AuthorsOpenReview ↗

Response to Reviewer f8HR (W2/Q2 (1))

W2/Q2: Joint-Action Optimization, Log-Density Scaling, and Grouping

1. Joint-action formulation

By joint-action optimization, we mean using terminal molecular feedback to update the policy for the complete sampled belief update. At each step, the action ai=(μi,yi)a_i=(\mu_i,y_i) consists of the sampled coordinate-belief means and sampled type logits, with softmax⁡(yi)\operatorname{softmax}(y_i) giving the next atom-type belief. Both action distributions depend on the shared coordinate and atom-type belief state [4]. The term joint refers to accounting for both components of this sampled action using the same reward-derived advantage.

Conditioned on the current state, pocket, and atom count, the joint action density factorizes into coordinate and type distributions. With the sampled state and action held fixed, its log-density gradient therefore contains both contributions. Omitting either term would generally omit part of this gradient. Computing the two log-density terms separately and summing them is an equivalent way to evaluate the joint policy.

2. Log-density scaling

We use log-density scaling to reduce the risk that accumulated changes across a high-dimensional joint action prematurely truncate reward-driven updates through PPO clipping. For one molecule at one sampling step, the joint log-density sums over natoms(3+13)n_{\mathrm{atoms}}(3+13) scalar components. Consequently, changes across individual components can accumulate and push the joint likelihood ratio beyond the clipping interval. We therefore use

\Delta=\log\pi_\theta(a\mid s)-\log\pi_{\mathrm{old}}(a\mid s),$$ with $D=24(3+13)=384$, where 24 is the approximate mean number of heavy atoms. This matters because PPO clipping directly affects whether a sampled action continues to contribute to learning [5]. Here, $A$ denotes the group-relative advantage. When $A>0$ and $r_D>1+\epsilon$, or $A<0$ and $r_D<1-\epsilon$, the selected surrogate branch is constant with respect to the ratio, and that sample contributes zero policy-surrogate gradient. Scaling brings the ratio closer to one for a given joint log-density change, reducing the risk of reaching these flat regions during repeated updates on a rollout batch. We compared $D=384$ and $D=1$ over the first 150 training iterations: | Ratio interval | Iterations containing out-of-range ratios: $D=384$ | $D=1$ | |---|---:|---:| | $[0.99,1.01]$, the training clipping interval | 1/150 | 150/150 | | $[0.8,1.2]$, a broader diagnostic interval | 0/150 | 38/150 | Thus, without scaling, out-of-range ratios occurred in every observed iteration under the configured clipping interval, and remained present in 38 iterations even when assessed against $[0.8,1.2]$. Scaling substantially reduced their occurrence. These iteration-level counts measure the occurrence of out-of-range ratios; they do not quantify the fraction of samples with zero surrogate gradient. To assess molecular performance, we also compared D=384 with D=1 using the same initialization, reward weights, training budget, and learning rate, changing only the log-density scaling factor. Vina entries report **mean / median**, and SE entries report the **25th / 50th / 75th percentiles**. Vina scores and SE are in kcal/mol. | Setting | Vina Score ↓ | Vina Min ↓ | Vina Dock ↓ | SE ↓ | SA ↑ | QED ↑ | Div ↑ | |---|---:|---:|---:|---:|---:|---:|---:| | D=1 | −5.75 / −6.83 | −7.01 / −7.19 | −8.02 / −8.21 | 61 / 162 / 318 | 0.75 | 0.52 | 0.76 | | GRAB-B, D=384 | −7.593 / −8.428 | −8.759 / −8.736 | −9.515 / −9.535 | 44 / 104 / 200 | 0.76 | 0.51 | 0.75 | D=384 achieves better results across all Vina metrics and SE quartiles. Mean Vina Dock improves from −8.02 to −9.515, and median SE decreases from 162 to 104. SA also increases from 0.75 to 0.76, while D=1 retains slightly higher QED and diversity, both by 0.01. Together with the ratio statistics above, this ablation supports the practical benefit of log-density scaling: under the same training settings, it reduces the occurrence of out-of-range ratios and improves docking performance and conformational strain. We have added the ratio statistics and complete molecular results to Appendix J, referenced from Section 4.2.
AuthorsOpenReview ↗

Response to Reviewer f8HR (W2/Q2 (2))

3. Grouping and comparison with a straightforward policy-gradient implementation

We form each group from molecules generated for the same pocket and with the same sampled atom count. Reward distributions can vary across pockets and ligand sizes, including across different atom counts within the same pocket. Grouping by both pocket and atom count allows us to compare candidates under the same conditions. Normalizing each reward channel within these groups then makes the advantage reflect relative molecular performance, reducing the influence of differences in reward levels and variability across pockets and ligand sizes.

We agree that a straightforward policy-gradient implementation is an important comparison. To our knowledge, a published REINFORCE baseline for the same MolCRAFT setting is not available. Following your suggestion, we therefore implemented REINFORCE [6] on the MolCRAFT backbone and compared it with GRAB-B using the same training data and reward weights: (0.35, 0.15, 0.15, 0.35) for Vina Min, QED, SA, and SE, respectively.

MethodVina Score Avg ↓Med ↓Vina Min Avg ↓Med ↓Vina Dock Avg ↓Med ↓SE 25% ↓50% ↓75% ↓SA ↑QED ↑Div ↑
MolCRAFT−6.307−6.842−7.089−7.106−7.808−7.907661533460.690.530.74
REINFORCE (our implementation)−6.546−7.143−7.381−7.365−8.116−8.189751714140.690.550.72
GRAB-B−7.593−8.428−8.759−8.736−9.515−9.535441042000.760.510.75

REINFORCE improves docking scores over the initial MolCRAFT model, but its strain quartiles increase and diversity decreases. GRAB-B achieves substantially better docking scores and lower strain than REINFORCE: mean Vina Dock improves from −8.116 to −9.515, and median SE decreases from 171 to 104. SA also increases from 0.69 to 0.76, and diversity from 0.72 to 0.75. REINFORCE achieves higher QED (0.55 versus 0.51).

Under the same reward weights and training data, the complete GRAB procedure with group-relative updates achieves stronger docking performance, lower strain, and higher SA and diversity than REINFORCE. This comparison supports the effectiveness of the complete procedure, although it does not establish the necessity of grouping alone. We thank the reviewer for suggesting this comparison. We have clarified the grouping rationale in Sec. 4.2 and added the REINFORCE comparison setting and complete results in Appendix H, with a summary in Sec. 5.2.

Reviewer_gQVLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce GRAB (Group Relative Alignment of Bayesian flow networks), a method that allows adaptation of GRPO (Group Relative Policy Optimization) to BFNs. The methodological contribution is to treat the samplers coordinate belief means and atom-type logits as a joint policy action, making the conditional action density Gaussian. The authors then compute group-relative advantages for each channel and combine them with fixed weights. The policy is then updated with a clipped surrogate objective. In their experiments, they fine-tune the MolCRAFT BFN backbone and optimize common molecular properties on CrossDocked 2020, demonstrating consistent improvements compared to the unoptimized BFN baseline for Vina docking scores, conformational strain energy, QED and SA across most test pockets.

Critical strengths
  • An explicit likelihood formulation for the given BFN sampler by modelling actions as a joint Gaussian over coordinates and atom type logits.
  • Consistent improvement on the optimised properties over the baseline BFN in a multi-pocket setup with reported gains for different reward configurations on the held-out test set.
  • Common geometric failure modes are addressed: Posebusters validity and clash metrics improve over the baseline, even though only indirectly optimized.
Critical weaknesses
  • Demonstrating improvement in using GRAB over other reward-based RL methods is not yet conclusive: The main baseline seems to be a DPO-finetuned version of MolCRAFT on the same starting checkpoint and preference pairs form the same objectives. However, this comparison used a vastly smaller preference pair pool (10K candidates versus around 77K sampling attempts in the GRAB paper and some setup details on how the preference pairs are set up are not specified, see questions). Other baselines include MolFORM-RL, that were not retrained and were initially setup on a slightly different reward. I don’t request retraining of every historical baseline, but I think doing one comparison (e.g. with BFN-DPO) properly with matched budgets and reward functions will help demonstrating the benefit of using GRAB better.
  • The paper only partly shows that diversity collapse is avoided, a common failure mode in RL: They report diversity as averaged Tanimoto distance over the sampling pool per pocket. A aggregated metric of number of PB-valid, unique and property passing molecules would be helpful. In Figure 3, success over generation attempts is already shown, but doesn’t include PB-valid or uniqueness.
  • The analysis focusses on a limited hand-picket set of reward configurations. A broader analysis of GRAB as a new RL technique for small molecule generation would help to identify strengths and failure modes, e.g. Pareto-frontier analysis of selected properties like docking scores vs. SA. A single-pocket optimization run with limited budget could assess its feasibility when concentrating just on a single target of interest.

With this in mind, I am leaning towards weak reject, but I am happy to reconsider my evaluation with matched baselines provided, and analysis of its failure modes.

Questions for authors
  • Was the BFN-DPO setup the same as in Lee et al. PMLR 2025? Was IPDiff+DPO trained on a matched setup?
  • How many unique PB-valid molecules are obtained per pocket as the number of generation attempts increases?
  • Section 6.1 and Table 2 compare five configurations that show only positive reward weights for each property. To understand the Pareto-frontier of GRAB better, a more complete exploration of weights would help, including just pushing a single property to its limits. Did the authors experiment with these reward weights?
  • Did the authors try a case study with a single pocket setup?
7 comments
AuthorsOpenReview ↗

Response to Reviewer gQVL (W1/Q1)

W1/Q1: BFN-DPO Implementation and Matched Experimental Protocol

To make GRAB and BFN-DPO [1] comparable, we repeated the comparison using the same pretrained MolCRAFT [2] checkpoint, sampler settings, training pockets, and reward weights.

We used our own implementation of BFN-DPO following Lee et al. [1]. The candidate pool and reward processing were adapted to match GRAB, so the experimental settings differ from those in the original study.

1. Matched comparison and training protocol

We matched the generation budget by using the same training pockets, the same number of groups per pocket, and eight generation attempts per group. GRAB runs for 1,200 rounds with eight sampled pockets per round, giving 1,200 × 8 × 8 = 76,800 generation attempts. For DPO, we generated the corresponding groups before training using the initial MolCRAFT checkpoint. The candidates were scored once, and the same offline pool and stored scores were reused for all three DPO reward-weight settings.

Both methods use the same scoring functions for Vina Min, QED, SA, and SE. For DPO, we selected the highest- and lowest-scoring valid candidates in each group to form one winner–loser pair. Groups with fewer than two valid candidates or no score variation were excluded. DPO uses AdamW with a learning rate of 2 × 10⁻⁵, batch size 2, and coefficients βx=βv=1.0\beta_x=\beta_v=1.0, with one pass over the preference pairs. GRAB uses Adam with a learning rate of 3 × 10⁻⁵ and two update epochs per rollout batch.

Both methods select checkpoints on the same 24 validation pockets, with 20 generation attempts per pocket, using the same selection rule. Test results are not used for checkpoint selection.

2. Results and trade-offs

The table reports the new BFN-DPO results alongside GRAB under each matched reward-weight setting, including equal weights and the GRAB-A/B weight settings. Weights are ordered as (Vina Min–based affinity, QED, SA, SE). Vina entries report mean / median, and SE entries report the 25th / 50th / 75th percentiles. Vina scores and SE are in kcal/mol.

MethodReward weightsVina Score ↓Vina Min ↓Vina Dock ↓SE ↓SA ↑QED ↑Div ↑
Reference—−6.36 / −6.46−6.71 / −6.49−7.35 / −7.2634 / 107 / 1910.730.48—
MolCRAFT—−6.31 / −6.84−7.09 / −7.11−7.81 / −7.9166 / 153 / 3460.690.530.74
DPO-1(.25, .25, .25, .25)−6.34 / −7.15−7.28 / -7.34−8.04 / −8.3258 / 135 / 3030.740.580.73
GRAB-Equal(.25, .25, .25, .25)−6.12 / −7.17−7.26 / −7.44−8.31 / −8.4151 / 125 / 2450.750.560.76
DPO-2(.40, .20, .20, .20)−6.67 / −7.13-7.40 / −7.58−8.15 / −8.4260 / 148 / 4000.720.590.72
GRAB-A(.40, .20, .20, .20)−6.08 / −7.00−7.51 / −7.60−8.53 / −8.6142 / 105 / 2230.770.550.76
DPO-3(.35, .15, .15, .35)−6.81 / −7.63−7.85 / −7.88−8.48 / −8.6450 / 120 / 2670.730.590.72
GRAB-B(.35, .15, .15, .35)−7.59 / −8.43−8.76 / −8.74−9.52 / −9.5444 / 104 / 2000.760.510.75

In the equal-weight comparison, GRAB achieves lower mean Vina Dock (−8.31 versus −8.04) and median SE (125 versus 135) than BFN-DPO, together with higher SA (0.75 versus 0.74) and diversity (0.76 versus 0.73). These results also improve on the shared MolCRAFT initialization, whose mean Vina Dock and median SE are −7.81 and 153, respectively.

The advantage is consistent across all three reward-weight settings: GRAB achieves lower mean and median Vina Dock, lower strain quartiles, and higher SA and diversity. These gains persist when the starting checkpoint, sampler, training pockets, reward weights, and generation budget are matched, directly addressing your concern about the sampling budget and reward settings. Together, the results support the effectiveness of GRAB's training procedure and its advantage over BFN-DPO in docking, strain, SA, and diversity, while BFN-DPO retains higher QED.

3. Other baselines

For IPDiff+DPO, we kept our original training setup and did not rerun it under the matched settings above. The MolFORM-RL result uses the authors' released checkpoint without retraining it with GRAB's reward settings. The new BFN-DPO experiment provides the matched comparison requested here.

Thank you for these suggestions. We have updated Table 1 and Section 5.2 with the matched BFN-DPO comparison and described the candidate generation, pair construction, and training settings in Appendix B.4. These changes are highlighted in blue.

AuthorsOpenReview ↗

Response to Reviewer gQVL (W2/Q2 (1))

W2/Q2: Sample Accounting and Unique PB-Valid Molecular Yield

With 100 generation attempts per pocket, GRAB-B produces an average of 85.75 unique connected PB-valid molecules, compared with 83.23 for MolCRAFT and 82.01 for BFN-DPO under the GRAB-B reward weights. Thank you for suggesting this analysis: it directly measures how many distinct valid candidates remain after filtering and deduplication.

1. Evaluation budget and sample accounting.

Each method uses 100 generation attempts per pocket across 100 held-out test pockets, totaling 10,000 attempts. Failed attempts remain included in this budget. The following table documents reconstruction, connectivity, scoring availability, and the PoseBusters evaluation pool.

MethodRecon.Conn.VinaSE-scoredScorableSc.connPB-eval
TargetDiff9,9069,2309,9069,8749,9019,2259,906
IPDiff9,8269,2599,8269,7999,8259,2589,826
IPDiff + DPO9,7829,0999,7829,7339,7739,0909,782
PAFlow9,4318,2448,7918,0488,7917,7309,431
MolFORM (Base)9,7878,3419,3409,7639,3317,9749,787
MolFORM-RL9,9959,4109,3809,8869,3798,8409,995
MolCRAFT9,9379,6879,9379,9309,9379,6879,937
BFN-DPO (DPO-1)9,9959,8989,9919,9879,9919,8989,995
BFN-DPO (DPO-2)9,9939,9489,9939,9849,9939,9489,993
BFN-DPO (DPO-3)9,9969,9279,9959,9949,9959,9279,996
GRAB-A9,9879,8739,9879,9829,9879,8739,987
GRAB-B9,9959,9319,9939,9829,9939,9319,995

Definitions. Recon. counts attempts producing an RDKit molecule with a conformer; Conn. requires a single molecular fragment. The Vina column counts molecules with successfully computed Vina Min scores in the existing evaluation records, and SE-scored counts molecules with successfully computed strain-energy scores. Scorable requires usable Vina Min, QED, SA, and atom-count values; SE is not required. Sc.conn is the intersection of scorable and connected samples, corresponding to the Retained count in Appendix B.5. These counts record the filtering outcome; each property summary uses its own successfully evaluated molecules. Missing results, NaN, and positive or negative infinity are excluded for the corresponding property only. PB-eval counts molecules evaluated with PoseBusters.

2. Unique connected PB-valid yield.

A qualifying sample must be connected and pass all checks in the PoseBusters dock configuration. For dup% and U(100)U(100), we deduplicate molecules within each pocket using canonical SMILES [3]: qualifying molecules with identical canonical SMILES strings are counted only once.

We report U(100)U(100), the mean number of distinct qualifying molecules obtained per pocket within 100 attempts. PB-valid (%) is calculated as the PB-valid count divided by the corresponding PB-eval count in the table above, multiplied by 100.

MethodPB-valid (%)PB-valid countdup%* ↓Unique connected PB-valid molecules per pocket, U(100)U(100) ↑
TargetDiff52.88%5,2381.0%51.84
IPDiff22.62%2,2230.8%22.06
IPDiff + DPO45.70%4,4700.5%44.47
PAFlow11.91%1,1230.5%11.17
MolFORM (Base)38.61%3,7790.7%37.51
MolFORM-RL42.37%4,2358.2%38.88
MolCRAFT85.60%8,5062.2%83.23
BFN-DPO (DPO-1)84.46%8,4426.4%79.02
BFN-DPO (DPO-2)83.87%8,3816.1%78.70
BFN-DPO (DPO-3)87.37%8,7346.1%82.01
GRAB-A84.07%8,3966.0%78.91
GRAB-B91.96%9,1916.7%85.75

The dup% column is calculated as 100[1−100U(100)/NPB-valid]100[1-100U(100)/N_{\mathrm{PB\text{-}valid}}]. It measures the reduction from PB-valid outputs to distinct connected PB-valid molecules, including connectivity filtering if any PB-valid outputs are disconnected.

AuthorsOpenReview ↗

Response to Reviewer gQVL (W2/Q2 (2))

3. Results and trade-offs

GRAB-B yields the most unique connected PB-valid molecules among the evaluated methods at this generation budget. Across the 100 test pockets, it provides 8,575 such molecules, compared with 8,323 for MolCRAFT and 8,201 for the reward-weight-matched BFN-DPO (DPO-3). This corresponds to 2.52 and 3.74 additional unique valid molecules per pocket, respectively.

We acknowledge that GRAB-B generates more repeated molecules than the pretrained MolCRAFT model, and we thank the reviewer for highlighting this limitation of reporting average diversity alone. The reduction after connectivity filtering and deduplication is 6.7% for GRAB-B versus 2.2% for MolCRAFT. However, GRAB-B also achieves a higher PB-valid fraction, 91.96% versus 85.60%, and still provides more unique connected PB-valid molecules after repeated molecules are removed. Thus, the increased repetition does not eliminate the gain in the number of distinct valid candidates.

This benefit depends on the reward setting: GRAB-A yields 78.91 unique connected PB-valid molecules per pocket, close to DPO-2's 78.70 but below MolCRAFT. The conclusion of increased unique valid yield therefore applies to GRAB-B. These results measure the yield at 100 attempts per pocket; they do not additionally require the four property thresholds in Figure 3.

4. Results on the five pockets in Figure 3

We have updated Figure 3 to show the results after PoseBusters checks and deduplication. The analysis uses the same 500 generation attempts per pocket, totaling 2,500 attempts per method. A molecule must be connected, PoseBusters-valid, and simultaneously outperform the reference ligand on Vina Dock, strain, QED, and SA. Each canonical SMILES is counted once per pocket. The revised curves show the cumulative number of these distinct qualifying molecules as generation attempts increase.

The stricter evaluation preserves our main conclusion that GRAB increases the number of useful molecular candidates. After requiring PoseBusters validity and deduplicating within each pocket, GRAB-B produces 1,293 distinct molecules satisfying all four property criteria, compared with 515 for MolCRAFT and 967 for GRAB-A. Its advantage over MolCRAFT changes from 2.68-fold under the original criterion to 2.51-fold under the stricter criterion, supporting the effectiveness of GRAB after accounting for invalid poses and repeated molecules.

The eNOS case nevertheless reveals a failure mode involving highly repeated generation of small ligands. Generated ligands in this case average 11.9 heavy atoms, compared with 27.0–28.6 in the other four pockets. For GRAB-B, 182 property-passing, PB-valid records reduce to only 44 distinct molecules after deduplication, corresponding to a repeat rate of 75.8%. Although this remains above MolCRAFT's 33 distinct molecules, the 1.33-fold improvement is substantially smaller than the 1.99–14.31-fold gains on the other four pockets, and the result is below GRAB-A's 87 molecules. This shows that improvements in molecular properties can coexist with excessive repetition and limited growth in the number of distinct candidates.

We thank the reviewer for suggesting this analysis. We have added the unique connected PB-valid results across 100 test pockets to Appendix D.1, updated Figure 3 and the results in Section 6.2, added the eNOS analysis to Appendix D.2, and discussed high repetition of small ligands as a limitation in Section 7. The new table, revised text, and figure caption are highlighted in blue. The overall gains remain substantial, while this analysis makes clear where the benefits of GRAB are limited.

Reviewer_BhgeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRAB fine-tunes a molecular Bayesian flow network using terminal molecular scores. It treats sampled coordinate-belief means and pre-softmax atom-type logits as policy actions, evaluates their joint Gaussian density, and applies a clipped group-relative update. Training uses Vina Min, strain energy, QED, and synthetic accessibility; the stated generation protocol retains the pretrained sampler without additional property-gradient guidance. Experiments use MolCRAFT and 100 CrossDocked2020 test pockets. The main contribution is a concrete adaptation of online policy optimization to this molecular BFN sampler.

Critical strengths
  1. The policy formulation is clear and mathematically consistent within its stated scope. The action density in Eq. (6) is defined on the sampled means and logits, and the appendix distinguishes the frozen-terminal objective from the implemented training surrogate. This provides a usable way to learn from scores without differentiating through reconstruction or molecular evaluation. The local-gradient and timestep-sampling statements retain their assumptions rather than claiming a guarantee for the complete generator (pp. 5–6, Appendix A.1–A.4).

  2. The comparison with the re-evaluated initialization shows a substantial improvement. With the same starting checkpoint and disclosed generation settings, GRAB-B changes mean Vina Dock from −7.81 to −9.52 kcal/mol and median as-generated strain from 153 to 104 kcal/mol (Appendix C, Table 3). SA and diversity also increase, while QED decreases from .53 to .51. Figure 2 shows the lowest pocket-median Dock score for GRAB-B on 85/100 pockets among MolCRAFT, GRAB-A, and GRAB-B. These results support the usefulness of the complete fine-tuning recipe.

  3. The pose checks and five-pocket joint-property test strengthen the empirical case. Appendix D evaluates reconstruction, completeness, clashes, and PoseBusters validity. GRAB-B improves the reported PB-valid fraction from 85.61% to 91.97% within the evaluated molecular pool, although GRAB-A has worse PB-validity and more clashes than the initialization. Figure 3 also counts failed attempts: on its five pockets, GRAB-B produces 1,490/2,500 molecules meeting all four reference-relative property criteria, versus 555/2,500 for MolCRAFT. This is useful evidence of joint computational-property gains; Dock uses redocking and strain uses the as-generated conformer (Appendix B.5).

Critical weaknesses
  1. The experiments leave the incremental benefit of the update mechanism unclear. BFN-DPO shares the checkpoint and sampler, which is a useful control. However, its initial preference pool, data refresh, and the main reward weights differ from GRAB, and scoring and optimization budgets are incompletely specified (Appendix B.2/B.4). Table 2 already supplies equal-weight GRAB: compared with BFN-DPO, it improves median strain (125 versus 256), SA, and diversity, but has weaker mean Dock (−8.31 versus −8.51) and QED (.56 versus .57). Thus the evidence establishes useful trade-offs more clearly than it explains the contribution of the proposed optimization procedure. This matters because the Gaussian belief representation and continuous-generator GRPO ingredients have close precedents in MolCRAFT v4, §4.2 (https://arxiv.org/html/2404.12141v4), MolJO v4, Appendix C.1 (https://arxiv.org/html/2411.13280v4), and Flow-GRPO v1, §4.1 (https://arxiv.org/html/2505.05470v1); the BFN-specific adaptation and its measured benefit carry much of the contribution.

  2. Training variability and uncertainty in the rankings are not quantified. Each reward configuration uses one fine-tuning run (p. 7, lines 368–372). Different weight settings do not measure repeatability of one setting, and the strain quartiles describe the molecule distribution. The broad improvement over initialization is persuasive, but smaller cross-method gaps, such as mean Dock −9.52 versus PAFlow's −9.29, are difficult to assess without paired-pocket uncertainty and repeated training.

  3. Some evaluation populations and baseline results cannot yet be traced precisely. Appendix B.5 states that property statistics use retained outputs, but does not give per-method retained/scored counts or the numerical PoseBusters pool sizes. Appendix B.4 identifies several historical rows and the BFN-DPO reimplementation; the remaining rows need clearer copied/recomputed/generated provenance, fixed versions, and evaluation settings. The MolFORM rows particularly need an explanation of their source relative to the cited versions. These gaps limit interpretation of the broad cross-generator comparison; their actual effect on the scores is unknown.

Questions for authors
  1. Could you document BFN-DPO pair construction, optimization and checkpoint selection, then compare it with equal-weight GRAB at comparable scored-candidate budgets, or provide a suitable learning curve? Generation attempts and initial preference candidates are different quantities, so their reported counts alone do not establish relative evaluation cost. Evidence that the favorable trade-off persists under this control would strengthen the case for the optimization contribution.

  2. Can you report paired-pocket effects or intervals for the main comparisons and repeat the central GRAB configuration across training seeds? Pocket resampling would address target uncertainty; repeated training would address optimization variability. Stable docking/strain gains and trade-offs across these checks would increase confidence in the empirical contribution even if small leaderboard differences change.

  3. Could you provide a compact per-method sample-flow and provenance table with attempted, reconstructed/connected, successfully Vina/SE-scored, and PB-evaluated counts? Please include the evaluator revision, non-finite-score handling, and the source/checkpoint or sample release for the main baselines. Please specify whether Vina/SE/QED/SA summaries pool retained molecules or weight pockets equally. Diversity pooling is already specified. This would clarify whether the compared numbers describe sufficiently comparable output populations.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Bhge (W1/Q1)

W1/Q1: Matched BFN-DPO Comparison and BFN-specific Contribution

To make GRAB and BFN-DPO [2] comparable, we repeated the experiments using the same pretrained MolCRAFT [1] checkpoint, sampler settings, training pockets, and reward weights.

1. Matched comparison and training protocol

We matched the generation budget by using the same training pockets, the same number of groups per pocket, and eight generation attempts per group. GRAB runs for 1,200 rounds with eight sampled pockets per round, giving 1,200 × 8 × 8 = 76,800 generation attempts. For DPO, we generated the corresponding groups before training using the initial MolCRAFT checkpoint. The candidates were scored once, and the same offline pool and stored scores were reused for all three DPO reward-weight settings.

Both methods use the same scoring functions for Vina Min, QED, SA, and SE. For DPO, we selected the highest- and lowest-scoring valid candidates in each group to form one winner–loser pair. Groups with fewer than two valid candidates or no score variation were excluded. DPO uses AdamW with a learning rate of 2 × 10⁻⁵, batch size 2, and coefficients βx=βv=1.0\beta_x=\beta_v=1.0, with one pass over the preference pairs. GRAB uses Adam with a learning rate of 3 × 10⁻⁵ and two update epochs per rollout batch.

Both methods select checkpoints on the same 24 validation pockets, with 20 generation attempts per pocket, using the same selection rule. Test results are not used for checkpoint selection.

2. Results and trade-offs

The table reports the new BFN-DPO results alongside GRAB under each matched reward-weight setting, including the equal-weight comparison requested in Q1. Weights are ordered as (Vina Min–based affinity, QED, SA, SE). Vina entries report mean / median, and SE entries report the 25th / 50th / 75th percentiles. Vina scores and SE are in kcal/mol.

MethodReward weightsVina Score ↓Vina Min ↓Vina Dock ↓SE ↓SA ↑QED ↑Div ↑
Reference—−6.36 / −6.46−6.71 / −6.49−7.35 / −7.2634 / 107 / 1910.730.48—
MolCRAFT—−6.31 / −6.84−7.09 / −7.11−7.81 / −7.9166 / 153 / 3460.690.530.74
DPO-1(.25, .25, .25, .25)−6.34 / −7.15−7.28 / -7.34−8.04 / −8.3258 / 135 / 3030.740.580.73
GRAB-Equal(.25, .25, .25, .25)−6.12 / −7.17−7.26 / −7.44−8.31 / −8.4151 / 125 / 2450.750.560.76
DPO-2(.40, .20, .20, .20)−6.67 / −7.13-7.40 / −7.58−8.15 / −8.4260 / 148 / 4000.720.590.72
GRAB-A(.40, .20, .20, .20)−6.08 / −7.00−7.51 / −7.60−8.53 / −8.6142 / 105 / 2230.770.550.76
DPO-3(.35, .15, .15, .35)−6.81 / −7.63−7.85 / −7.88−8.48 / −8.6450 / 120 / 2670.730.590.72
GRAB-B(.35, .15, .15, .35)−7.59 / −8.43−8.76 / −8.74−9.52 / −9.5444 / 104 / 2000.760.510.75

In the equal-weight comparison requested in Q1, GRAB achieves lower mean Vina Dock (−8.31 versus −8.04) and median SE (125 versus 135) than BFN-DPO, together with higher SA (0.75 versus 0.74) and diversity (0.76 versus 0.73). These results also improve on the shared MolCRAFT initialization, whose mean Vina Dock and median SE are −7.81 and 153, respectively.

The advantage is consistent across all three reward-weight settings: GRAB achieves lower mean and median Vina Dock, lower strain quartiles, and higher SA and diversity. These gains persist when the starting checkpoint, sampler, training pockets, reward weights, and generation budget are matched, directly addressing the concern in W1/Q1. Together, the results support the effectiveness of GRAB's training procedure and its advantage over BFN-DPO in docking, strain, SA, and diversity, while BFN-DPO retains higher QED.

AuthorsOpenReview ↗

Response to Reviewer Bhge (W2/Q2)

W2/Q2: Training Variability and Paired-Pocket Uncertainty

Thank you for suggesting these analyses. We have added three independent training runs of GRAB-B and a paired-pocket comparison between GRAB-A and MolCRAFT. All four GRAB-B runs improve docking and strain relative to the initialization, and the paired-pocket confidence intervals favor GRAB-A across all reported metrics. The complete analyses are now included in Appendix F.

1. Repeated training of GRAB-B using multiple seeds

We repeated GRAB-B fine-tuning with three additional random seeds, starting from the same pretrained MolCRAFT checkpoint and keeping the reward weights and training settings fixed. Together with the original run, this gives four independent training runs.

Vina entries report mean / median, and SE entries report the 25th / 50th / 75th percentiles. Vina scores and SE are in kcal/mol.

MethodVina Score ↓Vina Min ↓Vina Dock ↓SE ↓SA ↑QED ↑Div ↑
MolCRAFT−6.307 / −6.842−7.089 / −7.106−7.808 / −7.90766 / 153 / 3460.690.530.74
GRAB-B, original run−7.593 / −8.428−8.759 / −8.736−9.515 / −9.53544 / 104 / 2000.760.510.75
GRAB-B, seed 1−7.683 / −8.456−8.962 / −9.040−9.724 / −9.78954 / 123 / 2510.760.530.73
GRAB-B, seed 2−7.157 / −8.105−8.444 / −8.461−9.113 / −9.23542 / 116 / 2340.770.540.74
GRAB-B, seed 3−7.478 / −8.212−8.436 / −8.481−9.177 / −9.25963 / 138 / 2690.750.540.76

All four runs improve docking scores and reduce all three strain-energy quartiles relative to MolCRAFT. Across runs, mean Vina Dock is −9.382 ± 0.288, compared with −7.808 for MolCRAFT, and median SE is 120.25 ± 14.20, compared with 153. Here, ± denotes the standard deviation across the four training runs. These results show that the docking and strain improvements are reproduced across the tested training seeds.

2. Paired-pocket comparison of GRAB-A and MolCRAFT

For each of the 100 test pockets, we computed the difference between GRAB-A and the reevaluated MolCRAFT baseline. Vina Min and SE use the median within each pocket. We averaged the paired differences with equal weight for each pocket and computed 95% BCa confidence intervals by resampling the test pockets, preserving the pairing between methods. The table reports the mean paired difference, the confidence interval, and the number of pockets on which GRAB-A performs better. Differences are defined as GRAB-A − MolCRAFT.

MetricMean paired difference95% BCa confidence intervalPockets favoring GRAB-A
Vina Dock, pocket mean ↓−0.733[−0.899, −0.584]88 / 100
Vina Dock, pocket median ↓−0.743[−0.873, −0.624]90 / 100
Vina Min, pocket median ↓−0.420[−0.586, −0.230]74 / 100
SE, pocket median ↓−55.5[−76.1, −45.1]94 / 100
QED ↑+0.022[+0.003, +0.041]67 / 100
SA ↑+0.084[+0.076, +0.093]99 / 100
Diversity ↑+0.019[+0.012, +0.025]69 / 100

All reported confidence intervals exclude zero in the direction favoring GRAB-A. In particular, GRAB-A improves mean Vina Dock on 88 of 100 pockets and median SE on 94 of 100 pockets. The paired analysis therefore supports improvements across target pockets in docking, strain, and molecular properties.

The docking and strain gains persist across the four tested GRAB-B training seeds, while the paired analysis supports GRAB-A's improvements over MolCRAFT across test pockets. We have added the full training-run results and paired-pocket analysis to Appendices F.1 and F.2, clarified the number of training runs in Section 5.1, and summarized the findings in Section 5.2. These changes are highlighted in blue. Thank you for helping us strengthen the evaluation.

AuthorsOpenReview ↗

Response to Reviewer Bhge (W3/Q3)

W3/Q3: Sample Counts, Baseline Sources, and Evaluation Settings

1. Sample counts and definitions

Each method below was evaluated with 100 generation attempts on each of the 100 held-out test pockets, giving 10,000 attempts per method, including failed attempts.

MethodRecon.Conn.RetainedVinaSE-scoredPB-evalPB-valid (%)
TargetDiff9,9069,2309,2259,9069,8749,90652.88%
IPDiff9,8269,2599,2589,8269,7999,82622.62%
IPDiff + DPO9,7829,0999,0909,7829,7339,78245.70%
PAFlow9,4318,2447,7308,7918,0489,43111.91%
MolFORM (Base)9,7878,3417,9749,3409,7639,78738.61%
MolFORM-RL9,9959,4108,8409,3809,8869,99542.37%
MolCRAFT9,9379,6879,6879,9379,9309,93785.60%
BFN-DPO (DPO-1)9,9959,8989,8989,9919,9879,99584.46%
BFN-DPO (DPO-2)9,9939,9489,9489,9939,9849,99383.87%
BFN-DPO (DPO-3)9,9969,9279,9279,9959,9949,99687.37%
GRAB-A9,9879,8739,8739,9879,9829,98784.07%
GRAB-B9,9959,9319,9319,9939,9829,99591.96%

Recon. counts reconstructed RDKit molecules with a conformer; Conn. counts molecules with a single connected fragment. Retained counts connected molecules with valid Vina Min, QED, SA, and atom-count results; SE is not required. Vina reports successful Vina Min scoring, and SE-scored reports successful SE scoring. PB-eval counts molecules evaluated with PoseBusters; PB-valid is the percentage passing all checks, with failed checks included in the denominator. The Retained count records the filtering outcome; each property summary uses its own successfully evaluated molecules.

2. Sources of baseline results

The MolCRAFT row in the manuscript's main table is copied from the original paper [1]. The MolCRAFT results in this response and the counts above come from our reevaluation of the pretrained checkpoint used to initialize GRAB. These counts refer to our reevaluation, not the published row.

We generated and evaluated molecules using the released TargetDiff, IPDiff, and PAFlow checkpoints. IPDiff + DPO uses our DPO implementation on IPDiff. We also implemented BFN-DPO [2] using the same initial MolCRAFT checkpoint as GRAB.

For MolFORM (Base) [4] (arXiv v1), we used the author-released molform_base_model.pt. For MolFORM-RL [5] (arXiv v3), we used checkpoints/molform-rl/final.pt, identified as NFT-Vina-SA in the authors' release description. Both rows use molecules generated and evaluated by us. We have updated Appendix B.4 with these baseline sources and checkpoint links, including the corrected source of the TargetDiff results.

3. Evaluation settings and summary statistics

We follow MolCRAFT's default evaluation settings [1, 3], using Vina 1.2.2, PoseCheck commit 57a1938, and PoseBusters 0.3.1. SE is computed with UFF on reconstructed, as-generated conformers. PoseBusters uses the dock configuration on as-generated poses against full receptors.

For our evaluations, Vina, SE, QED, and SA statistics are calculated from all successful scores for each property across the 100 pockets. Each molecule contributes equally; pockets with more successful scores contribute more observations. Missing results, NaN, and positive or negative infinity are excluded for the corresponding property only. For example, the reevaluated MolCRAFT Vina Min and SE summaries use 9,937 and 9,930 molecules, respectively.

The counts show that the numbers of successfully evaluated molecules differ across methods and properties. Property statistics describe these successful evaluations, while PB-valid uses all PB-evaluated molecules. We have added the sample counts and clarified the baseline sources, evaluation settings, and summary statistics in Appendices B.4–B.5. Appendix D and the pose-quality table now use the same PB-valid definition and corrected percentages. These changes are highlighted in blue. Thank you for pointing out these reporting gaps.

AuthorsOpenReview ↗

Response to Reviewer Bhge (References)

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Revisions

Summary of Rebuttal and Revisions

Dear Area Chair and Reviewers,

Thank you for the constructive feedback. During the rebuttal, we added controlled baseline comparisons, repeated training runs, stricter molecular evaluation, and experiments examining design choices and failure modes. These additions strengthen the evidence for GRAB as an effective approach to fine-tuning molecular BFNs from terminal molecular rewards.

1. The advantage persists under matched BFN-DPO comparisons.

We repeated the comparison using the same pretrained checkpoint, sampler settings, pocket allocation, scoring functions, and reward weights, with 76,800 training generation attempts per configuration. Across all three matched reward settings, including equal weights, GRAB achieves better Vina Dock, strain, SA, and diversity. Under the GRAB-B weights, mean Vina Dock is −9.52 versus −8.48, and median SE is 104 versus 120. Our additional REINFORCE comparison also favors GRAB-B in docking, strain, SA, and diversity, while both baselines retain higher QED. These results support the complete GRAB training procedure. (Section 5.2; Appendices B.4 and H.)

2. The improvements are reproduced across training seeds and test pockets.

All four independent GRAB-B runs improve docking and strain over the re-evaluated MolCRAFT initialization. Separately, the paired-pocket analysis of GRAB-A shows a mean Vina Dock improvement of −0.733, with a 95% confidence interval of [−0.899, −0.584]. GRAB-A improves mean Vina Dock on 88 of 100 pockets and median SE on 94 of 100 pockets. These analyses support repeatability and improvements across targets. (Appendix F.)

3. The gains remain after validity checks and deduplication.

At 100 generation attempts per pocket, GRAB-B yields 85.75 unique connected PB-valid molecules on average, compared with 83.23 for MolCRAFT. In the five-pocket case study, requiring connectivity, PoseBusters validity, uniqueness, and simultaneous improvement over the reference ligand on all four properties gives 1,293 qualifying molecules for GRAB-B versus 515 for MolCRAFT, a 2.51-fold gain under the same 2,500-attempt budget. We also document high repetition when generating small ligands as a failure mode. (Figure 3; Appendix D.)

4. We clarified the BFN-specific contribution and added direct evidence for log-density scaling.

GRAB defines sampled coordinate-belief means and atom-type logits as joint policy actions, connecting terminal molecular evaluations to updates of both prediction branches while retaining the original sampler. We clarified its distinction from CByG, BFN-DPO, and FTDiff. Changing only the scaling factor from D=384 to D=1, with the learning rate and other settings unchanged, worsens mean Vina Dock from −9.515 to −8.02 and median SE from 104 to 162. This supports the practical benefit of scaling under the tested settings. (Sections 2 and 4; Appendix J.)

5. The additional experiments establish both useful applications and important trade-offs.

Four single-reward experiments show that improving one property can damage others, most clearly through severe strain under Vina-Min-only optimization. GRAB-A/B provide a more balanced combination of docking and molecular properties. Single-pocket optimization on three test pockets, with two seeds each, further improves GRAB-B’s docking performance. These pockets were excluded from both MolCRAFT pretraining and GRAB-B multi-pocket fine-tuning before the single-pocket experiments. (Appendices G and I.)

The revised manuscript includes these results together with per-method evaluation counts, baseline sources, and metric definitions. The evidence supports GRAB’s effectiveness on the studied MolCRAFT backbone through controlled comparisons, repeated training, and stricter measures of useful molecular output, while making the observed trade-offs and limitations explicit.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.