Use dyMEAN Online

dyMEAN: End-to-End Full-Atom Antibody Design with Equivariant Neural Networks

dyMEAN is an end-to-end, full-atom, E(3)-equivariant model for antibody design that takes an epitope and an incomplete antibody sequence and produces the missing sequence together with the 3D structure of the antibody-antigen complex. It was developed by Xiangzhe Kong, Wenbing Huang and Yang Liu and presented at ICML 2023.

Existing computational antibody design methods tackle only subtasks of the pipeline and tend to ignore either the framework regions or the side chains. As a result, they often depend on separate tools for structure prediction and docking.

dyMEAN handles CDR design, complex structure prediction and affinity optimization in a single model that works with full atoms, including side chains.

How dyMEAN Designs Antibodies End to End

1. Knowledge-guided initialization: well-conserved framework residues form a backbone template, and other residues are placed by interpolation or extrapolation.

2. Shadow paratope: a copy of the CDR-H3 paratope is placed around the epitope. It shares hidden states with the native paratope but keeps its own coordinates, so docking does not depend on the initial pose.

3. Adaptive multi-channel equivariant encoder: handles residues with different numbers of atoms, which makes full-atom modeling possible.

4. Progressive decoding: the model predicts sequence and coordinates over three iterations.

5. Docking by alignment: the native paratope is aligned to the shadow paratope with the Kabsch algorithm.

The paper also proves that the docked structure is E(3)-equivariant and the predicted sequence is invariant.

dyMEAN Benchmark Results for CDR-H3 Design and Complex Prediction

CDR-H3 design on the RAbD benchmark (60 complexes)

Other reported results

• Complex structure prediction: on the 51-complex IgFold test set, dyMEAN reaches DockQ 0.452 and RMSD 9.05.

• Affinity optimization: on SKEMPI-based evaluation, dyMEAN reaches an average delta delta G of -7.31 kcal/mol, compared with -6.48 for MEAN.

• Multi-CDR design: designing all six CDRs gives 60.07% amino acid recovery, and designing the full antibody gives 74.96%.

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 dyMEAN on Tamarind Bio

• CDR-H3 design: generate new binding loops for a given epitope.

• Antibody-antigen complex prediction: predict the structure of the complex from an incomplete antibody sequence.

• Affinity optimization: search for variants with improved predicted binding.

• Multi-CDR and full-antibody design: design several CDRs, or the whole antibody, in one run.

How to Use dyMEAN on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select dyMEAN.

2. Provide the antigen: upload the antigen structure and define the epitope, which the paper takes as the 48 antigen residues closest to the antibody.

3. Provide the antibody framework: supply an incomplete antibody sequence with the regions to design marked, if exposed.

4. Choose the design task: pick CDR-H3 design, multi-CDR design or affinity optimization, if exposed.

5. Set parameters: choose the number of candidates if exposed; the paper generates 100 candidates per antibody for affinity optimization.

6. Run the job: launch it and monitor progress from your dashboard.

7. Download outputs: retrieve the designed sequences and predicted complex structures.

8. Use results downstream: relax the structures, for example with OpenMM, and screen candidates before wet-lab testing.

Things to Keep in Mind

• The training and test sets are dominated by a CDR-H3 pattern (ARDG***DY) that can inflate metrics; amino acid recovery fell to 31.76% when flanking residues were removed.

• Affinity optimization relies on a learned delta delta G predictor, and the authors note that wet-lab validation is still needed.

• Generated side-chain angle distributions are smoother than reference ones, so relaxation may still be needed.

• The end-to-end setting performs poorly on general proteins (CATH amino acid recovery of 16.02%).

• Experiments focus mainly on CDR-H3, and adding CDR-L3 to the shadow paratope performed worse.

Source: Xiangzhe Kong, Wenbing Huang and Yang Liu, End-to-End Full-Atom Antibody Design, Proceedings of the 40th International Conference on Machine Learning (ICML 2023), PMLR 202. arXiv:2302.00203. Paper: https://arxiv.org/abs/2302.00203. Code: https://github.com/THUNLP-MT/dyMEAN