Bcell Epitope prediction ans
B-cell epitope prediction
Analysis Tips and conventions
- Experimental epitope definition: An HA amino-acid residue belongs to the antibody-binding footprint if any of its non-hydrogen protein atoms is strictly less than 5.0 Å from a non-hydrogen protein atom of the selected antibody. Exclude waters and other non-protein molecules. Count a residue only once, even if several of its atoms contact the antibody.
- One interface at a time: For 7K39, use antigen chains A/B and one Fab (G/H); do not combine all three equivalent antibody copies.
- Residue numbering: The chain identifier and PDB residue number are part of the residue identity. For the supplied 7K39 BepiPred outputs, chain A sequence position = PDB residue number + 16; chain B sequence position = PDB residue number. This was confirmed by comparing the amino-acid sequences.
- Prediction evaluation: Predictions were compared on the 490 resolved amino acids in 7K39 HA chains A/B (318 in A and 172 in B). The default scores/flags were used unless stated otherwise.
- Important distinction: A residue outside one antibody's footprint is not necessarily a biologically impossible epitope; a different antibody could bind there. Thus “false positive” below means “outside the experimentally observed interface being tested.”
Part A. Get your bearings in PyMOL (experimental epitopes)
Q1. List chain IDs for antigen versus antibody in each complex
| Experimental structure | Antigen HA1 | Antigen HA2 | Antibody heavy chain | Antibody light chain |
|---|---|---|---|---|
| 4FQI (H5N1, CR9114) | A | B | H | L |
| 7K39 (H3N2, head-binding Fab) | A, C, E | B, D, F | H, J, L | G, I, K |
For the 7K39 calculations below use A/B as the antigen and G/H as a single Fab (G is light, H is heavy).
Q2. Show the epitopes, count residues, and describe secondary structure
Using the experimental PDB coordinates, excluding waters, and applying the 5 Å cutoff
| Complex | HA1 (A) contact residues | HA2 (B) contact residues | Total | Qualitative structural location |
|---|---|---|---|---|
| 4FQI, stem-binding CR9114 | 4 | 14 | 18 | Mainly at the HA stem, including an HA2 helical region and adjacent loops/turns. |
| 7K39, head-binding Fab | 14 | 1 | 15 | Mainly on HA1 in the head, distributed among accessible loops and neighboring structural elements. |
Figure placeholders: Upload the two PyMOL images to MediaWiki before enabling the lines below.
To reproduce the selection in PyMOL, for 4FQI:
fetch 4fqi remove solvent select Ag, (chain A+B) select Fab, (chain H+L) select epitope, byres (Fab around 5) color red, epitope
For 7K39, replace the Fab selection with chain G+H
Q3. Which residues form the epitopes? Are they linear? Do they have a linear core?
4FQI (18 contacts):
- HA1, A: 291, 292, 293, 318.
- HA2, B: 18, 19, 20, 21, 36, 37, 38, 41, 42, 45, 46, 48, 49, 52, 56.
7K39 (15 contacts):
- HA1, A: 48, 49, 50, 90, 91, 92, 93, 269, 270, 271, 272, 273, 284, 285.
- HA2, B: 69.
Interpretation: Both are predominantly conformational (discontinuous) epitopes, because the contacting residues occur in sequence-distant regions that come together in three dimensions. They contain short contiguous stretches or potential linear cores, for example 4FQI B:18–21 and B:36–38, and 7K39 A:90–93 and A:269–273. These patches do not make the overall epitopes linear or show that the corresponding isolated peptides would bind the antibody.
The antibody complexes do not target the same region: CR9114(4FQI) binds the stem (involving HA1 and HA2), while the 7K39 antibody binds primarily the head (mostly HA1). HA1=head and HA2=stem are useful teaching approximations, but HA1 also contributes to the structural stem.
Part B. Sequence-based prediction: BepiPred-2.0 and BepiPred-3.0
Data limitation: The provided BepiPred-2.0 and BepiPred-3.0 CSV files contain predictions for the 7K39 H3N2 HA sequences. They were not run on the native 4FQI H5N1 HA sequences. The two structures contain homologous but non-identical antigens. The quantitative answers in Q4–Q5 therefore directly test the 7K39 antibody footprint only. To evaluate 4FQI independently, run both predictors on the 4FQI FASTA sequences.
The supplied prediction scores were thresholded at 0.50 (BepiPred-2.0) and 0.1512 (BepiPred-3.0 default, unsmoothed).
Q4. Do BepiPred-2.0 predictions overlap the actual epitopes?
Yes for 7K39, but the result is not very selective. All 15/15 experimentally observed 7K39 residues are predicted as epitope, giving 100% recall. However, 407/490 evaluated residues (83.1%) are predicted as positive, of which 392 are outside the particular observed antibody footprint. Precision is only 15/407 = 3.69%.
A BepiPred-2.0 score of 0.5 therefore recovers the experimental footprint by labeling nearly the whole antigen as a potential epitope. This is not good evidence that it can pinpoint where the specific 7K39 antibody binds. Its score-ranked average precision (AP) is 0.0306, approximately the random-ranking baseline 15/490 = 0.0306.
Q5. Does the BepiPred-3.0 prediction overlap now?
Only slightly for 7K39 at the default threshold. Using the unsmoothed BepiPred-3.0 output, 2/15 experimental contact residues are predicted (13.3% recall): A:92 and A:285. Their corresponding BepiPred sequence positions are 108 and 301. There are 128/490 predicted positive residues (26.1% coverage); precision is 2/128 = 1.56%, and AP = 0.0297.
The optionally smoothed BepiPred-3.0 score detects 0/15 experimental contacts at the same numerical threshold and labels 88/490 residues positive. Smoothing may emphasize contiguous patches but is not guaranteed to improve overlap with a discontinuous antibody footprint.
Conclusion for Q4–Q5: BepiPred-2.0 has higher recall at its default threshold but very poor localization, while BepiPred-3.0 makes fewer predictions but misses most of this particular antibody footprint. Neither is demonstrably good at ranking the 7K39 contact residues.
Q6. What is different in BepiPred-2.0 compared with BepiPred-3.0?
| Feature | BepiPred-2.0 | BepiPred-3.0 |
|---|---|---|
| Core method | Random Forest using sequence-derived features, trained with structural epitope annotations | Neural networks using protein language-model (ESM-2) representations, trained with updated structural epitope annotations |
| Input | Protein sequence | Protein sequence |
| Output focus | Mainly continuous/linear epitope prediction with smoothing | Residue-level prediction for potential linear and discontinuous epitope residues, with optional smoothing for linear epitope analysis |
| Default threshold used here | 0.50 | 0.1512 |
BepiPred-3.0 captures richer sequence context using learned embeddings, but neither tool receives the identity or structure of the specific antibody. Predictions are thus potential general B-cell epitope residues, not antibody-specific docking poses. The BepiPred-3.0 structural training annotation uses an antibody-contact cutoff of 4 Å, whereas this exercise uses 5 Å; this difference can also affect observed overlap.
See BepiPred-2.0 and BepiPred-3.0.
Part C. Structure-based prediction: DiscoTope-2.0 and DiscoTope-3.0
Data: The uploaded Discotope2pred.txt and file_A_discotope3.csv / file_B_discotope3.csv correspond to the 7K39 antigen A/B. At the supplied settings, DiscoTope-2.0 calls positive residues with combined score greater than approximately −3.7; DiscoTope-3.0 uses its calibrated score ≥ 0.90 (moderate confidence). The numerical scales are different and cannot be compared directly.
Input caveat: The DiscoTope-2.0 output includes other HA chains and Fab chains, indicating that the uploaded structure was not strictly antigen-only. DiscoTope-3.0 outputs were reported separately for HA chains A and B. The comparison is therefore useful for the supplied results, but not a fully controlled benchmark of the methods.
Q7. Does DiscoTope-2.0 overlap with the actual epitope?
No, not at the selected threshold for 7K39. DiscoTope-2.0 classifies 58/490 antigen residues as epitopes (34 in HA1 chain A; 24 in HA2 chain B). None of the 15 experimentally observed 7K39 contact residues is predicted positive, giving recall = 0% and precision = 0%. The ranking-based AP = 0.0274, slightly below the 0.0306 random-ranking baseline.
These findings are specific to the supplied output and settings; the method should ideally be rerun on the identical antigen-only structure supplied to the newer version.
Q8. Compare the results obtained using DiscoTope-2.0 and DiscoTope-3.0
| Measurement for 7K39 HA A/B | DiscoTope-2.0 | DiscoTope-3.0 |
|---|---|---|
| Model concept | Spatial epitope propensities plus residue exposure and neighborhood information | Learned structural representations, including inverse-folding features and exposure information |
| Threshold used | Combined score > −3.7 | Calibrated score ≥ 0.90 |
| Residues predicted positive | 58/490 | 77/490 |
| Experimental contact residues recovered | 0/15 | 4/15 |
| Recall | 0% | 26.7% |
| Precision | 0% | 5.19% |
| Average precision (AP) | 0.0274 | 0.0466 |
DiscoTope-3.0 correctly identifies experimental contacts A:91, A:93, A:271 and B:69. At the supplied cutoffs, the two versions share only 12 predicted-positive HA residues. DiscoTope-3.0 therefore agrees better with the known 7K39 footprint in the supplied results, although it still misses 11 of the 15 contacts and predicts 73 residues outside this one interface.
See DiscoTope-2.0 and DiscoTope-3.0.
Q9. Which prediction best matches the experimentally observed epitope?
To answer this, compare both recall and precision, rather than declaring any overlapping prediction correct.
Precision = TP / (TP + FP) Recall = TP / (TP + FN)
Here TP denotes predicted residues in the measured 5 Å antibody footprint; FP denotes predicted residues outside that specific footprint; and FN denotes measured contact residues missed by the method. A threshold-independent score-ranking summary is average precision (AP), calculated from a precision–recall curve. Its random-ranking baseline here is 15/490 = 3.06%.
| Method (7K39) | Correct contacts | All predicted positives | Recall | Precision | AP | ROC-AUC |
|---|---|---|---|---|---|---|
| BepiPred-2.0 | 15/15 | 407/490 | 100% | 3.69% | 0.0306 | 0.492 |
| BepiPred-3.0, unsmoothed | 2/15 | 128/490 | 13.3% | 1.56% | 0.0297 | 0.492 |
| DiscoTope-2.0 | 0/15 | 58/490 | 0% | 0% | 0.0274 | 0.413 |
| DiscoTope-3.0 | 4/15 | 77/490 | 26.7% | 5.19% | 0.0466 | 0.626 |
Best result for this experimental interface: DiscoTope-3.0, if “best” means highest precision and AP in the supplied data. BepiPred-2.0 has the highest recall, but almost no ability to discriminate the actual interface from the rest of HA at its default threshold. All AP values remain low in absolute terms. The table is not a general benchmark because only one measured antibody footprint has directly matched predictor outputs and the DiscoTope inputs were not identical.
Part D. Predicting where a known antibody binds (AlphaFold)
Q10. Are there tools that predict where an isolated antibody binds to an antigen? What do the provided AlphaFold models show?
Yes. Given antigen and antibody sequences, complex-modeling methods such as AlphaFold 3 can propose a three-dimensional antibody–antigen arrangement. Related approaches include Boltz and docking methods such as HADDOCK. Unlike BepiPred or DiscoTope, these methods use information about the specific antibody and aim to model its candidate binding pose. A predicted pose is a hypothesis, not proof of binding.
The provided ZIP archives contain five AlphaFold 3 models each for a hemagglutinin trimer plus antibody chains. They illustrate both successful and problematic complex prediction:
| Feature | 4FQI AlphaFold 3 archive | 7K39 AlphaFold 3 archive |
|---|---|---|
| Experimentally observed epitope | Stem | Head |
| Predicted binding location | Stem, consistent across the five models | Near the stem, not the experimental head site; the HA protomer contacted changes in some models |
| Example model 0 predicted 5 Å contact residues | 24 (8 in HA1 A, 16 in HA2 B) | 31 (12 in HA1 C, 19 in HA2 D) |
| Global complex ipTM, model 0 | 0.85 | 0.62 |
| Antibody–HA chain-pair ipTM, across supplied models | Approximately 0.49–0.70 | Only 0.08–0.13 |
| Interpretation | Plausible binding pose consistent with the known stem location | Poorly supported antibody–antigen arrangement, despite a moderately high global ipTM value |
Crucial caution: Global ipTM reflects multiple interfaces in a trimeric complex, including HA–HA interfaces. For determining whether the antibody is positioned correctly, examine antibody–antigen pairwise ipTM, interchain predicted aligned error (PAE), and the actual molecular contact geometry. In 7K39, the model-0 minimum antibody–HA pairwise PAE values are about 20–25 Å, indicating considerable uncertainty in the relative placement. This explains why an apparently folded antibody and HA trimer can nevertheless be docked incorrectly.
Conclusion: The AlphaFold 3 models offer a binding-site hypothesis, and the 4FQI prediction agrees qualitatively with experiment. The 7K39 prediction fails to place the antibody at its experimental head site and has very low antibody–antigen interface confidence. These models should be evaluated by comparison with experimental structures and, where appropriate, binding assays, mutagenesis or other validation.
For confidence-score interpretation see AlphaFold 3 output documentation.
Overall answer to the exercise
The known 4FQI stem and 7K39 head epitopes are spatially discontinuous and are associated with different antibody-binding regions of HA. Sequence-based BepiPred tools can recover candidate epitope residues but, for the supplied 7K39 predictions, do not discriminate the measured antibody footprint well. Of the four submitted predictions, DiscoTope-3.0 gives the strongest agreement with that particular 7K39 footprint (AP 0.0466 versus a random baseline of 0.0306), though the result is still modest. When the identity of an antibody is known, complex prediction can offer a more specific binding pose, but the 7K39 AlphaFold example demonstrates the risk of incorrect docking and the importance of interface confidence metrics. Predictions should be treated as testable hypotheses rather than definitive experimental epitopes.
Limitations and reproducibility
- All experimental epitope counts here use a strict 5 Å protein-heavy-atom cutoff without solvent; 4FQI = 23 and 7K39 = 15.
- BepiPred and DiscoTope numerical performance is directly measurable from the uploaded outputs for 7K39 only. Native-H5 predictions for 4FQI were not supplied, so a proper 4FQI precision–recall evaluation cannot be completed using these outputs.
- The BepiPred residue positions must be translated to PDB numbering (7K39 A: minus 16; 7K39 B: unchanged).
- The DiscoTope-2.0 input includes more chains than the DiscoTope-3.0 antigen-only result, so their numerical comparison is potentially confounded.
- AP is a precision–recall ranking summary; it is not the same quantity as ROC-AUC. The positive class is only 15/490 residues, so conventional classification accuracy is not a useful primary metric.
- Results against only one experimental antibody interface cannot establish that other exposed residues are never epitopes for other antibodies.