acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Rigid Pocket: Protein Flexibility-informed Ligand Generation for Structure-Based Drug Design

Abstract

Structure-based ligand generation commonly assumes a fixed target protein structure, overlooking conformational heterogeneity and the structural adjustments that accompany ligand binding. Modeling protein flexibility remains challenging: matched apo-holo structure pairs are scarce, and pocket-focused methods often omit structural context beyond the binding site. Here, we present FlexiDD (Flexibility-Informed Drug Design), a generative framework for de novo drug design that explicitly models full-protein conformational adaptation at atomic resolution during ligand generation. To address the scarcity of paired apo–holo data, FlexiDD introduces a transferable coordinate denoising strategy that uses coordinate perturbations to bring apo and holo conformations into overlapping noisy distributions. This enables training solely on holo complexes, while supporting both apo and holo inputs at inference. By modeling the full protein rather than isolated binding pockets, FlexiDD incorporates global structural context and constraints into local binding-site adaptation. Experiments on a leakage-audited benchmark demonstrate strong ligand generation performance, with favorable predicted binding affinities and higher PoseBusters validity rates across both holo and apo inputs. Under apo conditioning, FlexiDD achieves 91.6% PoseBusters validity, compared with less than 46.1% for the evaluated baselines that also account for target flexibility. These results demonstrate the potential of holo-only training and full-protein conformational modeling for ligand generation from apo structures, supporting FlexiDD as a practical approach to flexibility-informed structure-based drug design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
structure-based drug design, protein flexibility, de novo molecule generation, generative model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_z1HCOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FlexIDD to address problems from rigid-pocket assumptions and scarce paired apo–holo training data in structure-based drug design, which jointly generates ligands through transferable coordinate denoising strategies.

Critical strengths
  1. The research problem is interesting and important in drug design.

  2. The theoritical analysis in this paper in well-established.

Critical weaknesses
  1. Limited validation of practical efficacy: evaluation relies entirely on computational experiments, without binding assays or other wet experiments to establish actual biological activity.
Questions for authors
  1. How to guarantee the conformation consistency between the apo and holo through the denoising strategy in the paper?
1 comment
AuthorsOpenReview ↗

Response to Reviewer z1HC

We sincerely thank the reviewer for the thoughtful feedback and helpful question. We appreciate your recognition of the importance of the research problem in drug design and your positive assessment of our theoretical analysis. We have carefully considered your concerns and respond to both points below.

W1: Limited validation of practical efficacy: evaluation relies entirely on computational experiments, without binding assays or other wet experiments to establish actual biological activity.

Response: We acknowledge that wet-lab validation is valuable for establishing practical efficacy. Our submission is positioned as an AI methodology contribution, consistent with ICLR’s focus on advances in machine learning and their applications. Specifically, our work focuses on developing a new framework for flexible SBDD and evaluating it through comprehensive computational experiments. We use established computational metrics to assess complementary aspects of ligand generation, including binding affinity estimation (Vina Score/Min), physical plausibility (PB-validity), etc., which are standard metrics adopted across the SBDD literature. This computational evaluation approach is also consistent with prior SBDD studies published at machine learning venues (e.g. TargetDiff[1], ICLR 2023; DecompDiff[2], ICML 2023; IPDiff[3], ICLR 2024; DrugFlow[4], ICLR 2025). Wet-lab validation is beyond the scope of the present methodological study, but remains an important next step toward establishing its practical biological utility. We have clarified this in the manuscript (Apx. C) and identified experimental evaluation of selected generated compounds as future work.

Q1: How to guarantee the conformation consistency between the apo and holo through the denoising strategy in the paper?

Response: Apo and holo structures of the same protein can exhibit conformation differences (Apx. A.10.3). For an apo–holo pair, adding Gaussian noise at the same scale to both structures reduces the KL divergence between their corrupted coordinate distributions as the noise scale increases (Apx. A.1). Fig. 20 illustrates this effect for one pair: the overlap between the projected pocket-atom densities increases from 81% at σ=1\sigma=1 Å to 96% at the initial inference scale of σ=8\sigma=8 Å (Apx. B.9). The model trained on holo complexes thus learns to denoise from this shared noisy space into holo-like complexes, and is able to generalize to apo inputs at inference. Eq. 19 further provides a theoretical bound on the discrepancy between the resulting output distributions under the stated assumptions.

During sampling process, our method does not explicitly enforce identical outputs from paired apo and holo inputs. In both cases, denoising is guided by the same fixed protein sequence. Our goal is to generate plausible protein–ligand complexes from either apo or holo input conformation.

Reference

[1] "3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction", ICLR 2023

[2] "DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design", ICML 2023

[3] "Protein-ligand interaction prior for binding-aware 3d molecule diffusion models", ICLR 2024

[4] "Multi-domain distribution learning for de novo drug design." ICLR 2025

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

This paper focuses on Structure-based ligand generation and argues that existing works suffer from the scarcity of paired apo–holo data. To this end, the authors introduce a transferable coordinate denoising strategy that uses coordinate perturbations to bring apo and holo conformations into overlapping noisy distributions.

Critical strengths
  1. Jointly models full-protein conformational adaptation and ligand generation at all-atom resolution, and using full-protein modeling provides global structural context and geometric constraints for consistent local conformational adaptation, seems to be new to me.

  2. Extensive comparison with existing works.

Critical weaknesses
  1. PROTEIN-LIGAND NETWORK is a simple adaptation of existing work.
  2. Too many hyperparameters.
  3. The SAMPLING process is unclear.
  4. The advantage of the proposed method over existing work is not significant.
Questions for authors
  1. PROTEIN-LIGAND NETWORK is a simple adaptation of existing work.
  2. Too many hyperparameters.
  3. The SAMPLING process is unclear.
  4. The advantage of the proposed method over existing work is not significant.
2 comments
AuthorsOpenReview ↗

Response to Reviewer eMxP (1/2)

We sincerely thank the reviewer for the constructive comments and valuable suggestions. We appreciate your recognition of the novelty of jointly modeling full-protein conformational adaptation and ligand generation at atomic resolution, as well as our extensive comparisons with existing methods. We have carefully considered your concerns, and respond to each point below.

W1: PROTEIN-LIGAND NETWORK is a simple adaptation of existing work.

Response: We would like to clarify that FlexiDD’s primary methodological contribution lies at the framework level: enabling full-protein all-atom conformational adaptation during de novo ligand generation without requiring paired apo–holo training data. The PROTEIN-LIGAND NETWORK serves as the network architecture that implements this framework.

To the best of our knowledge, FlexiDD is the first generative framework to jointly model de novo ligand generation and full-protein conformational adaptation at all-atom resolution. Its design addresses two key challenges in flexible SBDD. First, to address the scarcity of paired apo–holo structures, we introduce a transferable protein denoising strategy that reduces the apo–holo distributional discrepancy in a shared noisy coordinate space, enabling training on holo complexes and generalization to both apo and holo inputs. Second, to address the restricted scope of protein flexibility modeling, FlexiDD models conformational adaptation across the full protein, incorporating global structural context and geometric constraints beyond the binding site.

The PROTEIN-LIGAND NETWORK integrates full-protein coordinate denoising, ligand geometry and chemical-state modeling, and protein–ligand information exchange to jointly update protein conformations and ligand structures during generation. Its role is to support these capabilities within the proposed framework, whose contribution centers on the combined modeling and training strategy for addressing the two challenges.

W2: Too many hyperparameters.

Response: The reported hyperparameters span network architecture, training objectives, and diffusion sampling. Many of these settings follow established practices in diffusion-based modeling, including EDM-style noise schedules and network configurations adopted from prior frameworks[1]. Related protein modeling methods, such as RFdiffusion[2] and Boltz-1[3], similarly involve configurable diffusion, architecture, and training settings.

For reproducibility, we provide the complete architecture configurations, loss weights, and training settings in Tab. 4, 5, and 8, together with the noise schedule and sampling details in Apx. A.5.

Reference

[1] Karras, Tero, et al. "Elucidating the design space of diffusion-based generative models." Advances in neural information processing systems 35 (2022): 26565-26577.

[2] Watson, Joseph L., et al. "De novo design of protein structure and function with RFdiffusion." Nature 620.7976 (2023): 1089-1100.

[3] Wohlwend, Jeremy, et al. "Boltz-1 democratizing biomolecular interaction modeling." BioRxiv (2025): 2024-11.

AuthorsOpenReview ↗

Response to Reviewer eMxP (2/2)

W3: The SAMPLING process is unclear.

Response: We outline the three stages of sampling process below; the complete algorithm is provided in Sec. 3.2.5 and Apx. A.5.

  1. Initialization: We perturb the input apo or holo protein coordinates with Gaussian noise at the initial noise level. Ligand coordinates are sampled from a pocket-centered Gaussian prior, while ligand atom types, charges, and bond types are initialized from uniform categorical priors.
  2. Joint iterative updates: Over 100 sampling steps, the network predicts protein and ligand states to guide their respective updates. Protein coordinates are updated using denoiser-based integration, ligand coordinates using endpoint-flow integration, and ligand atom types, charges, and bond types using categorical transitions.
  3. Final decoding: We decode the ligand chemical structure subject to valence-related constraints and return the resulting protein–ligand complex.

W4: The advantage of the proposed method over existing work is not significant.

Response: We would like to highlight FlexiDD’s methodological advances and substantial empirical improvements, particularly under apo conditioning.

Methodological advantages. Flexible SBDD is a significant and challenging problem, because binding pockets can adopt multiple conformations and undergo further adjustments upon ligand binding, particularly complicating design when only an apo structure is available. Our FlexiDD addresses two key limitations of flexible SBDD: the scarcity of paired apo–holo training data and the restricted scope of protein conformational adaptation. Our transferable protein perturbation strategy enables training on large-scale holo complexes without paired apo–holo supervision and generation from both apo and holo inputs. Full-protein all-atom modeling further incorporates structural context beyond the binding pocket during ligand generation.

Empirical Improvements. These advantages are supported by experimental results. In particular, under apo conditioning, FlexiDD achieves 91.6% PB-validity, compared with 46.1% for the best existing flexible baseline, while also achieving the best Vina Min and Holo Min scores (Tab. 3). Even in the holo setting, where rigid methods benefit from access to experimental structures, FlexiDD remains competitive or superior on structure-sensitive metrics.

Reviewer_DZudOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Existing flexible SBDD methods either (1) rely on paired apo-holo data or (2) model only limited protein flexibility. To address (1), the authors assume that noisy holo and apo structures have overlapping distributions and propose holo-only denoising to eliminate the need for paired data. To address (2), they adopt full-protein all-atom modeling. The method achieves promising empirical improvements, but its key claims lack sufficient validation.

Critical strengths
  • Important and practically relevant problem. The paper focuses on two important limitations of existing flexible SBDD methods: the scarcity of paired apo-holo data and the restricted modeling of protein flexibility.
  • Promising empirical improvements. FLEXIDD achieves competitive or improved performance across several ligand generation metrics. In particular, the improvement in Holo Min suggests that generated ligands may exhibit better compatibility with experimental holo structures.
Critical weaknesses
  • Insufficient validation of the proposed protein conformation adaptation mechanism.
    • A Gaussian noise scale of 8 Å is comparable to the size of a binding pocket and may severely disrupt structural continuity.
    • The noisy structure in Figure 2(a) does not appear consistent with independent Gaussian noise added to each atomic coordinate.
    • The paper lacks structural comparisons between noisy apo and noisy holo conformations at different noise scales, preferably with visualizations.
    • The paper lacks noise-scale ablations and a pure-noise initialization ablation to rule out improvements primarily driven by additional pretraining.
  • Lack of ablation studies to validate the benefits of full-protein all-atom modeling.
  • Limited apo-holo conformational differences in the benchmark. The paper lacks evaluations stratified by apo-holo conformational differences, making it difficult to assess whether the model can handle large conformational changes.
  • Lack of pocket RMSD evaluation, making it unclear whether the model can recover the correct holo conformation.
  • Potential evaluation bias in the apo setting. Flexible methods use generated protein structures, which may introduce bias in the PB-valid metric.
Questions for authors

see weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer DZud (1/3)

We sincerely thank the reviewer for the thoughtful assessment and constructive feedback. We appreciate your recognition of the importance and practical relevance of addressing paired apo–holo data scarcity and restricted protein flexibility modeling, as well as your positive assessment of FlexiDD’s promising results. We have carefully considered your concerns and provide our point-by-point responses below.

W1: A Gaussian noise scale of 8 Å is comparable to the size of a binding pocket and may severely disrupt structural continuity.

Response: We would like to clarify that 8 Å is the maximum standard deviation (σmax⁡\sigma_{\max}) in our noise schedule, not a fixed noise level applied throughout generation. The schedule decreases from σmax⁡=8\sigma_{\max}=8 Å at t=0t{=}0 to σmin⁡=0.05\sigma_{\min}=0.05 Å at t=1t{=}1 (Eq. 3; Apx. A.5), so the vast majority of denoising steps operate at smaller noise scales. This large initial perturbation is an intentional design choice that bridges the distributional gap between apo and holo conformations (Fig. 1), enabling training on abundant holo complexes without requiring paired apo–holo data. Formally, for an aligned apo–holo pair the KL divergence between their corrupted coordinate distributions decreases as σ−2\sigma^{-2} (Eq. 18, Apx. A.1.2), and Fig. 7 illustrates this scaling. We further visualize corrupted pocket-atom densities at different noise scales (Apx. B.9, Fig. 20). For the illustrated apo–holo pair, the projected atom densities exhibit substantial overlap (96%) at σ=8\sigma=8 Å.

Meanwhile, the corrupted coordinates are intermediate representations whose structural geometry need not be physically valid. Protein sequence, atom identities, and covalent connectivity remain available as fixed conditioning information to guide structural reconstruction. Structural continuity should therefore be evaluated on the final generated proteins, not the noisy inputs. Indeed, our generated structures maintain high global similarity to the reference holo conformations, with mean TM-scores of 0.960, 0.958, and 0.942 across three similarity subsets (Sec. 4.4, Fig. 4a).

We choose σmax⁡=8\sigma_{\max}=8 Å as a practical trade-off to balance reducing the apo–holo distributional gap against preserving protein structural information needed for reconstruction. Excessively large noise can obscure the input protein’s geometry and make reconstruction more difficult, whereas insufficient noise may leave a substantial apo–holo discrepancy. We have provided a noise-scale ablation in Apx. B.3.2, and the results are consistent with this reasoning.

W2: The noisy structure in Figure 2(a) does not appear consistent with independent Gaussian noise added to each atomic coordinate.

Response: Fig. 2a uses PyMOL’s cartoon representation[1], which depicts the protein backbone as an interpolated ribbon through sequential Cα positions, rather than directly displaying individual atomic coordinates. This representation can obscure atom-level perturbations; the apparent continuity of the ribbon therefore does not imply that local atomic geometry is preserved. The Gaussian perturbations in our method are sampled independently for every protein atom. We have clarified this visualization convention in the caption of Fig. 2a.

W3: The paper lacks structural comparisons between noisy apo and noisy holo conformations at different noise scales, preferably with visualizations.

Response: We have added the suggested comparison in Apx. B.9, Fig. 20, which visualizes apo–holo pocket structures at noise scales σ\sigma = 1, 2, 4, and 8 Å. For each scale, we show both noisy pocket atoms and projected atom-density contours, with the shared density region highlighted. The measured density overlap increases monotonically from 81.0% (σ\sigma = 1 Å) to 96.0% (σ\sigma = 8 Å), providing visual and quantitative evidence that larger noise scales effectively bridge the apo–holo distributional gap. This trend is consistent with the KL analysis in Apx. A.1.

Reference

[1] DeLano, Warren L. "Pymol: An open-source molecular graphics tool." CCP4 Newsl. protein crystallogr 40.1 (2002): 82-92.

AuthorsOpenReview ↗

Response to Reviewer DZud (2/3)

W4: The paper lacks noise-scale ablations and a pure-noise initialization ablation to rule out improvements primarily driven by additional pretraining.

Response: We provide two complementary ablation studies to examine whether FlexiDD’s performance gains are primarily driven by Stage 1 pretraining and to assess the contribution of noise-based bridging.

Pretraining Ablation. Without Stage 1 pretraining (with σmax⁡=8\sigma_{\max}=8 Å and all other settings unchanged), FlexiDD’s apo-input PB-validity reaches 0.878, still substantially surpassing all existing flexible baselines (e.g., FlexFlow: 0.461), and is comparable to holo inputs (Apx. B.3.1). The full results are shown in Tab. R2 in our response to W5. This demonstrates that FlexiDD’s performance advantage is not contingent on Stage 1 pretraining alone.

Maximum Noise Scale Ablation. With Stage 1 pretraining included, we evaluate σmax⁡∈{4,8,10}\sigma_{\max}\in\{4,8,10\} Å and a pure-noise baseline (σmax⁡=160\sigma_{\max}=160, following the Boltz-1[2] schedule); results are shown in Tab. R1 below.

Tab. R1: Ablation Study on Max Noise Scale

σmaxσ_{max}InputValidity↑PB-valid.↑Strain↓Vina↓Vina Min↓
4Apo0.9740.862108.6−4.68−5.92
4Holo0.9800.91875.6−5.26−6.59
8Apo0.9760.91659.4−5.02−6.42
8Holo0.9790.91676.8−5.39−6.67
10Apo0.9770.91474.3−5.08−6.44
10Holo0.9780.91579.1−5.20−6.53
160Pure Noise0.9670.780128.7−4.35−5.71

Increasing σmax⁡\sigma_{\max} from 4 to 8 Å consistently improves apo-input metrics (PB-validity: 0.862→0.916, Vina Min: −5.92→−6.42) and narrows the apo–holo Vina Min gap from 0.67 to 0.25, providing empirical support for noise-based apo–holo bridging consistent with the KL analysis in Apx. A.1. Further increasing to 10 Å does not consistently improve generation quality, suggesting 8 Å is an effective trade-off. The pure-noise setting (σmax⁡=160\sigma_{\max}=160) changes the protein component from adapting an existing conformation to reconstructing its 3D structure, which is a folding-like task. Compared with apo-conditioned initialization at σmax⁡=8\sigma_{\max}=8 Å, this setting reduces PB-validity from 91.6% to 78.0%. These results support the value of retaining input conformational information during initialization under the evaluated protocol. We have added these ablation studies in Apx. B.3.2.

Together, these ablations indicate that FlexiDD’s performance gains are not solely attributable to additional pretraining and provide evidence that both the noise scale and structure-conditioned initialization contribute to generation quality.

W5: Lack of ablation studies to validate the benefits of full-protein all-atom modeling.

Response: We have added an ablation study comparing full-protein and pocket-only modeling in Apx. B.3.1. To isolate the effect of full-protein modeling, we compare variants with and without full-protein input while holding other model components and evaluation protocol fixed. Since Stage 1 pretraining uses only protein monomers (without ligands or binding-pocket annotations), we control for its effect by comparing rows without Stage 1 (rows 2–3 and 5–6 in Tab. R2).

Tab. R2: Ablation of Model Components

InputStage 1Full ProteinValidity↑PB-valid.↑Strain↓Vina↓VinaMin↓QED↑SA↑LogPLipinski↑Div.2D↑Uniq.2D↑
Apo✓✓0.9760.91659.4-5.02-6.420.5450.6651.1244.8360.8730.985
Apo✗✓0.9790.87861.2-4.88-6.210.5290.6521.2884.7780.8640.989
Apo✗✗0.8950.648767.1-2.84-5.620.5430.6471.0464.5730.8570.972
Holo✓✓0.9790.91676.8-5.39-6.670.5500.6581.1854.7610.8710.980
Holo✗✓0.9820.88760.5-5.11-6.540.5350.6471.1414.7960.8670.997
Holo✗✗0.8910.6641228.4-3.00-5.700.5430.6491.0974.5810.8600.975

Removing full-protein modeling leads to consistent degradation across both input conditions: PB-validity drops by over 20 percentage points (apo: 0.878→0.648, holo: 0.887→0.664), ligand strain increases by more than an order of magnitude, and Vina Min worsens substantially. These results demonstrate the effectiveness of full-protein modeling under both apo and holo inputs.

Reference

[2] Wohlwend, Jeremy, et al. "Boltz-1 democratizing biomolecular interaction modeling." BioRxiv (2025): 2024-11.

AuthorsOpenReview ↗

Response to Reviewer DZud (3/3)

W6: Limited apo-holo conformational differences in the benchmark. The paper lacks evaluations stratified by apo-holo conformational differences, making it difficult to assess whether the model can handle large conformational changes.

Response: We have added two complementary analyses to evaluate FlexiDD's ability to handle large conformational changes across different magnitudes.

Pocket heavy-atom RMSD evaluation. We compare FlexiDD, FlexFlow, and Apo2Mol by measuring the heavy-atom RMSD between each generated pocket and its experimental holo reference. Table R3 reports the results stratified by the experimental apo–holo pocket RMSD (Apx. B.6).

Tab. R3: Pocket Heavy-atom RMSD (mean ± SD). "Apo" is the RMSD between experimental apo and holo pockets.

Subset (Å)ApoFlexFlowApo2MolFlexiDD
All1.13±0.861.22±0.781.51±0.791.15±0.35
0–10.62±0.250.89±0.301.12±0.411.09±0.40
1–21.33±0.241.21±0.281.61±0.281.19±0.26
>22.95±1.412.89±1.403.07±1.331.26±0.39

For cases with larger apo–holo differences (>2 Å), FlexiDD substantially outperforms two baselines. In the 1–2 Å range, FlexiDD also achieves the lowest RMSD. For small initial differences (<1 Å), all generative methods increase RMSD relative to the already near-holo apo inputs. FlexiDD’s advantage in larger-gap cases is consistent with its broader modeling scope: it supports full-protein modeling, whereas FlexFlow restricts adaptation only to side-chains and Apo2Mol models only the local pocket.

Pocket volume. As a complementary measure of pocket geometry, we compare apo→generated and apo→holo changes in binding-pocket volume (Apx. B.5). The generated and experimental structure volume changes show strong correlation (Pearson r = 0.889, Spearman ρ = 0.878), indicating that FlexiDD-generated pockets capture the experimental apo–holo volume change trends.

Together, these analyses support FlexiDD’s ability to model pocket conformational adaptation, including cases with relatively large apo–holo differences.

W7: Lack of pocket RMSD evaluation, making it unclear whether the model can recover the correct holo conformation.

Response: We have added a pocket heavy-atom RMSD evaluation against experimental holo structures in Apx. B.6 (see also Tab. R3 in our response to W6). We compare FlexiDD with two flexible baselines, FlexFlow and Apo2Mol, on the test set stratified by experimental apo–holo pocket RMSD.

Overall, FlexiDD achieves a mean pocket RMSD of 1.15 Å, the lowest among all methods. Its advantage is pronounced for cases with large initial conformational differences (>2 Å). These results show that FlexiDD generates pockets closer to the experimental holo reference than baselines, especially when the experimental apo–holo differences are large.

We also clarify that FlexiDD jointly generates novel ligands and adapts protein conformations. Because generated ligands differ from the original bound ligand, they may favor slightly different protein conformations, so exact recovery of the reference holo geometry is not necessarily expected.

W8: Potential evaluation bias in the apo setting. Flexible methods use generated protein structures, which may introduce bias in the PB-valid metric.

Response: PB-validity includes protein–ligand clash checks and therefore depends on receptor conformation, making direct comparisons between rigid and flexible methods inherently challenging. We complement this evaluation with the following analyses.

First, we compare FlexiDD with other flexible SBDD methods, all evaluated using their respective generated receptors and under the same protocol (Apx. B.1). FlexiDD achieves 91.6% PB-validity, compared with 46.1% for the strongest flexible baseline, indicating that FlexiDD’s performance advantage is not attributable to evaluation on model-generated protein structures alone. Second, we evaluate generated ligands from all methods against the same experimental holo structure using Vina Min after local minimization (Tab. 3). This receptor-controlled setting avoids differences from the generated receptors: FlexiDD achieves −6.53 versus −5.15 for the strongest rigid baseline (FLOWR.root), supporting the compatibility of ligands generated by FlexiDD from apo inputs with experimental holo pockets after local minimization. Third, FlexiDD-generated pockets can support independently trained SBDD ligand generators (such as FLOWR.root and PocketXMol) without retraining, achieving PB-validities comparable to those obtained using experimental holo pockets (Sec. 4.5; Apx. B.7).

Together, these evaluations provide complementary evidence that FlexiDD’s advantage is not solely attributable to evaluation on its own generated structures. We have added a clarification of PB-validity’s dependence on method-specific pocket conformations in Apx. B.1.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.