Research map / Probabilistic and scientific modeling / Molecules and proteins
AsEP: Benchmarking Deep Learning Methods for Antibody-specific Epitope Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the AsEP antibody-antigen epitope dataset with epitope-group splits and the WALLE hybrid model (PLM embeddings + GNN, bipartite link prediction), benchmarking it against general binding-site predictors.
Problem
Antibody-specific epitope prediction lacks a large, consistent benchmark, and general protein binding-site predictors perform poorly on it.
Approach
Reformulates epitope prediction as bipartite link prediction between antibody CDR graph and antigen surface graph, and combines pretrained protein language model embeddings (AntiBERTy, ESM2) with GNN encoders (GCN/GAT/GraphSAGE) and an edge decoder in WALLE.
Key results
- WALLE reaches MCC 0.305 on the epitope-ratio split versus 0.037 for the best baseline (MaSIF-site) and 0.029 for EpiPred.
- On the epitope-group split WALLE achieves MCC 0.152 versus 0.046 for MaSIF-site, but its performance drops considerably from the ratio split, indicating poor generalization to unseen epitopes.
- AlphaFold2-Multimer achieves MCC 0.262 on a 76-complex post-cutoff subset but averages 1.66 hours per pair.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WALLE instead of the strongest baseline, MaSIF-site | Epitope prediction · AsEP ratio split test | MCC | Helped |
| Use WALLE instead of the strongest baseline, MaSIF-site | Epitope prediction · AsEP epitope-group split test | MCC | Helped |
| Evaluate on an epitope-group split rather than a ratio split | WALLE · AsEP | MCC: 0.305 -> 0.152 | Hurt |
| Add graph neural network layers instead of fully connected layers (ablation) | WALLE · AsEP | MCC: approximately 39.8 % | Helped |
| Use protein language model embeddings instead of one-hot or BLOSUM62 embeddings (ablation) | WALLE · AsEP | MCC: over 62.4 % | Helped |
| Use AntiBERTy for antibody embeddings and ESM2 for antigen embeddings instead of ESM2 for both (ablation) | WALLE · AsEP | MCC | Helped |
Surprising findings
- General protein binding-site predictors and sequence-based BepiPred-3.0 perform poorly on antibody-specific epitope prediction, with MCC near zero for most baselines.
- WALLE's performance degrades substantially when evaluated on unseen epitope groups.
Limitations
- Performance degrades considerably on the epitope-group split, showing bias toward seen epitopes.
- Dataset is restricted to conventional VH/VL antibodies; nanobodies and other formats are excluded.
- Uses only a simple distance-matrix edge feature for structure representation.
Experimental setup
Datasets: AbDb (Antibody Database, 2022/09/26 snapshot) AsEP
Benchmarks: AsEP epitope ratio split AsEP epitope group split AsEP AlphaFold2.3-Multimer post-cutoff subset
Metrics: MCC precision recall AUCROC F1
Baselines: EpiPred ESMFold ESMBind MaSIF-site AlphaFold2-Multimer BepiPred-3.0
Builds on: PECAN EPMP ESM2 MaSIF
Closest papers by meaning
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.86)
- Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization (AAAI 2025 · similarity 0.86)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.86)
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.85)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.85)
- Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.