acceptodds
Under review as a conference paper at ICLR 2027

Atomistic Representation Alignment for 3D Molecule Generation in Protein Pockets

Abstract

Structure-based generative models have emerged as a powerful tool for creating novel 3D molecules within protein binding sites. However, these models often struggle to capture the fine-grained physical constraints of molecular systems. While typically trained on datasets of crystal structures and docked complexes, this spatial data provides only geometric supervision, lacking explicit information about the underlying energy landscape and atomic forces. In this work, we propose enhancing pocket-conditioned generative models by leveraging the rich internal representations of a machine-learned interatomic potential (MLIP). Because MLIPs are pre-trained on diverse quantum mechanical calculations, their latent spaces provide a highly structured, atom-aware target for our generator to learn from. By applying an equivariant projector to match the features of a denoiser layer to the MLIP, we expose the generator to a dense, physically grounded signal during training. We observe a significant improvement in the overall structural validity of the generated molecules. Evaluated on a challenging test set of out-of-distribution targets, we demonstrate that representation alignment transfers across architectures: when applied to two independent pocket-conditioned generators, it yields molecules with greater substructure overlap with native ligands and improves protein–ligand interaction fingerprint recovery in both, indicating more faithful reproduction of native binding modes. Furthermore, after a pocket-constrained relaxation in an independent computational force field, the fraction of generated molecules that are simultaneously valid, clash-free, low-strain, and well-docked increases. Together, these results indicate that aligning to an MLIP yields models whose outputs better recover native interaction patterns and converge to more physically favorable geometries.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative Models for Molecules, Structure Based Drug Discovery, Representation Alignment, Machine Learning Interatomic Potentials

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_G5tMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

REPA is a strategy for pocket conditional 3D generative models that allow to align the intermediate features of the 3D generative model (like FlowR) with a pre-trained MLIP model. The method require additional training cost with no additional inference cost overhead, reaching higher posebuster validity, better strains and interaction recovery on CrossDocked2020 and SPINDR.

Critical strengths
  • The physics aware supervision is well defined in the paper. Additionally, results seems solid for two different architectures two architectures (DrugFlow and FlowR) and two datasets (CrossDocked2020 and SPINDR).

  • The ablations provide results on different setup, ie encoder type, layers in the architecture and loss weights. Even though some relevant details like training cost overhead (ie running uma frozen) or discussions of MLIP size tradeoff seems to be missing.

  • In tables 1-5, the paper provide insights on several metrics tradeoff with direct comparison to the native ligand.

Critical weaknesses
  • The improvements seem not statistically significant and consistent in table 1 and 2 (ie see Vina and/or Strains). For instance, on FlowR strains are going from 92.4 -> 101.1 and doesn't seem to be clearly explained in the paper.

  • Since MLIP are known to yield DFT performances, metrics based on MMFF94s strain and Vina are both known to be less reliable proxies. The results could benefit from higher quality evaluation metrics like DFT single-points and FEP on very few samples.

  • Now results are mostly reported per-metric fixing the training budget for FlowR. It would be interesting to see the reported tradeoffs every K epochs during the training.

Questions for authors
  • Figure 1 in the caption miss: (1) what pretrained model was used and (1) how many molecules were generated. Moreover, it is not referenced in the introduction but it is in the results. It's not clear why is being placed in the front page.

  • The caption in tables 1 to 5 seem too synthetic. I would shortly add more explicit details, like how many molecules were sampled and for how many epoch both baseline and baseline+REPA were tuned. Additionally, I would add in the caption the number of denoising steps (NFE) for inference.

  • Why figure 6 in the appendix shows just the training curves for DrugFlow and not for FlowR?

  • It is unclear why FlowR's strain increase with REPA while DrugFlow's decreases.

  • It is unclear why the features of the pocket wouldn't benefit from this alignment. This is something the authors should better clarify in the paper.

  • Can you provide more details in the paper on how the model is sensitive to different protonation schemes?

  • Reporting xTB single-point energies on a subset would confirm that MMFF94s strain reflect a true improvement.

6 comments
AuthorsOpenReview ↗

Response to Reviewer G5tM - part 1/4

We appreciate the thoughtful comments of the reviewer. We address each of them below and support our answers by new experiments. Our response is split into four parts: (1) physics evaluation, (2) the FLOWR strain question, (3) FLOWR training curves and clarifications, (4) pocket alignment and protonation schemes.

Physics evaluation (W2, Q7)

W2: Since MLIP are known to yield DFT performances, metrics based on MMFF94s strain and Vina are both known to be less reliable proxies. The results could benefit from higher quality evaluation metrics like DFT single-points and FEP on very few samples.

Q7: Reporting xTB single-point energies on a subset would confirm that MMFF94s strain reflect a true improvement.

We thank the reviewer for suggesting additional energy evaluations. Extending our evaluation with higher-fidelity potentials moves beyond empirical proxies like Vina and MMFF94s, providing a more rigorous assessment of ligand strain and protein–ligand interaction energies.

For each model and pocket, we selected the PoseBusters-valid molecule with the best as generated Vina score. We then evaluated each selected candidate in three states: as generated, after RDKit bond/angle repair, and after in-pocket minimization with the protein held fixed using:

  • GFN2-xTB/ALPB water, a semiempirical quantum check independent of learned interatomic potentials,
  • MACE-OMol-0 (gas phase), an alternative MLIP to UMA, our alignment target, avoiding an overly redundant evaluation.

Net energy is the sum of ligand–pocket interaction energy and ligand deformation energy relative to a low-energy unbound reference conformer of the same ligand. Table D reports the number of SPINDR targets, out of 222, whose selected candidate has a net energy below the median for native ligands evaluated with the same energy method. To test whether these gains require large pose changes, the last two columns additionally require a heavy-atom RMSD ≤0.5 Å.

Table D: Number of SPINDR targets (out of 222) whose selected candidate passes the net-energy criterion. Candidates must fall below the native-ligand median: −16.94 kcal/mol for xTB or −97.45 kcal/mol for MACE. The last two columns additionally require a heavy-atom RMSD ≤0.5 Å from the generated pose. Missing or invalid results count as unsuccessful; higher is better.

ModelAs generated xTBAs generated MACEBond/angle repair xTBBond/angle repair MACEMinimized xTBMinimized MACEMinimized, RMSD ≤0.5 Å xTBMinimized, RMSD ≤0.5 Å MACE
DrugFlow7581062115766830
+ REPA12662069118886846
FLOWR2431958124728833
+ REPA2462361132899559

REPA increases the number of targets passing the criterion for DrugFlow across all settings, and for FLOWR after bond/angle repair and after in-pocket minimization. In the updated manuscript, we will include plots showing the number of successful candidates across all net-energy thresholds, not just the native-ligand median.

Enforcing pose preservation (last two columns of Table D) amplifies REPA's advantage under MACE, yielding statistically significant gains across both backbones (paired McNemar p=0.029 for DrugFlow; p=0.0004 for FLOWR) without requiring large conformational shifts; under xTB, DrugFlow is unchanged (p=1.00) and FLOWR increases modestly from 88 to 95 (p=0.470).

Large-scale DFT calculations were not feasible within the rebuttal period. We therefore performed two targeted case studies, one for FLOWR and one for DrugFlow, comparing a pair of baseline and REPA candidates with the same charge. We selected pairs where REPA showed a net-energy advantage after bond/angle repair. Using ωB97M-D3BJ/def2-SVP with C-PCM water on the full prepared, capped pockets, REPA retains its advantage in both cases: net energy −10.1 vs. −24.2 kcal/mol for FLOWR (baseline vs. REPA) and −32.7 vs −45.6 kcal/mol for DrugFlow (baseline vs. REPA). We will include the full protocol and example candidates in the updated manuscript.

Together, these results show that REPA improves candidate yield after structural refinement under xTB, MACE and, in these examples, DFT, extending the evidence beyond MMFF/Vina. We have not performed FEP; further binding free-energy calculations and experimental measurements would help assess whether these energetic improvements translate into higher binding affinity.

AuthorsOpenReview ↗

Response to Reviewer G5tM - part 2/4

Framing and the FLOWR strain question (W1, Q4)

W1: The improvements seem not statistically significant and consistent in table 1 and 2 (ie see Vina and/or Strains). For instance, on FlowR strains are going from 92.4 → 101.1 and doesn't seem to be clearly explained in the paper.

Q4: It is unclear why FlowR's strain increase with REPA while DrugFlow's decreases.

Thank you for pointing this out. Indeed, we do not have a definitive mechanistic explanation for the increased strain in molecules generated by FLOWR when trained with REPA. Empirically, however, we observe that compared to DrugFlow, even baseline FLOWR produces bond lengths with high variance—introducing stretched and compressed bonds that drive up the raw strain of generated molecules (Table 10). We hypothesize that this difference between generators stems from their architectures: DrugFlow is a heterogeneous GNN that explicitly models bond features, whereas FLOWR operates globally via an attention mechanism.

To investigate further, we added xTB and MACE evaluations that separately assess ligand deformation (strain) and ligand–pocket interaction energies:

Table E. Median energy components in kcal/mol (lower is better). Deformation is measured relative to a low-energy unbound conformer of the same ligand; "fixed" denotes values after RDKit bond/angle repair. Net fixed is the sum of repaired deformation and ligand–pocket interaction energies, computed for each candidate before taking the median. The component medians therefore need not sum to the median net energy. Results use a common set of pockets with valid calculations across all compared settings.

ModelEvaluatorStrain generatedStrain fixedInteraction fixedNet fixed
DrugFlowxTB33.430.0−25.25.7
+ REPAxTB30.823.6−23.63.4
FLOWRxTB42.022.4−23.7−1.4
+ REPAxTB53.123.7−25.9−0.5
DrugFlowMACE48.540.9−113.8−64.5
+ REPAMACE42.737.2−121.1−70.8
FLOWRMACE55.431.6−93.8−51.4
+ REPAMACE71.135.3−98.3−64.1

Both methods reproduce FLOWR's higher raw deformation with REPA, but local repairs involving approximately 0.05 Å heavy-atom movement substantially reduce the excess—from 11.1 to 1.3 kcal/mol under xTB. This suggests that much of the additional strain reflects small local geometric distortions. After repair, FLOWR's interaction energies improve under both methods, while its median net energy improves under MACE but not xTB. For DrugFlow, REPA lowers repaired xTB deformation (p=0.014) and repaired MACE net energy (p=0.024). We will clarify these architecture-dependent effects and the unresolved mechanism in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer G5tM - part 3/4

FLOWR training curves and clarifications (W3, Q3; Q1, Q2)

W3: Now results are mostly reported per-metric fixing the training budget for FlowR. It would be interesting to see the reported tradeoffs every K epochs during the training.

Q3: Why figure 6 in the appendix shows just the training curves for DrugFlow and not for FlowR?

We agree that FLOWR's training dynamics should also be reported. We now provide checkpoint results (Table J), evaluated on 68 pockets (SPINDR validation set) with 10 samples per pocket across three seeds. REPA yields higher PoseBusters validity and lower strain at early checkpoints. By epoch 300, it retains higher PoseBusters validity (0.874 versus 0.855), but has higher median strain (60 versus 53 kcal/mol), consistent with the trade-off observed in the final test-set results in Table 2. We will include these curves in the revised appendix. Evaluating intermediate checkpoints on the full test set would require additional sampling beyond our rebuttal compute budget; Vina scores were not recorded during training.

Table J. Validation performance of FLOWR and FLOWR + REPA at selected training epochs. Results are averaged over three seeds, with 10 molecules generated for each of 68 validation pockets per seed. Higher PoseBusters validity and lower median strain (kcal/mol) are better.

epochPoseBusters FLOWRPoseBusters FLOWR+REPAmedian strain FLOWRmedian strain FLOWR+REPA
200.2370.470697246
600.7490.80310792
1000.8190.8377672
2000.8580.8635864
3000.8550.8745360

Clarifications

Q1: Figure 1 in the caption miss: (1) what pretrained model was used and (1) how many molecules were generated. Moreover, it is not referenced in the introduction but it is in the results. It's not clear why is being placed in the front page.

We agree that this training comparison fits better with the experimental results. We will move Figure 1 to Section 5.1 and expand its caption to identify the pretrained UMA-S-1p1 teacher, the number of generated molecules, and the aggregation across seeds:

"Figure 1. Representation alignment improves PoseBusters validity and reduces strain for DrugFlowEMA. Training curves compare DrugFlowEMA with and without representation alignment to the frozen pretrained UMA-S teacher. At each checkpoint, each model generates 100 molecules for each of 100 CrossDocked2020 test pockets (10,000 molecules per seed). Lines show means and shading indicates ±1 standard deviation across three seeds. Higher PoseBusters validity and lower median MMFF94s strain are better."

Q2: The caption in tables 1 to 5 seem too synthetic. I would shortly add more explicit details, like how many molecules were sampled and for how many epoch both baseline and baseline+REPA were tuned. Additionally, I would add in the caption the number of denoising steps (NFE) for inference.

Thank you for this suggestion. We will expand the captions of Tables 1–5 to specify the number of test pockets, generated molecules, seeds, training epochs, and denoising steps. For our main comparisons, we generate 100 molecules per pocket per seed, using 100 CrossDocked2020 pockets or 222 SPINDR pockets, and average results across three seeds. Within each architecture, the baseline and REPA variant use the same training and sampling schedules: 600 epochs and 500 denoising steps (NFE) for DrugFlow, and 300 epochs and 100 denoising steps (NFE) for FLOWR. External baselines in Table 1 use released checkpoints and their default sampling settings, which we will distinguish explicitly.

Reviewer_AZ2tOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds representation alignment to pocket-conditioned molecular generators. Intermediate scalar and vector features are projected onto representations from a frozen interatomic potential, UMA, evaluated on clean protein–ligand complexes. The auxiliary loss is used only during training, leaving sampling unchanged. Experiments on DrugFlow and FLOWR show improved DrugFlow validity, greater similarity to native ligands, and higher candidate yield after restrained energy minimization. The main contribution is extending MLIP-based alignment to joint ligand identity and pose generation in protein pockets.

Critical strengths
  1. The comparisons include useful controls. DrugFlow is strengthened with EMA before adding alignment, and the main results are averaged over three seeds. Table 7 compares several MLIP teachers and includes both shuffled UMA features and a trained DrugFlow teacher. Table 8 tests scalar-only versus scalar-and-vector alignment. These experiments provide evidence that the gains depend on the teacher representations and their atomwise correspondence, beyond simply adding another training loss.

  2. The evaluation shows benefits beyond coordinate refinement. MCS and PLIF Tanimoto improve for both generators on SPINDR (Table 4), and the graph-based MCS improvement indicates a change in generated chemistry. After the same reported relaxation procedure, high-quality yield increases from 21.4% to 23.9% for DrugFlow and from 27.9% to 29.2% for FLOWR, with reported significance in both cases (Table 3). These are useful gains without additional sampling cost.

Critical weaknesses
  1. The claim of improved physical quality across architectures is stronger than the results support. For FLOWR, raw strain increases from 92.4 to 101.1 kcal/mol, while raw high-quality yield falls from 1.0% to 0.5% (Tables 2–3). Table 10 also shows increased raw stiff strain in both backbones. The authors acknowledge these results in Section 5.2 and Appendix D.1, but the conclusion still claims higher validity and lower strain across architectures. The evidence supports a narrower finding: alignment can improve the yield of a generation-plus-relaxation pipeline, despite making some aspects of the initial geometry worse. Likewise, the observation that relaxing native ligands worsens Vina does not establish that DrugFlow’s worse Vina scores reflect a harmless physical trade-off rather than poorer binding (Section 5.1). This matters because transferring useful physical constraints is the paper’s central scientific claim.

  2. The strongest shared benefit depends on a relaxation and evaluation protocol that is not sufficiently specified. Section 5.3 gives the force field, iteration count, and a 1 Å restraint, but does not clearly define the restraint function and force constant or the ligand–protein energy terms used during minimization. Appendix D.1 defines the separate vacuum minimization for measuring strain, which does not resolve these details. There is also an ambiguity in the reported results: Table 13 combines raw PoseBusters values with relaxed high-quality yields, without identifying the different pose states. The state used for PLIF in Tables 4, 13, and 14 is unclear. Consequently, I cannot determine whether the improved native interaction recovery survives the same relaxation that produces the high-quality yield gains. These are reporting gaps, not evidence of an incorrect implementation, but they limit confidence in the main practical result.

Questions for authors
  1. What is the exact objective used for in-pocket minimization, including ligand–protein interactions and the restraint function and strength? Do the high-quality yield gains persist under modest changes to the restraint settings and quality thresholds?

  2. Are the PLIF results in Tables 4, 13, and 14 computed before or after relaxation? Please identify the pose state for each metric and, if PLIF is currently measured before relaxation, show whether its improvement survives afterward.

  3. How much does the protein context in the teacher contribute? A comparison using UMA features of the ligand alone, while keeping the generator pocket-conditioned, would help distinguish the benefit of general ligand representations from the transfer of pocket-dependent information. The existing controls in Table 7 do not isolate this distinction.

5 comments
AuthorsOpenReview ↗

Response to Reviewer AZ2t - part 1/4

We appreciate the reviewer’s thoughtful comments. We address each point below with additional experiments and clarifications. Our response is split into four parts: (1,2) physical-quality claims and additional energy evaluation, (2) relaxation protocol and sensitivity to restraints and quality thresholds, (3) PLIF recovery before and after relaxation, and (4) the contribution of pocket context in the UMA teacher.

Scope of the physical-quality improvements and additional evaluations.

W1: The claim of improved physical quality across architectures is stronger than the results support. For FLOWR, raw strain increases from 92.4 to 101.1 kcal/mol, while raw high-quality yield falls from 1.0% to 0.5% (Tables 2–3). Table 10 also shows increased raw stiff strain in both backbones. The authors acknowledge these results in Section 5.2 and Appendix D.1, but the conclusion still claims higher validity and lower strain across architectures. The evidence supports a narrower finding: alignment can improve the yield of a generation-plus-relaxation pipeline, despite making some aspects of the initial geometry worse. Likewise, the observation that relaxing native ligands worsens Vina does not establish that DrugFlow’s worse Vina scores reflect a harmless physical trade-off rather than poorer binding (Section 5.1). This matters because transferring useful physical constraints is the paper’s central scientific claim.

We agree that our framing should more clearly distinguish universal improvements after relaxation from the architecture-dependent effects on as-generated geometry. We will revise the text accordingly, with revised passages highlighted in blue. For example, we will rephrase the second contribution in the Introduction as follows:

2. Improved Molecular Validity and Native Similarity: We demonstrate that this alignment acts as a transferable inductive bias across independent architectures and datasets. Across both evaluated generators, aligned models produce molecules with higher baseline PoseBusters validity, more faithfully recover native protein–ligand interaction fingerprints on out-of-distribution targets, and exhibit greater substructure similarity to native binders, while also reducing raw internal strain in DrugFlow.

Because our third contribution already specifically covers the post-minimization improvements across both architectures, this updated formulation accurately reflects our empirical findings.

We also agree that the native-ligand relaxation result does not explain the Vina difference between baseline and REPA-generated molecules. Motivated by these concerns about strain and Vina scores, we added an evaluation using two independent methods: semiempirical quantum mechanics method (GFN2-xTB/ALPB water) and an alternative MLIP (MACE-OMol-0, gas phase). These methods evaluate ligand strain, protein–ligand interaction energy and their sum as a net balance representing total favorability of the protein-ligand interaction represented by the modeled complex[3].

For each model and pocket, we selected the PoseBusters-valid molecule with the best initial Vina score. Each candidate is scored as generated, and after minimization inside the fixed protein pocket (same as in Section 5.3). For this rebuttal, we also include an intermediate setting: restrained RDKit bond/angle repair, which corrects local bond lengths and angles while limiting changes to the overall pose. Table D reports the number of SPINDR targets whose selected candidate has a net energy below the median for native ligands evaluated with the same energy method.

Table D: Number of SPINDR targets (out of 222) whose selected candidate passes the net-energy criterion. Candidates must fall below the native-ligand median: −16.94 kcal/mol for xTB or −97.45 kcal/mol for MACE. Missing or invalid results count as unsuccessful. Higher value is better.

ModelAs generated: xTBAs generated: MACEBond/angle repair: xTBBond/angle repair: MACEAfter minimization: xTBAfter minimization: MACE
DrugFlow758106211576
+ REPA1266206911888
FLOWR243195812472
+ REPA246236113289

REPA increases the number of targets passing the net-energy criterion for DrugFlow across all settings. For FLOWR, REPA increases the number of targets passing the net-energy criterion after bond/angle repair and after in-pocket minimization. This extends the evidence for a benefit after structural refinement beyond the original MMFF/Vina-based evaluation in Section 5.3. In the updated manuscript, we will include plots showing the number of successful candidates across all net-energy thresholds, not just the native-ligand median.

This response continues in part 2/4.

[3]Levine et al., The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models. https://arxiv.org/abs/2505.08762

AuthorsOpenReview ↗

Response to Reviewer AZ2t - part 2/4

To test whether gains in part 1 of our response require large pose changes, we added a heavy-atom RMSD ≤0.5 Å requirement, measured against generation in the protein frame without fitting (Table E).

Table E. Net energy successes with a pocket minimum within 0.5 Å of the generated pose. Unit is the number of pockets for which the candidate moves below 0.5 Å RMSD during minimization and obtains net energy below the median net energy of native ligands calculated using xTB (−16.94 kcal/mol) or MACE (−97.45 kcal/mol).

ModelxTBMACE
DrugFlow6830
+ REPA6846
FLOWR8833
+ REPA9559

Counts are out of 222 targets and require passing both energy and movement thresholds. Enforcing pose preservation amplifies REPA's advantage under MACE, yielding statistically significant gains across both backbones (paired McNemar p=0.029 for DrugFlow; p=0.0004 for FLOWR) without requiring large conformational shifts; under xTB, DrugFlow is unchanged (p=1.00) and FLOWR increases modestly from 88 to 95 (p=0.470).

Together, these results strengthen the evidence that REPA improves candidate yield after small structural refinement.

Relaxation and evaluation protocol

W2a: The strongest shared benefit depends on a relaxation and evaluation protocol that is not sufficiently specified. Section 5.3 gives the force field, iteration count, and a 1 Å restraint, but does not clearly define the restraint function and force constant or the ligand–protein energy terms used during minimization.

Q1: What is the exact objective used for in-pocket minimization, including ligand–protein interactions and the restraint function and strength? Do the high-quality yield gains persist under modest changes to the restraint settings and quality thresholds?

We thank the reviewer for raising this important point. We agree that the relaxation and evaluation protocol need to be specified clearly. We follow the relaxation procedure implemented in GenBench3D [4] with its default restraint settings, since it was also used in the original FLOWR publication. It minimizes MMFF94s ligand internal energy together with ligand–protein van der Waals and electrostatic interactions. All protein atoms remain fixed, while ligand hydrogens are optimized without positional restraints. Each ligand heavy atom incurs no restraint penalty within 1 Å of its starting position and a quadratic penalty beyond that distance, with force constant 999 kcal/mol/Ų. We allow up to 1,000 minimization steps and will specify these details in Section 5.3.

To test whether different restraint settings affect the observed improvement, we relaxed samples under several settings (Table F). Each number is REPA high quality (HQ) percentage minus baseline HQ percentage, in percentage points, using the original strain ≤20 and Vina ≤−6 cutoffs. Positive values mean more HQ molecules with REPA. For example, +1.15 corresponds to 28.74% with REPA versus 27.59% for baseline. The observed HQ advantage remains positive at every tested restraint setting.

Table F. Sensitivity of High Quality (HQ) improvement to relaxation restraints for SPINDR dataset. Entries report REPA minus baseline HQ fractions in percentage points; positive values favor REPA. Results use one seed, with the quality criteria held fixed across all restraint settings.

Restraint radius (Å)Force constant (kcal/mol/Ų)DrugFlow HQ gain (pp)FLOWR HQ gain (pp)
1.00 (default)999+2.42+1.15
0.75999+2.63+1.43
1.25999+2.43+1.09
1.00499.5+2.42+1.15
1.001998+2.44+1.18

Separately, we tested sensitivity to the strain and Vina thresholds using the existing poses relaxed under the default protocol; no additional minimization was performed. Table G reports REPA minus baseline HQ fractions in percentage points, with the validity and clash criteria unchanged.

Table G. Sensitivity of HQ improvement to strain and Vina thresholds for SPINDR dataset. Entries report REPA minus baseline HQ fractions in percentage points, evaluated on the same poses relaxed under the default protocol. Positive values favor REPA.

Strain ≤ (kcal/mol)Vina ≤ (kcal/mol)DrugFlow HQ gain (pp)FLOWR HQ gain (pp)
15−6.5+1.42+1.50
15−6.0+1.91+1.55
15−5.5+2.65+1.62
20−6.5+1.91+1.29
20−6.0+2.56+1.30
20−5.5+3.36+1.26
25−6.5+2.10+0.85
25−6.0+2.79+0.78
25−5.5+3.68+0.67

Both architectures retain positive HQ differences across all five restraint settings and all nine threshold combinations, supporting robustness to the tested changes in relaxation and evaluation.

[4] Baillif et al., Benchmarking structure-based three-dimensional molecular generative models using GenBench3D: ligand conformation quality matters. arXiv:2407.04424 (2024)

AuthorsOpenReview ↗

Response to Reviewer AZ2t - part 3/4

PLIFs before and after relaxation

W2b: There is also an ambiguity in the reported results: Table 13 combines raw PoseBusters values with relaxed high-quality yields, without identifying the different pose states. The state used for PLIF in Tables 4, 13, and 14 is unclear.

Q2: Are the PLIF results in Tables 4, 13, and 14 computed before or after relaxation? Please identify the pose state for each metric and, if PLIF is currently measured before relaxation, show whether its improvement survives afterward.

Thank you for raising this fair point. We will revise the manuscript to make the pose state explicit throughout.

All PLIF metrics in Tables 4, 13 and 14 are computed on the generated poses, before the pocket-constrained minimization of Section 5.3. MCS is computed on the molecular graph and is pose-independent. In Table 13, PoseBusters and PLIF recovery are likewise computed for molecules as generated, while High quality is the post-minimisation value; we will make the values consistent in all tables in the updated manuscript. We also explicitly note that by default, the fingerprinting step itself adds hydrogens and optimises them with MMFF94s, holding the pocket and all ligand heavy atoms fixed, since ProLIF requires explicit hydrogens to assign H-bonds; the reference ligand is treated identically and no heavy atom is moved.

Following the reviewer's suggestion we repeated the interaction analysis on the minimised poses, for all four de-novo models over three seeds:

Table H: PLIF based metrics before and after minimization for SPINDR dataset. Higher values are better.

MethodPLIF recoveryPLIF Tanimoto
as generated → after minimizationas generated → after minimization
DrugFlow0.412 → 0.4150.362 → 0.358
+ REPA0.423 → 0.4230.373 → 0.367
FLOWR0.452 → 0.4490.399 → 0.391
+ REPA0.475 → 0.4670.417 → 0.405

Relaxation slightly reduces fingerprint overlap for every model, which is the same trade-off we report for Vina in Section 5.3, and it affects both the baseline and the aligned models. The benefits of REPA are maintained: for FLOWR they remain significant in both states (recovery +0.023 → +0.018, Tanimoto +0.018 → +0.014, all p < 10⁻⁵), and for DrugFlow the Tanimoto advantage remains significant (+0.011 → +0.009, p < 0.05) while the recovery difference remains non-significant, as reported in the submission. The improvement is therefore not an artefact of the unrelaxed pose. We will add this comparison to the appendix of the updated manuscript.

Reviewer_j4GNOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Generative models for small molecules are often only trained to generate the geometry of the molecule. Beyond what is implied by the structure, there is therefore little supervision to learn physical and chemical properties that are important for the function of the generated molecule. This work attempts to address this by aligning the hidden layers of the generator to the hidden layers of a pretrained MLIP, based on the observation that such representation alignment has been beneficial for improving diffusion models for images. They apply this alignment to two different pocket-conditioned small molecule generators. Depending on the model and evaluation set, this alignment can lead to lower strain, higher PoseBusters pass rates, and better Vina scores, though the results are mixed. The alignment also leads to better recovery of native interactions with the pocket.

Critical strengths

The reliance on geometric supervision of current models is a timely issue, and looking for ways to teach the models about the physics that determine properties like binding affinity or steric strain is important. Representation alignment has also been relatively under-explored in the context of molecular generative models. The paper's approach and experiments are technically sound. It properly handles the equivariant features used by the denoiser and MLIP, and includes ablations of different MLIPs, injection and extraction layers.

Critical weaknesses

I see three weaknesses in the paper. First, it is not clear whether the improvement comes specifically from aligning to a model trained to predict physical properties (like the MLIPs used in the paper), or whether similar gains would be achieved from aligning to any pretrained molecular representations, including representations trained in a geometric way. While the authors include ablations with different MLIPs, exploring alignment to other non-MLIP representations could complete the picture. Second, it is unclear how far this approach will scale since aligning to fixed pretrained representations imposes a ceiling on performance. For image generators, more recent work has considered directly adding a self-supervised objective into the generator training algorithm (see SelfFlow [1] for example). In this case, it seems one could add auxiliary objectives like predicting energies from the model's predictions. Finally, a molecular generator samples both a molecular identity (the molecular graph) as well as a docked pose. Thus, if we consider the quality of the docked pose, this confounds the quality of the generated molecule (in its ideal docked pose) with the quality of the model's prediction of the pose. It is therefore unclear whether improving the quality of the generator's predicted pose actually improves the quality of the generated molecules, or just their predicted poses.

references: [1] Chefer, Hila, et al. "Self-supervised flow matching for scalable multi-modal synthesis." arXiv preprint arXiv:2603.06507(2026)

Questions for authors
  • How do the terms in the strain calculation compare to the energies modeled by UMA? One uses heuristic rules, while the other is trained on DFT energies. Did the authors also measure the predicted energies under UMA itself, or even using DFT (if feasible)?
  • Have the authors explored alignment to non-MLIP representations, or directly training the model with an auxiliary force/energy prediction objective for example?
  • Have the authors tried evaluating the designs by e.g. predicting the structures or redocking based on just the molecular identity of the designs (the molecular graph), while discarding the poses predicted by the model itself? The relaxations do this to an extent, but still start from the pose generated by the model. This would be more similar to e.g. the refolding evaluations done for protein binder design. This would help elucidate whether improving the quality of the poses generated by the model actually leads to finding molecules with better affinities/properties.
5 comments
AuthorsOpenReview ↗

Response to Reviewer j4GN - part 1/4

We appreciate the thoughtful comments of the reviewer. We address each of them below and support our answers by new experiments. Our response is split into four parts: (1) geometric baselines, (2) auxiliary energy/force objectives, (3) cofolding evaluation, (4) physics evaluation.

Geometric baselines (W1)

W1: It is not clear whether the improvement comes specifically from aligning to a model trained to predict physical properties (like the MLIPs used in the paper), or whether similar gains would be achieved from aligning to any pretrained molecular representations, including representations trained in a geometric way. While the authors include ablations with different MLIPs, exploring alignment to other non-MLIP representations could complete the picture.

We thank the reviewer for this question and agree that it is important to understand whether the improvements come specifically from aligning to a model trained on molecular energies and forces, or whether other molecular representations could provide similar benefits.

In our manuscript (Section C.2) we used a separately trained DrugFlow EMA model to provide the target representations. This serves as a control to test whether simply adding another training objective could explain the improvement. As shown in Table 7, this produced almost no improvement in PoseBusters validity (0.691 → 0.696), compared with 0.691 → 0.763 when aligning to UMA.

In Table A below, we extend our experiments to include alignment to two other pretrained molecular models, Uni-Mol2 [1] and TorchMD-Net [2]. Both models learn representations only from the 3D structures of molecules and don't require reference energies or forces during pretraining. Unlike UMA, Uni-Mol2 operates only on the ligand without the protein pocket. To make a fair comparison, we also show results of the UMA embeddings extracted without the pocket. For clarification, in the approach we propose in the paper we use the pocket as context for extracting the MLIP embedding (i.e. MLIP takes as input the whole protein-molecule complex), but we always align only the atom representations of the ligand.

Table A: Effect of encoder choice on physical and binding metrics on CrossDocked dataset. For UniMol2 and TorchMD-Net we used final layer encoder embeddings. All models were trained for 100 epochs, tested on 100 pockets with 100 samples per pocket.

MethodScalar/vector channelsPocketSizePB ↑Strain ↓Vina ↓
DrugFlowEMA-/--21.870.69181.6−5.86
+ UMA (Ours)128/128Yes21.820.76376.0−5.78
+ UMA128/128No21.670.76276.4−5.75
+ DrugFlowEMA128/32Yes21.400.69680.8−5.73
+ UniMol2768/0No21.880.70680.6−5.75
+ TorchMD-Net256/256Yes21.540.72878.1−5.72

Both models improve PoseBusters validity compared with the DrugFlow baseline (0.691). Uni-Mol2 increases it to 0.706, while TorchMD-Net reaches 0.728. However, neither achieves the improvement obtained with UMA (0.763). We observe a similar pattern for molecular strain: Uni-Mol2 and TorchMD-Net reduce it from 81.6 to 80.6 and 78.1 kcal/mol, respectively, while UMA reaches 76.0 kcal/mol.

Overall, these experiments show that alignment to other pretrained molecular representations can also improve generation. However, UMA consistently provides the largest improvements in PoseBusters validity and molecular strain, which supports the usefulness of representations learned from energy and force supervision. If you let us know which other models you would like to see, we will be happy to try them.

[1] Ji et al., Exploring Molecular Pretraining Model at Scale. NeurIPS (2024)

[2] Zaidi et al., Pre-training via Denoising for Molecular Property Prediction. ICLR (2023)

AuthorsOpenReview ↗

Response to Reviewer j4GN - part 2/4

Training setup: auxiliary energy and force objectives (W2, Q2)

W2: It is unclear how far this approach will scale since aligning to fixed pretrained representations imposes a ceiling on performance. For image generators, more recent work has considered directly adding a self-supervised objective into the generator training algorithm (see SelfFlow for example). In this case, it seems one could add auxiliary objectives like predicting energies from the model's predictions.

Q2: Have the authors explored … directly training the model with an auxiliary force/energy prediction objective for example?

We agree that directly incorporating additional objectives into the generator's training, as proposed in SelfFlow, is an interesting direction. However, our approach offers an orthogonal advantage: it uses existing representations learned from large-scale quantum-annotated datasets. These datasets cannot simply be added to the denoising model's training set, as they contain highly energetic conformations that are useful for learning molecular energetics but are not suitable as clean denoising targets.

To explore different objectives, we tried aligning the DrugFlow generator directly to atomic forces instead of UMA's intermediate representations. For each ligand heavy atom, we extracted the force predicted by UMA and used it as an alignment target. We used UMA rather than running DFT calculations because of the computational cost. The results are shown in Table B below.

Table B: Comparison between aligning to UMA intermediate embeddings and to UMA force predictions on CrossDocked dataset.

MethodSizePB ↑Strain ↓Vina ↓
DrugFlow EMA21.870.69181.6−5.86
+REPA (UMA embeddings, ours)21.820.76376.0−5.78
+REPA (align to forces)21.820.499103.3−5.04
+REPA (align to forces, time weighted)21.840.547103.1−5.12

As shown in the table, this approach performs poorly, reducing PoseBusters validity from 0.691 to 0.499 and increasing strain from 81.6 to 103.3 kcal/mol. We hypothesize that this is because predicting forces from highly noisy molecular coordinates might be too difficult for the generator. To test this, we introduce time-dependent weighting that gives lower weight to the REPA loss for samples with higher noise levels, where forces are harder to predict, using a quadratic weighting function (t²). This slightly improved PoseBusters validity to 0.547, but still performs considerably worse than the baseline.

We hypothesize that intermediate MLIP representations are easier to learn from than raw atomic forces because such embeddings offer a smoother alignment target while still capturing both molecular geometry and atomic interactions. Depending on the noise level along the generation trajectory, the generator can selectively leverage different components of this representation, rather than attempting to predict sharp local forces when the molecular structure is still heavily noised.

Although these experiments did not produce favorable results, predicting energies or forces at different stages of denoising, rather than matching forces calculated for the clean structure, remains an interesting direction for future work.

AuthorsOpenReview ↗

Response to Reviewer j4GN - part 3/4

Cofolding: pose vs. molecule identity (W3, Q3)

W3: Molecular generator samples both a molecular identity (the molecular graph) as well as a docked pose. Thus, if we consider the quality of the docked pose, this confounds the quality of the generated molecule (in its ideal docked pose) with the quality of the model's prediction of the pose. It is therefore unclear whether improving the quality of the generator's predicted pose actually improves the quality of the generated molecules, or just their predicted poses.

Q3: Have the authors tried evaluating the designs by e.g. predicting the structures or redocking based on just the molecular identity of the designs (the molecular graph), while discarding the poses predicted by the model itself? The relaxations do this to an extent, but still start from the pose generated by the model. This would be more similar to e.g. the refolding evaluations done for protein binder design. This would help elucidate whether improving the quality of the poses generated by the model actually leads to finding molecules with better affinities/properties.

We agree that pose-based metrics entangle molecule quality with the model's pose-prediction quality. In our submission we demonstrated one positive pose-independent result: molecules generated by models with REPA achieved better results in high maximal common substructure (MCS) with native ligands (see Table 4 in the submitted paper), suggesting that the generated molecules were topologically more native-like. But we are also very grateful for the co-folding suggestion, as it is another way of assessing the quality of the generated molecules independently of their pose. We therefore re-predict each design's complex from its molecular graph alone, discarding the model's pose, using Boltz-2 cofolding — the small-molecule analogue of protein-binder refolding.

For each model, we select the three PoseBusters-valid molecules with the best Vina scores per pocket. We give Boltz-2 only their SMILES and the full receptor, together with an MSA and a pocket constraint, so it predicts new poses without seeing the generated coordinates.

We compare each baseline with its REPA variant in Table C. We measure how closely the Boltz-2 poses match the generated poses using RMSD, and compare predicted affinity (log₁₀ IC50; lower indicates stronger binding) and binder probability.

Table C: Comparison between baseline and baseline+REPA metrics based on co-folding.

ModelMetricbase → REPApockets favouring REPAp
DrugFlowRefold RMSD (Å) ↓2.76 → 2.8151%0.90
DrugFlowBoltz affinity ↓−0.07 → −0.2360%0.001
DrugFlowBoltz affinity, refolds ≤2 Å ↓−0.14 → −0.2056%0.02
DrugFlowP(binder) ↑0.64 → 0.6249%0.62
FLOWRRefold RMSD (Å) ↓2.41 → 2.1353%0.12
FLOWRBoltz affinity ↓−0.08 → −0.1560%0.001
FLOWRBoltz affinity, refolds ≤2 Å ↓−0.11 → −0.1758%0.001
FLOWRP(binder) ↑0.65 → 0.6758%0.005

↓/↑ = lower/higher is better; values are medians over pockets; p = paired Wilcoxon.

Across all tested models, Boltz consistently reproduces the generated poses for a large fraction of designs (with the best design per pocket typically achieving <1 Å RMSD), demonstrating that the models reliably generate cofolding-consistent binding modes. Compared to its baseline, REPA achieves comparable or lower pose RMSDs while significantly improving Boltz-predicted binding affinity in the majority of pockets (p<0.05, across both models). Because Boltz computes affinity from its own refolded pose rather than the originally generated coordinates, this improvement reflects higher intrinsic ligand quality rather than artifacts of the generated pose. Importantly, this affinity gain persists even when restricting the analysis to designs where Boltz closely reproduces the generated conformation (≤2 Å RMSD), confirming that REPA yields higher-affinity ligands even if we demand consistency between generated and cofolded conformations.

Comments on the paper

1 comment

We sincerely thank the reviewers for their constructive feedback and productive discussion throughout the review period. Their suggestions helped us significantly strengthen the empirical evaluation and improve the clarity of the paper.

We have revised the manuscript to incorporate the new experiments, ablations, and clarifications discussed during the rebuttal and have uploaded the updated PDF, with changes highlighted in blue. Given the discussion-period timeline, a few minor additions (e.g., extending the protonation sensitivity pilot to FLOWR) are still in progress. Should the paper be accepted, we will include them in the camera-ready version and further polish the text so that all new results are fully integrated.

Summary of key additions during the rebuttal period:

1. Independent energy evaluations. We evaluated candidates using GFN2-xTB/ALPB and MACE-OMol-0 in three states: as generated, after bond and angle repair, and after in-pocket minimization with the respective evaluator. Across the 222 SPINDR targets, REPA increased the number of targets passing the native-ligand energy cutoff after minimization for both generators under both evaluators (Section 5.5, Table 5). We also separated ligand deformation from protein–ligand interaction energy and added DFT calculations for two selected candidate pairs (Appendix D.2, Tables 15 and 17).

2. Additional alignment baselines. We compared UMA representations with features from two molecular encoders pretrained without explicit energy or force supervision, Uni-Mol2 and TorchMD-Net. Both pretrained encoders improved PoseBusters validity and reduced strain, with UMA providing the largest gains among these three teachers (Appendix C.2, Table 8). Direct alignment to UMA-predicted atomic forces performed worse than the DrugFlow baseline, even with time-dependent weighting (Appendix C.7, Table 12).

3. Cofolding from molecular identity. We discarded the generated coordinates and used Boltz-2 to predict fresh poses from ligand SMILES and the target receptor. REPA improved Boltz-2-predicted binding affinity for both DrugFlow and FLOWR (Appendix D.4, Table 20).

4. Robustness and scope of the gains. We clarified that REPA reduces as-generated strain for DrugFlow but increases it for FLOWR (Section 5.2, Table 2). After relaxation, both generators maintained higher high-quality candidate yields than their respective baselines across all five restraint settings and all nine strain/Vina threshold combinations tested (Appendix D.3, Tables 18 and 19). The PLIF Tanimoto advantage also remained significant for both generators after relaxation (Appendix D.7, Table 24). Additional checks covered six protonation preparations for DrugFlow (Appendix D.5, Table 21), the contribution of pocket context to the UMA teacher, and alignment to pocket residue features (Appendix C.6, Table 11). We also measured 5.7% training overhead for DrugFlow, including teacher-embedding precomputation (Appendix B, “Computational cost,” p. 18).

We thank the reviewers again for their time and thoughtful engagement with our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.