acceptodds
Under review as a conference paper at ICLR 2027

FusedBFN: Probabilistic Composition for Dual-Target Molecular Generation

Abstract

Dual-target drug design aims to generate 3D molecules that can be compatible with two distinct protein binding pockets, yet the scarcity of dual-target structural data makes direct model training difficult. We study this problem in a zero-shot setting by employing pretrained single-target generative models and investigate where target-specific information should be composed within the generative process. We propose FusedBFN, which combines two target-aware sender likelihoods using a weighted Product of Experts before performing a shared Bayesian update. To examine composition at different probabilistic levels, we further introduce CompBFN and FusedDiff, which, together with CompDiff, instantiate distribution fusion and transition-kernel composition under both Bayesian flow and diffusion backbones. Our analysis reveals that seemingly different composition mechanisms can become equivalent under matched settings, while their differences otherwise depend on the induced uncertainty and generative dynamics. We also introduce a chemically weighted prior-based alignment method and a prior-free pocket alignment strategy to construct aligned dual-target contexts. Extensive experiments show that FusedBFN achieves strong dual-target generation performance, and further reveal that reducing sampling uncertainty around the fused signal can substantially improve average dual-target binding. These results clarify how the probabilistic formulation of composition interacts with the chosen generative framework in dual-target molecular design. The source code is provided at https://anonymous.4open.science/r/FusedBFN/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Dual-target molecular generation, Structure-based drug design, Bayesian flow networks, Diffusion models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9Gx2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FusedBFN, a zero-shot framework for dual-target molecular generation that reuses a pretrained single-target Bayesian Flow Network and combines target-specific information through a weighted Product of Experts (wPoE). The authors further introduce CompBFN and FusedDiff to compare distribution fusion with transition-kernel composition across Bayesian flow and diffusion frameworks. They also investigate the effects of sampling covariance and pocket alignment on generation performance.

Critical strengths
  1. The problem is practically relevant, and the proposed setting is well motivated. The scarcity of structural data for dual-target ligands makes training dedicated generative models difficult. Reusing pretrained single-target models without additional training or fine-tuning provides a practical approach to incorporating information from both binding pockets.
  2. The analysis of probabilistic composition mechanisms offers useful insights. Through FusedBFN, CompBFN, FusedDiff, and CompDiff, the paper examines how distribution fusion and transition-kernel composition behave under different generative frameworks. The equivalence analysis under matched parameter settings, together with controlled experiments on sampling covariance, helps clarify which factors contribute to the observed improvements.
  3. The experimental evaluation is broad and includes complementary assessments. The benchmark contains 12,917 target pairs. In addition to AutoDock Vina scores, the authors report GNINA-based evaluations, conformational strain energy, steric clashes, and molecular properties. Ablations on alignment strategies, fusion weights, and sampling covariance, together with statistical analyses accounting for repeated targets, provide complementary evidence for the main findings.
Critical weaknesses

1 The handling of incompatible target pairs remains unclear Appendix F.5 analyzes disagreement between the two target-conditioned experts and presents two examples in which retaining moderate sampling noise performs better than deterministic sampling. However, this does not establish whether both pockets admit a feasible ligand. Under specified constraints on molecular size, chemical validity, and binding affinity, could the requirements of the two pockets be sufficiently incompatible that no molecule satisfies both? How would the current method identify and handle such cases? Could fusion still produce compromise molecules that bind poorly to both targets? A controlled evaluation of incompatible cases, together with an analysis of the relationship between expert disagreement and dual-target generation success, would help clarify the method’s applicability.

2 The efficiency advantage over single target generation followed by docking based filtering has not been established Appendix F.11 and Table 21 report generation times of 113.68 seconds for FusedBFN, 57.27 seconds for MolCRAFT, and 1128.40 seconds for DualDiff when generating 10 molecules. These results demonstrate a generation-speed advantage over DualDiff, but they do not establish an advantage over generating candidates for one target and subsequently filtering them using Vina docking scores against the second target. The fixed-sample comparison with single-target models in Table 2 does not replace this comparison. Please compare the two pipelines under an equal computational budget or at the same number of qualifying dual-target candidates, including the costs of probe-ligand generation, pocket alignment, docking, and filtering. This would clarify whether joint generation reduces end-to-end runtime or computational resource requirements.

3 The alignment ablations do not fully address sensitivity to the initial alignment or the search for an optimal alignment Table 5 compares several alignment strategies, and Appendix F.2 shows that probe-ligand selection affects generation performance. Nevertheless, the alignment that maximizes geometric overlap or minimizes probe-ligand registration error may not maximize downstream generation quality. The current method treats alignment as fixed preprocessing and does not search over or update the relative rigid-body transformation between the pockets during generation. Could the authors compare multiple candidate alignments for the same target pair, examine the relationship between registration quality and generation performance, and assess sensitivity to initial rotations, translations, and probe-ligand selection? Can the model itself explore a more suitable alignment based on the generation objective?

4 It remains unclear whether the generated conformation satisfies the steric constraints of both pockets under the original alignment Appendix F.8 reports steric-clash and conformation metrics, but it should be clarified whether these results directly assess the original shared conformation produced during generation. Specifically, if the generated ligand coordinates and the pocket alignment used for generation are preserved, does the ligand exhibit substantial steric clashes with either pocket when each ligand–pocket configuration is evaluated separately? If severe clashes remain, would this indicate that the joint constraints were not adequately satisfied during generation, despite both pockets being provided as conditioning inputs? Independent re-docking can alleviate clashes by changing the ligand’s position or conformation, so favorable docking scores may not directly demonstrate compatibility of the original generated conformation with both pockets. Please report the fraction of original generated conformations exhibiting clashes with each pocket and the fraction passing steric-clash checks for both pockets, alongside the corresponding results after optimization or re-docking.

Questions for authors

The questions were provided in the Critical Weaknesses section.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 9Gx2 - W1

W1: The handling of incompatible target pairs remains unclear.

A to W1: We thank the reviewer for this valuable suggestion. To clarify the applicability of FusedBFN to potentially incompatible target pairs, we conduct two additional analyses: (1) examining the relationship between expert disagreement and dual-target generation performance, and (2) evaluating FusedBFN on challenging target pairs for which single-target generation fails to identify a shared feasible ligand.

Relationship between disagreement and generation performance. We evaluate 1,000 target pairs and compute the coordinate disagreement score DpairxD_{\mathrm{pair}}^x and categorical conflict score CpairvC_{\mathrm{pair}}^v, as defined in Appendix F.5. For each target pair, we calculate mean Max Vina Dock and Dual High Affinity (DHA), and examine their relationships with the two disagreement measures using Spearman correlation. Both disagreement measures show statistically significant correlations, indicating that stronger expert disagreement is associated with less favorable binding to the target.

DisagreementMax Vina DockDHA
DpairxD_{\mathrm{pair}}^xρ=0.217, p<0.05\rho=0.217,\ p<0.05ρ=−0.161, p<0.05\rho=-0.161,\ p<0.05
CpairvC_{\mathrm{pair}}^vρ=0.189, p<0.05\rho=0.189,\ p<0.05ρ=−0.150, p<0.05\rho=-0.150,\ p<0.05

For each target pair pp, we further compute Δp=MaxDock⁡p(c=0.25)−MaxDock⁡p(c=0)\Delta_p = \operatorname{MaxDock}_p(c=0.25) - \operatorname{MaxDock}_p(c=0), where MaxDock⁡p\operatorname{MaxDock}_p denotes the average Max Vine Dock score across generated molecules. A negative Δp\Delta_p indicates better performance with c=0.25c=0.25. We then calculate Spearman correlations between Δp\Delta_p and two disagreement measures. Both measures exhibit significant negative correlations with Δp\Delta_p. This suggests that moderately increasing the sampling variance can improve generation performance on target pairs with greater expert disagreement.

DisagreementSpearman ρ\rhop-value
DpairxD_{\mathrm{pair}}^x vs. Δp\Delta_p-0.117<0.05<0.05
CpairvC_{\mathrm{pair}}^v vs. Δp\Delta_p-0.104<0.05<0.05

Performance on potentially incompatible target pairs. To investigate whether FusedBFN can generate feasible dual-target ligands when the two targets are incompatible, we construct a candidate pool L\mathcal L from molecules generated by single-target model MolCRAFT. For each target pair, we define the feasible candidate sets: Fj={m∈L:Ej(m)≤Tj, m is valid and satisfies the molecular size constraint}\mathcal F_j=\{m\in\mathcal L:E_j(m)\leq T_j,\ m\text{ is valid and satisfies the molecular size constraint}\}, where Ej(m)E_j(m) denotes the Vina Dock score of molecule mm against target jj, and TjT_j is the Vina Dock score of the corresponding reference ligand. The molecular size constraint is defined using the size range of the reference ligands in the test set. Across all target pairs in the test set, we identify 260 challenging pairs satisfying

$\mathcal F_1\neq\varnothing,\qquad \mathcal F_2\neq\varnothing,\qquad \mathcal F_1\cap\mathcal F_2=\varnothing.$

For these pairs, the candidate pool contains feasible molecules for each target, but no molecule simultaneously satisfies the constraints for both targets. The two conflicting cases in Appendix F.5 are among the 260 pairs. We then evaluate FusedBFN on these 260 pairs under the same chemical validity, molecular size, and binding affinity constraints. A target pair is considered successfully addressed if at least one generated molecule satisfies the constraints for both targets simultaneously. FusedBFN achieves a success rate of 58.85%, demonstrating its ability to discover shared feasible candidates even when the two targets are incompatible under independent single-target generation.

Overall, expert disagreement is associated with reduced dual-target performance, while moderately increasing sampling variance can partially mitigate this effect for pairs with greater disagreement. More importantly, FusedBFN can discover shared feasible candidates for incompatible target pairs, highlighting its broad applicability to challenging dual-target molecular design.

AuthorsOpenReview ↗

Response to Reviewer 9Gx2 - W2

W2: The efficiency advantage over single target generation followed by docking based filtering has not been established.

A to W2: We thank the reviewer for raising this important concern. To address this, we conduct an additional pipeline-level comparison between (i) single-target generation using MolCRAFT followed by dual-target docking-based filtering and (ii) FusedBFN with the complete probe-preparation, dual-target generation, docking, and filtering procedure. Both pipelines are evaluated on the same set of 1,000 randomly selected target pairs.

For both pipelines, a molecule is considered qualifying if its Vina Dock scores against both targets are better than those of their respective reference ligands. We deduplicate the qualifying molecules within each target pair using canonical isomeric SMILES, preserving stereochemical differences. Both pipelines are evaluated under identical computational resources and runtime settings. Molecule generation is performed on one TITAN Xp GPU per target pair. Docking is performed on 64-core CPU nodes, with 16 CPU cores and 8 GB of memory allocated to each target pair. All remaining stages run on a single CPU core. We report the time required by each stage below.

StageMolCRAFT + Filtering (s)FusedBFN (s)
Probe-ligand preparation and pocket alignment–22,551
Molecule generation16,09199,270
Vina docking against P-11,1312,190
Vina docking against P-24,3601,922
Filtering13.7813.83
Total runtime21,595.78125,946.83

The resulting numbers of unique qualifying molecules and their coverage across target pairs are summarized below.

MetricMolCRAFT + FilteringFusedBFN
Qualifying molecule rate32.96%41.78%
Runtime per qualifying molecule (s)7.6635.22
Proportion with ≥1 qualifying molecule71.85%79.67%
Proportion with ≥3 qualifying molecules52.22%61.10%
Proportion with ≥5 qualifying molecules35.05%44.51%

The improvement of FusedBFN is most pronounced under the stricter requirement of ≥5 qualifying molecules per target pair, where coverage increases from 35.05% to 44.51%. FusedBFN requires 35.22 seconds per qualifying molecule, compared with 7.66 seconds for MolCRAFT followed by filtering. However, the results on incompatible target pairs in A to W1 demonstrate that FusedBFN can explore shared feasible regions that the MolCRAFT pipeline fails to identify. Computational efficiency is meaningful only when a method can successfully find valid solutions. Although FusedBFN incurs a higher computational cost, it can generate feasible molecules that single-target pipeline cannot discover. Therefore, FusedBFN offers substantial practical value despite its higher computational cost.

AuthorsOpenReview ↗

Response to Reviewer 9Gx2 - W3

W3: The alignment ablations do not fully address sensitivity to the initial alignment or the search for an optimal alignment.

A to W3: We thank the reviewer for this suggestion. We clarify that Table 5 already compares different alignment strategies on the same 1,000 target pairs, with Lig-ChemPrior achieving the best generation performance. Surface-RI serves as a practical alternative when a probe ligand is difficult to obtain or when a simpler alignment pipeline is preferred. To examine sensitivity to initial alignment and the relationship between registration quality and generation performance, we conduct additional experiments on 200 target pairs, evaluating variant alignments for each pair while keeping the generation settings unchanged.

Starting from the original chemically weighted alignment, we apply relative rotations of 10° and 20°, relative translations of 2 Å and 4 Å, and an alternative alignment obtained using an unrelated probe ligand. The original alignment achieves the best overall binding performance. Small rotations cause limited degradation, while translation perturbations have a more pronounced effect. Using an unrelated probe ligand further reduces performance. Meanwhile, QED, SA, and Diversity remain relatively stable across these settings. These results demonstrate that FusedBFN tolerates moderate alignment perturbations, although larger deviations and inappropriate probe selection can compromise dual-target binding performance.

AlignmentP-1 Vina Dock ↓P-2 Vina Dock ↓Max Vina Dock ↓Dual High Aff. ↑QED ↑SA ↑
Original-8.76-8.79-8.1052.9%0.540.67
Rotation 10°-8.75-8.78-8.0752.3%0.540.68
Rotation 20°-8.66-8.76-7.9852.0%0.540.68
Translation 2 Å-8.71-8.63-7.9953.2%0.540.68
Translation 4 Å-8.53-8.56-7.8650.0%0.550.69
Wrong Probe-8.58-8.57-7.8949.2%0.550.70

We further quantify registration quality using the nearest-neighbor RMSD between the two aligned pockets. We compute the Spearman correlation between RMSD and Max Vina Dock, obtaining a mean correlation of 0.197 (p<0.05p<0.05). This suggests that poorer geometric alignment tends to be associated with weaker binding performance. In FusedBFN, the relative transformation is determined during preprocessing and remains fixed throughout sampling. An interesting extension would be to search over candidate relative transformations using a generation-aware objective that jointly considers dual-target binding and molecular quality, which is a promising direction for future work.

Reviewer_rkaJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes FusedBFN, a framework that adapts a pretrained single-target BFN to dual-target molecular generation without additional training. It combines target-specific sender likelihoods before a shared Bayesian update and analyzes distribution fusion versus transition-kernel composition across BFN and diffusion backbones. Experiments examine the effects of sampling uncertainty and pocket alignment.

Critical strengths

Clear motivation and systematic comparisons. Reusing pretrained generators is practical given limited dual-target structural data. Comparing two composition strategies across two backbones directly addresses the research question.

Useful theoretical and empirical analysis. The paper establishes conditions under which different composition mechanisms coincide and explains their differences through covariance and categorical updates.

Extensive evaluation. The experiments include large-scale target-pair evaluation, concentration and alignment ablations, GNINA scoring, conformational analysis, and statistical tests accounting for recurring targets.

Critical weaknesses
  • The novelty of weighted PoE and weighted alignment is limited; the main contribution lies in their formulation and comparative analysis.

  • The baseline coverage is largely restricted to composition-based approaches. The paper does not evaluate alternatives such as CByG, adapted to favor binding to both targets through property gradients. Consequently, the performance and efficiency advantages over gradient-based guidance remain unclear.

  • The analysis of expert disagreement relies on a small number of cases, leaving the generality of covariance selection unresolved. Moreover, computational evaluations do not establish experimental dual-target activity.

Questions for authors

I currently lean toward Weak Reject. Although the analysis of probabilistic composition is valuable, the absence of comparisons with gradient-guided alternatives adapted to dual-target generation leaves the practical advantages insufficiently established. I would be willing to raise my score if the rebuttal convincingly addresses these concerns or provides additional comparative evidence.

3 comments
AuthorsOpenReview ↗

Response to Reviewer rkaJ — W1 and Q1

W1+Q1: (W1) The novelty of weighted PoE and weighted alignment is limited; the main contribution lies in their formulation and comparative analysis. (Q1) Given the equivalence between FusedBFN and CompBFN under matched settings, please clarify the methodological contribution separately from the analytical contribution.

A to W1+Q1: We thank the reviewer for this comment. We agree that wPoE and weighted alignment build upon established techniques. Our contributions lie in their formulation for zero-shot dual-target generation and, more importantly, in the systematic investigation of probabilistic composition mechanisms across generative frameworks.

Methodological contribution. When reusing a pretrained single-target model for zero-shot dual-target generation, there are two possible strategies: fusing target-specific information before or after a generative update. Prior work [1] has only explored the latter within diffusion models. We investigate both composition strategies under BFN and diffusion backbones. Specifically, we propose using wPoE to fuse target-aware sender likelihoods before a shared Bayesian update. Furthermore, we introduce FusedBFN, FusedDiff, and CompBFN, which, together with the existing CompDiff, establish the first unified 2×22\times2 framework for studying composition level and generative backbone in dual-target molecular generation. We also improve the existing alignment by incorporating the chemical semantics of probe-ligand atoms, and introduce a prior-free alignment strategy that provides a practical alternative when probe ligands are difficult to obtain.

Analytical contribution. Beyond constructing these methods, we theoretically characterize the relationships between the two composition mechanisms, which was not systematically studied in prior work. Under their default settings, FusedBFN uses w1=w2=0.5w_1=w_2=0.5, supported by the experiments in Appendix F.3, while CompBFN adopts η=0.5\eta=0.5, following the CompDiff setting in [1]. Tables 2 and 29 show that FusedBFN significantly outperforms CompBFN under the same alignment strategy. To understand this performance gap, we examine their probabilistic formulations and find that FusedBFN and CompBFN share the same fused mean but differ in sampling covariance. Specifically, their covariance scales are w12+w22w_1^2+w_2^2 (when w1+w2=1w_1+w_2=1) and 1/(2η)1/(2\eta), respectively. Their covariances coincide when w12+w22=1/(2η)w_1^2+w_2^2=1/(2\eta).

Motivated by this finding, we introduce a covariance scale cc to further investigate the effect of sampling concentration. Here, cc serves as an explicit experimental control and is not an additional parameter derived from wPoE. Table 4 shows that reducing cc consistently improves affinity-related metrics, with the best performance achieved at c=0c=0. Based on this observation, we evaluate FusedBFN-Det (c=0c=0) on the full benchmark, where it achieves the strongest overall binding performance (Table 2). These results identify sampling covariance as a key factor affecting dual-target generation performance when the sampling center is reliable. Moreover, CompBFN uses a fixed sampling center, which may be suboptimal. In contrast, the experiments in Appendix F.3 demonstrate that FusedBFN can adjust the sampling center toward more favorable regions through w1w_1 and w2w_2, highlighting its advantage in adapting to different target pairs.

We also analyze FusedDiff and CompDiff. Under their default settings, their coordinate sampling distributions are identical, while their categorical distributions differ. Section 5.3 shows that FusedDiff yields softer categorical distributions with lower confidence in the most probable atom types. Furthermore, the two methods apply composition at different stages relative to forward corruption, as analyzed in Appendix C.4 and further explored in Appendix F.4. These differences lead to distinct categorical sampling behaviors and help explain their performance differences.

Conclusion. Methodologically, we propose using wPoE to fuse target-aware likelihoods before a shared generative update. We develop FusedBFN and two variants, FusedDiff and CompBFN, which, together with CompDiff, establish a unified 2×22\times2 framework. Analytically, we investigate how distribution fusion and transition-kernel composition behave across different generative backbones through theoretical derivations and empirical experiments, and identify the factors underlying their differences. Our work provides a foundation for future research on dual-target molecular generation, offering insights and guidance for the development of new composition methods.

[1] Zhou X, Guan J, Zhang Y, et al. Reprogramming pretrained target-specific diffusion models for dual-target drug design[J]. Advances in Neural Information Processing Systems, 2024, 37: 87255-87281.

AuthorsOpenReview ↗

Response to Reviewer rkaJ — W2, Q2 and Q3

W2+Q2+Q3: (W2) The baseline coverage is largely restricted to composition-based approaches. The paper does not evaluate alternatives such as CByG, adapted to favor binding to both targets through property gradients. Consequently, the performance and efficiency advantages over gradient-based guidance remain unclear. (Q2): A comparison with a gradient-guided method such as CByG, adapted to favor binding to both targets, would strengthen the evaluation. (Q3) Please explain its omission and, if feasible, provide results comparing dual-target affinity, molecular quality, and generation cost under comparable conditions.

A to W2+Q2+Q3: Thank you for this suggestion. CByG [1] is originally developed for single-target property-guided molecular generation and does not directly support dual-target generation. Moreover, the official generated molecules, complete implementation, and required checkpoints are not available to us. We contact the authors to request the missing resources but have not received a response. To address this concern, we reproduce CByG based on its paper and the available partial code, with AI-assisted implementation(Grok 4.7) and manual verification. The code and checkpoints are available at https://anonymous.4open.science/r/CByG_code/.

We further implement two dual-target variants: CByG-Dual retains the original single-target generator and replaces single-target property guidance with joint gradient guidance toward both targets. CByG-Fused integrates the wPoE fusion mechanism of FusedBFN into CByG. We evaluate these methods under the same experimental protocol, using binding affinity and molecular quality metrics consistent with our main experiments.

MethodsP-1 Vina Dock ↓P-2 Vina Dock ↓Max Vina Dock ↓Dual High Aff. ↑QED ↑SA ↑Diversity ↑
CByG-9.38-7.95-7.6357.3%0.540.700.70
MolCRAFT-9.03-7.66-7.3848.1%0.500.660.70
CByG-Dual-9.09-7.63-7.3850.7%0.500.660.70
CByG-Fused-9.32-9.34-8.5267.1%0.590.770.72
FusedBFN-8.60-8.63-7.9556.5%0.530.660.69

CByG improves upon MolCRAFT through property-gradient guidance, particularly on P-1. However, its performance remains imbalanced between the two targets. CByG-Dual does not resolve this imbalance, achieving a Dual High Affinity of 50.7% and a Max Vina Dock score of -7.38. This suggests that directly combining target-specific property gradients is insufficient to achieve balanced dual-target binding. In contrast, CByG-Fused achieves the strongest overall performance, with a Max Vina Dock score of -8.52 and a Dual High Affinity of 67.1%, while also improving molecular properties. These results indicate that our probabilistic fusion mechanism is complementary to gradient-based guidance and can be effectively extended to methods that use BFNs as their generative framework.

We additionally measure the average generation time for 10 molecules on an NVIDIA RTX 4090 under the same runtime settings. FusedBFN offers a favorable balance between generation cost and dual-target affinity, while combining probabilistic fusion with gradient guidance achieves further improvements at additional computational cost. We will incorporate these comparisons into the revised manuscript.

MethodMolCRAFTCByGCByG-FusedCByG-DualFusedBFN
Time (s)19376810640

[1] Choi S, Kim H, Park C, et al. Controllable 3d molecular generation for structure-based drug design through bayesian flow networks and gradient integration[J]. Advances in Neural Information Processing Systems, 2026, 38: 125612-125647.

AuthorsOpenReview ↗

Response to Reviewer rkaJ — W3 and Q4

W3.1+Q4.2: (W3.1) The analysis of expert disagreement relies on a small number of cases, leaving the generality of covariance selection unresolved. (Q4.2) To what extent do the expert-disagreement results support a general principle for selecting covariance scales across target pairs?

A to W3.1+Q4.2: We thank the reviewer for this comment. Appendix F.5 presents two cases where moderate covariance improves generation under strong expert disagreement. To examine whether this observation generalizes across target pairs, we further analyze the relationship between disagreement and the benefit of introducing sampling covariance. For each target pair pp, we compute Δp=MaxDockp(c=0.25)−MaxDockp(c=0)\Delta_p = MaxDock_p(c=0.25) - MaxDock_p(c=0), where MaxDockpMaxDock_p denotes the average Max Vina Dock score across generated molecules. A negative Δp\Delta_p indicates better performance with c=0.25c=0.25. We then calculate Spearman correlations between Δp\Delta_p and two expert-disagreement measures: mean coordinate RMSD ($D_{\mathrm{pair}}^x )andcategoricalexpertconflictscore() and categorical expert conflict score ( C_{ \mathrm{pair}} ^v $).

Disagreement measureSpearman ρ\rhop-value
DpairxD_{\mathrm{pair}}^x vs. Δp\Delta_p-0.117<0.05<0.05
CpairvC_{\mathrm{pair}}^v vs. Δp\Delta_p-0.104<0.05<0.05

Both measures exhibit statistically significant negative correlations with Δp\Delta_p. This suggests that stronger expert disagreement is associated with a greater benefit from introducing covariance. These findings indicate a practical guideline for covariance selection. Since strong expert disagreement is uncommon (Appendix F.5) and c=0c=0 achieves the best overall benchmark performance, we suggest using c=0c=0 in most cases. For target pairs exhibiting substantial disagreement, moderate covariance within c∈(0,0.5]c\in(0,0.5] may be worth exploring. Since disagreement alone is insufficient to reliably determine the optimal cc for an individual target pair, we recommend evaluating a small set of cc using docking scores to select promising candidates when computational resources permit.

Overall, our results provide preliminary evidence for covariance selection, while a reliable adaptive selection criterion remains an open direction for future work.

W3.2: Moreover, computational evaluations do not establish experimental dual-target activity.

A to W3.2: We acknowledge the limitations of docking-based evaluation. Our work focuses on the probabilistic composition of pretrained single-target generators, aiming to understand how different fusion mechanisms affect dual-target molecular generation. Evaluations across 12,917 target pairs provide computational evidence of improved predicted binding affinity toward both targets, while biochemical activity remains to be experimentally verified. Such validation is beyond the scope of the present methodological study and remains an important direction for future work, as acknowledged in Section 6.

Q4.1: Please clarify which sources of stochasticity are removed in FusedBFN-Det and which remain, including final atom-type sampling.

A to Q4.1: In FusedBFN-Det, setting c=0c=0 removes the additive Gaussian perturbations from both the coordinate updates and the atom-type messages, but does not eliminate all sources of randomness. The process is initialized with fixed zero coordinate parameters and uniform atom-type probabilities, while the number of atoms is sampled from the pocket priors. At each intermediate step, atom types are sampled from the predicted categorical distribution, and these stochastic draws influence both subsequent type and coordinate predictions. In the final step, the coordinates are averaged across experts, and the atom types are sampled from the averaged expert probabilities.

Reviewer_eZjjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies zero-shot dual-target 3D molecular generation: instead of training on scarce dual-target complex data, it reuses pretrained single-target generative models and asks at which probabilistic level the two target-specific signals should be combined. The proposed FusedBFN fuses two target-aware sender likelihoods with a weighted product of experts (wPoE) and then performs a single shared Bayesian update, building on a pretrained BFN (MolCRAFT). To compare composition levels, the authors additionally introduce CompBFN (transition-kernel composition ported to the BFN backbone) and FusedDiff (wPoE plus a diffusion backbone), which — together with the existing CompDiff — give a 2x2 grid of {distribution fusion, kernel composition} × {BFN, diffusion}. A theoretical analysis argues that under matched settings FusedBFN and CompBFN can become exactly equivalent, that their default configurations share the same fused center and differ mainly in covariance, and that under diffusion the coordinate transitions can coincide while categorical updates remain distinct. The paper also contributes two pocket-alignment strategies: a chemically weighted ligand-prior alignment and a prior-free surface-based one. Validation uses the dual-target benchmark of Zhou et al. (2024a) with 12,917 target pairs, 10 molecules per pair, evaluated with Vina docking scores on each pocket, dual high-affinity rate, QED, SA and diversity, plus ablations on sampling concentration, alignment strategy, and composition variants.

Critical strengths
  1. The composition-level question is well posed and the 2x2 design is genuinely useful. Rather than proposing one more fusion heuristic, the paper separates where fusion happens (distribution vs. transition kernel) from which backbone carries it (BFN vs. diffusion), and instantiates all four cells with the same pretrained checkpoints. The derived equivalence between FusedBFN and CompBFN under matched settings, and the observation that their default configurations differ primarily in induced covariance, is a clean and non-obvious clarification of why prior dual-target methods could not be compared on equal footing. This is the kind of negative/structural result that helps the subfield calibrate future claims.

  2. The empirical scale and the honesty of the analysis are above the norm for this task. Evaluation over 12,917 target pairs with per-pocket docking, dual high-affinity rate, and property/diversity metrics is far broader than the small case studies common in dual-target papers, and the paper reports the ablation on sampling concentration c rather than hiding it — including the finding that performance improves monotonically as c decreases, and that moderate covariance can help when the two experts disagree (Appendix F.5). Reporting this, even though it complicates the headline story, is a strength.

Critical weaknesses
  1. The reported gains appear to be driven mostly by sampling-variance suppression, not by the composition operator the paper is about. Section 5.4 states that "all affinity-related metrics improve as c decreases", and the best configuration, FusedBFN-Det, is simply the one that sets c = 0 (i.e., removes the fused covariance entirely). But c = 0 is orthogonal to the paper's central claim about composition level: it is a knob on sampling noise that could presumably be applied to CompDiff or DualDiff as well. As written, the paper's strongest numbers therefore do not isolate the contribution it advertises. To support the headline claim, the paper should either (a) apply the same c-sweep to the diffusion baselines so that the composition comparison is made at matched sampling noise, or (b) report the composition-level comparison only at fixed c, and state explicitly how much of the FusedBFN-over-CompBFN gap survives.

  2. The "zero-shot" framing is in tension with the reliance on per-target reference ligands, and the practical claim is therefore overstated. Both alignment strategies need an interaction prior or probe ligand per pocket (and the chemically weighted variant, which the paper shows is the best, is explicitly ligand-prior based; Surface-RI is described as the fallback "when reliable probe ligands are unavailable"). In the realistic dual-target scenario that motivates the paper — a new target pair with little structural data — such ligands are precisely what is missing. Either the abstract/introduction should be qualified to "zero-shot with respect to dual-target complexes, but assuming a reference ligand per target", or the paper should quantify degradation when the ligand prior is absent or mispaired. As it stands, a reader could reasonably conclude the method needs no per-target annotations at all.

Questions for authors
  1. What is the composition-level comparison (FusedBFN vs. CompBFN vs. FusedDiff vs. CompDiff) at a fixed c, e.g. the default c used by CompDiff? Put differently: how much of FusedBFN's advantage over CompBFN survives once FusedBFN-Det's c = 0 trick is applied to both? A table with rows = {4 composition variants} and columns = {c = default, c = 0} would settle this directly.

  2. How sensitive are the results to the alignment prior? Please report performance with (i) a deliberately mismatched ligand prior, and (ii) Surface-RI only, on the same target-pair subset, and state whether the ranking of the composition variants changes.

  3. The equivalence result between FusedBFN and CompBFN is stated for "matched settings". What exactly are the matching conditions, and are they reachable by the default hyperparameters of the two methods or only in a limiting case? If only a limiting case, what is the practical implication for a practitioner choosing between them?

  4. When the two target experts disagree strongly (Appendix F.5), how should c be chosen at inference time without access to ground-truth affinity? Is there a data-driven criterion, or is it a per-pair hyperparameter search?

  5. "Dual High Affinity" is defined relative to the reference ligands of each pair. How are pairs handled where the two reference ligands have very different affinity levels, and could this metric favour molecules that simply resemble one of the two references?

8 comments
AuthorsOpenReview ↗

Response to Reviewer eZjj — W1 and Q1 — Part 1/3

W1+Q1: (W1) The reported gains appear to be driven mostly by sampling-variance suppression, not by the composition operator the paper is about. Section 5.4 states that "all affinity-related metrics improve as c decreases", and the best configuration, FusedBFN-Det, is simply the one that sets c = 0 (i.e., removes the fused covariance entirely). But c = 0 is orthogonal to the paper's central claim about composition level: it is a knob on sampling noise that could presumably be applied to CompDiff or DualDiff as well. As written, the paper's strongest numbers therefore do not isolate the contribution it advertises. To support the headline claim, the paper should either (a) apply the same c-sweep to the diffusion baselines so that the composition comparison is made at matched sampling noise, or (b) report the composition-level comparison only at fixed c, and state explicitly how much of the FusedBFN-over-CompBFN gap survives. (Q1) What is the composition-level comparison (FusedBFN vs. CompBFN vs. FusedDiff vs. CompDiff) at a fixed c, e.g. the default c used by CompDiff? Put differently: how much of FusedBFN's advantage over CompBFN survives once FusedBFN-Det's c = 0 trick is applied to both? A table with rows = {4 composition variants} and columns = {c = default, c = 0} would settle this directly.

A to W1+Q1: We thank the reviewer for this comment. In fact, identifying sampling covariance as a key factor underlying the performance differences is an important finding of our analysis of probabilistic composition, which emerged from comparing the two composition mechanisms.

Default composition comparison. Our investigation begins with FusedBFN and CompBFN under their respective default settings. CompBFN adopts η=0.5\eta=0.5, following the CompDiff configuration [1], while FusedBFN uses w1=w2=1/2w_1=w_2=1/2, supported by the weight ablation in Appendix F.3. Under the same pocket alignment, Table 2 shows that FusedBFN outperforms CompBFN in dual-target binding, with statistically significant improvements confirmed by the paired tests in Table 29.

Why covariance matters. To understand this difference, we analyze the two composition formulations. When w1+w2=1w_1+w_2=1, FusedBFN gives

μ∼N(βx(τ)(w1x^M(1)+w2x^M(2))1+βx(τ),βx(τ)(w12+w22)[1+βx(τ)]2I),Y~v[n]∼N(βv(τ)[K(w1v^M(1)[n]+w2v^M(2)[n])−1K],(w12+w22)βv(τ)KIK),(A⁡1.1)\boldsymbol{\mu}\sim \mathcal{N}\left(\frac{\beta^x(\tau)(w_1\hat{\mathbf{x}}_M^{(1)}+w_2\hat{\mathbf{x}}_M^{(2)})}{1+\beta^x(\tau)},\frac{\beta^x(\tau)(w_1^2+w_2^2)}{[1+\beta^x(\tau)]^2}\mathbf{I}\right), \widetilde{\mathbf{Y}}^v[n]\sim\mathcal{N} \left( \beta^v(\tau)[K(w_1\hat{\mathbf{v}}_M^{(1)}[n]+w_2\hat{\mathbf{v}}_M^{(2)}[n])-\mathbf{1}_K],(w_1^2+w_2^2)\beta^v(\tau)K\mathbf{I}_K \right), (\operatorname{A}1.1)

while CompBFN gives

μ∼N(βx(τ)1+βx(τ)x^M(1)+x^M(2)2,βx(τ)2η[1+βx(τ)]2I),Yv[n]∼N(βv(τ)[Kv^M(1)[n]+v^M(2)[n]2−1K],βv(τ)K2ηIK).(A⁡1.2)\boldsymbol{\mu}\sim\mathcal{N}\left(\frac{\beta^x(\tau)}{1+\beta^x(\tau)}\frac{\hat{\mathbf{x}}_M^{(1)}+\hat{\mathbf{x}}_M^{(2)}}{2},\frac{\beta^x(\tau)}{2\eta[1+\beta^x(\tau)]^2}\mathbf{I}\right), \mathbf Y^v[n]\sim\mathcal N\left(\beta^v(\tau)[K\dfrac{\hat{\mathbf v}_M^{(1)}[n]+\hat{\mathbf v}_M^{(2)}[n]}{2}-\mathbf 1_K],\dfrac{\beta^v(\tau)K}{2\eta} \mathbf I_K\right). (\operatorname{A}1.2)

Under w1=w2=1/2w_1=w_2=1/2 and η=1/2\eta=1/2, the two formulations yield identical fused means but different covariance scales:

w12+w22=1/2w_1^2+w_2^2=1/2 for FusedBFN and 1/(2η)=11/(2\eta)=1 for CompBFN.

Furthermore, Eqs. (A1.1) and (A1.2) show that FusedBFN and CompBFN yield identical sampling distributions when w1w_1, w2w_2, and η\eta are chosen such that their means and variances coincide, as in the case of w1=w2=1/2w_1=w_2=1/2 and η=1\eta=1.

These theoretical observations motivate the controlled experiment in Table 4, where we explicitly introduce cc to vary covariance while keeping the fused means computed from x^M(1)+x^M(2)2\frac{\hat{\mathbf{x}}_M^{(1)}+\hat{\mathbf{x}}_M^{(2)}}{2} and v^M(1)[n]+v^M(2)[n]2\frac{\hat{\mathbf v}_M^{(1)}[n]+\hat{\mathbf v}_M^{(2)}[n]}{2}. Both FusedBFN and CompBFN can then be expressed as

μ∼N(βx(τ)1+βx(τ)x^M(1)+x^M(2)2,cβx(τ)[1+βx(τ)]2I),Y~v[n]∼N(βv(τ)[Kv^M(1)[n]+v^M(2)[n]2−1K],cβv(τ)KIK).(A⁡1.3)\boldsymbol{\mu}\sim \mathcal{N}\left(\frac{\beta^x(\tau)}{1+\beta^x(\tau)}\frac{\hat{\mathbf{x}}_M^{(1)}+\hat{\mathbf{x}}_M^{(2)}}{2},\frac{c\beta^x(\tau)}{[1+\beta^x(\tau)]^2}\mathbf{I}\right), \widetilde{\mathbf{Y}}^v[n]\sim\mathcal{N} \left( \beta^v(\tau)[K\dfrac{\hat{\mathbf v}_M^{(1)}[n]+\hat{\mathbf v}_M^{(2)}[n]}{2}-\mathbf{1}_K],c\beta^v(\tau)K\mathbf{I}_K \right). (\operatorname{A}1.3)

Importantly, cc is a manually introduced parameter, not an additional parameter derived from either composition formulation. The covariances in Eqs. (A1.1) and (A1.2) are mathematically determined by w1,w2w_1,w_2 and η\eta, respectively, whereas cc is introduced solely to examine how changing the covariance affects generation performance under the same fused mean. It has no independent theoretical interpretation. The results in Table 4 reveal that reducing sampling covariance generally improves dual-target binding, with the best average performance achieved at c=0c=0. FusedBFN-Det is subsequently introduced to evaluate this deterministic setting on the full benchmark.

AuthorsOpenReview ↗

Response to Reviewer eZjj — W1 and Q1 — Part 2/3

Effect of sampling center. However, deterministic sampling performs well only when the sampling center is accurate. To examine this, we fix the sampling covariance at zero and randomly select w1w_1 and w2w_2, subject to w1+w2=1w_1+w_2=1, to compute the coordinate mean βx(τ)(w1x^M(1)+w2x^M(2))1+βx(τ)\frac{\beta^x(\tau)(w_1\hat{\mathbf{x}}_M^{(1)}+w_2\hat{\mathbf{x}}_M^{(2)})}{1+\beta^x(\tau)}. The atom-type mean is computed using the same weighting scheme. We compare this setting with deterministic sampling using fused means computed with w1=w2=1/2w_1=w_2=1/2, as shown below.

MethodsP-1 Vina Dock ↓P-2 Vina Dock ↓Max Vina Dock ↓Dual High Aff. ↑QED ↑SA ↑
Random weights-8.67-8.58-7.9851.5%0.580.70
w1=w2=1/2w_1=w_2=1/2-9.03-8.96-8.2959.5%0.600.72

These results demonstrate that, beyond sampling covariance, the accuracy of the sampling center also plays an important role in generation quality. When the sampling center is inaccurate, zero variance concentrates sampling at a suboptimal location, limiting the achievable performance.

Moreover, using the average expert predictions (x^M(1)+x^M(2)2\frac{\hat{\mathbf{x}}_M^{(1)}+\hat{\mathbf{x}}_M^{(2)}}{2} and v^M(1)[n]+v^M(2)[n]2\frac{\hat{\mathbf{v}}_M^{(1)}[n]+\hat{\mathbf{v}}_M^{(2)}[n]}{2}) does not guarantee an optimal sampling mean for every target pair. Table 12 provides an example where shifting the center toward P-2 improves the dual-target binding affinity. In such cases, FusedBFN allows flexible adjustment of the sampling center through w1w_1 and w2w_2, while the covariance changes accordingly through w12+w22w_1^2+w_2^2. In contrast, CompBFN uses a fixed sampling center, with η\eta controlling only the covariance.

We extend Table 12 by evaluating additional weight combinations and find that (w1,w2)=(0.25,0.75)(w_1,w_2)=(0.25,0.75) achieves the best performance. We then set η=0.8\eta=0.8 for CompBFN to match the sampling variance of FusedBFN (0.25,0.75)(0.25,0.75). As shown in the table below, CompBFN underperforms FusedBFN despite having the same sampling variance. This is because CompBFN computes its sampling means using the average expert predictions x^M(1)+x^M(2)2\frac{\hat{\mathbf{x}}_M^{(1)}+\hat{\mathbf{x}}_M^{(2)}}{2} and v^M(1)[n]+v^M(2)[n]2\frac{\hat{\mathbf{v}}_M^{(1)}[n]+\hat{\mathbf{v}}_M^{(2)}[n]}{2}, which leads to a suboptimal sampling center. These results demonstrate the fundamental difference between FusedBFN and CompBFN and highlight the superiority of FusedBFN. The optimal choice of w1w_1 and w2w_2 may depend on the structural complexity of the two targets and how challenging they are for the network. Adaptively selecting optimal weights for each target pair is therefore a promising direction for future research.

Methodsw1w_1w2w_2η\etaP-1 Vina Dock ↓P-2 Vina Dock ↓Max Vina Dock ↓QED ↑SA ↑
FusedBFN0.90.1–-11.221-9.70-9.640.610.73
FusedBFN0.750.25–-9.09-7.47-7.470.280.57
FusedBFN0.60.4–-9.63-9.55-9.240.650.75
FusedBFN0.50.5–-8.30-7.48-7.430.290.59
FusedBFN0.40.6–-8.45-8.67-7.850.500.61
FusedBFN0.250.75–-10.09-10.17-9.810.530.74
FusedBFN0.10.9–-8.28-8.35-7.980.310.64
CompBFN––0.8-9.40-9.27-8.880.620.71
AuthorsOpenReview ↗

Response to Reviewer eZjj — W1 and Q1 — Part 3/3

Additional matched-variance experiments. To investigate the effect of sampling variance in diffusion-based methods, we introduce a noise multiplier λ\lambda into the coordinate reverse transition: xt−1=μt(xt,x^0)+λσtϵ,ϵ∼N(0,I)x_{t-1}=\mu_t(x_t,\hat{x}_0)+\sqrt{\lambda}\sigma_t\epsilon, \epsilon\sim\mathcal N(0,I). Following the reviewer's suggestion, we additionally compare all four variants under their default settings and with sampling variance removed (c=0c=0 for BFN and λ=0\lambda=0 for diffusion), as reported in the table below. Under the default settings, FusedBFN outperforms CompBFN. When c=0c=0 is applied to both BFN variants, their performance becomes nearly identical because their sampling distributions have the same mean and zero variance, consistent with our theoretical analysis. The additional diffusion experiments also show improved binding performance with reduced variance, indicating that the benefit of sampling concentration extends to the diffusion framework.

MethodDefault Max Vina ↓Default Dual High Aff. ↑Zero-noise Max Vina ↓Zero-noise Dual High Aff. ↑
CompDiff-7.5049.6%-7.5753.0%
FusedDiff-7.6148.6%-7.6851.0%
CompBFN-7.9450.8%-8.3059.3%
FusedBFN-8.1456.2%-8.2959.5%
λ\lambdaMethodsP-1 Vina Dock ↓P-2 Vina Dock ↓Max Vina Dock ↓Dual High Aff. ↑
1CompDiff-8.52-8.33-7.5049.6%
0.5CompDiff-8.48-8.47-7.5250.8%
0CompDiff-8.67-8.47-7.5753.0%
1FusedDiff-8.44-8.34-7.6148.6%
0.5FusedDiff-8.48-8.40-7.6748.1%
0FusedDiff-8.56-8.42-7.6851.0%

Prior work introduced CompDiff with η=0.5\eta=0.5, without systematically investigating how composition strength affects sampling covariance and downstream performance. Our contribution is not limited to introducing a new composition formulation. By theoretically comparing distribution fusion and transition-kernel composition, we reveal the importance of sampling center for generation quality and identify sampling noise as a key factor underlying performance differences when the sampling center is accurate. These findings have not been reported in prior work and provide theoretical insights and practical guidance for future dual-target methods.

[1] Zhou X, Guan J, Zhang Y, et al. Reprogramming pretrained target-specific diffusion models for dual-target drug design[J]. Advances in Neural Information Processing Systems, 2024, 37: 87255-87281.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.