This is the repo of AntiBMPNN project, for antibody sequence design.

Protocol

Installation

Requirements:

Hardware Requirement:

AntiBMPNN can run on both CPU and GPU. For optimal performance, using an NVIDIA graphics card is recommended.

System Requirement: Linux & Conda

A Linux based operating system is required.
We have tested our code on both Ubuntu system and Windows Subsystem for Linux (Click the link for installation instructions for Windows Subsystem for Linux (WSL).).

An Anaconda python environment or Miniconda python environment is required.(Click the link for installation instructions for Anaconda/Miniconda.)

Python >= 3.11
numpy
pandas
torch
peptides
scikit-learn
tqdm

1. Obtaining AntiBMPNN Source Code

Download the AntiBMPNN repository:

git clone https://github.com/zeysun/AntiBMPNN

2. One-step Installation (recommended way)

We provided an auto-install bash script. This script will automatically build a python environment via conda and download packages and model weights.

cd AntiBMPNN  
bash Initialize.sh

Manual Installation

If you have completed the one-step installation, you can skip the manual installation step. This section is provided for reference, just in case.

conda create -n mlfold python=3.11
conda activate mlfold
pip install numpy pandas torch peptides scikit-learn tqdm

3. Run Sample Script

Then user can go to example folder and run example_scripts.sh to check if everything goes well.

cd example/  
bash example_scripts.sh

4. Make New Design

1. Prepare Input Structures

There are several ways to obtain an antibody 3D structure, such as searching on PDB or modeling via AlphaFold3, among others.
AntiBMPNN requires only the variable regions of the heavy or light chain in PDB file format as input.
To proceed, create a new directory at ~/AntiBMPNN/input and copy the PDB file into that folder.

2. Modification of Example Script

There are some necessary variables that must be changed, make sure to modify them before running:

pdb_file_path
CHAINS_TO_DESIGN
DESIGN_ONLY_POSITIONS
THEME

You can also try to modify the following parameters to meet different design purposes. Please find and modify the corresponding variable values in the script:

    --model_name 
    --num_seq_per_target 
    --sampling_temp 
    --batch_size 
    --backbone_noise

3.In Silico Design Output

The output should appear in the ~/AntiBMPNN/example folder, where the subfolder seqs contains the output raw sequence fasta file and the csv file contains the parsed and clustered sequence.

4.Validation

Users can apply AlphaFold3 to model the designed sequence. By integrating frequency information, they can determine which sequence to proceed with for experimental validation.

Sequence Judgement

<img width="250" src="https://github.com/zeysun/AntiBMPNN/blob/main/figures/AntiBMPNN_4.jpg"> A typical testing workflow involves designing a sufficiently large number of sequences, followed by clustering using the built-in scripts and analyzing the frequency associated with each unique sequence. Subsequently, AlphaFold3 is employed for structural modeling. Experimental validation is then prioritized based on a combination of modeling scores and the design frequency of each sequence.

AntiBMPNN Performance

AntiBMPNN has a better sequence recovery rate.

<img width="750" src="https://github.com/zeysun/AntiBMPNN/blob/main/figures/AntiBMPNN_3re.jpg"> ## Example output Detailed output file can be found after running the `example_scripts.sh`, along with the sequence files.

| Positions(94-110) | Recovery | Frequency | PD1_volume | PD2_hydro | PD3_charge | Changes | |--------------------|----------|-----------|------------|-----------|------------|------------------------------------------------------------------------------------------| | ARSKSTYLSRDSSGYDY | 0.67 | 6495 | -0.4271 | 0.3253 | 0.9881 | [('I', 101, 'L'), ('Y', 103, 'R'), ('N', 104, 'D'), ('N', 106, 'S')] | | ARSKSTYLSYNSSGYDY | 0.83 | 5722 | -0.3929 | 0.2118 | 0.9858 | [('I', 101, 'L'), ('N', 106, 'S')] | | ARSKSTYLSYDSSGYDY | 0.75 | 5055 | -0.4071 | 0.2376 | -0.014 | [('I', 101, 'L'), ('N', 104, 'D'), ('N', 106, 'S')] |

Positions(start-end) - Amino acid position range of your design.
Recovery - The sequence recovery rate of designed sequence.
Frequency - The number of times the sequence was designed。
PD1_volume - The physical descriptors of residue volume.
PD2_hydro - The physical descriptors of hydrophilicity.
PD3_charge - The charge properties at pH 7.4.
Changes - Summary of the difference between input and output sequences.

Model Weigts and Training Sets

Link for AntiBMPNN model weights: https://zenodo.org/records/13387792/files/model_weights.zip Link for AntiBMPNN training sets: https://zenodo.org/records/13387792/files/training_set.zip

Training performance:

AntiBMPNN

Install / Use

README