Use AntiFold Online
Commercially Available AntiFold No-Code Web Server
AntiFold: Antibody-Specific Inverse Folding for Structure-Based Sequence Design
AntiFold is an antibody-specific inverse folding model that predicts amino acid sequences compatible with an antibody variable-domain structure. It was developed by Magnus Haraldson Høie, Alissa Hummer, Tobias H. Olsen, Broncio Aguilar-Sanjuan, Morten Nielsen and Charlotte M. Deane, and is fine-tuned from ESM-IF1 on solved and predicted antibody structures. The code is available under a BSD 3-Clause license.
General-purpose inverse folding models such as ProteinMPNN and ESM-IF1 are trained on diverse proteins, yet antibodies have distinctive features: conserved frameworks and hypervariable CDR loops that are hard to model. Antibody-specific training data helps the model learn these patterns.
AntiFold uses this to recover CDR sequences more accurately, to propose structure-compatible designs, and to score mutations, including a zero-shot signal for binding affinity.
Property | Detail from the paper |
|---|---|
Base model | ESM-IF1 fine-tuned, 142M parameters |
Training data | 2,074 solved antibody complexes from SAbDab plus 147,458 paired OAS sequences modeled with ABodyBuilder2 |
Input | Heavy- and light-chain variable domain structure, optional IMGT region to design |
Output | Per-position tolerance (perplexity), amino acid probabilities, designed sequences |
Availability | pip-installable package, web server and BSD 3-Clause license |
How AntiFold Designs Antibody Sequences from Structure
Input a backbone: heavy and light chain backbones are concatenated with a 10-position gap, and a solved or predicted structure can be used.
Score every position: the model outputs amino acid probabilities and a perplexity indicating how tolerant each position is to mutation.
Sample designs: by default 10 sequences are sampled, controlled by a temperature parameter, for chosen regions such as CDR3.
Add antigen context: optional antigen information can improve affinity-related scoring in CDRs.
AntiFold Benchmarks: Sequence Recovery, Refolding and Affinity
Model | CDRH3 recovery |
|---|---|
AntiFold | 60% |
AbMPNN | 56% |
ESM-IF1 | 43% |
ProteinMPNN | 35% |
For other CDRs, AntiFold recovery was 75 to 84% versus 63 to 76% for AbMPNN. Recovery on predicted structures changed by under 0.5%, versus a 2.7% drop for AbMPNN.
Refolding and binding affinity
In refolding tests, mean CDR backbone RMSD was 0.95 angstroms for AntiFold, against 0.98 for AbMPNN, 1.01 for ESM-IF1 and 1.03 for ProteinMPNN. On an anti-lysozyme deep mutational scan of 2,209 variants, Spearman correlation was 0.418 for AntiFold versus 0.334 for ESM-IF1. In the Hie et al. affinity maturation set, improved variants had a median rank of 80% for AntiFold. A log-likelihood threshold of -11 removed about 40% of lower-affinity variants while losing about 5% of improved ones.
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 AntiFold on Tamarind Bio
CDR redesign: generate alternative CDR sequences that remain compatible with the antibody backbone.
Affinity maturation guidance: flag mutations unlikely to disrupt structure or antigen binding.
Mutation filtering: combine with protein language model scores to prune variant libraries.
Predicted-structure workflows: design on models from tools such as ABodyBuilder2.
How to Use AntiFold on Tamarind Bio
Log in and open the tool: sign in to tamarind.bio and select AntiFold.
Upload a structure: provide a PDB or CIF file of the antibody variable domain.
Identify chains: specify heavy, light and optional antigen chains, or the nanobody chain, if exposed.
Choose regions: select IMGT regions such as CDR3, or specific positions, to design.
Set sampling: adjust number of sequences and sampling temperature, if exposed.
Run: submit the job.
Download outputs: retrieve the per-residue log-probability CSV and FASTA of designed sequences.
Interpret: use perplexity and design scores to prioritize variants.
Things to Keep in Mind
Recovery drops for long CDRH3 loops: median 71% for 6 to 9 residues versus 48% for 16 to 28 residues.
Refolding accuracy was assessed on backbone atoms only.
Affinity performance appears mainly driven by removing loss-of-binding variants.
AlphaFold generalization was tested on a single structure.
Source: Høie M. H., Hummer A., Olsen T. H., Aguilar-Sanjuan B., Nielsen M., Deane C. M. "AntiFold: Improved antibody structure-based design using inverse folding." arXiv:2405.03370 (2024), https://arxiv.org/abs/2405.03370. Code: https://github.com/oxpig/AntiFold.