acceptodds
Under review as a conference paper at ICLR 2027

Molweaver: Scalable Foundation Model for Structure-based Drug Discovery

141 views

Abstract

Structure-based drug discovery requires designing molecules for target protein pockets and predicting ligand bound poses, both of which require learning molecular connectivity, three-dimensional structure, and ligand–protein interactions. We introduce Molweaver, a family of molecular foundation models ranging from 60M to 500M parameters, pretrained on 100 million SELFIES representations, 1 billion molecular conformers, 3.2 million protein pockets, and 167,000 redocked pocket–ligand complexes with associated properties. Molweaver is a unified transformer framework that learns molecular representations across 2D structure, 3D geometry, and chemical properties. It generates molecular sequences through autoregressive prediction and 3D conformations through flow matching. We adapt Molweaver to downstream structure-based tasks using supervised fine-tuning and reinforcement learning. For pose prediction, Molweaver places 80.2% of ligands within 2 Å RMSD of the reference pose while also passing all applicable PoseBusters checks, comparable to equivariant structure-based models despite using no equivariant inductive biases in the architecture. For molecule generation, docking-based feedback and molecular property optimization steer the model toward compounds with favorable interactions within CrossDock2020 target pockets, yielding substantially improved docking scores, full connectivity, and the lowest clash counts among compared methods. We also show that these metrics improve with model size. These results highlight the potential of large-scale joint pretraining and task-specific adaptation for developing general-purpose models for structure-based drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure-based Drug Discovery, Protein-ligand Docking, 3D Molecule Generation, Flow-matching

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_HBMtOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Molweaver, which jointly models target pockets, molecular sequences, and 3D coordinates. This considers autoregressive-based generation for molecular sequences and flow matching-based generation for 3D coordinates. This further studies PPO and Flow-DPO for target-specific online optimization. The authors evaluate Molweaver on ligand pose prediction and pocket-conditioned molecule generation, showing competitive performance and consistent improvements with model scaling from 60M to 500M parameters.

Critical strengths
  • Practical target-specific optimization. The paper considers an online fine-tuning method given a target using relatively cheap labels. This can be a new practical recipe for adapting binder generative models to new targets.

  • Strong performance and consistent scaling benefits. The model achieves competitive or superior performance against models with expensive architectures, e.g., Pairformer in UniMol2, with consistent improvements as model size increases, supporting the potential of scalable approaches for SBDD.

Critical weaknesses

None

Questions for authors
  • During inference, the model is optimized for each target pocket using PPO and DPO. How much computational budget does this require? Can the authors comment on the computational cost with respect to the baselines?

  • How does the model perform without PPO and DPO? I was wondering about some ablations that show how much they contribute to the performance.

2 comments
AuthorsOpenReview ↗

Response to reviewer `HBMt`

We sincerely thank the reviewer HBMt for the positive assessment of our work and for recognizing the practical value of target-specific optimization and the consistent improvements with model scaling. We have also made the edits to the paper with the changes highlighted in blue. We address the questions below.

Q1: Computational cost of target-specific PPO and Flow-DPO optimization

We thank the reviewer for asking us to clarify the computational budget. We have added the optimization protocol in Section 5.2 and added computational cost estimates in Appendix B.5. PPO and Flow-DPO each run for 200 optimizer steps with batch size 8, using examples shuffled across all 100 evaluation pockets. Within each stage, model parameters are shared across pockets. Each stage therefore involves 1,600 pocket-example appearances, averaging 16 appearances per pocket.

For Molweaver-500M, the estimated model-optimization cost across both stages is 11.4 EFLOPs in total, corresponding to an amortized cost of 114 PFLOPs per pocket over the 100-pocket evaluation. These estimates exclude reward evaluation and inference-time sampling. CPU-based evaluation of Vina Score and Vina Dock rewards remains the main computational bottleneck.

Q2: Performance without PPO and Flow-DPO

We have added ablation experiments in Appendix B.5 (Table 7) to examine the individual contributions of PPO, Flow-DPO, and auxiliary-property pretraining using Molweaver-60M.

ConfigurationPB-valid ↑Vina Score ↓Vina Dock ↓QED ↑SA ↑Strain ↓Clashes ↓
No PPO / No Flow-DPO67.2%-3.38-6.720.480.61360.146.37
PPO only66.9%-4.73-8.640.520.73330.033.46
PPO + Flow-DPO (without property pretraining)74.5%-6.38-8.490.410.58232.411.61
PPO + Flow-DPO75.7%-6.43-8.880.520.73226.531.66

Table: Ablation of PPO, Flow-DPO, and property-aware pretraining for Molweaver-60M (Appendix B.5, Table 7). Vina Score and Vina Dock are reported as mean values, and strain as the median.

The ablation results show that PPO primarily improves molecular generation by steering the model toward molecules with more favorable docking scores, drug-likeness, and synthetic accessibility. Adding Flow-DPO further improves pose quality and physical plausibility, leading to higher PB validity, lower ligand strain, and fewer clashes, while also improving direct Vina scores.

We additionally observe that including auxiliary-property prediction during pretraining improves QED, SA, and Vina Dock scores compared with pretraining without this objective. These results demonstrate the complementary roles of PPO in optimizing molecular properties, Flow-DPO in refining ligand poses, and property-aware pretraining in improving downstream molecular-generation performance.

We hope these clarifications and additional ablations address the reviewer's questions.

Thank you for detailed clarification. I'd like to maintain my positive assessment.

Reviewer_echqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Molweaver, a family of scalable molecular foundation models for structure-based drug discovery. The proposed framework aims to jointly model molecular connectivity, three-dimensional geometry, molecular properties, and protein–ligand interactions within a unified Transformer architecture. Specifically, Molweaver employs autoregressive prediction over SELFIES representations to generate molecular graphs and flow matching to model continuous atomic coordinates. Molecular property prediction and AutoDock Vina score prediction are additionally incorporated as auxiliary pretraining objectives. The pretrained models are adapted to two downstream tasks including protein-ligand docking and structure-based drug design.

Critical strengths
  1. This paper combines autoregressive prediction and flow matching for generation of discrete and continuous modalities within a unified Transformer architecture, which is novel in the field of structure-based drug design.
Critical weaknesses
  1. The improvement is marginal or overclaimed. In the docked-pose prediction experiment, though with large-scale pre-training, the improvement of Molweaver-500M over PocketXMol is relatively small. Besides, in pocket-constrained molecule generation, the authors explicitly state that both PPO and Flow-DPO are applied online to each CrossDocked test pocket. In contrast, the baselines in Table 1 are evaluated in a zero-shot generation setting without the same target-specific online optimization. Therefore, the comparison is unfair.
  2. Insufficient ablations and data-split analysis limit the evidence for the claimed benefits of large-scale joint pre-training.
Questions for authors
  1. Could the authors at least provide ablation results comparing downstream task training from random initialization on a representative model scale?
  2. The motivation of this paper is unclear, making the overall methodology seem to be a combination of different existing techniques. It would be better to further elaborate the motivation.
7 comments
AuthorsOpenReview ↗

Response to reviewer `echq` - Part 1/2

We sincerely thank the reviewer echq for their comments and constructive feedback. We address the concerns and questions point by point below. We have also made the edits to the paper with the changes highlighted in blue.

W1: Marginal improvement in docked-pose prediction and fairness of the molecule-generation comparison

We have updated Figure 5 to report the results of both Molweaver and PocketXMol using a matched sampling budget of N=250N=250, along with the variability across five sampling runs with different seeds. Molweaver-500M achieves a success rate of 80.20%(+/-1.84%), compared with 79.4%(+/-1.95%) for PocketXMol. We do note that the performance gain with our model is very minor and is statistically overlapping, but it is also important to note that Molweaver can match PocketXMol's result with no inductive bias components like equivariance or pariformer, and we also we also show that our results consistently scales with model size.

Additionally, we have included a new Figure 9 in Appendix B.4 comparing Molweaver and PocketXMol across different sampling budgets, evaluated using the joint PB-validity and RMSD below 2 Å criterion. The results show that PocketXMol's performance improves until N=100 and then plateaus, whereas Molweaver continues to improve until N=250 before reaching a plateau. Notably, increasing PocketXMol's sampling budget from 100 to 250 does not yield further improvements. Thus, Molweaver closely matches PocketXMol’s results even when both methods are evaluated with the same number of candidate poses.

Secondly, Molweaver is explicitly evaluated as an online optimization method for pocket-constrained molecule generation, as described in Section 5.2. To make this comparison clearer, we have updated Table 1 to separately report offline generation methods and online optimization methods. In particular, we include DecompOpt (Zhou et al., 2024b) and TAGMol (Dorna et al., 2024), two works which also performs target-specific optimization on CrossDock2020 like Molweaver. The evaluations uses the same type of docking, molecular-property metrics and optimizes on test pockets. We summarize the online optimization comparisons baselines below.

MethodPB-valid ↑Vina Score ↓Vina Dock ↓QED ↑SA ↑Diversity ↑Connected ↑
DecompOpt68.7%-5.87-8.980.480.650.6071.6%
TAGMol56.0%-7.02-8.590.550.560.6992.0%
Molweaver-500M79.6%-7.22-9.590.570.840.71100.0%

Table: Comparison with online optimization baselines on CrossDock2020 (Section 5.2, Table 1).

Molweaver-500M achieves better mean Vina Score and Vina Dock values, higher QED and SA, and greater diversity than these two online optimization baselines, while also achieving higher PB validity and full molecular connectivity. Our intention is therefore to demonstrate the effectiveness of Molweaver as an online optimization framework that achieves better results than previous methods on the pocket-constrained molecule generation task, rather than to claim superior zero-shot generation.

AuthorsOpenReview ↗

Response to reviewer `echq` - Part 2/2

W2 and Q1: Ablations on pretraining, and different components. Finetuning on random initialization

We have added ablation experiments in Appendix B.4 (Table 6) and Appendix B.5 (Table 7) to better understand the contributions of pretraining, fine-tuning, pretraining data scale, flow matching, and auxiliary-property supervision.

Table 6 (Appendix B.4) compares different pretraining and fine-tuning configurations of the 500M model, with the same downstream fine-tuning and inference settings of the original fine-tuned model.

ConfigurationMean RMSD (Å) ↓RMSD < 2 Å and PB-valid ↑
Fine-tuned from random initialization7.571.40 ± 0.20%
Pretrained without fine-tuning9.680.00 ± 0.00%
Fully autoregressive (10% pretraining data) + fine-tuning4.8739.25 ± 1.84%
Molweaver (10% pretraining data) + fine-tuning2.0158.88 ± 1.94%
Molweaver (full pretraining data) + fine-tuning1.2180.20 ± 1.84%

These results show that large-scale pretraining followed by supervised fine-tuning substantially improves docked-pose prediction compared with training from random initialization or using the pretrained model alone. Increasing the pretraining data scale further improves performance. Moreover, under the same reduced-data setting, flow matching outperforms fully autoregressive coordinate generation, providing empirical support for our choice of flow matching for modeling continuous 3D coordinates.

Secondly, we have added ablations in Table 7 (Appendix B.5) using Molweaver-60M to examine the individual contributions of PPO, Flow-DPO, and auxiliary-property supervision to pocket-constrained molecule generation.

ConfigurationPB-valid ↑Vina Score ↓Vina Dock ↓QED ↑SA ↑Strain ↓Clashes ↓
No PPO / No Flow-DPO67.2%-3.38-6.720.480.61360.146.37
PPO only66.9%-4.73-8.640.520.73330.033.46
PPO + Flow-DPO (without property pretraining)74.5%-6.38-8.490.410.58232.411.61
PPO + Flow-DPO75.7%-6.43-8.880.520.73226.531.66

These results show that PPO primarily improves docking scores, drug-likeness, and synthetic accessibility, while adding Flow-DPO further improves PB validity, ligand strain, and clashes, along with direct Vina scores. We also observe that including auxiliary-property prediction during pretraining improves downstream molecular properties and docking performance compared with pretraining without this objective.

Together, these ablations provide additional evidence for the benefits of large-scale pretraining and help clarify the contributions of the different training objectives and design choices in Molweaver.

Q2: Motivation and combination of techniques

Our main motivation is to investigate whether a single, scalable Transformer can jointly model molecular connectivity, 3D geometry, and protein–ligand interactions through large-scale pretraining, without relying on specialized geometric inductive biases such as equivariance or pairformer modules.

Molecular connectivity is discrete and naturally suited to autoregressive SELFIES generation, while atomic coordinates are continuous and require modeling their joint spatial arrangement, motivating our use of flow matching within the same backbone. Our additional ablations in Appendix B.4 (Table 6) and in Appendix B.5 (Table 7), as discussed above, further support these design choices. Large-scale pretraining followed by fine-tuning substantially outperforms training from random initialization, increasing the pretraining data scale improves downstream performance, and flow matching outperforms autoregressive coordinate modeling under the same reduced-data setting.

While the individual techniques are established, our work integrates them into a unified framework and provides empirical evidence for the importance of pretraining and the choice of coordinate-generation objective. Together with the competitive performance achieved without explicit geometric inductive biases, these findings provide useful guidance for designing future scalable molecular foundation models.

We hope these clarifications and additional ablations address the reviewer's concerns and make the experimental setting and contributions clearer.

References:

[1] Xiangxin Zhou, Xiwei Cheng, Yuwei Yang, Yu Bao, Liang Wang, and Quanquan Gu. Decompopt: Controllable and decomposed diffusion models for structure-based molecular optimization. In The Twelfth International Conference on Learning Representations, 2024b. URL https://openreview.net/forum?id=Y3BbxvAQS9.

[2] Vineeth Dorna, D. Subhalingam, Keshav Kolluru, Shreshth Tuli, Mrityunjay Singh, Saurabh Singal, N. M. Anoop Krishnan, and Sayan Ranu. Tagmol: Target-aware gradient-guided molecule generation. arXiv preprint arXiv:2406.01650, 2024.

Dear Reviewer echq,

This is a gentle reminder to review the authors’ response at your earliest convenience.

Thank you for your time and consideration.

AC

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

Molweaver pretrains 60M–500M transformers on molecular strings, conformers, and protein pockets for pose prediction and pocket-conditioned molecule generation. Its shared backbone combines autoregressive SELFIES generation with flow matching for coordinates, followed by pose fine-tuning and PPO/Flow-DPO adaptation (Section 4). Experiments cover 428 PoseBusters complexes and 100 CrossDock2020 test pockets (Section 5). The reported results show competitive pose recovery and favorable docking scores for generated molecules, with generation evaluated after per-target optimization.

Critical strengths
  1. Combining molecular sequences with continuous coordinate generation provides a useful framework for both molecular design and pose prediction (Sections 4.1–4.4). Representing connectivity explicitly is valuable for molecular generation, and Table 1 reports 100% connected molecules across model sizes.

  2. The 80.2% PoseBusters result requires both RMSD below 2 Å and physical validity, making it more informative than RMSD alone (Section 5.1). The consistent improvement from 60M to 500M in Figure 5 also provides useful evidence that a transformer without explicit equivariant layers can perform well in this setting.

Critical weaknesses
  1. Molweaver performs PPO/Flow-DPO optimization on each test pocket and selects among 250 poses per molecule, while the baselines generate zero-shot. Without matched scoring and search budgets, the comparison cannot separate model quality from additional optimization (Section 5.2, Tables 1–2).

  2. Better docking scores come with a tradeoff in physical validity: Molweaver achieves 79.6% PB validity, compared with MolPilot's 95.9% (Table 1).

  3. Pose success is 80.2% versus PocketXMol's 79.4%, but Molweaver uses 250 rather than 100 candidates and requires 10,000 network evaluations per complex. Comparisons at matched budgets, runtime measurements, and variability across the five seeds are missing, making this small lead difficult to assess (Section 5.1, Figure 5).

  4. The rule for combining the two confidence models to select the final pose is unclear, although this selection directly affects the headline result and its reproducibility (Section 4.3).

Questions for authors
  1. With the final sampling budget fixed, how do generation results change before target-specific adaptation, after PPO, and after PPO plus Flow-DPO? How does the full method compare with sampling and reranking the frozen model under the same scoring budget?

  2. For Figure 5, can you report joint RMSD/PB success, runtime, and uncertainty at matching candidate counts, and specify how the confidence scores are combined? Figure 9 measures RMSD alone and therefore does not answer the budget question.

  3. Can focused controls at a smaller model size establish the contributions of joint flow matching, pretraining, and auxiliary-property supervision while keeping downstream training and inference fixed?

  4. How does Molweaver compare with 3DMolFormer under matched input information and target-specific optimization budgets?

3 comments
AuthorsOpenReview ↗

Response to reviewer `XxTi` - Part 1/2

We thank the reviewer XxTi for their comments and constructive feedback. We address the concerns and questions point by point below. We have also made the edits to the paper addressing the questions and weaknesses with the changes highlighted in blue.

W1: Concerns about comparison with zero shot methods

Molweaver is explicitly evaluated as an online optimization method for pocket-constrained molecule generation, as described in Section 5.2. To make this comparison clearer, we have updated Table 1 to separately report offline generation methods and online optimization methods. In particular, we include DecompOpt (Zhou et al., 2024b) and TAGMol (Dorna et al., 2024), two works which also performs target-specific optimization on CrossDock2020 like Molweaver. The evaluations uses the same type of docking, molecular-property metrics and optimizes on test pockets. We summarize the online optimization comparisons baselines below.

MethodPB-valid ↑Vina Score ↓Vina Dock ↓QED ↑SA ↑Diversity ↑Connected ↑
DecompOpt68.7%-5.87-8.980.480.650.6071.6%
TAGMol56.0%-7.02-8.590.550.560.6992.0%
Molweaver-500M79.6%-7.22-9.590.570.840.71100.0%

Table: Comparison with online optimization baselines on CrossDock2020 (Section 5.2, Table 1).

Molweaver-500M achieves better mean Vina Score and Vina Dock values, higher QED and SA, and greater diversity than these two online optimization baselines. It also achieves higher PB validity and full molecular connectivity. Our intention is therefore to demonstrate the effectiveness of Molweaver as an online optimization framework that surpasses previous methods on the pocket constrained molecule generation task rather than to claim superior zero-shot generation.

W2: Trade-off between docking scores and physical validity

We agree with the reviewer that MolPilot achieves higher PB validity than Molweaver-500M. However, Molweaver significantly achieves better mean Vina Score, Vina Dock, and synthetic accessibility, while maintaining 100% molecular connectivity. These results highlight a trade-off between docking performance and physical validity rather than an improvement across all metrics.

Furthermore, we observe an overall improvement in PB validity with model scaling, from 75.7% at 60M to 79.6% at 500M, suggesting that scaling may help further improve physical plausibility. While MolPilot achieves higher PB validity, our results demonstrate encouraging improvements in both docking performance and physical validity as the model scales without inductive biases like equivariance unlike MolPilot.

W3 and Q2: Pose sampling budget vs RMSD & PB-validity, variability, and inference cost of Molweaver and PocketXMol.

We have updated Figure 5 to report the results of both Molweaver and PocketXMol using a matched sampling budget of N=250N=250, along with the variability across five sampling runs with different seeds. Molweaver-500M achieves a success rate of 80.20% (+/-1.84%), compared with 79.4%(+/-1.95%) for PocketXMol.

Additionally, we have included a new Figure 9 in Appendix B.4 comparing Molweaver and PocketXMol across different sampling budgets, evaluated using the joint PB-validity and RMSD below 2 Å criterion. The results show that PocketXMol's performance improves until N=100 and then plateaus, whereas Molweaver continues to improve until N=250 before reaching a plateau. Notably, increasing PocketXMol's sampling budget from 100 to 250 does not yield further improvements. Thus, Molweaver closely matches PocketXMol’s results even when both methods are evaluated with the same number of candidate poses. Also it is worth noting that PocketXMol has equivariance(inductive bias) built into the model.

W4, Q2: Combining the two confidence scores

We have updated the section 4.3 to include about this detail on how we combine and use the two confidence scores. To combine the two confidence models, we first convert their raw scores which have different scales, into percentile ranks within each complex and average the resulting ranks to obtain a single score for pose selection.

AuthorsOpenReview ↗

Response to reviewer `XxTi` - Part 2/2

Q1: Generation results change before target-specific adaptation, after PPO, and after PPO plus Flow-DPO

We have added Table 7(Appendix B.5) showcasing ablations performed on Molwever-60M with No PPO + No Flow-DPO, PPO Only, PPO + Flow-DPO with Molweaver-60M pretrained without property objective, PPO + Flow-DPO with Molweaver-60M full pretrained. The table is as follows:

ConfigurationPB-valid ↑Vina Score ↓Vina Dock ↓QED ↑SA ↑Strain ↓Clashes ↓
No PPO / No Flow-DPO67.2%-3.38-6.720.480.61360.146.37
PPO only66.9%-4.73-8.640.520.73330.033.46
PPO + Flow-DPO (without property pretraining)74.5%-6.38-8.490.410.58232.411.61
PPO + Flow-DPO75.7%-6.43-8.880.520.73226.531.66

The ablation results show that PPO primarily improves molecular generation by steering the model toward molecules with more favorable docking scores, drug-likeness, and synthetic accessibility. Adding Flow-DPO further improves pose quality and physical plausibility, leading to better PB validity, lower ligand strain, and fewer clashes, while also improving the direct Vina scores. Together, these results highlight the complementary roles of PPO in optimizing molecular properties and Flow-DPO in refining ligand poses

We additionally ablate with the auxiliary-property prediction task removed during pretraining. These results provide evidence that auxiliary-property supervision contributes to downstream molecular-generation performance.

Q3: Contributions of flow matching, pretraining, and auxiliary-property supervision

We thank the reviewer for this suggestion. We have added ablation experiments in Table 6 (Appendix B.4) to examine the contributions of pretraining, fine-tuning, pretraining data scale, and the coordinate-generation objective. These experiments use the 500M model under the same downstream fine-tuning and inference settings of the original fine-tuned model. The table is as follows:

ConfigurationMean RMSD (Å) ↓RMSD < 2 Å and PB-valid ↑
Fine-tuned from random initialization7.571.40 ± 0.20%
Pretrained without fine-tuning9.680.00 ± 0.00%
Fully autoregressive (10% pretraining data) + fine-tuning4.8739.25 ± 1.84%
Molweaver (10% pretraining data) + fine-tuning2.0158.88 ± 1.94%
Molweaver (full pretraining data) + fine-tuning1.2180.20 ± 1.84%

The results show that large-scale pretraining followed by supervised fine-tuning substantially improves docked-pose prediction compared with training from random initialization or using the pretrained model alone. Increasing the pretraining data scale further improves performance. Moreover, under the same reduced-data setting, flow matching outperforms fully autoregressive coordinate generation, providing empirical support for our choice of flow-matching for coordinate generation objective. We have also included an ablation as mentioned before for auxiliary-property supervision using Molweaver-500M in Table 6 (Appendix B.5), whose results suggests that including in property prediction as a task while pretraining improves performance in downstream tasks.

Q4: Comparison with 3DMolformer

Unfortunately, the pretrained weights of 3DMolFormer are not publicly available, which prevents us from conducting a direct comparison under the same evaluation setting. Furthermore, we also trained a fully autoregressive model similar to 3DMolFormer, but required it to predict all ligand coordinates without providing any reference coordinates. We report this comparison in Appendix B.4 (Table 6) with the reduced-data pretraining setting, Molweaver's flow-matching approach achieves better docked-pose prediction performance than fully autoregressive coordinate generation.

We hope these clarifications and additional ablations address the reviewer's questions.

References:

[1] Xiangxin Zhou, Xiwei Cheng, Yuwei Yang, Yu Bao, Liang Wang, and Quanquan Gu. Decompopt: Controllable and decomposed diffusion models for structure-based molecular optimization. In The Twelfth International Conference on Learning Representations, 2024b. URL https://openreview.net/forum?id=Y3BbxvAQS9.

[2] Vineeth Dorna, D. Subhalingam, Keshav Kolluru, Shreshth Tuli, Mrityunjay Singh, Saurabh Singal, N. M. Anoop Krishnan, and Sayan Ranu. Tagmol: Target-aware gradient-guided molecule generation. arXiv preprint arXiv:2406.01650, 2024.

Thank you for the detailed response and additional experiments. The new ablations and clarification of the confidence scores address part of my concerns. I will maintain my score! Hope you will get accepted! Best luck!

Comments on the paper

1 comment

We appreciate the constructive feedback and the recognition of Molweaver's potential as a scalable framework for structure-based drug discovery from all the reviewers. Based on the concerns and suggestions raised during the review process, we have revised the paper to clarify our contributions, strengthen the experimental analysis, and provide additional details on our evaluation methodology. All changes in the revised manuscript are highlighted in blue. We summarize the main revisions below.

1. Clarification of online optimization and comparison with baselines (XxTi, echq). We have clarified that Molweaver performs target-specific online optimization using PPO and Flow-DPO for pocket-constrained molecule generation, rather than zero-shot generation. We further added new online optimization baselines (DecompOpt and TAGMol) for comparison and explicitly separated offline generation baselines from online optimization methods in Table 1. Molweaver achieves better docking scores, molecular properties, diversity, and PB validity than these online optimization baselines.

2. Additional ablations on PPO, Flow-DPO, and property-aware pretraining (XxTi, echq, HBMt). We have added ablation experiments in Appendix B.5 (Table 7) to examine the contributions of the different optimization stages. The results show that PPO primarily improves docking scores, drug-likeness, and synthetic accessibility, while Flow-DPO further improves pose quality, PB validity, ligand strain, and clashes. We also ablate auxiliary-property supervision during pretraining and observe improvements in downstream molecular properties and docking scores when this objective is included. These experiments help clarify the complementary roles of the different training objectives.

3. Additional ablations on pretraining, fine-tuning, and coordinate generation (XxTi, echq). We have added experiments in Appendix B.4 (Table 6) comparing training from random initialization, pretraining without fine-tuning, reduced-data pretraining, and full-data pretraining followed by fine-tuning. The results demonstrate the advantage of combining large-scale pretraining with supervised fine-tuning, as well as the benefits of increasing pretraining data scale. We also compare flow matching with fully autoregressive coordinate generation under the reduced-data setting and observe better pose prediction performance with flow matching. These findings provide further empirical support for our architectural and training choices.

4. Clarification of inference-time sampling and pose selection (XxTi, echq). We have updated Figure 5 to report results for both Molweaver and PocketXMol under a matched sampling budget of N=250, including variability across five runs with different seeds. We have also added a sampling ablation in Appendix B.4 (Figure 9), comparing the joint RMSD and PoseBusters success rates across different sampling budgets. The results show that PocketXMol's performance plateaus at N=100, while Molweaver continues to improve until N=250. Thus, Molweaver closely matches the performance of pocketXMol even with matched candidate counts. Additionally, Section 4.3 now explicitly describes how the two confidence models are combined using averaged percentile ranks for final pose selection, improving the clarity and reproducibility of the evaluation procedure.

5. Clarification of motivation and scope of the contributions (echq). We have refined the discussion of our motivation and design choices. Our goal is to investigate whether large-scale joint pretraining, autoregressive molecular sequence generation, and flow matching for continuous coordinates can be combined within a general-purpose Transformer without explicit geometric inductive biases such as equivariance or pairformer modules. The additional ablations provide empirical evidence supporting the role of pretraining and the choice of coordinate-generation objective, while the scaling results demonstrate competitive performance across model sizes. We believe these findings provide useful guidance for future work on scalable molecular foundation models.

6. Clarification of computational cost for target-specific optimization (HBMt). We have clarified the optimization budget in Section 5.2. Molweaver performs approximately 16 PPO updates and 16 Flow-DPO updates per target pocket. We have also included the computation budget in FLOPs under Appendix section B.5. The main computational bottleneck is the CPU-based evaluation of Vina Score and Vina Dock rewards, rather than the optimization updates themselves. While this introduces additional computational cost compared with zero-shot generation, it enables target-specific adaptation using computational feedback.

Overall, these revisions aim to provide a clearer understanding of Molweaver's contributions. We hope the additional experiments and clarifications help address the main concerns raised during the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.