acceptodds
Under review as a conference paper at ICLR 2027

Robust Biomolecular Complex Design Across Protein Conformational Landscapes

Abstract

Proteins populate conformational ensembles, yet structure-based biomolecular design typically optimizes candidates against a single target conformation. Consequently, a candidate that fits one state can lose favorable interactions or develop steric clashes when the target adopts another. We introduce FlexEvo, a model-agnostic evolutionary framework that adapts candidates once at inference time from a single target conformation to improve compatibility with alternative natural conformations unseen during adaptation, without retraining the source model or requiring a conformational ensemble. FlexEvo casts cross-state adaptation as geometry-constrained bi-objective optimization, balancing preservation of input-state interactions against robustness to plausible conformational perturbations. To limit the search space and reduce invalid structural edits, geometry-derived FlexBoxes define protected anchor regions, adaptable regions for local exploration, and forbidden regions for clash avoidance. A unified all-atom representation supports topology-preserving adaptation across diverse binder categories, while Pareto selection preserves nondominated candidates across the two objectives. We evaluate FlexEvo across multiple generation baselines and nine representative binder categories spanning diverse molecular sizes and structural topologies. FlexEvo reduces the category-balanced mean relative performance degradation from 47.8% to 4.4%, while adding only 1.4–3.1 minutes of adaptation per sample. These results establish single-state inference-time adaptation as a practical route toward robust biomolecular complex design across protein conformational landscapes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Biomolecular Complex Design, Protein Conformational Flexibility, Inference-Time Adaptation

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_x3jDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

To address the problem that structure-based bimolecular design typically optimizes candidates against a single static target conformation, this paper proposed FLEXEVO, which is an evolutionary optimization algorithm to adapts conformation generated by design models without fine-tuning base model or providing conformation ensembles. The model formalizes cross-conformation design as a geometric constrained bi-objective optimization problem and requires no training, covering small molecules, peptides, antibodies, mRNA, etc., reducing performance degradation significantly from 47.8% to 4.4% with low computational cost.

Critical strengths
  1. FLEXEVO used simple local geometric perturbation instead of expensive conformations generated by MD simulations or ensemble-predicting models as stress tests for unknown conformations, serving as a plug-and-play inference-time plugin that can be used in any models with low computational cost.
  2. The experiment design covered 9 drug molecules classes ranging from small molecules to large mRNAs, convincingly demonstrated the extrapolation capacity of the method.
Critical weaknesses
  1. The method is not positioned as a "de novo design" framework like RFDiffusion models. All experiments are conducted under graph preserving mode, which allows no atom addtions, deletions or changes of chemical bonds. So it only serves as a dihedral fine-tuning algorithm.
  2. The algorithm relies on local perturbations and lacks theoretical coverage guarantees, but achieves significant improvements under drastic changes even in propeptide-associated. The mechanistic explanation is lacking.
Questions for authors
  1. In experiment, baselines only used first candidate, while FLEXEVO selects from 64. Does the gain of performance comes from candidate ranking or geometric editing? Please compare FLEXEVO with re-rank-only and a standard local relaxation baseline under same computational budgets.
  2. How do authors explain the highly similar data across different modalities? For example, in tab.6 and tab.10, all 18 DockQ means for proteins and long chain mRNA matches within 1e-3 error, in fig.25, subfigure (a-c small molecules) are exactly the same as subfigure (d-f macrocycles).
Reviewer_78hgOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose an evolutionary search algorithm named FlexEvo for complex design by considering the multiple conformations of protein structures. This method leverages a spatial prior information, i.e. FlexBoxes, to identify stable, flexible and forbidden receptor region to guide the pose of candidate molecules. FlexEvo is validated on both of de novo ligand generation and bimolecular complex design, and show improvements in Vina and DockQ respectively.

Critical strengths
  • This paper addresses an important problem on the complex design.
  • The method is well-motivated to solve the problem.
  • The experimental results seem sufficient to show its complex design capability.
Critical weaknesses
  • The method introduces lots of hyper-parameters but lacks of detailed analysis on how these hyper-parameters affect the final performance.
Questions for authors
  • Please explain how these hyper-parameters are chosen.
  • Please demonstrate how a complex evolves during optimization process.
  • Please explain how these reference, matched, and divergent protein are selected.
Reviewer_USHnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces FlexEvo, an inference-time framework for adapting existing binder candidates using a single target conformation. An outer evolutionary search optimizes residue-level spatial constraints. An inner procedure refines candidate coordinates through topology-preserving operations. Selection balances interaction preservation and robustness to synthetic target perturbations. The evaluation covers nine binder categories and multiple source models, with candidates frozen before evaluation on alternative natural target conformations. The authors report a reduction in category-balanced relative performance degradation from 47.8% to 4.4%, alongside reranking controls, component ablations, and runtime analyses. The evidence is computational, based mainly on docking and structural-agreement metrics.

Critical strengths
  1. The paper addresses a practically relevant single-state adaptation problem and includes a rerank-only control to assess gains beyond source pool selection.
  2. The theoretical analysis clearly acknowledges that robustness to synthetic probes does not establish robustness to unseen natural conformations.
Critical weaknesses
  1. The evaluation does not establish robustness of binding across receptor conformations. Appendix B.1 keeps the binder's coordinates fixed when changing the receptor structure. However, the same molecule may bind different receptor conformations by adjusting its pose. A worse score under this frozen-coordinate test therefore does not necessarily indicate weaker binding. The paper does not establish whether improving these scores identifies molecules that actually bind more robustly across conformations.
  2. The robustness mechanism is insufficiently justified. The synthetic receptor perturbations used during optimization depend on each genome's residue annotations. Because a genome partly determines the perturbations used to score it, higher fitness may reflect easier perturbations rather than greater robustness. The internal-geometry term also becomes uninformative for candidates with ≥46\ge 46 atoms, and the ablations do not isolate the effect of perturbation-based optimization.
  3. The experimental protocol does not consistently support the claimed generalization across natural conformations. The manuscript describes Divergent cases as experimentally observed alternative conformations, whereas the supplementary materials define them as synthetic perturbations derived from Matched structures. It is therefore unclear which protocol produced the headline results and whether those results demonstrate transfer to unseen natural conformations or only robustness to constructed stress tests.
  4. The central results cannot be reproduced from the supplied materials. Core adaptation and evaluation implementations, a complete benchmark manifest, and records connecting the released examples to the main tables are unavailable. The archive provides only Reference measurements.
Questions for authors
  1. How is perturbation difficulty kept comparable across genomes when label assignments dictate perturbed residues and amplitudes? Do gains persist under a fixed, genome-independent probe set?
  2. Eq. 20 yields g=0g=0 for any candidate with ≥46\ge 46 atoms. Is this the exact formula used in the codebase, and what geometry checks (with pass rates) were enforced in reported runs?
  3. What specific evidence proves perturbation-based optimization drives gains? Can you provide matched-budget ablations isolating ordinary structure refinement, removal of frobf_{\mathrm{rob}}, and removal of all perturbation terms?
  4. Which tasks performed genuine binder design versus structure prediction? Can you provide candidate ancestry, sequences, and atom mappings for the supplied examples to delineate candidate selection from coordinate editing?
  5. Was Table 1 generated using one natural structure per state or three Matched structures with synthetic Divergent derivatives? Please reconcile the text with configs/protocol.json, dataset.py, and docs/protocol.md.
  6. Alongside relative degradation (Eq. 30), what are the paired absolute Reference–Divergent score gaps and their uncertainties? How much reduction stems from smaller gaps versus shifts in Reference baseline scores?
  7. Can you provide the full benchmark manifest, actual target counts, eligible paired counts, failure rates, and bootstrap confidence intervals?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.