acceptodds
Withdrawn from ICLR 2027

EpiFormer: Learning Antigen-Antibody Interactions for Epitope Prediction via Geometric Deep Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Antibodies neutralize foreign antigens by binding to specific surface regions called epitopes. Computational epitope prediction is critical for understanding immune recognition and guiding antibody engineering. However, existing methods face three fundamental challenges: antibody-aware models encode each chain independently and combine them only at a late stage, failing to capture co-dependent structural features that define binding interfaces, whereas severe class imbalance and scarcity of known antibody-antigen complexes render standard training objectives ineffective. We propose EpiFormer, a general encoder-decoder framework that addresses these challenges jointly. Our key design principle is interleaved cross-attention within GNN encoding layers, enabling bidirectional antigen-antibody information flow throughout representation learning rather than only at the output. This early-fusion principle is backbone-agnostic, providing consistent gains across GNN architectures from simple GCNs to equivariant models. We further show that sparsity-aware objectives are effective when paired with early-fusion architectures for the epitope prediction task. EpiFormer improves over the previous best method by over 40% in F1 score on standard benchmarks, demonstrating generalizability and cross-dataset transferability. Notably, two behaviors of EpiFormer emerge from end-to-end training alone, where the learned cross-attention gates favor antigen-to-antibody information flow, consistent with the asymmetric roles of the two chains at the binding interface, and the model's preference for geometric over evolutionary features aligns with the established finding that epitope residues are not evolutionarily conserved.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Epitope prediction, protein-protein interaction, antibody design, graph neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Yg2ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EpiFormer, an encoder–decoder framework for antibody-specific epitope prediction. The proposed claim of the paper is that antigen and antibody information should be fused early, and the paper does so via bidirectional multi-head cross-attention interleaved in the GNN encoder block rather than applied only at the decoder. It pairs this with a sparsity-aware joint objective and with EGNN-R. On AsEP, it reports 0.924 AUC / 0.482 F1 against many baselines, with a new family-based split, backbone swaps, and structural-perturbation tests.

Critical strengths
  • Baseline comparison is very broad
  • The robustness suite is interesting and goes well beyond this task.
Critical weaknesses
  • The α=0.0 ablation contradicts the co-design claim. That configuration is a late-fusion model (decoder MHCA only) trained with the full sparsity-aware objective, and it reaches 0.454 F1, within 0.042 of the full model and +0.117 over the best baseline. The objective plainly does not require early fusion to work.
  • The headline mechanism is the smallest effect in the paper. Interleaved cross-attention contributes +0.042 F1; the sparsity/count regularizer contributes +0.132 and L_geo +0.035 (Table 16).
  • Labels are a geometric function of the same bound coordinates supplied as input, so part of the task is recovering a function of the input rather than predicting binding; the donor-conformation test is too weak to separate these.
  • Cross-attention operates purely on E(3)-invariant scalars with no inter-chain geometry, so “early fusion” is residue-descriptor compatibility matching, not geometric reasoning
Questions for authors
  • In Table 3, the three methods that improve all lack cross-attention entirely, and the three that degrade all have late cross-attention. Can you distinguish your “early fusion is required” reading from “these losses conflict with an existing late cross-attention module”?
  • Missing A.6 EXTERNAL DATASETS AND DISTRIBUTION SHIFT
  • Table 11 contradicts A.3. If a rigid transform of one chain moves outputs by <10⁻⁵, per-chain rotation cannot cost 0.106 AUC and global rotation cannot gain 0.010.
  • What is the computational complexity of the method compared to the baselines?
Reviewer_QmqZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EpiFormer, an encoder–decoder framework for antibody-aware epitope prediction. The central design is interleaved bidirectional cross-attention inside GNN encoding layers, so antigen and antibody representations exchange information at every encoder block rather than only at the output. This early-fusion principle is backbone-agnostic: the authors show gains across GCN, GIN, GAT, RGCN, EGNN, and their proposed EGNN-R, a multi-relational E(3)-equivariant GNN with relation-specific message functions and learned coordinate updates. The second core contribution is an architecture–loss co-design: sparsity-aware objectives, including Dice loss, per-graph epitope-count regularization, bipartite edge prediction, and auxiliary inter-chain distance classification, are developed to handle severe class imbalance and data scarcity. Validation is extensive. On the AsEP benchmark (1,721 complexes), EpiFormer reports 0.924 AUC / 0.482 F1 on the epitope-ratio split and 0.826 AUC / 0.305 F1 on the harder epitope-group split, outperforming 19 baselines spanning dedicated epitope predictors, binding-site predictors, docking/interaction models, structure/affinity predictors, and molecular property models.

Critical strengths
  1. The early-fusion framework is convincingly shown to be backbone-agnostic and co-designed with the objective. The strongest evidence is Table 2: replacing EGNN-R with a plain GCN while keeping interleaved cross-attention and the joint losses still reaches 0.447 F1, already surpassing all 19 baselines. This isolates the framework rather than the equivariant backbone as the primary driver. The architecture–loss co-design analysis (Table 3 and Appendix F.4) further shows that the auxiliary losses alone on late-fusion baselines reach only 0.332 F1, EpiFormer with plain BCE reaches 0.336 F1, and the two together reach 0.482 F1.

  2. The paper goes beyond the two official AsEP splits to a stricter family-based split, external zero-shot and LODO transfer, antigen donor-conformation substitution, injected correlated backbone error, and PLM-feature removal controls. The family-based split is particularly informative: early fusion holds a 0.211 F1 and 0.232 MCC advantage over an otherwise identical late-fusion model when antigen families and antibody lineages are removed.

Critical weaknesses
  1. Appendix A.4 reports that the official splits share antigens and antibody lineages, with median test-to-train antigen identity 1.00 and 0.95. The main table therefore measures performance under considerable similarity to training data. The authors deserve credit for constructing a family-based split and reporting EpiFormer’s 0.866 AUC / 0.407 F1 there, but that table compares only EpiFormer against an internal late-fusion control, not against the 19 baselines. Since the paper’s central claim is state-of-the-art antibody-aware epitope prediction, the absence of a full baseline comparison under the stricter family-based split is a consequential gap: it leaves open whether the margin over MIPE, WALLE, and others persists when antigen families and antibody lineages are truly held out.

  2. The structural perturbation experiments (Appendix E) perturb the antigen: rigid transforms, donor antigen conformations, and injected backbone error. The antibody chains are always extracted from solved bound complexes. In practice, a user may have an antigen structure but only a predicted or unbound antibody structure, and CDR loops are flexible. The paper does not test EpiFormer with predicted antibody structures, unbound antibody conformations, or injected CDR backbone error.

Questions for authors
  1. Can the authors report the main baselines (at least MIPE, WALLE, DiscoTope3, PECAN, and Boltz-2) under the family-based split used in Appendix E.4? The current family-split table compares only EpiFormer against an internal late-fusion control. A full baseline comparison would materially clarify whether the claimed advantage survives when antigen families and antibody lineages are removed.

  2. The sparsity regularizer enforces epitope-count matching during training, but at inference the ground-truth epitope count is unavailable. How is the decision threshold chosen for the threshold-dependent metrics? Is it fixed at 0.5 for all methods, tuned on validation data, or calibrated using the predicted count? If the threshold is selected per method or per complex, please state the protocol. This directly affects the interpretation of the large F1 and MCC gains.

  3. Can the authors evaluate EpiFormer with predicted antibody structures (e.g., from NanoBodyBuilder2, IgFold, or AlphaFold) or with injected correlated error in the antibody backbone/CDR loops?

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

This paper proposes EpiFormer, an antibody-aware epitope prediction framework that integrates interleaved cross-attention into geometric GNN encoders to enable early antigen–antibody information exchange. It also introduces EGNN-R and a joint objective combining multiple sparsity-aware and geometric supervision terms. Extensive experiments demonstrate substantial improvements over existing baselines.

The paper presents a clear motivation and well-defined contributions, supported by extensive experiments and analyses. However, the second main claim regarding Architecture–Loss Co-Design is not sufficiently supported by the current evidence. Given the generality of this claim, the authors should either provide further validation across different architectures or narrow the scope of their conclusion.

Critical strengths

S1. The paper presents a clear motivation, and the proposed architecture and training objectives are well aligned with the identified challenges in antibody-aware epitope prediction.

S2. EpiFormer achieves substantial improvements over existing methods on both AsEP evaluation splits, with comparisons against 19 baselines covering diverse architectures.

S3. The paper provides extensive ablation studies, cross-dataset evaluations, and robustness analyses. In particular, the additional family-based split helps assess generalization beyond the highly similar training and test samples in the original benchmark.

Critical weaknesses

W1. The paper makes a strong and general claim that sparsity-aware objectives are effective only with early-fusion architectures. However, Table 3 compares different architectures and loss configurations, making it difficult to establish this dependency. Notably, WALLE improves by +0.129 F1 without early fusion. Given the generality of this claim, further validation on other baselines with controlled early/late fusion and loss configurations, is needed.

W2.The main components largely build on existing techniques:

  • Early Fusion: Inserting cross-attention into each GNN encoder block.
  • EGNN-R: Combining EGNN with multi-relational message passing.
  • Joint Objective: Combining existing BCE, Dice, sparsity, edge, and distance losses. Although effective, the contribution appears closer to an integration of existing techniques than to novel modeling mechanisms.

W3. Frame randomization experiment needs more clarification. Appendix E.1 attributes the performance degradation under per-chain frame randomization to the loss of inter-chain orientation information. However, the randomization procedure and affected geometric features are not clearly specified. Since the features in Appendix B.1 are largely defined in local coordinate frames, the degradation might instead result from disrupted intra-chain geometry. Further clarification is needed to support the authors' interpretation.

Questions for authors

Please see Weakness.

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

Related papers

Discussion (0)

Sign in to comment.