- 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
- 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.
- 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
- 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.
- 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
- 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?
- 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.
- 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 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 / score | Mean Spearman correlation | 95% CI |
|---|---|---|
| Coordinate | 0.120 | [0.116, 0.124] |
| Topology | 0.695 | [0.674, 0.714] |
| Orientation | 0.204 | [0.195, 0.214] |
| Surface | 0.661 | [0.636, 0.684] |
| Combined reliability score | 0.252 | [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 condition | Variant | AAR (%) ↑ | CDR RMSD (Å) ↓ | DockQ ↑ |
|---|---|---|---|---|
| Clean | Full | 68.52 | 0.57 | 0.76 |
| Clean | Shuffled-scalar | 66.02 | 0.78 | 0.75 |
| Clean | Shuffled-field | 65.41 | 0.83 | 0.73 |
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 . Its components are combined into the scalar score , 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.
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
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.
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
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:
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
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.
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:
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.