EpiFormer

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

Mansoor Ahmed^1,2,*, Huirong Chai¹, Haoxin Wang¹, Hemanth Venkateswara¹, Murray Patterson^1,*

¹Georgia State University ²Georgia Institute of Technology

^*Corresponding authors

Paper arXiv Code BibTeX

Where the antibody enters. Three ways to predict an epitope, unrolled over the layers of a graph network from the input residues (bottom) to the predicted epitope (top). The color of a node shows whose information its representation carries. (a) An antigen-only predictor does not see the antibody and returns the same answer for any antibody. (b) Late fusion encodes the two chains apart and mixes them once, after the last layer. (c) EpiFormer interleaves cross-attention with message passing, and the antigen and the antibody therefore shape each other's representation at each layer.

TL;DRAn epitope depends on the antibody that binds it, yet antibody-aware predictors encode the antigen and the antibody separately and fuse them only at the end. EpiFormer lets the two chains attend to each other inside each layer of a geometric graph encoder and trains with objectives built for sparse labels. On the AsEP benchmark, it raises F1 from 0.337 to 0.482 over the strongest of 19 baselines. On the harder epitope-group split, whose test epitopes are absent from training, it raises F1 from 0.254 to 0.305.

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.

The epitope depends on the antibody

An epitope is the region of an antigen's surface that a given antibody binds. The AsEP benchmark labels an antigen surface residue as epitope when it lies within 4.5 Å of the antibody, and asks a model to predict these labels from the structures of the two chains. Different antibodies bind different regions of the same antigen. Hence, a predictor that scores the antigen alone can at best find surface that tends to be bound by some antibody, not the epitope of the antibody at hand.

Antibody-aware methods condition on the antibody, but they encode the two chains independently and combine them only after encoding, a design we call late fusion. The structural features that define a binding interface depend on both chains, since the complementarity-determining regions (CDRs) of the antibody adapt their conformation to the epitope. The labels are also sparse. In the median complex of the epitope-ratio test split, 9% of the antigen surface residues belong to the epitope, and AsEP holds 1,721 complexes in total.

Late fusion encodes each chain without knowledge of the other. EpiFormer lets the two chains exchange information in each encoder block and pairs this early fusion with objectives designed for sparse labels.

Interleaved cross-attention in a geometric encoder

The EpiFormer architecture on the test complex 7KFY — **EpiFormer on a held-out complex (PDB 7KFY).** The surface residues of the antigen and the CDR residues of the antibody become residue graphs with four edge relations. In each encoder block, EGNN-R passes messages within each chain while cross-attention exchanges information between the chains, and learned gates, drawn to scale, add it back into each stream. The decoder scores each pair of an antigen and an antibody residue, the mean of the two largest scores of an antigen residue is its epitope probability, and training compares these outputs with the true epitope and contacts. The attention weights, the gates, the interaction map, and the prediction are the released model's outputs on this complex.

We represent each chain as a residue graph with one node per residue, placed at its C$\alpha$ atom. Each node carries geometric features that encode its residue type and the local geometry of its backbone. Edges carry four relations, namely sequential neighbors ($\rho_1$), short-range sequence coupling ($\rho_2$), $K$-nearest neighbors ($\rho_3$), and contacts within 8 Å ($\rho_4$). EpiFormer takes the antigen graph $\mathcal G_{\text{ag}}$ and the antibody graph $\mathcal G_{\text{ab}}$ and returns, for each antigen residue, the probability that it belongs to the epitope.

Encoder. Two parallel encoders process the antigen and the antibody. Within each block, a relation-aware E(3)-equivariant graph neural network (EGNN-R) passes messages along the edges of one chain, with a separate message function for each relation, and updates both the residue features and their coordinates,

$$m_{ij}^{\rho}=\phi_m^{\rho}\big(\mathbf h_i^{\ell},\mathbf h_j^{\ell},\gamma(d_{ij}),\mathbf f_{ij}\big),\qquad \mathbf h_i^{\ell+1}=\mathbf h_i^{\ell}+\phi_h\Big(\mathbf h_i^{\ell},\sum_{j\in\mathcal N(i)}\sum_{\rho\in\mathbf r_{ij}}m_{ij}^{\rho}\Big),\qquad \mathbf x_i^{\ell+1}=\mathbf x_i^{\ell}+\sum_{j\in\mathcal N(i)}\sum_{\rho\in\mathbf r_{ij}}\frac{\boldsymbol\delta_{ij}}{\sqrt{d_{ij}+\varepsilon}}\,\phi_x^{\rho}\big(m_{ij}^{\rho}\big),$$

where $d_{ij}$ is the squared distance between residues $i$ and $j$, $\boldsymbol\delta_{ij}$ their displacement, $\mathbf f_{ij}$ the edge features, and $\gamma$ a radial basis expansion. In parallel, multi-head cross-attention (MHCA) lets each chain attend to the residues of the other, and a learnable gate $\alpha_c$ balances the geometric update $\mathbf H_c^{\text{intra}}$ with the cross-chain context,

$$\mathbf H_c^{\ell+1}=\mathbf H_c^{\ell}+\mathbf H_c^{\text{intra}}+\alpha_c\,\operatorname{FFN}\big(\operatorname{MHCA}(\mathbf H_c^{\ell},\mathbf H_{\bar c}^{\ell})\big),\qquad c\in\{\text{ag},\text{ab}\},$$

where $\bar c$ denotes the other chain and FFN is a two-layer feed-forward network.

Decoder. The decoder refines the two embeddings with $J$ further layers of bidirectional cross-attention and then scores each pair of an antigen and an antibody residue in both directions,

$$\mathbf S_{\text{ag}\to\text{ab}}=\frac{(\mathbf H_{\text{ag}}^{J}\mathbf W_Q)(\mathbf H_{\text{ab}}^{J}\mathbf W_K)^{\top}}{\sqrt{d_k}},\qquad \mathbf S_{\text{ab}\to\text{ag}}=\frac{(\mathbf H_{\text{ab}}^{J}\mathbf W'_Q)(\mathbf H_{\text{ag}}^{J}\mathbf W'_K)^{\top}}{\sqrt{d_k}}.$$

A learnable combination of the two score maps yields the interaction map $\hat{\mathcal E}_{\text{bg}}=\sigma(\mathbf Z)\in[0,1]^{n\times m}$ between the $n$ antigen and $m$ antibody residues, and the mean of the two largest entries in each row yields the epitope probability of each antigen residue,

$$(\hat y_{\text{ag}})_i=\frac{1}{2}\sum_{j\in\operatorname{top}_2(i)}\big(\hat{\mathcal E}_{\text{bg}}\big)_{ij},$$

where $\operatorname{top}_2(i)$ holds the two antibody residues with the highest interaction probability for antigen residue $i$.

Joint objective. We train EpiFormer with a joint objective. The node term combines a class-reweighted binary cross-entropy, a Dice loss that measures the overlap between the predicted and the true epitope, and a sparsity term that matches the predicted epitope size to the true one. Two auxiliary terms supervise the interaction map with the true contacts and classify inter-chain distances into four bins,

$$\mathcal L=\lambda_{\text{node}}\big(\beta_{\text{BCE}}\mathcal L_{\text{BCE}}+\beta_{\text{Dice}}\mathcal L_{\text{Dice}}+\beta_{\text{sparsity}}\mathcal L_{\text{sparsity}}\big)+\lambda_{\text{edge}}\mathcal L_{\text{edge}}+\lambda_{\text{geo}}\mathcal L_{\text{geo}}.$$

Early fusion

The antigen and the antibody exchange information in each encoder block, rather than only in the decoder.

Backbone-agnostic

With a plain GCN in place of EGNN-R, the framework still reaches 0.447 F1 and outperforms all 19 baselines.

Built for sparse labels

The sparsity term alone raises F1 from 0.329 to 0.461, since it keeps the model from labeling most residues as epitope.

The loss terms

With $y_i$ the label of antigen residue $i$, $\hat y_i$ its predicted probability and $\pi_{\text{epi}}>1$ a weight on the positive class, the node terms are

$$\mathcal L_{\text{BCE}}=-\frac{1}{n}\sum_{i=1}^{n}\big[\pi_{\text{epi}}\,y_i\log\hat y_i+(1-y_i)\log(1-\hat y_i)\big],\qquad \mathcal L_{\text{Dice}}=1-\frac{2\sum_i\hat y_iy_i+\epsilon}{\sum_i\hat y_i+\sum_iy_i+\epsilon},\qquad \mathcal L_{\text{sparsity}}=\big\|\hat{\mathbf y}-\mathbf y\big\|_1.$$

The edge term applies a positive-class-reweighted binary cross-entropy to all antigen-antibody residue pairs, with $E_{ij}=1$ when the two residues lie within 4.5 Å,

$$\mathcal L_{\text{edge}}=-\frac{1}{nm}\sum_{i,j}\big[\pi_{\text{edge}}E_{ij}\log\hat E_{ij}+(1-E_{ij})\log(1-\hat E_{ij})\big].$$

The geometric term $\mathcal L_{\text{geo}}$ classifies the distance between near-contact residue pairs into the bins 0–4, 4–8, 8–16, and 16–32 Å with a class-balanced cross-entropy. Distances enter only as training targets and do not reach the model at inference.

Results on AsEP

We compare EpiFormer with 19 baselines from five families, namely epitope and paratope predictors, protein binding-site predictors, protein interaction and docking models, structure and affinity predictors, and a molecular property model. We retrained each baseline on AsEP with its published configuration. AsEP comes with two splits. The epitope-ratio split balances the fraction of epitope residues across the partitions, and the epitope-group split keeps the test epitopes out of training.

Method	Epitope-ratio split			Epitope-group split
Method	AUC	F1	MCC	AUC	F1	MCC
Epitope and paratope prediction
EpiGraph	.790	.064	.089	.779	.056	.096
MIPE	.827	.337	.356	.740	.172	.206
WALLE	.808	.203	.208	.713	.170	.143
DiscoTope3	.821	.251	.247	.763	.210	.189
GraphBepi	.783	.079	.118	.781	.144	.161
PECAN	.740	.229	.182	.729	.254	.195
Protein binding-site prediction
EquiPocket	.761	.202	.183	.711	.194	.159
Protein interaction and docking
ATProt	.798	.252	.237	.710	.200	.160
DiffDock-PP	.735	.192	.169	.650	.165	.120
Structure and affinity prediction
Boltz-2	.798	.255	.232	.702	.220	.156
AlphaFold3 (no MSA)	.562	.153	.071	.570	.155	.079
GearBind	.799	.244	.236	.734	.206	.169
Molecular property prediction
EquiformerV2	.827	.260	.262	.729	.212	.169
EpiFormer (ours)	.924	.482	.464	.826	.305	.290

AUC is the area under the ROC curve, and F1 and the Matthews correlation coefficient (MCC) use a threshold of 0.3. Epitope-ratio values are means over three seeds. Bold marks the best value in each column and underline the second best. The paper reports all 19 baselines and three further metrics.

EpiFormer leads on both splits. On the epitope-ratio split, it improves F1 from 0.337 to 0.482 and MCC from 0.356 to 0.464 over MIPE, the strongest baseline. Models from neighboring tasks such as Boltz-2, GearBind, and EquiformerV2 reach competitive AUC but low F1, because they have no mechanism for the class imbalance of epitope prediction.

Per-complex AUROC, AUPRC, F1, MCC, precision and recall of EpiFormer, MIPE and WALLE on the epitope-ratio test split — **Per-complex metrics.** Distributions of six metrics over the 170 complexes of the epitope-ratio test split for EpiFormer, MIPE, and WALLE.

Other datasets and inputs

We also evaluated EpiFormer on three external benchmarks without retraining. Its AUC averages 0.786 across SAbDab, CoV-AbDab, and ANABAG, against 0.688 for MIPE. Fine-tuning on the two other external datasets (LODO, for leave-one-dataset-out) raises it further.

Dataset	Method	Zero-shot		LODO
Dataset	Method	AUC	F1	AUC	F1
SAbDab (494)	EpiFormer	.782	.297	.858	.364
	MIPE	.614	.140	.746	.190
CoV-AbDab (170)	EpiFormer	.819	.358	.890	.359
	MIPE	.730	.175	.813	.212
ANABAG (499)	EpiFormer	.758	.301	.837	.367
	MIPE	.719	.220	.772	.217

The number of complexes is given in parentheses. Bold marks the better model on each dataset.

Before a complex is solved, a practitioner has at best a predicted antigen structure or one bound by a different antibody. Replacing the antigen with the conformation it adopts against a different antibody costs 0.010 AUC on the epitope-ratio split, and 2 Å of correlated backbone error costs 0.007 AUC. On a stricter family-based split, which keeps antigen families and antibody lineages apart, EpiFormer reaches 0.407 F1, while the same model with late fusion only reaches 0.196.

Predictions on held-out complexes

We ran the released EpiFormer checkpoints on the test complexes of AsEP. Each video shows the antigen surface twice, colored by the true epitope on the left and by the probability that EpiFormer assigns to each residue on the right. The variable domains of the antibody appear as a cartoon ( heavy chain, light chain), and lift off midway through each loop to expose the epitope.

SARS-CoV-2 spike receptor-binding domain with antibody C1A-F10 (PDB 7KFY)

The receptor-binding domain (RBD) of the SARS-CoV-2 spike, bound by the human antibody C1A-F10. EpiFormer assigns a probability above 0.3 to 30 of the 32 epitope residues and to 14 of the other 145 surface residues.

F1 0.79MCC 0.75AUC 0.98epitope-ratio test split

Hen egg-white lysozyme with antibody HyHEL-63 (PDB 1DQJ)

Lysozyme is a model antigen of structural immunology, here bound by the mouse antibody HyHEL-63. EpiFormer recovers all 21 epitope residues and assigns a probability above 0.3 to 10 of the other 94 surface residues.

F1 0.81MCC 0.78AUC 0.99epitope-ratio test split

SARS-CoV-2 spike glycoprotein with antibody C105 (PDB 6XCM)

One protomer of the spike, with 898 surface residues, of which 25 form the epitope of the human antibody C105 on the receptor-binding domain. EpiFormer assigns a probability above 0.3 to 18 of these 25 residues and to 7 of the other 873.

F1 0.72MCC 0.71AUC 0.99epitope-ratio test split

Cadherin-6 with an anti-CDH6 Fab (PDB 5VEB)

A complex from the epitope-group split, whose test epitopes are absent from training. Its antigen, the fifth extracellular domain of cadherin-6, shares at most 28% sequence identity with any training antigen. EpiFormer recovers 24 of the 25 epitope residues and assigns a probability above 0.3 to 24 of the other 74 surface residues.

F1 0.66MCC 0.55AUC 0.90epitope-group test split

These four complexes rank among the 15 best of their test split by F1. Across all 170 complexes of the epitope-ratio test split, the released checkpoint reaches a median F1 of 0.51, and across the 169 complexes of the epitope-group split a median of 0.36.

What drives the gains

The framework, more than the backbone

We replaced EGNN-R with other graph networks while keeping the cross-attention, the decoder, and the objective fixed. Each backbone reaches between 0.408 and 0.482 F1, and even a plain GCN outperforms all 19 baselines. Equivariant message passing, multi-relational edges, and learned coordinate updates add smaller but consistent gains on top.

Encoder	Equivariant	Multi-relational	Coordinate updates	AUC	AUPRC	F1	MCC
GCN	no	no	no	.905	.397	.447	.424
GIN	no	no	no	.890	.349	.408	.369
GAT	no	no	no	.908	.399	.452	.424
RGCN	no	yes	no	.910	.405	.454	.422
EGNN	yes	no	yes	.912	.421	.470	.455
EGNN-R, coordinates frozen	partly	yes	no	.908	.412	.465	.441
EGNN-R (ours)	yes	yes	yes	.924	.493	.482	.464

Epitope-ratio split. All encoders share the cross-attention decoder and the joint objective.

Architecture and objective, designed together

The sparsity-aware objective helps only when the two chains are fused early. Added to six baselines, it raises F1 for WALLE, EquiformerV2, and DiscoTope3 but lowers it for PECAN, MIPE, and CheapNet. EpiFormer with a plain cross-entropy loss reaches 0.336 F1, and our objective in the best late-fusion baseline reaches 0.332, while the two together reach 0.482.

Geometric features over language model embeddings

EpiFormer reaches 0.924 AUC with geometric features and 0.889 with embeddings from protein language models (PLMs). The baselines move the other way, since removing their PLM embeddings costs DiscoTope3 0.133 AUC, WALLE 0.122, and MIPE 0.088. Epitope residues are not evolutionarily conserved, which limits the value of the conservation signal that PLMs encode.

An asymmetric exchange between the chains

The cross-attention gates are learned without supervision. With the gates initialized at 0 or 0.05, the antibody gate exceeds the antigen gate in all four encoder blocks. Hence, the antibody stream draws more on the antigen than the antigen stream draws on the antibody. This direction matches the roles of the two chains, since the CDR loops of the antibody adapt to a given epitope. Initializing the gates at zero instead of 0.05 lowers F1 from 0.496 to 0.454.

BibTeX

@article{ahmed2026epiformer,
  title   = {EpiFormer: Learning Antigen-Antibody Interactions for Epitope Prediction via Geometric Deep Learning},
  author  = {Ahmed, Mansoor and Chai, Huirong and Wang, Haoxin and Venkateswara, Hemanth and Patterson, Murray},
  journal = {arXiv preprint},
  year    = {2026}
}