acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic Evaluation of Induced-Fit Effects in Fragment-Based Drug Optimization

Abstract

Fragment-based lead optimization often involves relatively small chemical modifications, yet these changes can induce substantial rearrangements in ligand and receptor conformations and alter the preferred bound state among competing conformational minima. Accurately capturing and evaluating these ligand-dependent structural responses therefore remains a central challenge in structure-guided optimization. Here, we present a generation-and-evaluation framework for the probabilistic evaluation of a given bound structure. It first generates an ensemble of conformations around the input using flow matching and then evaluates their conditional log-probability scores. The generator learns to softly restrain scaffold positions near the initial ligand-fragment positions, focusing on sampling around the known binding context while allowing ligand and pocket conformations to adapt. Benchmarking on lead optimization tasks demonstrates that our generator provides more robust sampling than existing methods, producing valid and diverse conformations while accurately reproducing key interactions and induced-fit effects observed in experimental structures. The evaluator shows a meaningful correlation with the activity trends of congeneric ligand series without explicit affinity supervision of the generative model. Overall, these results suggest that the proposed strategy can serve as a competitive evaluation module in structure-based drug optimization workflows, replacing or complementing widely used docking and binding-affinity estimation methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure-based drug design, Fragment-based lead optimization, Protein–ligand conformational ensemble, Generative model

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_W7wtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a flow-matching framework for local protein–ligand conformational ensemble generation and probabilistic evaluation. It introduces learned soft scaffold restraints to preserve binding context while allowing ligand and protein side-chain flexibility. The framework further uses token-level log-probability estimates for conformer selection and ligand ranking. Experiments demonstrate competitive structural reconstruction performance, although improvements in downstream activity prediction remain limited.

Critical strengths
  • The paper addresses local conformational changes and induced-fit effects in fragment-based drug optimization, providing a useful perspective beyond single-pose evaluation.

  • The proposed method achieves strong structural validity and interaction recovery on PLINDER and Q-BioLiP benchmarks, outproming several establieded baselines.

Critical weaknesses
  • The paper does not sufficiently demonstrate the independent benefit of likelihood-based scoring over conventional scoring methods. Moreoveer, activity-ranking improvements over the BSA-only baseline are marginal (Spearman: 0.280 vs. 0.317), and the method underperforms Boltz-2.

  • The side-chain-fixed variant performs slightly better than the flexible model on all five structural metrics in the full Q-BioLiP benchmark. Improvements on the induced-fit-required subset are modest and evaluated on only 13 tasks, limiting the strength of the claimed benefits.

Questions for authors
  1. can the authors compare the proposed probabilistic evaluator against conventional scoring methods using the same generated conformers?

  2. Are the provements from likelihood-derived features over the BSA-only baseline statistically significant?

  3. Why does the side-chain-fixed variant outperform the flexible model on the full Q-BioLiP benchmark, despite the emphasis on induced-fit modeling?

3 comments

Response to Question 1 and Concern 1 — Independent benefit over conventional scoring

Can the proposed evaluator be compared with conventional scores using the same generated conformers?

We thank the reviewer for suggesting this direct comparison. We compared the contact-focused plausibility score with Vina on the same 32 conformers for each of 242 held-out PLINDER targets. Vina gave higher pairwise accuracy for all four tested quality metrics. For example, the accuracies were 55.4% versus 53.0% for lDDT-PLI and 69.7% versus 56.3% for clash count. The narrow quality range within the generated ensembles contributes to the weak ordering but does not fully explain it. Among pairs separated by at least 0.05 lDDT-PLI, Vina still reached 61.7% compared with 56.6% for our score. Thus, the current results do not establish an advantage over conventional whole-complex scoring.

At the residue level, the evaluated residue's own token score ranks the correct χ1\chi_1 state with 58.2% pairwise accuracy, compared with 50.3% for the mean score of the other residues. Residue-specific selection improves correctness by 5.21 percentage points over the conformer-pool mean. This indicates a modest local signal, not a global scoring advantage. We will revise the framing accordingly and present token decomposition as a local analysis tool rather than a replacement for conventional scoring.

Response to Question 2 and Concern 2 — Statistical significance of the likelihood-derived affinity features

Are the improvements over BSA-only statistically significant?

We thank the reviewer for requesting this comparison. The improvement over the BSA-only baseline is not statistically significant on the 39 held-out congeneric series. The full model increases target-mean Spearman correlation by 0.037 with a 95% paired target-bootstrap CI of -0.036 to 0.124 and pairwise accuracy by 0.012 with a CI of -0.022 to 0.052. We therefore do not claim that the divergence-derived features improve affinity ranking over BSA alone. We will remove affinity ranking from the Abstract and main Methods and Results and present it only as an exploratory supplementary analysis.

Response to Question 3 and Concern 3 — Why fixed side chains are slightly better on the full benchmark

Why does the fixed variant outperform the flexible model on all five full-set ligand-centered metrics despite the emphasis on induced fit?

We thank the reviewer for pointing out this apparent contradiction. The full-set average combines two different receptor regimes. Many substitutions occur in receptor pockets that are already preorganized for the new fragment. Keeping those side chains fixed preserves an appropriate pocket state, while sampling them introduces additional variance. Previous flexible-docking studies have similarly found that unnecessary receptor degrees of freedom can increase the search burden and scoring noise [1, 2]. The full benchmark is therefore not expected to show a universal benefit from flexibility.

Our residue-level analysis is consistent with this explanation. Flexibility is more useful for side chains that actually rearrange after substitution but can disrupt residues that are already preorganized. However, this analysis does not fully resolve the performance gap. Even on the induced-fit-required subset, improvements in the aggregate ligand-centered metrics are small or inconsistent. This suggests that the current sampler does not yet reliably focus the additional flexibility on the residues and conformations where it is needed.

We will revise the manuscript to make this limitation explicit and avoid implying that receptor flexibility is universally beneficial. We will also investigate whether the remaining gap arises primarily from flexible-residue selection, conformational sampling, or scoring of the generated states.

References

[1] Armen, R. S., Chen, J., and Brooks, C. L. III. ``An Evaluation of Explicit Receptor Flexibility in Molecular Docking Using Molecular Dynamics and Torsion Angle Molecular Dynamics.'' Journal of Chemical Theory and Computation 5, 2909--2923 (2009). https://doi.org/10.1021/ct900262t

[2] Park, H., Zhou, G., Baek, M., Baker, D., and DiMaio, F. ``Force Field Optimization Guided by Small Molecule Crystal Lattice Data Enables Consistent Sub-Angstrom Protein--Ligand Docking.'' Journal of Chemical Theory and Computation 17, 2000--2010 (2021). https://doi.org/10.1021/acs.jctc.0c01184

Thank you for your response. I have no further questions.

Reviewer_tSE6OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies local ligand and receptor conformational changes following fragment substitution. It uses coarse ligand anchors to constrain flow-matching sampling, jointly updates the ligand and nearby side chains, and decomposes the continuous-flow log density into atomic contributions for conformer scoring. Downstream ligand ranking combines ensemble-derived features with a separately trained ranking model. Validation covers local structure recovery, induced-fit examples, and ranking within congeneric ligand series.

Critical strengths

The coarse anchors provide a sensible structural constraint for fragment optimization. They preserve information from the existing binding mode while leaving room for local ligand changes and side-chain rearrangements, which suits the comparative nature of this task.

Critical weaknesses

The introduction motivates changes in state populations, while the method produces per-conformer plausibility scores. The central question in the introduction is how chemical substitutions change the relative populations of competing conformational states. The method computes pointwise densities conditioned on anchors, then clips, rescales, and averages their atomic contributions. A high density at one conformation does not imply a high total probability for the conformational state containing it. Appendix D further applies softmax weighting to generated samples using these transformed scores, without establishing a connection between the weights and state populations. The authors explicitly describe the scores as learned structural plausibility. This leaves a central question unresolved: what effect of chemical substitution does the evaluator capture, and what does it add beyond local structure generation and screening?

The log-density derivation does not cover the sampler described in the paper. Appendix A assumes that the vector field depends on the current coordinates and fixed conditioning. Appendix B.1 states that the default sampler carries the previous endpoint prediction across steps as an evolving self-conditioning input. This introduces an auxiliary state that depends on sampling history. The paper does not explain whether this mechanism is disabled during density calculation, or provide the state definition and change-of-variables derivation when it is retained. It is therefore unclear which generated distribution the reported log densities correspond to.

Questions for authors

Which statistic in the method corresponds to the relative populations of conformational states discussed in the introduction? How does it account for the conformational region belonging to a state, beyond comparing densities at individual conformations? Is self-conditioning enabled during log-density calculation? If so, how does the endpoint prediction carried across steps enter the flow state and its divergence calculation? Appendix D applies softmax weighting to conformations already sampled from the generator. What distribution is the weighted ensemble intended to represent? Do these weights serve structural selection, or are they intended to represent state probabilities?

2 comments

Response to Concern 1 — Pointwise plausibility and conformational-state populations

A high density at one conformation does not imply high integrated probability for the state containing it. Which method statistic corresponds to the relative state populations motivated in the introduction?

We thank the reviewer for raising this important distinction. None of the statistics in the current method corresponds to a relative thermodynamic state population. The raw structural score is a pointwise, pathwise quantity, and a softmax over finitely sampled conformers cannot replace the required basin-volume integral. Unweighted contact frequencies likewise describe the generated sample profile rather than thermodynamic populations. We will therefore remove the state-population interpretation and restrict the scope to local structural ensemble generation and evaluation around an assumed binding mode.

Although we cannot estimate state populations, we can evaluate which residue contacts change after a fragment substitution. We use each residue's unweighted contact frequency across 32 generated conformers and compare it with the contact change observed between the experimental source and target complexes.

Evaluation setTasks / groupsEnsemble-profile APStatic-pose APΔAP (95% CI)
All tasks with defined AP109 / 460.7440.6390.105 (0.052 to 0.156)
Induced-fit-required subset12 / 100.8450.6920.153 (0.068 to 0.268)

The 32-conformer profile improves AP from 0.639 to 0.744 on the full evaluable set and from 0.692 to 0.845 on the induced-fit-required subset. A single generated conformer does not improve over the static pose, indicating that the gain comes from persistent contact patterns across the ensemble. We will describe these outputs as ensemble-derived structural profiles, not as thermodynamic state probabilities.

Response to Concern 2 — Sampling-time self-conditioning in the practical score

Is self-conditioning enabled during density calculation? If so, how does the carried endpoint prediction enter the state and divergence?

We thank the reviewer for identifying this issue. Sampling-time self-conditioning is enabled. At step jj, the network receives the current coordinates xjx_j and the cached endpoint prediction hjh_j from the preceding step. We differentiate with respect to xjx_j while holding hjh_j fixed and accumulate this partial divergence along the sampled path. Because hjh_j itself depends on earlier steps, the accumulator omits derivative paths through the sampling history and is not the exact change-of-variables term for the complete sampler.

We will revise Appendix A to distinguish the standard memoryless-flow identity from this practical score. We will refer to the latter as a raw structural plausibility score, not an exact normalized log density. Its node terms remain additive contributions to the computed accumulator, but not to the exact density of the history-dependent generator.

We retain self-conditioning because a paired ablation over all 909 PLINDER test targets shows a clear sampling benefit.

MetricSelf-conditioning enabledSelf-conditioning disabled
Ligand RMSD (Å) ↓1.5351.662
lDDT-PLI ↑0.8170.793
χ₁ accuracy ↑0.8320.798
Clash count ↓6.5107.617

All paired bootstrap intervals exclude zero. This ablation justifies retaining self-conditioning for sampling quality but does not restore an exact-density interpretation of the accumulated score.

Response to Concern 3 — Meaning of the softmax-weighted ensemble

What distribution is the weighted ensemble intended to represent? Is it a structural-selection device or a state-population estimate?

We thank the reviewer for asking us to clarify this interpretation. The softmax-weighted ensemble does not represent a Boltzmann distribution, a corrected sampling distribution, or a conformational-state population. The weights are used only as a heuristic feature-aggregation rule in the exploratory ligand-ranking analysis. Conformers with higher model-derived structural plausibility contribute more to the averaged bound--reference features. Because there is no calibrated target distribution or importance correction, this should not be interpreted as probabilistic reweighting.

We will revise Appendix D to state this limitation where the weights are defined and remove any population-level interpretation. The structural fragment-substitution analyses, including contact-change profiling, use unweighted conformer occurrence frequencies.

Reviewer_tSE6OpenReview ↗

Post-rebuttal assessment: concessions confirm structural gaps in the evaluator claim

I thank the authors for the candid and substantive response. The acknowledgments on all three points are appreciated.

However, the three concessions together remove the theoretical foundation of the paper’s central claim. The title, abstract, Section 3.2 heading (“Token-Level Decomposition of Conditional Log-Probability”), and the narrative from Introduction through Discussion are built around probabilistic evaluation of conformational states. After the rebuttal:

The method does not estimate state populations (Concern 1, confirmed). The computed quantity is not an exact log-density of the generator (Concern 2, confirmed). The softmax weighting is a heuristic with no probabilistic interpretation (Concern 3, confirmed).

These are not presentation issues fixable by terminology changes. They affect what the method provides and what conclusions can be drawn from it. The proposed contact-frequency AP analysis is a useful addition, but it evaluates the generator’s ensemble quality, not the evaluator. The self-conditioning ablation demonstrates sampling improvement but does not validate the approximate score as a ranking signal. A comparison against simple structural quality proxies (e.g., clash count, contact preservation) would be needed for that. The BSA-only ablation in Table 7 (pairwise accuracy 0.607 vs. full model 0.619) suggests that the fitted ranking signal is dominated by a geometric feature rather than the probabilistic score.

The generator results (Tables 1 and 2) are strong and represent genuine progress for local conformer sampling. I encourage the authors to reframe the work around this strength. My score remains unchanged.

Reviewer_Mp5VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a local protein–ligand conformer sampler based on flow matching. It conditions on an initial complex and coarse ligand anchors, updates ligand and nearby side-chain coordinates while holding the backbone fixed, and estimates node-wise contributions to conditional log density. Stabilized ligand and contact-residue scores are used to rank generated conformers. A separate fitted linear model combines ensemble-derived score differences and buried-surface-area features to rank ligands within congeneric series. The structural benchmarks show strong recovery of scaffold and contact geometry, while the affinity-ranking result is useful but moderate and is weaker than Boltz-2.

Critical strengths
  1. The method addresses a relevant setting between rigid docking and unrestricted complex generation: a known scaffold pose is retained softly while a substituted fragment and nearby side chains may adapt.

  2. The paper separates all-sample structural quality from score-based selection. On PLINDER, the full model reports 98.0% target-mean PB-valid poses, 0.82 lDDT-PLI, and 0.65 interaction-fingerprint similarity. On Q-BioLiP cross-docking, it improves fragment metrics over the listed baselines.

  3. The appendices specify the node-wise divergence calculation, prior decomposition, clipping used for practical scores, training losses, benchmark construction, and per-target affinity-ranking failures. This helps distinguish a model likelihood from the final engineered ranking score.

Critical weaknesses
  1. The novelty is relatively limited. The main components—flow matching for conformer generation, anchor-based local restraints, side-chain updates, and continuous-flow likelihood estimation—are established techniques.

  2. The title and discussion emphasize “induced-fit effects,” but the measured benefit of updating side chains is limited. In Table 2, flexible versus fixed side chains changes most metrics only slightly; on the induced-fit-required Q-BioLiP subset, flexible sampling raises near-native fragment sampling success from 81.82% to 90.91%, yet top-1 and top-5 ranking success are unchanged (Table 10). The subset has only 13 tasks from 11 groups. A paired analysis focused on experimentally observed side-chain displacements is needed to show when flexible sampling materially improves the decision a user would make.

  3. The evaluator’s operational score is not a calibrated probability of binding or a free energy. The raw node scores have an additive log-density identity, but the contact-focused score clips each node value to an empirical interval and averages ligand and selected residue terms (Appendix A.2). The ligand-ranking model then fits five features using affinity labels from 53 training series (Appendix D.2). Claims about “probabilistic evaluation” should consistently distinguish raw density, heuristic conformer score, and supervised ligand ranking.

  4. The incremental contribution of learned likelihood features to activity ranking is modest. The full feature set reaches target-mean Spearman 0.317 and pairwise accuracy 0.619, compared with 0.280 and 0.607 using buried-surface-area features alone. Please provide uncertainty for these paired differences and a direct comparison with simpler contact or geometry descriptors. The main text notes that Boltz-2 is stronger on both ranking metrics.

  5. The conformer benchmarks start from known experimental binding contexts and the model uses coarse anchors from the input ligand. This is appropriate for local lead optimization, but it makes the results sensitive to the quality of the initial pose. It would help to quantify degradation under realistic pose error and under larger induced-fit changes. The existing PLACER perturbation control tests only modest local perturbations.

Questions for authors
  1. On the induced-fit-required subsets, which individual tasks benefit from side-chain motion, and are the fixed-versus-flexible differences significant under a paired group-level analysis?

  2. What is the test-set uncertainty for the full versus BSA-only ranking comparison in Table 7? How does a contact-geometry baseline perform with the same training split?

  3. How do conformer selection and ligand ranking change with the number of Hutchinson probes, ODE steps, and the empirical score-clipping interval?

  4. How robust are the results when the input scaffold pose is predicted or deliberately perturbed beyond the small local errors already tested?

4 comments

Response to Concern 1 — Limited novelty of the individual components

Flow matching, anchor-based local restraints, side-chain updates, and continuous-flow scoring are established techniques.

We thank the reviewer for raising this point. We agree that flow matching, local side-chain modeling, and divergence-based scoring are established individually. Our contribution is their integration for a specific lead-optimization task. The method generates local ligand--pocket ensembles around a source complex while retaining the shared scaffold and allowing the substituted fragment and nearby side chains to adapt.

PLACER is closest in structural scope but does not preserve the shared scaffold across a source ligand and its analog. rDock uses a hard scaffold restraint but keeps the pocket fixed. Our learned soft restraint occupies the intermediate regime, as supported by the Q-BioLiP results.

MethodScaffold restraintPocketScaffold RMSDFragment RMSDFragment lDDT-PLI
rDockHardFixed1.923.430.63
PLACERNoneFlexible3.334.670.57
OursSoft, learnedFlexible1.222.620.70

These methods differ beyond their restraint mechanisms, so we do not attribute the full performance difference to anchor conditioning alone. The ensemble also improves substitution-induced contact-change AP over a constructed static pose from 0.639 to 0.744 (paired difference 0.105, 95% CI, 0.052–0.156). We will revise the manuscript to make clear that the contribution is the fragment-substitution task and integrated generation-and-analysis workflow, not a new flow-matching or likelihood formulation.

Response to Concern 2 and Question 1 — When side-chain flexibility helps

The flexible and fixed variants differ only slightly on the aggregate ligand-centered metrics. A paired analysis focused on experimentally observed side-chain displacement is needed to determine when flexibility matters.

We thank the reviewer for requesting a more focused analysis. We agree that the current results do not establish a general benefit from enabling side-chain flexibility. Our residue-level analysis shows mixed behavior. Flexible sampling better recovers side chains that actually rearrange but can perturb side chains that are already preorganized. This tradeoff persists even on the predefined induced-fit-required subset, where the aggregate ligand metrics show only small or inconsistent gains.

Thus, the current analysis reveals a remaining performance gap rather than fully explaining when flexibility helps. We will revise our claims accordingly and investigate whether the degradation arises from flexible-residue selection, sampling, or scoring, with the goal of preventing unnecessary flexibility from disrupting otherwise correct local structures.

Response to Concern 3 and Question 3 — Score interpretation and numerical sensitivity

The operational score is not a calibrated binding probability or free energy. Clipping and aggregation break the additive raw-score identity, and the final ligand-ranking model is supervised.

We thank the reviewer for highlighting this ambiguity. We agree that these quantities must be distinguished more clearly. Because the cached self-conditioning input depends on earlier sampling steps but is held fixed when computing the divergence, the accumulated score is not the exact log density of the complete sampling map. We therefore interpret it only as a raw structural plausibility score. Its raw node terms are additive, but subsequent clipping, scaling, and contact-focused aggregation do not preserve that identity. The ligand-ranking score is a separate supervised model and is neither a probability nor a free-energy estimate.

We will revise the terminology and presentation accordingly, remove probability and thermodynamic-population interpretations, and present the standard change-of-variables identity only as motivation for the practical score. We will also clearly label the ligand-ranking analysis as exploratory and supervised and move it to the supplement.

Numerical sensitivity

We also thank the reviewer for asking about the numerical score settings. We evaluated sensitivity on a predefined 96-target PLINDER panel with 32 conformers per target using a fixed seed. At 25 ODE steps, conformer rankings were stable across Hutchinson probe counts. Using the reported three-probe setting as the reference, we computed the Spearman correlation between the 32 conformer rankings within each target and then averaged across targets.

Hutchinson probesMean within-target Spearman ρ with paper settingMean top-ranked-conformer lDDT-PLI
10.9100.820
20.9690.825
3 (paper setting)1.0000.825
40.9820.826
80.9620.820

All tested settings had a mean rank correlation of at least 0.910, while the top-ranked conformer's mean lDDT-PLI remained between 0.820 and 0.826. The correlations measure ranking stability relative to the paper setting, not convergence to an exact trace. No tested probe count materially improved selected-structure quality.

We evaluated clipping post hoc on identical raw node scores. Only 0.131% of PLINDER node values fell outside [-15, 1]. The unclipped, [-20, 1], and [-10, 1] scores had within-target rank correlations of 0.997, 0.9999, and 0.993 with the paper setting, respectively, and top-ranked-conformer lDDT-PLI remained 0.825 in all cases. Thus, clipping suppresses isolated trace outliers without materially determining conformer ranking.

Increasing the Euler integration from 25 to 50 or 100 steps also did not improve generation. Pool-level lDDT-PLI was 0.812, 0.808, and 0.803. Top-1 lDDT-PLI was 0.825, 0.815, and 0.813. Pool-level ligand RMSD was 1.612, 1.635, and 1.673 Å, respectively. We therefore retain 25 steps as the tested accuracy–cost setting.

Response to Concern 4 and Question 2 — Incremental activity-ranking benefit over BSA

Please provide uncertainty for the full-versus-BSA-only difference and compare with a simpler contact or geometry descriptor on the same split.

We thank the reviewer for requesting these controlled comparisons. We performed them on the same training and test splits. Relative to BSA alone, the full ranking model changed target-mean Spearman correlation by +0.037 (95% paired-bootstrap CI, -0.036 to 0.124) and pairwise accuracy by +0.012 (-0.022 to 0.052). Adding the divergence-derived token features to the same generated-conformer contact-and-BSA baseline likewise gave no resolved improvement. The changes were -0.001 (-0.042 to 0.041) in Spearman correlation and +0.003 (-0.017 to 0.023) in pairwise accuracy. A simple input-pose contact-and-BSA model also performed at least as well as the ensemble variants.

Thus, these controlled comparisons do not show an incremental affinity-ranking signal from either ensemble generation or the token features. We will reconsider the paper's central framing and focus it on the structural results that are directly supported, rather than treating ligand ranking as a primary contribution. A stronger affinity claim would require a clear improvement over these simpler baselines on held-out congeneric series.

Response to Concern 5 and Question 4 — Dependence on the starting pose

The method begins from a known experimental binding context. Please quantify degradation under realistic pose error and larger induced-fit changes.

We thank the reviewer for raising the method's dependence on the starting pose. We agree that this is an important limitation of a local-refinement method. To quantify the dependence, we evaluated all 135 Q-BioLiP benchmark tasks after perturbing the input ligand to scaffold RMSDs of 1, 2, or 3 Å. We first perturbed internal ligand torsions and then applied a rigid-body rotation and translation to reach the target RMSD. Coarse anchors were reconstructed from each perturbed pose, and no unperturbed ligand coordinates were supplied to the model. We generated 32 conformers per task and condition.

The table reports group-balanced paired changes from the unperturbed condition. Positive RMSD changes and negative IFP changes indicate degradation. Scaffold metrics include all 135 tasks from 48 congeneric groups, while fragment metrics apply to the 113 fragment-addition tasks.

Input-pose perturbationScaffold tasks / groupsScaffold RMSD change (95% CI)Scaffold IFP change (95% CI)Fragment RMSD change (95% CI)Fragment IFP change (95% CI)
1 Å135 / 48+0.13 (0.09 to 0.18)-0.008 (-0.016 to 0.000)+0.13 (-0.01 to 0.27)-0.009 (-0.035 to 0.017)
2 Å135 / 48+0.51 (0.40 to 0.62)-0.070 (-0.089 to -0.051)+0.31 (0.07 to 0.56)-0.034 (-0.068 to -0.002)
3 Å135 / 48+1.20 (1.08 to 1.32)-0.145 (-0.163 to -0.127)+0.84 (0.47 to 1.22)-0.095 (-0.151 to -0.037)

At 1 Å, interaction recovery is essentially unchanged and the RMSD increase is small. Degradation becomes measurable at 2 Å and more pronounced at 3 Å. At 3 Å, generated scaffold RMSD increases by 1.20 Å, from 1.22 to 2.42 Å, while scaffold IFP decreases from 0.84 to 0.70. These results indicate tolerance to small starting-pose errors but increasing dependence beyond approximately 1 Å. This experiment tests ligand-pose error rather than large receptor rearrangements, which remain outside the present local-refinement setting. We will revise the manuscript to include this perturbation analysis and clarify the method's practical range of applicability.

Thank you for the thoughtful rebuttal. I appreciated the new starting-pose experiments and the comparisons with simpler activity-ranking baselines, both of which clarified where the method is useful.

My remaining concern is whether modeling side-chain flexibility leads to better predictions when induced fit matters. Given how central this question is to the paper, I lean toward maintaining my original score. I would raise my score if the authors could provide more direct evidence of that benefit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.