acceptodds
Withdrawn from ICLR 2027

UniTact: Sequence-Only Unified Antibody–Antigen Interaction Modeling for Contact and Interface Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Identifying antibody–antigen interfaces is essential for understanding immune recognition and the effects of antigen mutations. Given a binding antibody–antigen pair, however, their relative binding geometry is unknown from sequences or monomer structures alone, making the joint prediction of antibody-side paratopes, antigen-side epitopes, and residue-level contacts challenging. Existing approaches either rely on complex structures or focus on individual interface tasks, and residue-pair predictions can be dominated by marginal residue preferences rather than partner-specific interactions. We introduce UniTact, a unified framework that predicts paratopes, epitopes, and residue-level contacts from a shared antibody–antigen contact surface. The model takes sequences as its only raw input and incorporates separately predicted monomer structures to construct interaction representations through intrachain encoding, cross-attention, and pair updates. We further introduce a double-centering strategy on pair scores to reduce additive row and column effects, making the readouts more sensitive to pair-specific variation. On SAbDab2, our method achieves an epitope AUPRC of 0.2232, improving over the sequence-based baseline SPPIDER-seq (0.2034), while obtaining paratope AUROC of 0.9561 and contact AUROC of 0.9203. Further score decomposition, controlled experiments, and split-based analyses help characterize the sources of residue-level bias and clarify the remaining challenges of pair-specific contact prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody-Antigen Interaction, Protein Language Models, Molecular Representation Learning, Attention Mechanisms

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_KidROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

UniTact predicts antibody binding residues (paratopes), antigen binding residues (epitopes) and the contacts between them. It takes sequences as input and uses separately predicted structures, without a known complex pose. Double centering removes broad residue preferences from pair scores to help identify specific interactions. The model is evaluated on SAbDab2 and two AsEP splits. The strongest results are for paratopes; contact detection remains weak.

Critical strengths

Jointly predicting binding sites (paratope + epitope) and contacts without a known complex pose is a useful practical goal. A reliable model in this setting could help guide follow-up structural studies.

Critical weaknesses
  1. The paper does not show whether its main idea improves prediction. Section 5.4 and Appendix C describe removing centering, changing the interface readouts and varying pooling, but give no numerical results. Section 5.3 claims that antibody-side effects dominate the final contact scores without showing the final-model analysis. Table 4 concerns an earlier model and cannot support this claim. I therefore cannot judge whether centering adds value. Deriving interface labels from contacts is a natural definition, and centering is a standard operation; the contribution needs a demonstrated benefit.
  2. The comparisons and reported metrics do not establish a clear advantage. Key comparators are missing: WALLE for AsEP, Paragraph and ParaSurf for paratopes, and Epi4Ab and EpiScan for epitopes. Since UniTact uses predicted structures, relevant structural baselines should also be tested on those same predicted antibody and antigen structures, where applicable, for a fairer comparison. A co-folding baseline would show how UniTact compares with predicting a complex directly. Results are also missing on established benchmark datasets such as PECAN, Paragraph-expanded and MIPE. Evaluation on these datasets would help assess whether the findings generalize across benchmarks. The abstract highlights contact AUROC of 0.9203, but AUPRC is 0.0208 and recall is 0.0058: only 0.58% of true contacts are recovered at the chosen threshold. Contact AUPRC is also below ProTact-best-antibody (0.0302), a structure-only baseline in the same table, so the model loses the contact AUPRC column to a baseline while the abstract reports the AUROC. The epitope gain over SPPIDER-seq is modest, and paratope AUPRC (0.7278) is below Paraplume (0.7319). No uncertainty estimates are shown despite ten runs. Precision, F1 and MCC are promised but missing, making the recall comparisons difficult to interpret.
Questions for authors
  1. What are the matched ablation results with and without centering? Please report the final model's contact performance using the full scores, the additive terms alone and the centered interaction alone. How much useful information comes from the antibody partner?
  2. Were the baselines retrained on the same splits, and were labels and residue masks matched? Please provide WALLE and a co-folding comparison, or explain any concrete evaluation limitations.
  3. What exact clustering thresholds, coverage rules and split files were used? Does the AsEP group split separate antigens or only epitope groups? Please also clarify whether contact labels use heavy atoms and exclude hydrogens.
  4. Can you provide uncertainty estimates across runs, the omitted precision, F1 and MCC, and contact prevalence? These are needed to assess the reported gains and interpret contact AUPRC.
Reviewer_FZmfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents UniTact, a unified framework for antibody–antigen interaction prediction, covering paratope, epitope, and residue-level contact prediction. The model integrates sequence representations and predicted monomer structures through cross-attention and pairwise interaction modeling. It further introduces double centering to separate pair-specific signals from additive residue-level effects. Experiments on SAbDab2 and AsEP demonstrate competitive performance on several metrics, while additional analyses investigate residue-level biases in contact prediction and the impact of dataset splits.

Critical strengths

The paper addresses antibody–antigen interaction prediction without requiring experimentally determined complex structures, which is an important problem with potential applications in co-folding and binder design.

Critical weaknesses

W1. Severely incomplete experimental evaluation.

The current experimental evaluation is severely incomplete in several aspects:

  1. Task difficulty: Given the substantial performance differences across the three tasks, the authors should provide more context on their relative difficulty, such as class imbalance statistics and random-baseline performance.
  2. Structure-based comparisons: UniTact already incorporates predicted monomer structures, so it should be more directly compared with structure-based methods rather than evaluated primarily within the sequence-only category.
  3. Missing baselines: The evaluation should include docking and co-folding baselines, as these approaches can also produce complex structures from which contact, paratope, and epitope predictions can be derived.

W2. Insufficient experimental evidence supporting the proposed contributions.

Despite achieving a contact AUROC of 0.9203, the model obtains an AUPRC of only 0.0208 and a recall of 0.0058. Moreover, the authors report that antibody-side residue effects still dominate the final contact scores. These findings raise questions about whether the proposed double-centering operation meaningfully improves pair-specific contact prediction.

More importantly, although several ablation experiments are described, their quantitative results are not reported. In particular, the effectiveness of double centering, one of the paper's main methodological contributions, remains unverified. Ablations of the unified multi-task formulation are also needed to establish the benefits of the proposed architecture.

W3. Insufficient empirical support for the motivation.

The paper argues that residue-pair contact predictions can be dominated by residue-level preferences rather than pair-specific interactions. While this observation is interesting, some of the motivating claims lack sufficient empirical evidence, particularly regarding the limitations of existing methods. Additional analyses of representative baselines would help establish the generality of this problem.

W4. Limited methodological novelty.

The methodological contributions are relatively straightforward. The first contribution formulates three existing prediction tasks using labels derived from a shared contact matrix and standard multi-task learning. The second applies double centering, a well-established mathematical operation, to pairwise scores. Overall, the proposed framework largely integrates existing techniques rather than introducing substantially new modeling principles.

W5. Presentation quality.

Some figures have insufficient resolution, making their details difficult to read. Replacing them with vector graphics (e.g., PDF) would improve readability.

Questions for authors

Please see Weakness

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

UniTact is a unified framework for predicting antibody-side paratopes, antigen-side epitopes, and residue-level antibody-antigen contacts from sequence inputs. Frozen ESM-C representations and separately predicted monomer structures are processed by chain-specific encoders, bidirectional cross-attention, and pair updates. A mask-aware double-centering operation removes additive antibody-row and antigen-column effects from the learned pair scores before the shared interaction representation is used by the three prediction heads.

The model is evaluated on SAbDab2 and two AsEP splits using multiple random seeds. On SAbDab2, it reports contact AUROC and AUPRC of 0.9203 and 0.0208, epitope AUROC and AUPRC of 0.6371 and 0.2232, and paratope AUROC and AUPRC of 0.9561 and 0.7278. The paper also analyzes additive effects in pair scores, partner-dependent epitope labels, and the gap between ranking metrics and thresholded contact detection.

Critical strengths
  1. The unified formulation is conceptually valuable. Contact, paratope, and epitope prediction are derived from a common antibody-antigen contact matrix, while the model requires only sequences as raw inputs and predicted monomer structures rather than a known complex pose. This connects three interface tasks that are often treated separately and makes their performance directly comparable under a shared label definition.

  2. The paper goes beyond headline AUROC values by examining additive row and column effects, severe class imbalance, partner-dependent epitope labels, and split sensitivity. The multi-seed evaluation and the explicit discussion of the gap between contact ranking and thresholded detection provide a useful and unusually candid view of the remaining difficulty of pair-specific contact prediction.

Critical weaknesses
  1. The empirical evidence does not yet fully support the paper's central mechanistic claims. Section 5.4 and Appendix C describe removing double centering, changing the interface readout, and varying the epitope top-k value, but the corresponding quantitative results are not reported. In addition, Table 4 decomposes scores from an earlier uncentered three-block configuration rather than the final model, and the partner-dependence analysis characterizes label variation without directly testing whether the trained model uses antibody identity to change its predictions. These omissions make it difficult to determine what the proposed centering operation improves and how much partner-specific signal the final model actually learns.

  2. The evaluation and baseline protocols need more detail. Table 2 does not clearly state which baselines were retrained on the same data and which used published weights, how MIPE scores were converted into contact predictions, or how every method was aligned to the common 4.5 angstrom labels. The evaluation section lists precision, F1, and MCC, but the main result tables report AUROC, AUPRC, and recall; without precision and threshold analysis, the contact recall of 0.0058 at a fixed threshold of 0.5 is difficult to interpret operationally. Minor presentation problems also remain: all three Figure 1 output branches are labeled Contact Prediction, the figure text is small, and the Table 4 caption is separated from the table across pages 12 and 13.

Questions for authors
  1. What are the matched-seed numerical results for removing double centering, replacing the centered interface readouts, and varying the epitope top-k value? Please report the affected metrics and indicate whether the effects are consistent across datasets or splits.

  2. Can the final model's contact scores be decomposed in a way that directly measures the predictive contribution of the complete score, additive row and column terms, and pair-specific residual? On antigens paired with multiple antibodies, do paired predictions or an antibody-shuffling control show that the predicted epitope changes with antibody identity?

  3. For each Table 2 baseline, was the method retrained on the current data or evaluated from released weights, and how were its outputs mapped to the shared 4.5 angstrom contact labels? Please also clarify the precision, F1, and MCC values at the fixed threshold and whether the conclusions change under a validation-selected threshold.

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

Related papers

Discussion (0)

Sign in to comment.