acceptodds
Under review as a conference paper at ICLR 2027

ApoFlowMol: Geometry-Constrained Flow Matching for 3D Molecular Generation in Flexible Protein Pockets

11 views

Abstract

Deep generative models have advanced structure-based drug design (SBDD), but most methods treat protein binding pockets as rigid, overlooking ligand-associated conformational changes. Existing flexible-pocket generators address this limitation, yet residue-wise coordinate or frame updates do not inherently preserve peptide-chain geometry. We introduce ApoFlowMol, a geometry-constrained full-atom flow-matching framework that jointly generates 3D ligands and adapts apo pockets to the evolving ligand state. ApoFlowMol employs a hierarchical SE(3)-equivariant complex network to model ligand coordinates, atom and bond identities, and protein conformational changes. Protein motion is represented either by fixed-backbone local torsion updates or by chain-linked backbone-and-side-chain kinematics. Both representations preserve sequence-defined peptide links recognized among modeled pocket residues: the former leaves the backbone unchanged, whereas the latter permits backbone motion while preserving backbone bond lengths and peptide connectivity within the pocket by construction throughout sampling. ApoFlowMol generates high-affinity molecules with performance comparable to leading flexible-pocket generative models. The resulting adapted pocket conformations provide more effective conditioning than unadapted apo pockets for established fixed-pocket SBDD methods. Surprisingly, pocket conformations selected through lead-guided side-chain-flexible docking similarly enable a fixed-pocket generator to achieve performance comparable to end-to-end flexible-pocket models. Together, these results highlight the potential of geometry-constrained pocket adaptation for flexible-pocket molecular generation and for improving existing fixed-pocket SBDD methods.

Primary area:
generative models
Keywords:
Structure-Based Drug Design, Flow Matching, Molecular Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3CXnOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ApoFlowMol is a model based on flow matching for jointly designing novel 3D ligands while evolving the apo pocket of the protein. The authors introduce ApoFlowMol$\chi$ fixing N–C$\alpha$–C updating side chains torsions and ApoFlowMol-Link to preserve chain-linked backbone and side-chain peptide kinematics connectivity by construction. They evaluated the model on the Apo2Mol benchmark. Additionally, they show that lead guided side-chain flexible docking can directly produce adapted pockets that improve downstream for fixed pocket generators (like FLOWR) without any retraining.

Critical strengths

ApoFlowMol provides principled geometric guarantee to preserve the peptide C-N connectivity, trying to address a failure mode in flexible pocket generators.

Even though the general concept is known, they provide insights in how this pocket adaptation guided by sidechain flexible docking, can directly offer advantages using fixed-pocket generator (FLOWR) without any retraining of the model.

Critical weaknesses
  • In table 1, the improvements on the vina scores are small −7,90 vs −7,74 kcal/mol. I see that this can be a limitation with the data itself.

  • In table 1, ApoFlowMol-Link-SDE shows QED 0.39, SA 0.32 and LogP −1.51, mismatching the data distribution QED 0.60 SA 0.77 LogP 2.05, and getting worse scores compared to Apo2Mol QED 0.54 SA 0.54 LogP 3.60. This suggest that the small improvements over docking scores estimates cause significantly worse scores on metrics such as drug likeness.

  • In table 2, FlexDock seems to outperform the main contribution on some metrics. I think that limiting the results to 100 systems can reduce the statistical power of the results.

  • The kinematic reconstruction for C-N is not learnt for the peptide-link geometry (sec 4.2 line 376). I would better frame this aspect as an architectural guarantee rather than empirical result in the paper.

  • No comparison of sampling cost vs baselines. An ablation showing inference time costs and scaling the model at different sizes is missing. This makes it hard to compare to the existing alternatives.

  • The comparisons omit several recent flexible pocket methods despite being discussed in related works: DynamicFlow, FlexSBDD, DrugFlow.

  • The RdKit molecular validity and PoseBuster valid on the generated molecule are not reported anywhere.

Questions for authors
  • Will code, trained weights and data be released?

  • I see that the references sometimes follow different formats, ie they sometimes present or not present the url. Can you uniform that?

  • I think that table 5 in the appendix has an inconsistent abbreviation Ave/Med. In the rest of the paper you use "Avg." "Med.".

  • Can you provide matched budget comparisons for FlexDock vs ApoFlowMol-Link?

  • Why does SDE sampling produce such poor LogP and SA values providing slight improvements on Vina docking scores?

  • Can you provide more insights into how does the hierarchical architecture scale with pocket size? and what is the inference time per complex comparing your method to other recent methods like Apo2Mol, YuelDesign, and FLOWR?

  • Would it be possible for completeness to report the results on the full test set of 478 systems? This would also contribute to validate how reliable it is your initial subset of 100 systems.

  • How sensitive are results to t>0.65 threshold for activating interaction losses? Did you try to tune it in your setting? It would be interesting to see an ablation on this parameter.

1 comment
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for the constructive feedback. We address each weakness and question below.

Responses to Weaknesses

W1. Small docking improvement.

We agree that the Vina improvement is modest. Our main contribution addresses geometrically constrained pocket adaptation alongside ligand generation, further evaluated in Table 2 (Page 8).

W2. Molecular-property trade-offs.

We acknowledge the lower QED/SA and shifted LogP. Considering ligand properties and docking performance together, we regard ODE as the stronger overall configuration. SDE's molecular-property performance remains a limitation.

W3. FlexDock and evaluation size.

FlexDock is explicitly presented as a separately budgeted analysis (Table 3, Page 9). We consider the 100-system evaluation sufficient for the comparisons reported in this study.

W4. Architectural guarantee.

We agree. Peptide-link preservation is described as a property of the kinematic construction (Pages 4–5), distinct from empirical pocket-quality checks.

W5. Sampling cost.

We added runtime comparisons with Apo2Mol, YuelDesign, and FLOWR (Appendix G.10, Pages 33–34; Table 19A, Page 36).

W6. Additional baselines.

At submission, our search did not identify publicly available code and trained weights enabling reproducible comparisons with DynamicFlow and FlexSBDD.

W7. Molecular validity.

Table 2 (Page 8) evaluates 500 original attempts per method. Link-ODE achieves 82.0% RDKit-valid and connected outputs and 64.6% confirmed PoseBusters passes, versus 87.4%/26.4% for Apo2Mol and 39.8%/13.0% for YuelDesign.

Responses to Questions

Q1. Resource release.

Yes. We plan to release code, trained weights, and evaluation resources through GitHub and an open-access repository such as Zenodo after acceptance.

Q2. Reference formatting.

Thank you. We have standardized the reference formatting, including URL presentation.

Q3. Abbreviations.

We have standardized the table abbreviations to "Avg." and "Med."

Q4. Matched-budget FlexDock comparison.

Table 19B (Page 36) now reports key FlexDock timings: docking takes 8.98–17.24 seconds for one previously selected lead per pocket, and pocket export takes 0.76–1.05 milliseconds. Ranking, preparation, rescue searches, and FLOWR generation are excluded. The initial protocol evaluates three leads per interaction center, each with three independent docking seeds (Table 19B, Page 36). Total cost therefore depends on screening success; this is not a matched-total-budget comparison.

Q5. SDE molecular properties.

We have not yet determined the cause of the lower LogP and SA values in Link-GR-SDE. Our results show an empirical trade-off between docking scores and molecular properties. Considering these metrics together, we regard Link-GR-ODE as the stronger overall configuration for direct generation.

Q6. Scaling and inference time.

Table 19A (Page 36) covers 92–445 pocket heavy atoms and 12–34 ligand heavy atoms. Link-GR-ODE takes 1.17–2.57 seconds per molecule and is fastest in seven of nine settings under native sampling settings, excluding external postprocessing.

Q7. Full test-set evaluation.

Our current computational resources do not permit completing the 478-system evaluation within the discussion period. We acknowledge this limitation.

Q8. Interaction-loss threshold.

We adopted t>0.65t>0.65 from DynamicFlow without tuning or a threshold ablation. Its sensitivity remains unquantified.

Reviewer_3f7NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ApoFlowMol, a geometry-constrained flow-matching framework that jointly generates ligands and adapts apo pockets. It uses SE(3)-equivariant modeling and kinematic representations to preserve peptide connectivity during conformational updates. Experiments show competitive performance and demonstrate that adapted pockets improve existing fixed-pocket generators. Lead-guided side-chain-flexible docking also enables a fixed-pocket generator to achieve performance comparable to end-to-end flexible-pocket models.

Critical strengths
  1. Conditioning molecular generation on apo protein structures is practically valuable, as ligand-bound holo structures are often difficult to obtain.

  2. The SE(3)-equivariant architecture is well motivated, as it respects the rotational and translational symmetries of molecular systems.

  3. Relaxing the rigid-pocket assumption while preserving peptide-chain geometry addresses an important limitation of existing SBDD methods.

Critical weaknesses
  1. The motivation is insufficiently developed. Although the authors emphasize the importance of apo-conditioned molecular generation, they do not clearly distinguish their approach from Apo2Mol or explain which limitations of that method they address.

  2. The methodological novelty appears limited. SE(3)-equivariant networks are already widely used in SBDD, and applying such architectures to apo-conditioned molecular generation alone does not establish a substantial contribution.

  3. The empirical results do not convincingly demonstrate the advantages of the proposed method. In particular, the results in Table 1 do not show a clear performance improvement over Apo2Mol.

Questions for authors

The symbols L1\mathcal{L}_1 and S1m\mathcal{S}_1^m in line 150 do not appear to be defined. Please clarify their meanings.

1 comment
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for recognizing the practical value of apo-conditioned generation and preserving peptide-chain geometry. We address the concerns below.

1. Motivation and distinction from Apo2Mol.

We clarify in the Introduction (Page 2) that SE(3) equivariance alone does not preserve covalent connections between neighboring residues. Our chain-linked representation addresses this limitation by preserving template backbone bond lengths and recognized peptide C–N connections throughout pocket adaptation. New Table 2 (Page 8) provides supporting evidence: the evaluated baselines exhibit substantial peptide-link errors and backbone strain, showing that existing flexible-pocket generation pipelines do not necessarily produce geometrically plausible pockets.

2. Methodological contribution.

Our contribution is the integration of chain-linked protein kinematics with joint ligand–pocket flow matching. To our knowledge, this is the first application of this combination to flexible-protein SBDD. The model jointly updates ligand coordinates, atom types, bond types, and protein-motion states, reconstructing the pocket at each integration step. The design rationale and motion representations are described on Pages 3–5.

3. Empirical advantages.

We agree that Table 1 does not establish uniformly better ligand quality. QED and SA remain limitations. Our model addresses a complementary question: can flexible-protein SBDD generate plausible ligands while also maintaining geometrically plausible protein pockets? New Table 2 (Page 8) evaluates five original attempts per system without replacing failures. Link-ODE achieves a C–N MAE of 0.0066 Å and a Rosetta backbone penalty of 0.045 RU/term, compared with 0.8007 Å/25.913 for Apo2Mol and 1.1334 Å/92.382 for YuelDesign. Its confirmed PoseBusters pass rate is 64.6%, compared with 26.4% and 13.0%, respectively. These results support improved pocket-internal geometry alongside ligand and pocket-context validity in the evaluated pipelines. Table 3 (Page 9) additionally demonstrates the utility of adapted pockets for FLOWR generation without retraining. Considering ligand properties and docking performance together, we regard Link-GR-ODE as the stronger overall configuration, while Link-GR-SDE may provide more useful pockets for downstream FLOWR generation.

4. Symbol definitions.

We now define both symbols immediately after Equation (1) (Page 3). The subscript 11 denotes the flow endpoint t=1t=1; L1=(X1,A1,E1)\mathcal{L}_1=(\mathbf{X}_1,\mathbf{A}_1,\mathbf{E}_1) contains terminal ligand coordinates, atom types, and bond types. S1m\mathcal{S}_1^m denotes the terminal protein-motion state under representation m∈{χ,Link}m\in\{\chi,\mathrm{Link}\}, which the reconstruction map KmK_m combines with the apo template to obtain the full-atom pocket.

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

ApoFlowMol jointly generates 3D ligands and adjusts protein pockets, starting from unbound structures. Its main contribution is a chain-based internal-coordinate representation that allows backbone motion while preserving bond lengths at identified peptide connections. The model includes a χ variant with a fixed backbone, a Link variant with a flexible backbone, and ODE/SDE sampling options. Across 100 test systems, Link-GR-SDE achieves a mean Vina score of −7.90, compared with −7.74 for Apo2Mol. Its adapted pockets also improve results from the independent generator FLOWR. Here, HA means the fraction of generated ligands with better docking scores than the reference ligand—not experimentally confirmed high affinity.

Critical strengths

Addresses a concrete structural problem. SE(3) equivariance alone does not preserve covalent connections between residues. Encoding these constraints in the kinematic representation has clear value.

Critical weaknesses

The main configuration was selected through exploration on the same test subset. Appendix G.2 states that Link-GR was chosen from three Link variants based on their SDE Vina results. Its reported advantage therefore needs confirmation on data not used for configuration selection. The authors acknowledge this. The issue is configuration selection, not post hoc selection of the 100-system subset.

Better docking scores come with substantially worse molecular properties. Link-GR-SDE has QED/normalized SA scores of 0.39/0.32, compared with 0.54/0.54 for Apo2Mol. Its direct Vina difference relative to ODE is −0.17, but the 95% CI [−0.43, 0.10] includes zero. These results do not establish better overall generation quality.

Plausible bond lengths do not establish plausible conformations. For Link-GR-SDE, mean RMSD is 2.72 Å from apo to generated pockets and 2.84 Å from generated to holo pockets, compared with an apo-to-holo reference of 0.81 Å. Different generated ligands may induce different valid conformations, so these values do not prove the structures are wrong. They do, however, call for independent checks of protein strain, peptide planarity, clashes, and compatibility with the full protein.

The sources of the performance gains are not fully isolated. χ and Link differ in degrees of freedom, supervision, and parts of the processing pipeline, as well as chain constraints. A direct comparison of independent-residue and chain-based motion under the same network and training setup is missing. The source of the peptide-bond improvement is clear; the source of the docking gains is less clear.

Questions for authors

None.

3 comments
AuthorsOpenReview ↗

Response to Reviewer

We respectfully disagree with the plagiarism concern raised in this review. The reviewer did not provide specific evidence or clear justification to support this concern. Our figures were independently created by us and illustrate scientific concepts that are substantively different from those presented in the work cited by the reviewer.

We thank the reviewer for recognizing the value of explicitly preserving peptide connectivity. We have expanded the evaluation and clarified the scope of our conclusions.

1. Configuration Selection

We added a separate validation-set comparison of four ODE configurations, with three original attempts per system (Appendix G.2, Page 20). Link-GR-ODE achieves the lowest backbone-overlap count and Rosetta bonded penalty and the highest RDKit-valid and connected fraction among these configurations. These results support its use in the additional evaluation. We retain the complete test-set comparisons and acknowledge that the original configuration exploration used the test subset.

2. Docking and Molecular Quality

We agree that docking scores alone do not establish better overall generation quality. We do not claim that SDE is a better overall generation model. Our observation is that pockets generated by Link-GR-SDE yield better docking scores when used to condition FLOWR generation. We additionally report pocket geometry and molecular validity in Table 2 (Page 8). On the fixed five-attempt evaluation cohort, the Link-GR-ODE pipeline achieves a PoseBusters pass rate of 64.6%, compared with 26.4% for Apo2Mol and 13.0% for YuelDesign.

3. Pocket Conformational Quality

We agree that plausible C–N lengths are insufficient. Table 2 (Page 8) adds backbone bond-angle errors, backbone-heavy overlaps, and native Rosetta backbone bonded penalties. Link-ODE achieves a C–N MAE of 0.0066 Å, an angle MAE of 1.30°, 3.14 overlap pairs per structure, and a Rosetta penalty of 0.045 RU/term. Corresponding Apo2Mol/YuelDesign values are 0.8007/1.1334 Å, 5.49°/1.62°, 42.60/44.17 overlaps, and 25.913/92.382 RU/term. Supplementary Tables 21–24 (Page 37) additionally report peptide planarity, ω\omega torsions, and Ramachandran statistics with interpretability coverage.

4. Attribution of Improvements

The χ\chi and Link models share the core network architecture. We have kept the comparison as controlled as practical, while their different motion representations require different motion variables, prediction heads, reconstruction, and supervision. We therefore attribute peptide-connectivity preservation to the chain-linked construction, without attributing all docking differences solely to this constraint.

  1. Would you like to replace the configuration based on the test subset? Using the validation set makes more sense to me.

  2. As the author acknowledged, "docking scores alone do not establish better overall generation quality." My concern is that this method is not practical for generating high-quality molecules.

  3. Thanks for adding those pocket conformational quality metrics.

  4. Thanks for the justification. But it doesn't fully convince me.

  5. "The reviewer did not provide specific evidence or clear justification to support this concern." I think the visual similarity is quite high between [1]'s figures and this paper's. Hope this is clear enough to justify this concern.

AuthorsOpenReview ↗

Response to Reviewer

Our manuscript already provides examples of high-quality molecules generated by our method, as shown in Figure 4.

We disagree with the reviewer’s plagiarism concern and consider the justification insufficient and unsupported by specific evidence. As we have clearly stated, all figures were created by us and illustrate scientific concepts that are substantively different from those presented in the cited work. Our models differ from those in the work cited by the reviewer, and it is clear that these are distinct studies.

The reviewer has still not identified which specific figures, panels, or graphical elements are allegedly copied or highly similar. The statement that “visual similarity is quite high” provides no specific, verifiable basis for the concern. For example, visualizations of protein or compound structures produced independently by different models might look similar. Such visual resemblance alone does not establish that one figure was copied from another.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.