Use ImmuneBuilder Online
Commercially Available ImmuneBuilder No-Code Web Server
ImmuneBuilder: Deep-Learning Structure Prediction for Antibodies, Nanobodies and T-Cell Receptors
ImmuneBuilder is a set of three deep-learning models that predict the 3D structures of immune receptor proteins: ABodyBuilder2 for antibodies, NanoBodyBuilder2 for nanobodies and TCRBuilder2 for T-cell receptors. It was developed by Brennan Abanades, Wing Ki Wong, Fergus Boyles, Guy Georges, Alexander Bujotzek and colleagues, with Charlotte M. Deane (University of Oxford) as final author, and published in Communications Biology in 2023. The code, trained weights and inference script are open source on GitHub.
Immune repertoires from next-generation sequencing contain huge numbers of antibody and TCR sequences, and experimental structures cannot keep pace. General tools such as AlphaFold-Multimer are accurate but slow, taking about 3 hours on five CPUs, while faster methods can be less accurate or less physically plausible.
ImmuneBuilder is tailored to immune proteins to give accurate, fast and physically sensible models.
Item | Detail from the paper |
|---|---|
Models | ABodyBuilder2, NanoBodyBuilder2, TCRBuilder2 |
Antibody input | Heavy and light chain sequences |
Speed | About 5 seconds per antibody on a Tesla P100 GPU |
Training data | 7,084 SAbDab structures at 3.5 A resolution or better |
Test set | 34 non-redundant Fvs with X-ray resolution better than 2.3 A |
How ImmuneBuilder Predicts Immune Protein Structures
Antibody-specific structure module: an adapted AlphaFold-Multimer structure module with eight update blocks, Invariant Point Attention layers and backbone updates, predicting torsion angles to build an all-atom structure.
Ensemble of four models: four separately trained models each predict a structure, and the one closest to the average is selected.
Physical refinement: restrained energy minimization with OpenMM (AMBER14) removes clashes and stereochemical errors, including cis-peptide bonds and D-amino acids.
Error estimation: disagreement among the four predictions gives a per-residue confidence estimate.
Accuracy, Speed and Error Estimates
CDR-H3 loop accuracy
On the 34-structure benchmark, ABodyBuilder2 reached a CDR-H3 RMSD of 2.81 A versus 2.90 A for AlphaFold-Multimer, while running over 100 times faster. EquiFold, IgFold and ABlooper were around 10% less accurate on this loop. NanoBodyBuilder2 reached an average CDR-H3 RMSD of 2.89 A, 0.55 A better than AlphaFold2, and TCRBuilder2 was comparable to AlphaFold-Multimer.
Model | CDR-H3 RMSD | Comparison |
|---|---|---|
ABodyBuilder2 | 2.81 A | AlphaFold-Multimer 2.90 A |
NanoBodyBuilder2 | 2.89 A | 0.55 A better than AlphaFold2 |
Filtering unreliable models
Removing structures with predicted error above about 1 A would drop five structures with an average RMSD of 4.46 A and bring the remaining set's average CDR-H3 RMSD to 2.53 A.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with ImmuneBuilder on Tamarind Bio
Repertoire modeling: build structures for antibody and TCR sequences from sequencing data at high throughput.
Quality filtering: use the per-residue error estimate to exclude likely inaccurate models.
Downstream analysis: identify antibodies with similar binding targets, model antibody-antigen interactions through side-chain placement, and support structure-based virtual screening.
How to Use ImmuneBuilder on Tamarind Bio
Open the tool: log in to tamarind.bio and open ImmuneBuilder.
Choose the receptor type: antibody, nanobody or TCR, if exposed.
Provide sequences: enter heavy and light chain sequences for antibodies, or the corresponding chains for the other types.
Run the job: submit and wait; the paper reports about 5 seconds per antibody on a GPU.
Download the model: retrieve the refined all-atom structure.
Check confidence: review the per-residue error estimate, if provided, and discard high-error models.
Use downstream: take the structure into docking, virtual screening or repertoire analysis.
Things to Keep in Mind
CDR-H3 remains the hardest loop for every method tested, and the authors suggest multi-conformation models may be needed.
A low predicted error does not guarantee accuracy, although a high predicted error is a good filter.
The main benchmark has only 34 antibody Fvs.
Long side chains are predicted less accurately.
Models were trained only on structures at 3.5 A resolution or better.
Source: Abanades B, Wong WK, Boyles F, Georges G, Bujotzek A and colleagues, Deane CM. ImmuneBuilder: Deep-Learning models for predicting the structures of immune proteins. Communications Biology 6, 575 (2023). DOI: 10.1038/s42003-023-04927-7. Article: nature.com/articles/s42003-023-04927-7. Code: github.com/oxpig/ImmuneBuilder.