Use DiffAb Online
Commercially Available DiffAb No-Code Web Server
DiffAb: Diffusion-Based Antibody CDR Sequence and Structure Co-Design
DiffAb is a diffusion-based generative model that jointly designs the sequences and 3D structures of antibody complementarity-determining regions (CDRs) conditioned on an antigen. It was introduced by Shitong Luo, Yufeng Su, Xingang Peng, Sheng Wang, Jian Peng and Jianzhu Ma in "Antigen-Specific Antibody Design and Optimization with Diffusion-Based Generative Models for Protein Structures" (NeurIPS 2022). The authors' repository is released under an Apache-2.0 license.
Antibodies recognize antigens mainly through their CDR loops, so designing these loops is central to antibody engineering. Traditional computational approaches such as Rosetta-based design rely on sampling and energy functions, and many machine learning approaches handle only sequence or only structure.
DiffAb treats the problem as generative modeling: given an antigen structure and the antibody framework, it samples CDR sequences and structures together, and can be used for several related tasks with the same model.
Property | Detail from the paper |
|---|---|
Representation | Per residue: amino acid type, C-alpha coordinate and orientation in SO(3) |
Diffusion | Multinomial for amino acid types, Gaussian for coordinates, perturb-and-denoise on SO(3) for orientations |
Network | MLP embeddings followed by invariant point attention and three denoising heads |
Training data | SAbDab; structures worse than 4 angstroms removed; test set of five CDR-H3 clusters (19 complexes) |
Post-processing | Side-chain packing with Rosetta and AMBER99 refinement in OpenMM |
How DiffAb Designs Antibody CDRs
Start from noise: sampling begins from uniform amino acid types, standard-normal positions and uniform SO(3) orientations.
Condition on context: the model sees the antigen structure and the antibody framework, and denoises types, positions and orientations step by step.
Equivariant network: the design is equivariant to rotation and translation of the overall structure.
Build and refine: full-atom structures are built from ideal local coordinates, side chains are packed with Rosetta and the result is relaxed.
DiffAb Tasks and Reported Performance
Sequence-structure co-design
With 100 samples per CDR, DiffAb was compared against RosettaAntibodyDesign. The authors report higher amino acid recovery and more structurally diverse CDR-H3s, with a binding-energy improvement rate on par with Rosetta for CDR-H3 and lower for other CDRs.
Structure prediction and fixed-backbone design
With the sequence fixed, the model predicts CDR structures, with RMSD of 1.5 angstroms or less for H1, H2, L1, L2 and L3, while accuracy drops for CDR-H3 longer than about 10 residues. In fixed-backbone sequence design it achieves higher amino acid recovery than a Rosetta baseline for all CDRs.
Antibody optimization
By perturbing a CDR for a limited number of steps and denoising it, DiffAb can improve Rosetta binding energy while keeping the CDR similar to the original.
Design without a bound framework
As a demonstration, the authors docked a human antibody template (PDB 3qhf) with CDR-H3 removed to the SARS-CoV-2 Omicron RBD (PDB 7wvn) using HDOCK, then designed CDR-H3s. They note effectiveness is hard to confirm without a reference antibody.
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 DiffAb on Tamarind Bio
De novo CDR design: generate candidate CDR sequences and structures for a given antigen-antibody complex.
Antibody optimization: refine existing CDRs to improve predicted binding energy while preserving similarity.
CDR structure prediction: predict loop conformations when the sequence is known.
Fixed-backbone sequence design: propose alternative CDR sequences for a given backbone.
How to Use DiffAb on Tamarind Bio
Log in and open the tool: sign in to tamarind.bio and select DiffAb.
Provide the complex: upload an antibody-antigen complex PDB file, with heavy and light chain IDs if exposed.
Choose the design mode: the repository defines single-CDR co-design, multi-CDR co-design, optimization, fixed-backbone design and structure prediction.
Select the CDR: pick which CDR or CDRs to design, if exposed.
Run the job: submit and wait for the sampled designs.
Download outputs: retrieve the generated structures and sequences.
Evaluate downstream: rank designs with your own binding or developability filters before experimental testing.
Things to Keep in Mind
The model requires an antibody framework already bound to the target antigen; designing without bound structures is left to future work.
It is unclear whether generated antibodies can be produced in the lab or actually bind the target, so experimental validation is needed.
CDR-H3 prediction is less accurate than other CDRs, especially for longer loops.
The original repository has been archived by its owner and is read-only.
Source: Luo S., Su Y., Peng X., Wang S., Peng J., Ma J. "Antigen-Specific Antibody Design and Optimization with Diffusion-Based Generative Models for Protein Structures." NeurIPS 2022; preprint bioRxiv DOI 10.1101/2022.07.10.499510, https://www.biorxiv.org/content/10.1101/2022.07.10.499510v5. Code: https://github.com/luost26/diffab.