OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Mean

MEAN (Multi-channel Equivariant Attention Network) is Python research code for conditional antibody design as 3D equivariant graph translation. Its repository contains scripts for structure-data preparation, model training and evaluation, CDR redesign, and affinity optimization.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Mean — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use MEAN when you need research code for structure-conditioned antibody CDR redesign or for reproducing its documented antibody-model experiments, and can provide the required PDB structures, dataset summaries and computing environment.

Who it is for

  • Computational biology researchers studying antibody design or redesign.
  • Researchers evaluating graph-based antibody models on SAbDab, RAbD or SKEMPI_V2 data.

Documented uses

  • Prepare antibody structure summaries and generate dataset splits for research experiments.
  • Train and evaluate models in the documented SAbDab k-fold workflow.
  • Generate a CDR-H3 redesign from an IMGT-renumbered PDB with heavy chain, light chain and antigen.
  • Run the repository’s documented affinity-optimization workflow using SKEMPI_V2 data.
Before you choose

Requirements & limits

The repository is public and uses the MIT license. The documented primary use is running local Python and shell scripts.

  • Input structures for the design script must contain heavy chain, light chain and antigen, and be renumbered using IMGT.
  • The design script saves a generated PDB and a CDR-H3 sequence summary to ./results.
  • The README describes mode 100 as heavy-chain-only context and mode 111 as heavy chain, light chain and epitope context.
  • Data preparation requires external structure files; the README points to SAbDab downloads and specifies IMGT-renumbered structures.
  • The repository includes checkpoints identified for CDR-H3 redesign and optimization.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare a structure and generate a CDR-H3 design

Use the repository’s documented single-CDR script to produce a design from an IMGT-renumbered PDB containing antibody and antigen chains.

Before you start

  • A local checkout of the MEAN repository and dependencies installed with its setup script.
  • A PDB containing heavy chain, light chain and antigen, renumbered in the IMGT system.
  • The heavy- and light-chain identifiers in the PDB; a checkpoint is supplied by default, with an option to provide another checkpoint.

Steps

  1. Install the dependencies listed in scripts/setup.sh by running the command below.

    Documented command
    bash scripts/setup.sh

  2. Renumber the input structure with the repository’s ImmunoPDB script, following the documented example: the command below.

    Documented command
    python ./data/ImmunoPDB.py -i data/1ic7.pdb -o 1ic7.pdb -s imgt

  3. Run the command below, replacing the example file and chain identifiers with those for the input structure.

    Documented command
    python ./scripts/design.py --pdb 1ic7.pdb --heavy_chain H --light_chain L

  4. For a different checkpoint, pass its path with --ckpt; the README names the provided CDR-H3 redesign and optimization checkpoints.

Expected outcome

The script saves a generated PDB and CDR-H3 sequence summary in ./results.

Guide limitations

  • The described input must contain heavy chain, light chain and antigen and be IMGT-renumbered first.
  • The README documents single and multiple PDB inputs for design; it does not specify a performance or biological validation guarantee.
Documentation 1 ↗
Source-based practical guide

Prepare and run SAbDab k-fold evaluation

Follow the repository’s listed data-preparation, training and evaluation stages for its SAbDab k-fold workflow.

Before you start

  • Dependencies installed using scripts/setup.sh.
  • SAbDab summary file and downloaded structure files, with structures arranged as IMGT-renumbered PDBs in the referenced folder.
  • A GPU identifier for the documented training and evaluation commands.

Steps

  1. Prepare the data with the command below.

    Documented command
    bash scripts/prepare_data_kfold.sh summaries/sabdab_summary.tsv all_structures/imgt

  2. Train with GPU=0 bash scripts/k_fold_train.sh summaries 111 mean 9901.

  3. Evaluate with GPU=0 bash scripts/k_fold_eval.sh 111 mean 0, using the version identifier corresponding to the saved training checkpoint.

Expected outcome

The documented workflow prepares data splits, trains MEAN and runs evaluation on the selected mode and checkpoint version.

Guide limitations

  • Data preparation is stated to require about 5 GB of space.
  • Mode 111 includes heavy chain, light chain and epitope; mode 100 uses only the heavy chain.
  • The example assumes the repository’s example paths and a usable GPU; the README recommends absolute paths if path problems occur.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
18 October 2026
Original website title
THUNLP-MT/MEAN · Official homepage title · retained in its source language
How we review sourcesSuggest a correction