acceptodds
Under review as a conference paper at ICLR 2027

When Frozen Features Help or Harm: Correspondence and Auxiliary Supervision in Antibody–Antigen Prediction

Abstract

Frozen protein representations and relational auxiliary losses provide different routes for augmenting an interaction predictor. We separate the benefit of an added representation from sensitivity to its residue correspondence, and distinguish contact supervision from partner-ranking supervision. On a SAbDab-derived development set, aligned features outperform features permuted throughout downstream training and evaluation by 0.00820 macro pair AUPRC for ESM-IF1 and 0.00319 for ProteinMPNN, with positive differences in all ten paired seeds. A prespecified evaluation of the frozen checkpoints on a 217-complex test partition yields a mixed confirmation: ProteinMPNN retains a 0.00359 effect (95% seed-bootstrap interval [0.00156, 0.00588]; nine positive seeds), whereas ESM-IF1 falls to 0.00226 (interval [-0.000007, 0.004901]; six positive seeds). A prespecified 185-complex sensitivity subset excludes prior exposure through an unrelated baseline evaluation without replacing the primary partition. Thus, strong development-set correspondence sensitivity need not transfer with the same magnitude. Separately, a four-arm AsEP development study shows that removing partner ranking improves pair AUPRC by 0.0214 and epitope MCC by 0.0369, both in all ten seeds. The node advantage appears at matched epochs before learning rates diverge; gradient diagnostics do not identify a unique mechanism. These results distinguish correspondence sensitivity, baseline-relative benefit, and auxiliary-objective utility, while making the different evidential status of development findings and held-out confirmation explicit.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody-antigen prediction, protein representation learning, frozen representations, inverse folding, auxiliary supervision, transfer learning, biological machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_KJM9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether including or excluding certain features derived from the representation spaces of different models can enhance performance or adversely affect predictions.

The authors frame their study within partner-conditioned predictions and analyze the effects of adding or removing frozen components in data representation pipelines.

The first setting and task presented involves producing pair scores and separate antigen- and antibody-node probabilities to predict which residues participate in antibody-antigen interactions and which residue pairs are in contact.

The authors quantify the effect of additional features on the correct residues, the effect of shuffling features across residues, and the correspondence sensitivity.

Subsequently, a retraining experiment is conducted, examining both training and evaluation with shuffled features, as well as training with aligned features while validating on shuffled ones.

The evaluation considers the sensitivity measure—a relative comparison between aligned and shuffled features when applied to the correct interaction residues—focusing specifically on whether additional structural features improve performance.

The analysis of ESM-IF1 and ProteinMPNN yields inconclusive generalization results, as the technique is highly beneficial in the former case but provides only negligible improvements in the latter.

Further analysis compares aligned versus shuffled features to measure the difference in performance advantage. Here too, the results are not entirely conclusive; while the advantage was substantial during development, it failed to generalize to the test set.

The final experimental setting considers shuffling features across all residues rather than just the key interacting ones. In this scenario, the models consistently degrade in performance.

The authors then provide insights into potentially adding a contact loss and a partner-ranking loss in antibody-antigen prediction. The results indicate that supervision of the actual contact points between chains is the primary driver of improvement, whereas the ranking component harms model performance. Even with scheduling or gradient inspections, the authors do not clearly address this issue.

Critical strengths

Overall, the paper is clear, and the underlying idea is interesting. Given the increasing prevalence of complex architectures that merge intermediate representations from different networks—often without a clear understanding of each component's contribution—this is an important question to address. The experimental section is thorough and includes training losses. In the discussion, the authors also analyze limitations and offer observations on the method.

Critical weaknesses

The methodology lacks clear overall organization. The authors' intent is difficult to follow, and the manuscript lacks a cohesive narrative that would better motivate and frame the underlying problem. Some sections are oversimplified, lacking formal notation as well as helpful tables or visual representations that would render the analysis clearer. The figures are uninformative and lack clarity, even when reading the captions.

The authors should include additional benchmarks and incorporate more models into the analysis to better validate the method's generalization capabilities. Moreover, the gains obtained by applying the method across the evaluated cases are not consistently substantial. This highlights a lack of compelling evidence for the method's effectiveness and an absence of analysis explaining the reasons behind the reported results.

Questions for authors

a. What is the motivation that led the authors to pick this specific task? The evidence is fairly limited due to sticking with a single task and a small number of chosen models.

b. For the training procedure, there is a lack of continuous ablation of the loss component weights. The absence of such interpolation greatly limits the conclusions that can be drawn regarding the proposed losses.

c. The analysis regarding the insertion of regularization later in training is not sufficiently formalized or sound. It lacks clear criteria for insertion timing as well as ablation of the approach.

d. The analysis of the gradients lacks an explanation of the underlying intuition behind this approach.

e. The related work section lacks clarity; it could be expanded and revised to include more baselines for comparison.

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

This paper investigates how frozen pretrained representations and auxiliary objectives affect antibody–antigen contact and interface prediction. Its first contribution is to distinguish the benefit of adding correctly aligned features from the performance gap between aligned and residue-permuted features. Using a common partner-conditioned predictor, it evaluates ESM-IF1 and ProteinMPNN features against sequence-only inputs and other controls. Development experiments show positive aligned–permuted differences for both encoders. A subsequent evaluation of frozen checkpoints on 217 test complexes supports the ProteinMPNN contrast, while the ESM-IF1 effect is smaller and its interval includes zero. Its second contribution is a four-arm study separating node, contact-edge, and partner-ranking supervision. On AsEP validation data, removing ranking from the full objective improves contact AUPRC and epitope MCC in all ten paired seeds. Common-epoch analyses show that the node advantage appears before learning rates diverge. The paper presents these as bounded empirical findings and explicitly separates development evidence, held-out confirmation, and exploratory gradient diagnostics.

Critical strengths

The paper distinguishes whether adding features improves prediction from whether assigning those features to the correct residues matters. This is a useful distinction when interpreting ablations. Separating contact supervision from partner ranking also helps clarify which part of the training objective affects performance.

Critical weaknesses
  1. The contribution is limited. The paper shows that residue correspondence matters and that an auxiliary loss can hurt prediction performance. These observations are useful, but the study offers little explanation of when or why they occur. All experiments use the same downstream predictor, and neither finding is independently replicated across both datasets. I am therefore not convinced that the results provide a sufficiently substantial insight beyond this particular setup.
  2. The baseline evaluation is incomplete. WALLE, PEPNet, and LF3DRoPE appear only as published reference numbers in Table 11, so their performance is not directly comparable under the authors’ setup. More importantly, the held-out experiment omits the ESM-2 baseline. It therefore cannot establish whether adding the frozen features improves prediction on unseen complexes. The ESM-IF1 random-projection control also differs in input dimensions and parameter count, which limits what that comparison can establish.
  3. The ranking-loss result needs stronger support. The authors test one nonzero ranking weight and one negative-sampling configuration, with results reported only on validation data. The sampled substitutions are not verified non-binders, and partner retrieval performance is not reported. It remains unclear whether ranking is unhelpful in this setting, whether its weight is too large, or whether it improves retrieval at some cost to contact prediction.
  4. Checkpoint selection complicates the interpretation. The representation experiments select checkpoints using ranking loss, although contact AUPRC is the main evaluation metric. In the objective study, removing a loss also changes the quantity used for scheduling and checkpoint selection. The common-epoch analysis helps address this concern for node prediction, but the corresponding contact-prediction results are not shown. These choices make it difficult to separate the effects of the features and losses from those of the training procedure.
  5. Generalization to new biological examples is not sufficiently assessed. The SAbDab2 splits are disjoint by instance identifier, but the paper does not adequately report antigen and antibody similarity across splits. The confidence intervals measure variation across training seeds on the same complexes, so they provide limited information about uncertainty on new antigen families. The exposure-exclusion analysis does not address these biological dependencies.
  6. The paper is difficult to follow. The abstract introduces detailed statistics and exposure qualifications before clearly explaining the task and motivation. Important choices about inputs, losses, checkpoint selection, and evaluation are scattered across sections and appendices. Repeated discussion of protocols and numerical reconstruction interrupts the main argument. I often had to look back through the paper to determine which experiment supported a conclusion, making the contribution harder to assess.
Questions for authors
  • What do you consider the main new insight beyond showing that correspondence matters and auxiliary losses can hurt? Which results support its relevance beyond the chosen predictor?

  • Can you provide held-out results for the ESM-2 baseline alongside the aligned and permuted conditions? A comparison with an established predictor under the same setup would also help establish the practical relevance of the findings.

  • How were the ranking weight and sampled negatives chosen? Do smaller weights avoid the performance drop, and does the ranking objective improve partner retrieval?

  • Why was ranking loss used for checkpoint selection in the representation study? Do the conclusions hold when selection uses the intended prediction endpoint, and what happens to contact AUPRC at matched epochs?

  • How similar are the antigens and antibodies across the SAbDab2 splits? Do the observed effects remain consistent when results are examined by antigen family or sequence-similarity group?

  • Could you provide a single summary of the main experiments, specifying their inputs, training objectives, checkpoint-selection rules, and evaluation splits? This would help clarify how the separate studies support the paper’s overall conclusion.

Reviewer_bDJ3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

It is unclear to me what the paper is trying to do. I found it impossible to comprehend the text.

Critical strengths

I did not understand the manuscript and have no comments on the strengths.

Critical weaknesses

I did not understand the manuscript and have no comments on the weaknesses.

Questions for authors

I have no questions for the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.