EPIC: Joint Epitope, Paratope, and Residue Contact Prediction | Clickmab Research Accepted at the NeurIPS 2026 AI4DD Workshop

NeurIPS is one of the world's leading international conferences on artificial intelligence and machine learning. AI4DD (AI for Drug Discovery) is a workshop held as part of NeurIPS 2026. It focuses on new AI methods for drug discovery, validation of their reliability, and translation into real-world R&D, bringing AI researchers and drug discovery practitioners together.

In antibody R&D, we can often obtain the amino acid sequences of candidate antibodies and target proteins, yet reliable information about their interactions is frequently missing.

Which region of the target does the antibody bind?
Which antibody residues are involved in recognition?
Which amino acids on the two molecules are close to each other?

EPIC brings three closely related prediction tasks into a single model. Epitope, paratope, and residue contact information reinforce one another, enabling a more comprehensive prediction of the antibody–antigen binding interface.

E. Epitope
Predicts the region of the antigen bound by an antibody, informing epitope-specific antibody design.

P. Paratope
Identifies key antibody residues involved in binding to support affinity maturation and targeted sequence optimization.

C. Contact
Predicts residue contacts between the antigen and antibody, providing potential constraints for complex structure modeling.

EPIC takes antibody and antigen amino acid sequences as its initial input. It extracts sequence features using protein language models and obtains three-dimensional geometry from predicted monomer structures, enabling binding interface prediction without requiring an experimentally determined complex structure.

The model has three key design innovations:

1. Learning Sequence Features and Three-Dimensional Relationships Together
It combines both types of information, considering amino acid sequence features alongside the spatial arrangement of residues in protein structures.

2. Enabling Bidirectional Antibody–Antigen Information Exchange Around Candidate Contacts
By predicting potential residue contacts, it allows antigen and antibody representations to inform one another, improving the identification of potential binding regions for a specific pair.

3. Sharing Interface Information Across Contact, Epitope, and Paratope Prediction
It unifies the three tasks in a single model so that each prediction provides clues for the others, yielding a more comprehensive prediction of the antibody–antigen binding interface.

Using a test set from the public antibody structure database SAbDab2, the team systematically compared EPIC with several existing prediction methods on 455 antibody–antigen complexes.

EPIC showed strong discrimination in residue contact prediction, identified more true binding sites in antigen epitope prediction, and maintained performance comparable to a dedicated model for antibody paratope prediction. These results demonstrate the potential of joint modeling for antibody–antigen interface prediction.

1. Residue Contact Prediction: More Accurately Identifying Where Contacts May Occur
EPIC achieved an AUROC of 0.9531 for distinguishing contacting from non-contacting antibody–antigen residue pairs, substantially higher than the 0.7374 achieved by the comparison method ImaPEp. An AUROC closer to 1 indicates better discrimination. This suggests that EPIC can identify and rank potential antibody–antigen contact positions more effectively.

2. Antigen Epitope Prediction: Finding More True Binding Sites
An antigen epitope comprises the key sites on an antigen recognized and bound by an antibody. For this task, EPIC achieved a recall of 60.97% for true epitope residues, compared with 55.23% for MIPE.

On the same test set, EPIC therefore identified a higher proportion of true epitope residues, providing more clues for locating potential antigen-binding regions.

3. Antibody Paratope Prediction: Performance Comparable to a Dedicated Model
An antibody paratope is the key region of an antibody directly involved in antigen binding. EPIC achieved an AUROC of 0.9540, comparable to 0.9514 for the dedicated prediction model Paraplume, with similar recall. This shows that, even while handling all three tasks, EPIC remains competitive in identifying and distinguishing antibody paratope residues.

Predicting the antibody–antigen binding interface is an important foundation for precise antibody design and engineering.

EPIC's capabilities could be incorporated into Clickmab's Click.mAb. AI antibody design platform to support several stages of R&D:

1. Epitope-Specific De Novo Antibody Design
Generate and screen candidate antibodies for a target epitope.

2. Complex Structure Prediction
Use contact information to guide complex structure modeling.

3. Antibody Affinity Maturation
Identify key residues to guide the selection of mutation sites.

4. CDR Diversity Engineering
Identify regions for modification and design more focused sequence libraries.

Paper: EPIC: Epitope–Paratope Interleaved Contact Learning for Antibody–Antigen Interface Prediction

Accepted at: NeurIPS 2026 Workshop on AI for Drug Discovery (AI4DD)

Clickmab will continue integrating interface prediction, antibody generation, and sequence optimization to further improve the accuracy and efficiency of AI antibody design.