DTU Health Tech

Department of Health Technology

Research
Education
Collaboration
Services and Products
News
About

SHARE ON

AbEpiTope - 1.0

Accurately scoring and predicting antibody targets from structures

Submission

1. Upload a single structure file (.pdb/.cif) or a .zip file.



A .zip file may contain a maximum of 120 structure files. Each structure must contain an antibody-antigen complex, that includes an antibody chain (light, heavy or both) along with any number of antigen chains. Structure files where this is not detected will not produce a score.

Example Files

We provide five example .zip files containing modeled structures of antibodies targeting antigens from a range of diseases: SARS (PDB: 7VYR), HIV (PDB: 3LEV), Bacteria, (Meningococcal, PDB: 5O14), Cancer, (PD-1 receptor, PDB: 7E9B), and Autoimmune (grass pollen, PDB: 5OTJ). Each file includes 30 structures in PDB format generated using AlphaFold-2.3. Additionally, a sixth example .zip file contains modeled structures for four different antibodies and a HIV antigen (PDB: 3LHP). Each antibody and the HIV antigen was modeled separately, not simultaneously. One of these antibodies has been experimentally confirmed to target the HIV antigen, while the others are known to target different antigens; angiopetin 2 (PDB: 4ZFG), arrestin-2 (PDB: 7DFA), interleukin-1 beta (PDB: 7Z4T). The .zip file includes a total of 120 individually modeled structure files, all in PDB format.

SARS HIV Bacteria Cancer Autoimmune Antibody Target Prediction

2. Set antibody-antigen interface distance (default: 4Å):

Instructions

AbEpiTope-1.0 is a tool that features two scores. The first is AbEpiScore-1.0, which is desgined to evaluate the accuracy of modelled antibody-antigen interfaces. The second is, AbEpiTarget-1.0, a score designed to distinguish AbAg complexes modelled with the correct antibody, from those modelled with incorrect or "swapped" antibodies.

Input

1. Users can upload a single structure file (pdb/cif) or a zip file containing pdb/cif files.

Each structure file must include a light and heavy chain or a single-chain variable fragment (scFv), along with one or more antigen chains. Due to computational resource limits on the web server, we restrict uploads to a maximum of 30 files per submission. For larger batches, we recommend using the local installation (see Versions or Download)
Note: Scores will not be produced for antibody-antigen structures where this is not detected.

2. Users can set a custom Angstrom (Å) distance for defining antibody-antigen interfaces.

The default is 4 Å. The antibody-antigen interface is made up of epitope and paratope residues. Epitope residues are any residues with at least one heavy atom (main-chain or side-chain) at a distance of 4 Å or less to any light or heavy chain. The corresponding interacting residues on the light or heavy chain are the paratope residues.
Note: Scores will not be produced for antibody-antigen structures if no epitope and paratope residues are detected at the set Å distance.

Output

See Output format

Guidance on using AbEpiTope-1.0 scores for predicting modelled AbAg accuracy
and antibody screening


Figure: We illustrate how AbEpiScore-1.0 can be applied to predict the modelled AbAg accuracy and how AbEpiTarget-1.0 can be used for antibody screening A) AbEpiScore-1.0 scores for 51.900 predicted antibody-antigen structures plotted against corresponding DockQ values (y-axis) in hexagonal bins. Color scales capped at 50 structures, show the structure count per bin and orange indicates single structure bins. A red dashed line indicates a linear fit computed across all AbAg structures. B) Same as A), but plotting AbEpiScore-1.0 against AbAgIoU values. C) The PPV values for predicting whether 51.900 predicted antibody-antigen complexes have acceptable, medium or high DockQ (y-axis) as a function of X AbEpiScore-1.0 (x-axis) was computed in the range (0.0-0.55). PPV values for selected AbEpiScore are indicated with white dots. D) Evaluation of AbEpiTarget-1.0, AbEpiTarget-1.0 Δ, and AlphaFold 2.3 confidence scores for identifying correct AbAg pairs. For AbEpiTarget-1.0 and AlphaFold 2.3, the 1730 antigens groups were ranked by their maximum score (from either the true AbAg or one of the three swapped AbAgs). For AbEpiTarget-1.0 Δ, by the score gap between the top two pairs. The y-axis shows average True Rank Scores (0 = worst, 1 = perfect ranking of true AbAg) as more antigen groups are included (x-axis). White dots indicate score cutoffs corresponding to expected True Rank Scores of 0.85, 0.90, and 0.95.

  • Predicting modelled AbAg Accuracy with AbEpiScore-1.0, A) and B): A modelled antibody-antigen structure with an AbEpiScore-1.0 of 0.3, has an expected interface accuracy 0.3 ≈ 0.400 DockQ or ≈0.373 AbAgIoU.
  • Predicting modelled AbAg Accuracy with AbEpiScore-1.0, C): A modelled antibody-antigen structure with an AbEpiScore-1.0 of 0.3, has a 91.2%, 83.9%, and 22.1% probability of having acceptable (≥0.23), medium (≥0.49) and high (≥0.8) DockQ interface accuracy respectively.
  • High confidence antibody screening with AbEpiTarget-1.0, D): For an antigen group (antibodies modelled to the same antigen), only predict the highest scoring antibody-antigen pair as the correct antibody, if its AbEpiTarget-1.0 score is 0.1544 more than the second best scoring antibody-antigen pair.
  • Medium confidence antibody screening with AbEpiTarget-1.0, D): For an antigen group (antibodies modelled to the same antigen), only predict the highest scoring antibody-antigen pair as the correct antibody, if its AbEpiTarget-1.0 score is 0.0803 more than the second best scoring antibody-antigen pair.
  • Low confidence antibody screening with AbEpiTarget-1.0, D): For an antigen group (antibodies modelled to the same antigen), only predict the highest scoring antibody-antigen pair as the correct antibody, if its AbEpiTarget-1.0 score is 0.0296 more than the second best scoring antibody-antigen pair.

Output format

The tool generates two CSV files.

1. The first, output.csv, lists each input structure file along with its AbEpiScore-1.0 and AbEpiTarget-1.0 scores.

2. The second, interface.csv, lists each input structure file along with epitope and paratope residues used to compute these scores.

Note: If a row contains "None" in any column, it indicates that no antibody was identified, or no AbAg interface was detected within the specified Å distance.

3. The third, abag_sequence_data.fasta, is a fasta formmatted file containing the sequences in each each antibody-antigen complex. The header >FILENAME_CHAINNAMES and the sequences of each abag are joined with ':'.

4. The fourth, failed_files.csv, is an error file that only appears if an error occurs for one or more of the files in the zip file upload. Each row contains filename and reason for the error.

Abstract

AbEpiTope-1.0: Improved antibody target prediction by use of AlphaFold and inverse folding
Authors: Joakim Clifford, Eve Richardson, Bjoern Peters, Morten Nielsen
Publication:

Abstract

B-cell epitope prediction tools are crucial for the design of vaccines and disease diagnostics. However, predicting which antigens a specific antibody will bind to, and their exact binding sites (epitopes), remains challenging. Here, we present AbEpiTope-1.0, a computational tool for antibody-specific B-cell epitope prediction, utilising AlphaFold-2.3 for structural modelling and inverse folding for the machine learning models. AbEpiTope-1.0 outperforms AlphaFold’s confidence ranking in predicting the accuracy of modelled antibody-antigen interfaces. Most importantly, we show that the predicted accuracy is sensitive to antibody input, offering a reliable metric for selecting antibodies most likely to bind a given antigen. Furthermore, a variant of our model trained specifically for this task shows a significant performance improvement. The tool can evaluate hundreds of antibody-antigen structures in minutes, providing researchers with a valuable resource for antibody screening and B-cell epitope prediction. AbEpiTope-1.0 is freely available as a web server and standalone package at https://services.healthtech.dtu.dk/services/AbEpiTope-1.0.

Predicting Antibody-Antigen Interface Accuracy

This data is related to predicting the interface accuracy of modelled AbAg structures. We first tested AbAg interface scores for AlphaFold-2.3 and inverse folding GVP-Transformers, ESMIF1 and AntiFold, on 1,730 AbAgs without fine-tuning, creating 30 structures for each using AlphaFold-2.3 multimer, totalling 51,900 structures. AbAgIoU was used to measure the match between predicted epitope and paratope residues and the corresponding ground truth crystal structures. Later, we created finetuned models: OneHot-AbAgIoU, ESM2-AbAgIoU, AntiFold-AbAgIoU and AbEpiScore-1.0.

Downloads

AlphaFold-2.3 (ColabFold) AbAg Fasta Inputs:

A .zip file with containing all input fasta files for modelling 1730 antibody-antigen complexes with AlphaFold-2.3 colabfold version. We used 6 seeds, generating 30 structures per antibody-antigen complex and 51900 structures in total.
File Format: .zip
Download: abag_fastafiles.zip

AbAg Interface Scores:

A .csv file with AbAgIoU and DockQ scores for all 51900 structures, as well as corresponding AbAg interface model scores done in nested cross-validation. These models were: Random, Onehot-AbAgIoU, ESM2-AbAgIoU, AlphaFold-2.3, AntiFold, AntiFold-AbAgIoU, ESMIF1, AbEpiScore-1.0. Nested cross-validation was done b creating 5 data partitions of antibody-antigen complexes not sharing more 65% or 95% antigen or antibody sequence identity. Data partitions are indicated by PartitionNum
File Format: .csv
Header: StructureNames,AbAgIoU,AgIoU,DockQ, AbAg Interface Scores..., PartitionNum
Download: abag_interface_scores.csv

Antibody Target Prediction

This data is related to predicting the antigen target of a given antibody, distinguishing modelled true AbAg complexes from those modelled with incorrect or "swapped" antibodies. All modelled AbAg complexex were made with AlphaFold-2.3. We created 1,730 groups of antibody-antigen complexes, each containing one true antibody-anitgen complex and three swapped antibody-antigen complexes, all modelled with the same antigen. To avoid data leakage, antibodies for constructing swap antibody-antigen complexes were taken from other antibody-antigen complexes within the same data partition and only if the antibody was targeting a different antigen.

AlphaFold-2.3 (ColabFold) Swap AbAg Fasta Inputs:

A .zip file with containing all input fasta files for modelling 1730x3 = 5190 swapped antibody-antigen complexes with AlphaFold-2.3 colabfold version. We used 6 seeds, generating 30 structures per antibody-antigen complex and 51900 structures in total.
File Format: .zip
Download: swap_abag_fastafiles.zip

Antibody Target Prediction Scores:

A .csv file with AgIoU scores, measuring the match between predicted epitope residues and the ground truth crystal structure epitopes, for the 51900 antibody-antigen complexes and the 155700 swapped antibody-antigen structures. We also supply the corresponding model scores all evaluated in nested cross-validation. These include AbAg interface score models as well as models made specifically for antibody target predcition. These models were: Random, AlphaFold-2.3, Onehot-AbAgIoU, AbEpiTarget-1.0 (OneHot), AbEpiTarget-1.0 (OneHot+AlphaFold-2.3), AbEpiTarget-1.0 (OneHot+AbEpiScore-1.0), ESM2-AbAgIoU, AbEpiTarget-1.0 (ESM2+AlphaFold-2.3), AbEpiTarget-1.0 (ESM2+AbEpiScore-1.0) ESM2-AbAgIoU, AntiFold, AntiFold-AbAgIoU, ESMIF1, AbEpiScore-1.0, AbEpiTarget-1.0 AbEpiTarget-1.0 (+AlphaFold-2.3) and AbEpiTarget-1.0 (+AbEpiScore-1.0). Data partitions are indicated by PartitionNum
File Format: .csv
Header: StructureNames,AgIoU,AbAg Target Scores...,PartitionNum
Download: abag_abtarget_scores.csv

Antibody Target Prediction - 17x17 swapped antibody-antigen complex dataset

This data is also related to predicting the antigen target of a given antibody. We assesed the model performances of antibody target prediction for scenarios with more than three swapped AbAgs. Here, we created 17 groups, each consisting of one true AbAg - featuring the correct antibody and antigen- as well as 16 swapped AbAgs, where the true antigen was paired with incorrect antibodies. /net/urban/var/www/services/suppl/immunology/AbEpiTope-1.0/17x17_benchmark/abag_abtarget_scores.csv

AlphaFold-2.3 (ColabFold) True AbAg Fasta Inputs:

A .zip file with containing all input fasta files for modelling 17 antibody-antigen complexes with AlphaFold-2.3 colabfold version. We used 6 seeds, generating 30 structures per antibody-antigen complex and 510 structures in total.
File Format: .zip
Download: true_abag_fastafiles.zip

AlphaFold-2.3 (ColabFold) Swap AbAg Fasta Inputs:

A .zip file with containing all input fasta files for modelling 17 antibody-antigen complexes and 17x16= 272 swapped antibody-antigen complexes with AlphaFold-2.3 colabfold version. We used 6 seeds, generating 30 structures per antibody-antigen complex and 8160 structures in total.
File Format: .zip
Download: swap_abag_fastafiles.zip

Antibody Target Prediction Scores:

A .csv file with AgIoU scores, measuring the match between predicted epitope residues and the ground truth crystal structure epitopes, for the 510 antibody-antigen complexes and the 8160 swapped antibody-antigen structures. We also supply the corresponding AlphaFold-2.3 and AbEpiTarget-1.0 scores done in nested cross-validation.
File Format: .csv
Header: StructureNames,AgIoU,AlphaFold-2.3,AbEpiTarget-1.0
Download: abag_abtarget_scores.csv

GitHub Please visit our GitHub repository for a local installment of current version

The code and data can be used freely by academic groups for non-commercial purposes.
If you plan to use these tools for any for-profit application, you are required to obtain a separate license (contact Morten Nielsen, morni@dtu.dk)

This service offers no downloadable software

See a list of available software

GitHub Please visit our GitHub repository for a local installment of current version

The code and data can be used freely by academic groups for non-commercial purposes.
If you plan to use these tools for any for-profit application, you are required to obtain a separate license (contact Morten Nielsen, morni@dtu.dk)

GETTING HELP

If you need help regarding technical issues (e.g. errors or missing results) contact Technical Support. Please include the name of the service and version (e.g. NetPhos-4.0) and the options you have selected. If the error occurs after the job has started running, please include the JOB ID (the long code that you see while the job is running).

If you have scientific questions (e.g. how the method works or how to interpret results), contact Correspondence.

Correspondence: Technical Support: