Can predicted protein shapes help find where antibodies bind?
Feeding AlphaFold2-predicted 3D structures into a graph neural network made antibody-binding-site prediction clearly better than sequence-only and earlier structure-based tools.
Source
Identifying B-cell epitopes using AlphaFold2 predicted structures and pretrained language model
Study at a glance
- Design
- Computational / modelling — Supervised residue-level classifier trained on antigen chains from the Protein Data Bank; 10-fold cross-validation plus a time-split independent test set of antigens deposited after January 2021.
- N
- N=633 · 633 non-redundant antigen sequences: 577 for training and 56 for the independent test (the test set holds 1393 epitope residues).
- Population
- Antigen protein chains from antibody-antigen complexes in the Protein Data Bank
- Outcome
- Per-residue epitope prediction scored by AUC, AUPR, F1 and MCC
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built GraphBepi, which takes only an antigen's amino-acid sequence, predicts its 3D shape with AlphaFold2, and turns that shape into a graph of residues linked by spatial and sequence neighbours. An edge-enhanced graph neural network reads the graph while a bidirectional LSTM reads the sequence (represented by the ESM-2 protein language model); the two are combined to score each residue as epitope or not. They trained on 577 antigens and tested on 56 later-deposited antigens, comparing against two sequence-based and four structure-based published tools and running ablations.
What they found
On the independent test GraphBepi reached an AUC of 0.751 and AUPR of 0.261, beating the next-best tool (ScanNet with transfer learning) by 44.0% in AUPR. Removing the graph module caused the biggest drop, and the advantage over a transformer baseline grew for residues with many long-range contacts. Using real experimental structures instead of AlphaFold2 predictions barely changed results, but antigens with poorly predicted structures got much worse F1 (0.241 versus 0.406 for the best-predicted fifth).
The limits
What it doesn't show
Absolute performance is still modest: an AUPR around a quarter means most predicted epitope residues are wrong, so the tool narrows the search rather than replacing experiments. The test set is small (56 antigens) and competitors were run with default settings rather than retrained on the same data. The authors note the model depends on AlphaFold2 quality and is a black box, and features were chosen using cross-validation results, which can slightly flatter the final design.
Key terms
- B-cell epitope
- The patch of an antigen's surface that an antibody physically binds to.
- Conformational epitope
- An epitope made of residues that are far apart in the sequence but brought together when the protein folds.
- Graph neural network
- A neural network that updates each node's representation using its neighbours in a graph, here residues linked by distance or sequence position.
- Protein language model
- A transformer trained on millions of protein sequences by predicting masked amino acids, giving each residue a learned feature vector.
- AUPR
- Area under the precision-recall curve; more informative than AUC when positives (epitope residues) are rare.
- Ablation study
- Removing one component of a model at a time to measure how much it contributes.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What does GraphBepi use to obtain 3D structural information?
Common questions
Why not just use the experimental structure?
Most antigens have no solved structure. The paper shows AlphaFold2 predictions work almost as well as real structures, so the method can be used from sequence alone.
Why report AUPR as well as AUC?
Epitope residues are only a small minority, so a model can get a decent AUC while still producing many false positives; AUPR exposes that.
What does the ablation tell us?
Dropping the graph module hurt performance the most, which suggests the spatial neighbourhood information, not just the sequence features, drives the gains.
More on Graph neural networks