Use ABodyBuilder3 Online

ABodyBuilder3: Fast, Accurate Antibody Structure Prediction with Uncertainty Estimates

ABodyBuilder3 is a deep learning model for predicting the 3D structure of antibodies from their heavy- and light-chain variable-region sequences. It was developed by Henry Kenlay, Frédéric A. Dreyer, Daniel Cutting, Daniel Nissley and Charlotte M. Deane (Exscientia and the University of Oxford) as an improved, scalable successor to ABodyBuilder2. The authors release the code under an Apache-2.0 license and archive the model weights on Zenodo.

Antibodies are central to modern therapeutics, and their binding behavior is dictated largely by the complementarity-determining region (CDR) loops. Experimental structure determination is slow and expensive, so accurate computational models are needed to triage large numbers of candidate sequences. The hardest region is CDRH3, which is highly variable in length and conformation.

ABodyBuilder3 addresses this by building on the ABodyBuilder2 architecture, adding protein language model embeddings, a faster implementation and a per-residue confidence estimate, so that predictions can be produced at scale and filtered by reliability.

How ABodyBuilder3 Predicts Antibody Structures

The model embeds the variable-region sequence and refines the structure through a stack of eight structure modules. Each module combines an invariant point attention layer with a backbone update layer.

1. Language model embeddings: in the ABB3-LM variant, heavy and light chains are embedded separately with ProtT5 and concatenated, giving the network richer sequence context than one-hot encoding.

2. Scalable implementation: the code is vectorised using OpenFold optimisations and uses bfloat16 weights with mixed precision, which the authors report makes it more than three times faster.

3. Confidence prediction: a pLDDT head predicts per-residue lDDT-C-alpha in 50 bins, helping users decide which predicted loops to trust.

4. Optional relaxation: the authors show that energy minimisation, in particular YASARA2 minimisation in explicit water, improves accuracy, especially in framework regions.

ABodyBuilder3 Accuracy and Benchmark Results

On a 100-structure test set, which includes the 34-structure ABodyBuilder2 test set, the authors report mean RMSD in angstroms for each region.

The language model variant gives a small reduction in error, mainly in CDRH3 and CDRL3. ABodyBuilder3 achieves accuracy comparable to the ABodyBuilder2 ensemble with a single, faster model.

Uncertainty estimation

The pLDDT-based confidence correlates with actual error. The Pearson correlation between predicted uncertainty and CDRH3 RMSD is 0.57 for ABodyBuilder2, 0.61 for ABodyBuilder3 and 0.73 for ABodyBuilder3-LM. Applying a CDRH3 pLDDT threshold of 85 retains about 32% of structures, and over 80% of those have CDRH3 RMSD below 2 angstroms.

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 ABodyBuilder3 on Tamarind Bio

• Antibody candidate triage: model many candidate sequences quickly and use the confidence scores to prioritize those with reliable CDR loops.

• CDR loop modelling: obtain structures for the loops that drive antigen recognition, including the challenging CDRH3.

• Large-scale structure generation: the scalable implementation suits generating structures for large antibody sequence sets, which the authors suggest could feed future self-distillation.

• Structure-based design inputs: use predicted structures as starting points for downstream antibody engineering workflows.

How to Use ABodyBuilder3 on Tamarind Bio

1. Log in and open the tool: sign in to tamarind.bio and select ABodyBuilder3.

2. Provide the sequences: enter the heavy-chain and light-chain variable-region sequences of your antibody.

3. Choose model options: if exposed, select between the standard and language model variants and any relaxation setting.

4. Run the prediction: submit the job and wait for the structure to be generated.

5. Download the outputs: retrieve the predicted structure file and any per-residue confidence values, if exposed.

6. Interpret confidence: treat regions with high predicted pLDDT as more reliable, and be cautious with low-confidence CDRH3 loops.

7. Use downstream: take the structure into docking, design or developability analysis.

Things to Keep in Mind

• CDRH3 remains the hardest region, with a mean RMSD of about 2.40 angstroms for the language model variant.

• Training and validation excluded nanobodies and ultra-long CDRH3 loops, so performance on those may differ.

• The improvement from language model embeddings is described by the authors as small.

• The paper compares quantitatively only against ABodyBuilder2 and its own variants.

Source: Kenlay H., Dreyer F. A., Cutting D., Nissley D., Deane C. M. "ABodyBuilder3: Improved and scalable antibody structure predictions." arXiv:2405.20863 (2024), https://arxiv.org/abs/2405.20863. Code: https://github.com/Exscientia/ABodyBuilder3.