acceptodds
Under review as a conference paper at ICLR 2027

TrustCDR: Reliability-Guided CDR Co-Design under Local Antigen Uncertainty

Abstract

Target-specific antibody CDR sequence–structure design uses antigen geometry to guide binding-interface generation. Recent antigen-aware generators encode this geometry but generally do not estimate its local reliability and may therefore assign comparable trust to geometrically heterogeneous cues. Antigen structures used for design may be experimentally reconstructed, predicted, docked, or preprocessed and may also contain conformationally variable regions. Such structures can be globally plausible while their local geometric cues differ in reliability. Consequently, lower-reliability cues may exert excessive influence on CDR–antigen relation updates and degrade interface quality. We introduce TrustCDR, which estimates a reference-free, perturbation-based relative geometric reliability signal and converts it into relative trust weights that modulate residual CDR–antigen pair updates. Starting from a single observed antigen, TrustCDR constructs a reference-free perturbation ensemble and evaluates residue-level stability across coordinate, topology, orientation, and surface cues. It preserves the base antigen condition and uses down-only raw weights with mean-preserving normalization to reduce the relative influence of lower-reliability regions. On the RAbD test set, TrustCDR achieves 68.52% AAR, 0.57 RMSD, and 0.76 DockQ under clean inputs, while maintaining 0.75 DockQ under a 1.5 epitope perturbation. It also achieves IMP rates of 18.50% and 20.40% under clean and epitope-perturbed inputs, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody design, CDR sequence–structure co-design, Antigen uncertainty, Reliability-guided generation, Flow matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_m3HmOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
3
Presentation
3
Soundness
2(−1)
Read review
Paper summary

This paper proposes TrustCDR, a reliability-guided method for joint CDR sequence–structure generation under local antigen structural uncertainty.

I think this paper introduces a meaningful research perspective: existing antigen-conditioned generative models typically use antigen geometric information directly, but the local geometric stability of different regions may vary. Therefore, these geometric cues should not necessarily be assigned the same level of trust.

Based on this idea, the authors propose a reference-free method for estimating local geometric reliability. By perturbing the antigen structure, the method estimates residue-level geometric stability from four aspects: coordinates, topology, orientation, and surface features. These estimates are then converted into relative trust weights and incorporated into the Flow Matching-based CDR generation process.

Experimentally, TrustCDR achieves strong performance, outperforming all compared methods on the main RAbD benchmark. The ablation studies also support the contribution of the complete reliability-guided framework to generation performance. In addition, the antigen-condition diagnostics are particularly interesting, as they provide evidence that TrustCDR's outputs are more sensitive to the supplied antigen conditions than those of the other baselines.

Overall, I think this is a valuable work with a clearly defined research problem and convincing experimental results.

However, I believe the paper has not yet sufficiently demonstrated that the proposed definition of Reliability has a clear, task-relevant structural meaning, rather than merely serving as an effective geometric heuristic. If the authors could further establish the mechanistic connection between reliability estimation, local geometric properties, and the generation process, it would significantly strengthen the scientific value of this work and make the proposed perspective more potentially generalizable to other structure-conditioned generation tasks.

Critical strengths
  1. A clearly defined research problem with a meaningful and potentially generalizable perspective.

The authors focus on the local geometric reliability of different antigen regions and adjust the influence of geometric information during CDR generation accordingly. I think this is a worthwhile research problem. Compared with directly using all geometric conditions, the idea of treating information differently according to its estimated reliability offers a certain degree of novelty and may provide a useful perspective for other generative tasks involving uncertain structural inputs. Although its generalizability has not yet been experimentally established, I believe this idea has promising potential.

  1. The antigen-condition diagnostics provide an interesting experimental design.

I particularly appreciate the authors' use of incorrect antigen inputs and extreme translations to examine the model's dependence on antigen conditions. The experiments show that some other generative models produce nearly unchanged sequence and structure predictions even when the antigen condition changes significantly, whereas TrustCDR is more sensitive to these changes. I believe this provides valuable evidence that the model's outputs are sensitive to and dependent on the supplied antigen conditions. It also introduces an interesting diagnostic approach that could be useful for evaluating other conditional generative models.

Critical weaknesses
  1. The mathematical definition of Coordinate Stability may lack a direct connection to structural reliability. In Section 3.2, the authors define Coordinate Stability as the average distance between the original residue coordinates and their perturbed counterparts:$ q_j^{\mathrm{coord}}=\frac{1}{K}\sum_{k=1}^{K}\left|x_j^{(k)}-x_j^{(0)}\right|_2 $

However, under the additive coordinate perturbation setting adopted by the authors: $ x_j^{(k)}=x_j^{(0)}+\epsilon_j^{(k)} $

This expression can be simplified to: $ q_j^{\mathrm{coord}}=\frac{1}{K}\sum_{k=1}^{K}\left|\epsilon_j^{(k)}\right|_2 $

This means that the original coordinates cancel out mathematically, and the resulting metric may primarily reflect the magnitude of the artificially injected perturbations rather than the actual local structural quality.

If the perturbation process does not depend on the original geometry, it remains unclear how much meaningful structural reliability information this descriptor can provide.

  1. The mechanistic interpretation of the proposed Reliability definition remains insufficient.

The authors propose estimating Reliability from the stability of local geometry under artificial perturbations and incorporating it into the generation process. Although the experiments demonstrate the effectiveness of the complete framework, they do not yet sufficiently establish whether the estimated relative geometric stability provides meaningful, task-relevant reliability information for CDR design. In particular, the existing ablations do not fully distinguish whether the improvements arise from meaningful reliability-guided trust allocation or from the additional geometric descriptors and conditioning pathways.

I believe the authors should further clarify whether the proposed Reliability provides a meaningful basis for residue-level trust allocation, beyond serving as an effective geometric heuristic. More convincing mechanistic evidence would significantly strengthen the scientific value and potential generalizability of this work.

Questions for authors
  1. Performance changes under distant perturbations in Tables 1 and 9.

Under the 3.0 Å Distant Perturbation condition in Table 1, TrustCDR's AAR decreases from 68.5% to 64.1%, whereas dyMEAN's performance remains almost unchanged.

Furthermore, Table 9 shows that this decline is particularly pronounced for certain CDRs. For example, H3 AAR decreases from 37.23% to 27.15%, while L3 AAR decreases from 65.85% to 57.18%.

I would like the authors to explain why antigen perturbations far from the binding interface can noticeably affect sequence recovery in these CDRs. In particular, could the Reliability Estimation, relative weight normalization, or antigen feature propagation introduce non-local effects that indirectly influence CDR generation?

  1. Could the authors provide further experiments to validate the Reliability mechanism?

I think the existing ablation studies have already demonstrated the effectiveness of the complete framework. However, more targeted experiments could help establish a clearer understanding of Reliability itself.

For example:

  • Reliability Consistency: Repeatedly sample different perturbation ensembles for the same antigen to examine whether the residue-level reliability estimates remain stable.
  • Reliability Shuffling: Randomly shuffle the reliability weights assigned to different antigen residues while preserving the remaining geometric features and model architecture as much as possible, and examine whether the model's performance is significantly affected.

These experiments could help determine whether the performance improvements truly arise from meaningful residue-level reliability information, rather than simply introducing an additional geometric conditioning pathway.

  1. Interpretation of the Coordinate-only Reliability ablation results.

In Table 2, Coordinate-only Reliability still leads to performance improvements. Compared with No Reliability, clean AAR increases from 64.3% to 65.3%, while RMSD decreases from 0.75 Å to 0.70 Å.

However, according to the mathematical definition of Coordinate Stability, the original coordinates cancel out under additive perturbations, suggesting that this descriptor may primarily capture the statistical properties of the artificially injected perturbations.

Therefore, I would like the authors to explain: What meaningful structural information does Coordinate Stability actually capture, and why does it still lead to performance improvements?

Could these improvements arise from perturbation statistics or the additional conditioning pathway rather than from local structural reliability itself?

4 comments

Response to Reviewer m3Hm - Part 1

We thank you for your thoughtful comments. Because of the response-length limit, this response is split into two parts; this is Part 1. Changes are highlighted in red in the main paper; further details are in the supplement.

Critical Weakness 1: Coordinate Stability

“The mathematical definition of Coordinate Stability may lack a direct connection to structural reliability.”

Response: The coordinate channel records the realized displacement magnitude, not a local-accuracy label. Removing it from both pathways changes clean AAR/RMSD from 68.52%/0.57 Å to 67.32%/0.78 Å, supporting a contribution to generation; the channel alone does not establish structural accuracy.

xjk=xj0+εjk,qjcoord=(1/K)Σk‖εjk‖2.x_j^k = x_j^0 + ε_j^k, q_j^coord = (1/K) Σ_k ‖ε_j^k‖_2.

Together with the other channels, it informs corrections, gates, and relative weights; it is not divided by displacement magnitude or used alone. Q3 separately tests Coordinate-only.

Critical Weakness 2: Mechanistic interpretation of Reliability

“The mechanistic interpretation of the proposed Reliability definition remains insufficient.”

“In particular, the existing ablations do not fully distinguish whether the improvements arise from meaningful reliability-guided trust allocation or from the additional geometric descriptors and conditioning pathways.”

Response: We added pathway ablations to separate descriptor conditioning from scalar trust weighting. Descriptor-only improves AAR, Scalar-only improves AAR and RMSD, and Full has the highest AAR with lower or equal RMSD than either single pathway; this supports contributions from both pathways. Descriptor-only fixes weights to one, Scalar-only removes descriptor conditioning and its gate, and No Reliability removes both:

Antigen conditionVariantAAR (%) ↑CDR RMSD (Å) ↓DockQ ↑
CleanFull68.520.570.76
CleanDescriptor-only67.280.750.75
CleanScalar-only66.000.730.75
CleanNo Reliability64.300.750.76
CleanNo-coordinate67.320.780.75

Results for epitope and distant 1.5 Å inputs are in the manuscript appendix, “Pathway and Residue-Assignment Controls.”

These trends hold in all three conditions, supporting both uses rather than a descriptor-only explanation.

Question 1: Distant perturbations

“Performance changes under distant perturbations in Tables 1 and 9.”

“In particular, could the Reliability Estimation, relative weight normalization, or antigen feature propagation introduce non-local effects that indirectly influence CDR generation?”

Response: With the reliability field and participating set fixed, we tested distant geometry and WrongAg. Geometry still changes generation through pair features and gates: overall AAR changes little at 1.5 Å but falls to 64.1% at 3.0 Å, while WrongAg degrades TrustCDR and leaves dyMEAN nearly unchanged. Thus the model uses antigen geometry, but 3.0 Å remains a limitation.

For each valid CDR–antigen pair:

z′=z+αgwΔz.z′ = z + αgwΔz.

Antigen distances and orientations enter z, Δz, and g, so geometry can change predictions. The field remains at its clean value, so re-estimation is not the source. With the field and participating set fixed, the weights remain fixed. Normalization alone cannot make generation invariant to distant geometry.

The original results distinguish moderate perturbation from the stronger stress test: overall AAR is 68.5% for both clean and distant 1.5 Å inputs at the main table's precision, and 64.1% for distant 3.0 Å. The H3 and L3 results you highlighted are:

CDRClean AAR (%)Distant 3.0 Å AAR (%)
H337.2327.15
L365.8557.18

These unequal responses are consistent with reported loop-dependent conformational changes and L3 recognition roles (Liu et al., 2024; Persson et al., 2013), but those studies provide context rather than a causal explanation.

The WrongAg diagnostic is complementary: clean-to-wrong-antigen AAR/RMSD changes from 68.52%/0.57 Å to 44.50%/2.09 Å for TrustCDR and 65.47%/0.68 Å to 66.70%/0.66 Å for dyMEAN. Unchanged recovery alone therefore does not establish robust antigen use; distant 3.0 Å remains a limitation.

Response to Reviewer m3Hm - Part 2

This is the continuation of the response.

Question 2: Reliability consistency and shuffling

“Could the authors provide further experiments to validate the Reliability mechanism?”

“Reliability Consistency: Repeatedly sample different perturbation ensembles for the same antigen to examine whether the residue-level reliability estimates remain stable.”

“Reliability Shuffling: Randomly shuffle the reliability weights assigned to different antigen residues while preserving the remaining geometric features and model architecture as much as possible, and examine whether the model's performance is significantly affected.”

Response: We added both requested tests: 20 independent K=8 ensembles per RAbD antigen for ranking consistency, and scalar/field shuffling for generation utility. Low-score overlap is above random, but the full combined ranking is only moderately repeatable; shuffling reduces generation quality.

For the combined score, the lowest-reliability quintile had mean overlap 0.514 (95% CI 0.491–0.537) versus an approximately 0.20 random reference. Overlap is the fraction of selected residues shared across repeats; Spearman measures agreement over the full ranking:

Channel / scoreMean Spearman correlation95% CI
Coordinate0.120[0.116, 0.124]
Topology0.695[0.674, 0.714]
Orientation0.204[0.195, 0.214]
Surface0.661[0.636, 0.684]
Combined reliability score0.252[0.238, 0.266]

Topology and surface are more repeatable than coordinate and orientation. The combined ranking is only moderately consistent, while low-score overlap exceeds random. Coordinate correlation 0.120 reflects variable sampled displacement rankings, not generator use.

Residue shuffling. Shuffled-scalar permutes scalar scores before weighting, preserving geometry and descriptors. Shuffled-field jointly permutes each eight-dimensional entry across descriptor and weighting uses. Network components are retained.

Antigen conditionVariantAAR (%) ↑CDR RMSD (Å) ↓DockQ ↑
CleanFull68.520.570.76
CleanShuffled-scalar66.020.780.75
CleanShuffled-field65.410.830.73

Results for epitope and distant 1.5 Å inputs are in the manuscript appendix, “Pathway and Residue-Assignment Controls.”

Scalar shuffling lowers AAR and increases RMSD despite preserving descriptors; field shuffling causes larger losses and lowers DockQ in all conditions. Thus residue assignments help generation; these are mean differences, not a statistical-significance claim.

Question 3: Coordinate-only Reliability

“Interpretation of the Coordinate-only Reliability ablation results.”

“Could these improvements arise from perturbation statistics or the additional conditioning pathway rather than from local structural reliability itself?”

Response: Coordinate-only improves generation relative to No Reliability; Full performs better than Coordinate-only, and No-coordinate measures coordinate's contribution alongside the other channels. These comparisons support task utility, while Coordinate-only is not a structural-accuracy score.

Coordinate-only versus No Reliability. Coordinate-only supplies realized per-residue displacement magnitudes to descriptor conditioning and scalar weighting. The generator retains base CDR–antigen geometry. Topology, orientation, and surface responses are absent, so the joint magnitude/response interpretation of Full does not apply.

Clean AAR/RMSD is 65.3%/0.70 Å versus 64.3%/0.75 Å. This supports empirical utility, but does not distinguish the two operations, establish structural accuracy estimation, or identify stochastic regularization as the cause.

Full versus Coordinate-only. Clean AAR/RMSD is 68.52%/0.57 Å versus 65.3%/0.70 Å, supporting the complete representation without isolating any individual added channel.

Full versus No-coordinate. Removing coordinate from both pathways while retaining the other channels changes clean AAR/RMSD from 68.52%/0.57 Å to 67.32%/0.78 Å. The same direction holds for epitope and distant 1.5 Å inputs (complete results in “Pathway and Residue-Assignment Controls”). This measures its joint contribution; No-coordinate still exceeds No Reliability in AAR.

Together with the pathway and shuffling controls, these results support a task-useful perturbation descriptor, not a standalone structural-accuracy score. Low ensemble repeatability does not imply negligible influence on generation.

Thank you for the detailed response and additional experiments. I would still appreciate further clarification of the perturbation evaluation protocol, particularly the use of the reliability field computed from the clean antigen. Overall, the additional results strengthen my assessment, and I maintain my original rating of 3 (Weak Accept).

Thank you for taking the time to consider our response and examine the additional experiments. We are glad that the new results strengthened your assessment. We appreciate your follow-up about the clean antigen, because the choice of structure used to compute the reliability field affects how the perturbation results should be interpreted.

There are two uses of perturbation in the protocol. The first creates the test input. For example, in the distant 3 Å condition, we displace a selected region of the antigen and rescale the displacements to a target Cα RMSD of 3 Å. The generator receives the resulting antigen coordinates together with the antibody framework and designs the six masked CDRs. Recovery is evaluated against the native antibody. For DockQ, the generated antibody is paired with the original antigen so that all conditions are scored against the same native complex. This use of the original antigen for scoring occurs after generation.

The second use of perturbation estimates the reliability field. As described in Reliability Estimation, we start from one antigen structure and sample eight nearby conformers using small, locally smoothed Cα displacements. At each residue, we measure the applied displacement and the changes in its neighbors, local orientation, and surface exposure. These measurements form the residue descriptor UA,jU_{A,j}. Its components are combined into the scalar score ρj\rho_j, which controls how strongly the corresponding geometric correction influences a CDR–antigen interaction. Once computed, the field is reused across generation steps and candidate antibodies.

By “the field computed from the clean antigen,” we mean that the eight conformers are sampled around the antigen before the test displacement. In the 3 Å example, the generator therefore sees distances and orientations from the displaced coordinates, while its reliability information comes from the original antigen. Here the clean antigen contributes information during generation, in addition to its separate role in evaluation. This setting tests how generation responds to altered geometry when the original reliability information is retained.

The supplementary section Leakage Control and Perturbation Evaluation describes computing a fresh field for each perturbed input. Under that protocol, the eight conformers in the same example would be sampled around the already displaced antigen. This is also how TrustCDR can be used when only a predicted antigen structure is available: that single structure supplies both the geometry for generation and the starting coordinates for reliability estimation. A result obtained with a retained clean field therefore answers the fixed-field sensitivity question. Assessing performance when only the perturbed structure is available requires computing the field from that structure.

Thank you again for raising this question. We hope this explanation addresses your concern about how the clean antigen is used, and we would be happy to answer any further questions.

Reviewer_827ZOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TrustCDR for antibody CDR sequence–structure co-design. It derives stability scores from synthetic antigen perturbations and uses them to reweight residual CDR–antigen pair updates in a flow-matching model.

Critical strengths
  1. The method explicitly considers the reliability of local antigen geometry to guide CDR–antigen interactions.
  2. The method estimates residue-level geometric reliability from a single antigen structure without requiring reference structures or paired conformations.
Critical weaknesses
  1. The proposed reliability scores are derived from synthetic perturbations without validation against experimentally observed structural uncertainty, raising concerns about their biological relevance.
  2. The method is evaluated under known antibody–antigen complex poses, and its applicability to scenarios with unknown binding orientations remains unexplored.
Questions for authors
  1. Can the authors provide experimental evidence demonstrating that the proposed reliability scores reflect biologically meaningful structural uncertainty?
  2. Local geometric instability may arise from molecular flexibility. Why do the authors rely on synthetic perturbations rather than considering experimental electron-density maps, cryo-EM density maps, or MD simulations to estimate or validate local structural reliability?
  3. How can the authors demonstrate that the proposed perturbation-based reliability scores accurately distinguish structurally reliable and unreliable antigen regions?
  4. The proposed reliability estimation approach should be compared with existing structural confidence measures, such as pLDDT or experimental B-factors, to demonstrate its advantages?
  5. Can the authors introduce random or shuffled reliability weights as additional controls to demonstrate that the improvements arise from meaningful reliability estimation rather than the additional model components?
  6. The paper does not provide publicly available code or implementation details.
7 comments
AuthorsOpenReview ↗

This is Part 1 of our response to the reviewer.

Response to Reviewer 827Z - Part 1

We thank you for your thoughtful comments. Because of the response-length limit, this response is split into two parts; this is Part 1. Changes are highlighted in red in the paper; further details are in the supplement.

Critical Weakness 1 and Question 1: external structural evidence

Critical Weakness 1

The proposed reliability scores are derived from synthetic perturbations without validation against experimentally observed structural uncertainty, raising concerns about their biological relevance.

Question 1

Can the authors provide experimental evidence demonstrating that the proposed reliability scores reflect biologically meaningful structural uncertainty?

Response: Thanks for raising this critical concern. Our score measures local response to perturbations. Larger responses mean lower reliability. We tested local accuracy using predicted coordinates and experimental-reference lDDT-Cα. We evaluated 14 whole-chain cases (3,370 residues) and 16 expanded-epitope cases (983 residues). We then averaged the case-wise Spearman correlations:

RegionCasesResidues in these casesMean Spearman correlation
Whole antigen chain143,370+0.2568
Expanded epitope16983+0.3086

Both mean correlations were positive, with a stronger association in expanded epitopes. Examples include 4lvn/A (+0.2553), 3cx5/E (+0.2550), and 2xwt/C (+0.1770). Question 4 compares pLDDT on the same cases. These results support a link with local accuracy, not a general measure of experimental uncertainty.

Critical Weakness 2: known and approximate binding poses

The method is evaluated under known antibody–antigen complex poses, and its applicability to scenarios with unknown binding orientations remains unexplored.

Response: TrustCDR designs CDRs from an antigen and a pre-positioned framework, which specifies an approximate binding region and orientation. Although searching over an unknown orientation is outside our search scope, we nevertheless followed your suggestion and added controlled pose-error experiments. Sequence recovery and CDR RMSD stayed close under modest pose errors, while DockQ declined with interface misalignment; this supports tolerance to modest errors, not recovery from an unknown orientation.

Across eight RAbD complexes, we applied two pose perturbations per complex: a 5° antigen rotation and 1 Å translation. Mean interface displacements were 0.85–2.11 Å:

Input poseAAR (%) ↑CDR RMSD (Å) ↓DockQ ↑
Original56.280.6150.754
Perturbed pose 156.810.6120.690
Perturbed pose 256.470.6210.701

Question 2: synthetic perturbations, density maps, and MD

Local geometric instability may arise from molecular flexibility. Why do the authors rely on synthetic perturbations rather than considering experimental electron-density maps, cryo-EM density maps, or MD simulations to estimate or validate local structural reliability?

Response: Thanks for suggesting this external validation. The method takes one antigen coordinate model. Perturbing coordinates tests the stability of the distances, orientations, and surface features used by the generator. Density maps and MD are not generation prerequisites. Following your suggestion, we ran two independent 5 ns explicit-solvent trajectories for complete 146-residue 3rkd/A, selected by completeness, length, and resolution before inspecting scores. Reliability came from the starting structure. After Cα alignment, we compared RMSF between quintiles:

Reliability groupTrajectory A RMSF (Å)Trajectory B RMSF (Å)
Lowest 20%0.6540.585
Highest 20%0.5200.475

The lowest-reliability quintile moved more in both trajectories, supporting short-time mobility separation for this antigen. Direct density-map analysis was not performed.

Question 3: distinguishing reliable and unreliable regions

How can the authors demonstrate that the proposed perturbation-based reliability scores accurately distinguish structurally reliable and unreliable antigen regions?

Response: Thanks for requesting this regional test. To test whether the score distinguishes regions with different structural accuracy, we compared the highest- and lowest-reliability quintiles against the same experimental reference using lDDT-Cα:

AntigenLow-reliability group lDDT-CαHigh-reliability group lDDT-Cα
2xwt/C0.90280.9358
3k2u/A0.88740.9376
5mes/A0.92740.9528

High-reliability regions have better local agreement in all three examples and higher mean lDDT-Cα in 11 of 16 epitope cases. The MD comparison in Question 2 provides complementary motion evidence. These results support regional discrimination in the evaluated cases.

AuthorsOpenReview ↗

This is Part 2 of our response to the reviewer.

Response to Reviewer 827Z - Part 2

This is the continuation of the response.

Question 4: comparison with pLDDT and B-factors

The proposed reliability estimation approach should be compared with existing structural confidence measures, such as pLDDT or experimental B-factors, to demonstrate its advantages?

Response: We added the requested pLDDT and experimental B-factor analyses. These signals address different questions: pLDDT estimates local accuracy, B-factors reflect experimental displacement or disorder, and ρ measures how antigen features respond to coordinate perturbations and weights CDR–antigen updates. The comparison shows that pLDDT is stronger for local-accuracy ranking, while ρ provides a different signal for generation.

On the same 14 whole-chain and 16 expanded-epitope cases as Question 1, we correlated both scores with lDDT-Cα over identical residues and averaged case-wise correlations:

RegionCasesMean Spearman(ρ, lDDT-Cα)Mean Spearman(pLDDT, lDDT-Cα)
Whole antigen chain14+0.2568+0.5769
Expanded epitope16+0.3086+0.5701

pLDDT correlates more strongly with lDDT-Cα on these cases. Across 58 X-ray antigens, unreliability (1−ρ) has a weak correlation with Cα B-factors (Fisher-z mean Spearman −0.060; −0.011 after neighbor-count adjustment). Thus, ρ is neither a replacement for pLDDT nor a B-factor proxy; its generation role is tested by the residue-shuffling controls in Question 5.

Question 5: shuffled reliability controls

Can the authors introduce random or shuffled reliability weights as additional controls to demonstrate that the improvements arise from meaningful reliability estimation rather than the additional model components?

Response: Thanks for requesting these controls. We added two controls that retain the model components but reassign reliability information across antigen residues. Shuffled-scalar permutes scalar scores ρ before computing weights and retains the descriptor inputs. Shuffled-field permutes complete eight-dimensional vectors, so descriptors and derived scores are reassigned together.

Antigen conditionVariantAAR (%) ↑CDR RMSD (Å) ↓DockQ ↑
CleanFull68.520.570.76
CleanShuffled-scalar66.020.780.75
CleanShuffled-field65.410.830.73

Results for epitope and distant 1.5 Å inputs are in the manuscript appendix, “Pathway and Residue-Assignment Controls.”

Under clean inputs, both shuffled controls reduce AAR and worsen RMSD relative to Full, and field shuffling also lowers DockQ. The same direction holds under epitope and distant 1.5 Å inputs. Thus residue assignment matters beyond including additional components.

Question 6: code and implementation details

The paper does not provide publicly available code or implementation details.

Response: Thanks for noting the implementation concern. Code was included with our submission at TrustCDR implementation. Implementation details are in the supplement’s reliability, leakage-control, training, and flow-backbone sections.

Ethics concerns: related manuscript and reference correction

First, Section 2 cites Energy-tilted interface potentials for antibody design (Anonymous, 2027), but I could not locate this paper in public academic databases or through online searches. No URL, DOI, or submission identifier is provided.

Response: Thanks for raising this disclosure concern. Energy-Tilted Interface Potentials for Antibody Design is a concurrent manuscript by a different author team in our laboratory. We included an anonymized copy with the original submission so the committee could assess its relationship to TrustCDR. It learns a label-free paratope–epitope potential for training-time calibration. TrustCDR estimates antigen-only geometric reliability to guide CDR generation. The contributions are distinct.

Second, the reference to Conformation-aware structure prediction of antigen-recognizing immune proteins incorrectly attributes authorship to “Henri Dreyer, Henry Kenlay, and Charlotte M. Deane.” However, the actual authors include Frédéric A. Dreyer, Jan Ludwiczak, Karolis Martinkus, and et al. Notably, the incorrectly cited names appear to have been taken from the paper's acknowledgments rather than its author list, raising concerns about possible LLM-generated citations.

Response: Thank you for identifying the error. We initially considered ABodyBuilder3 (Kenlay et al., 2024), then changed the intended citation to Conformation-aware structure prediction of antigen-recognizing immune proteins but failed to update its authors. We apologize, corrected the authors and DOI, checked the other references, and updated the AI-use statement.

My primary concern is that the proposed reliability estimation is not sufficiently grounded in the underlying physical properties of molecular structures. From a first-principles perspective, local structural instability is closely related to molecular flexibility, as flexible regions can adopt different conformations and exhibit varying degrees of structural variation. However, TrustCDR estimates reliability solely through coordinate perturbations of a single static antigen structure. Such perturbation sensitivity does not necessarily reflect the intrinsic conformational flexibility of the molecule.

Although the authors have added comparisons between their reliability scores and experimental-reference lDDT-Cα, these results remain insufficient to establish the biological relevance of the proposed metric. First, lDDT-Cα measures local structural accuracy rather than directly quantifying molecular flexibility. Second, the validation involves only 14–16 cases, with relatively weak mean Spearman correlations of approximately 0.26–0.31. Furthermore, the additional MD validation is limited to a single antigen, which is insufficient to establish a general relationship between the proposed reliability scores and molecular flexibility. Notably, the authors' own comparison shows that pLDDT correlates substantially better with local structural accuracy, while the proposed unreliability score exhibits almost no correlation with experimental B-factors, even after neighbor-count adjustment.

Therefore, it remains unclear to what extent the proposed reliability scores actually capture molecular flexibility, which is central to the biological motivation of this work. The additional experiments in the rebuttal do not adequately resolve this fundamental concern.

Additionally, the framework figures contain numerous garbled text elements, which significantly affects the clarity and presentation of the proposed methodology.

Overall, I do not believe the rebuttal sufficiently addresses the fundamental concerns regarding the biological validity of the proposed reliability estimation. I therefore lower my rating by one point, from Weak Reject to Clear Reject.

Reviewer_FF3MOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TrustCDR, a reliability-guided framework for antigen-conditioned antibody CDR sequence–structure co-design. Given a single antigen structure, the method constructs a small perturbation ensemble and measures residue-level changes in coordinates, neighborhood topology, local orientation, and surface exposure to derive relative geometric reliability scores. These scores guide learned residual corrections to CDR–antigen pair representations through applicability gates and normalized trust weights, while preserving the original pair features. The refined representations condition a flow-matching model that jointly generates CDR amino-acid identities, positions, and orientations. Experiments on the RAbD benchmark report improved overall performance under clean inputs and favorable interface quality under synthetic antigen perturbations, supported by mechanism ablations and computational affinity screening.

Critical strengths
  1. The paper is generally well written and easy to follow. It clearly motivates the problem and explains how perturbation-based reliability estimation, residual pair refinement, and joint CDR generation fit together.
  2. The paper’s research on antibody CDR sequence–structure co-design under local antigen uncertainty is meaningful, as it addresses the influence of potentially unreliable antigen geometry on target-specific antibody design.
  3. The experiments cover clean inputs, multiple perturbation strengths, and both epitope-local and distant perturbations. Baseline comparisons, mechanism ablations, computational affinity screening, and an external case study provide complementary evidence for the method’s utility.
Critical weaknesses
  1. The paper interprets local geometric stability under synthetic perturbations as residue-level geometric reliability, but lacks direct validation against actual structural errors or experimentally observed antigen variability.
  2. Table 5 reports wall-clock training time for 200,000 iterations, but does not clarify whether it includes the preprocessing costs of constructing the antigen-only perturbation ensemble and computing the reliability field. Generation time is also not reported. Training time alone is insufficient to assess computational efficiency.
  3. The multi-channel reliability descriptor guides both the residual correction and the pair-applicability gate, while the scalar reliability score produces relative trust weights. However, the “No reliability” variant in Table 2 removes reliability estimation and all reliability-derived guidance simultaneously, making it difficult to identify which pathway primarily contributes to the performance gains.
  4. Table 4 specifies that the perturbation ensemble uses Cα Gaussian offsets, whereas the orientation response in Equation (4) requires local backbone frames in the observed and perturbed structures. The paper does not clarify whether these frames are constructed from neighboring Cα atoms or from backbone atoms such as N, Cα, and C.
Questions for authors
  1. The paper interprets local geometric stability under synthetic perturbations as residue-level geometric reliability, but lacks direct validation against actual structural errors or experimentally observed antigen variability.
  2. Table 5 reports wall-clock training time for 200,000 iterations, but does not clarify whether it includes the preprocessing costs of constructing the antigen-only perturbation ensemble and computing the reliability field. Generation time is also not reported. Training time alone is insufficient to assess computational efficiency.
  3. The multi-channel reliability descriptor guides both the residual correction and the pair-applicability gate, while the scalar reliability score produces relative trust weights. However, the “No reliability” variant in Table 2 removes reliability estimation and all reliability-derived guidance simultaneously, making it difficult to identify which pathway primarily contributes to the performance gains.
  4. Table 4 specifies that the perturbation ensemble uses Cα Gaussian offsets, whereas the orientation response in Equation (4) requires local backbone frames in the observed and perturbed structures. The paper does not clarify whether these frames are constructed from neighboring Cα atoms or from backbone atoms such as N, Cα, and C.
2 comments
AuthorsOpenReview ↗

This is Part 1 of our response to the reviewer.

Response to Reviewer FF3M - Part 1

We thank you for your thoughtful comments. Because of the response-length limit, the response is split into two parts. We address geometric validation with experimental-reference and MD comparisons. We also discuss pathway contributions, computational cost, and orientation construction. Changes are highlighted in red in the main paper. Further details are in the supplementary material.

CW1 / Q1: Validation of geometric reliability

“The paper interprets local geometric stability under synthetic perturbations as residue-level geometric reliability, but lacks direct validation against actual structural errors or experimentally observed antigen variability.”

Response: Thanks for suggesting these validations. We performed two complementary tests. Prediction–experimental comparisons found positive associations between reliability and local accuracy. Two independent explicit-solvent trajectories also showed greater short-time motion in the lowest-reliability regions. Together, these results support a relationship between the score, local accuracy, and short-time mobility. Detailed settings and values follow.

RegionCasesResidues in these casesMean Spearman correlation between ρ and lDDT-Cα
Whole antigen chain143,370+0.2568
Expanded epitope16983+0.3086

Across 16 expanded-epitope cases, the high-reliability quintile had a mean lDDT-Cα advantage of 0.0148, positive in 11 cases. Whole-chain and epitope evaluation sets overlap.

For an independent mobility check, we selected 146-residue 3rkd/A by completeness, length, and resolution before inspecting scores and ran two 5 ns explicit-solvent trajectories. Reliability came from the starting structure. After Cα alignment, we compared RMSF (motion around each residue’s mean position) between reliability quintiles:

TrajectoryLowest-ρ quintile mean RMSF (Å)Highest-ρ quintile mean RMSF (Å)
A0.6540.520
B0.5850.475

The lowest-reliability quintile moved more in both trajectories. The two tests support a relationship to local accuracy and short-time motion; pathway and shuffling controls separately test generation utility.

CW2 / Q2: Preprocessing, training, and generation costs

“Table 5 reports wall-clock training time for 200,000 iterations, but does not clearly clarify whether it includes the preprocessing costs of constructing the antigen-only perturbation ensemble and computing the reliability field. Generation time is also not reported. Training time alone is insufficient to assess computational efficiency.”

Response: Thanks for asking us to separate these costs. Training time excludes reliability preprocessing. The reported 30.58 h covers 200,000 iterations, including validation and checkpoint saving.

On an RTX 4090, we measured the added preprocessing and sampling costs:

OperationMeasured cost
Reliability precomputation for 58 test complexes58.3 s total
Cached generation, five candidates per batch7.33 s per batch
Generation without cached reliability, five candidates per batch9.77 s per batch
Cached generation, amortized per candidate1.47 s

Reliability is computed once per antigen and reused across candidates. The summary is in Experimental Setup and the detailed breakdown in Training Details.

AuthorsOpenReview ↗

This is Part 2 of our response to the reviewer.

Response to Reviewer FF3M - Part 2

This is the continuation of the response.

CW3 / Q3: Contributions of descriptors and scalar weights

“The multi-channel reliability descriptor guides both the residual correction and the pair-applicability gate, while the scalar reliability score produces relative trust weights. However, the "No reliability" variant in Table 2 removes reliability estimation and all reliability-derived guidance simultaneously, making it difficult to identify which pathway primarily contributes to the performance gains.”

Response: Thanks for identifying the pathway ambiguity. We added pathway ablations and residue-shuffling controls. Descriptor-only and Scalar-only isolate descriptor conditioning and scalar weighting, while Full tests their combination. The ablations show that both pathways contribute. Shuffling the scalar or complete field reduces performance, showing that reliability must remain assigned to the corresponding residues. Detailed definitions and results follow.

Antigen conditionVariantAAR (%) ↑CDR RMSD (Å) ↓DockQ ↑
CleanFull68.520.570.76
CleanDescriptor-only67.280.750.75
CleanScalar-only66.000.730.75
CleanNo Reliability64.300.750.76

The epitope and distant 1.5 Å rows are reported in the revised manuscript and supplementary material.

Descriptor-only improves AAR; Scalar-only improves AAR and RMSD despite lacking descriptor conditioning. Full has the highest AAR. These trends hold in all three conditions, supporting contributions from both pathways.

Shuffled-scalar permutes scalar scores while preserving descriptors. Shuffled-field permutes complete eight-dimensional entries across both uses. Both retain the model components and test residue assignments.

Antigen conditionVariantAAR (%) ↑CDR RMSD (Å) ↓DockQ ↑
CleanFull68.520.570.76
CleanShuffled-scalar66.020.780.75
CleanShuffled-field65.410.830.73

The epitope and distant 1.5 Å rows are reported in the revised manuscript and supplementary material.

Scalar shuffling worsens AAR and RMSD despite retaining descriptors; field shuffling causes larger losses and also lowers DockQ in all conditions. Thus residue assignments matter beyond including the components.

CW4 / Q4: Construction of orientation frames

“Table 4 specifies that the perturbation ensemble uses Cα Gaussian offsets, whereas the orientation response in Equation (4) requires local backbone frames in the observed and perturbed structures. The paper does not clarify whether these frames are constructed from neighboring Cα atoms or from backbone atoms such as N, Cα, and C.”

Response: Thanks for requesting the frame details. The orientation frames use neighboring Cα coordinates and are recomputed after each perturbation. N and C backbone atoms are not required. For an internal position jj, we define the incoming and outgoing vectors:

nj=(xj−xj−1)/(‖xj−xj−1‖+εf),cj=(xj+1−xj)/(‖xj+1−xj‖+εf).n_j = (x_j − x_{j−1}) / (‖x_j − x_{j−1}‖ + ε_f), c_j = (x_{j+1} − x_j) / (‖x_{j+1} − x_j‖ + ε_f).

We define the frame axes as:

e3,j=(nj+cj)/(‖nj+cj‖+εf),e1,j=(nj×cj)/(‖nj×cj‖+εf),e2,j=e3,j×e1,j.e₃,j = (n_j + c_j) / (‖n_j + c_j‖ + ε_f), e₁,j = (n_j × c_j) / (‖n_j × c_j‖ + ε_f), e₂,j = e₃,j × e₁,j.

We assemble these axes as rows of OjO_j, with ε_f = 10⁻⁸. We recompute the frame for each perturbed conformer and average

arccos(clip((tr(Ojk(Oj0)T)−1)/2,−1+εf,1−εf)).arccos( clip( (tr(O_jᵏ (O_j⁰)ᵀ) − 1) / 2, −1 + ε_f, 1 − ε_f ) ). over the K=8K=8 conformers, in radians. The bisector tracks the local Cα path. The cross-product axis tracks its plane. Recomputing both after perturbation measures the local orientation change.

The first and last entries of the concatenated antigen array use identity frames. Internal chain boundaries and missing-residue gaps are not separately segmented; near-collinear or opposing vectors use epsilon stabilization without a separate fallback.

We have added this construction to Methods and the full equations and boundary conventions to the manuscript’s appendix, under “C-alpha Frame Construction.”

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.