MAGE

Generation of novel paired heavy-light chain antibodies using large language models

Monoclonal Antibody GEnerator (MAGE) - a fine-tuned LLM for generating paired heavy-light antibody variable sequences with predicted binding specificity to antigen prompt.

This repository contains Python scripts to accompany Wasdin et al., including antibody generation and follow-up analyses presented in the manuscript. All analyses were initially ran in Linux Red Hat 8.4, but have also been tested in Ubuntu 22.04. For training, 4 V100s were used and for antibody generation, an Nvidia A6000 was used. The model weights are hosted on HuggingFace at https://huggingface.co/perrywasdin/MAGE_V1.

The following libraries are needed, we recommened installing these within a Conda environment with Python 3.11. Note that other versions are likely compatible, but we used the following.

• Numpy 1.26
• Pandas 2.1.1
• Scikit-learn 1.3.0
• Matplotlib 3.8.1
• Seaborn 0.13.1
• Jupterlab 3.6.3

Model training and generation require the following libraries:

• PyTorch 2.1.0 with pytorch-cuda 11.8
• Transformers 4.32.1

The Progen2 repository must cloned from: https://github.com/enijkamp/progen2

Once downloaded, move to the 'progen' directory here in order to import within the training/generation scripts.

Follow-up Analyses used these additional libraries:

• python-Levenshtein 0.25.0
• pandarallel 1.6.4 (to speed up Pandas functions)
• Abnumber 0.3.2

Overview of directories

Data cleaning: notebooks for cleaning data, including a variety of different sources.

Output_analysis: notebooks for recreating figures in manuscript

• Include selection scripts for RBD, RSV-A, and H5/TX/24

Antibody_generation: script for generating antibody sequences against RBD. This yields a CSV file with raw sequences which can be analyzed using the notebooks in the previous directory.

Fine_tuning: script and example subset dataset (n=1000) for fine-tuning Progen2.

Tutorial for antibody generation

• Antibody sequences can be generated using the generate_antibodies.py Python script
• An antigen prompt should be provided as an amino acid sequence with any signal peptides or transmembrane regions removed, example: SARS-CoV-2 index strain RBD: RVQPTESIVRFPNITNLCPFGEVFNATRFASVYAWNRKRISNCVADYSVLYNSASFSTFKCYGVSPTKLNDLCFTNVYADSFVIRGDEVRQIAPGQTGKIADYNYKLPDDFTGCVIAWNSNNLDSKVGGNYNYLYRLFRKSNLKPFERDISTEIYQAGSTPCNGVEGFNCYFPLQSYGFQPTNGVGYQPYRVVVLSFELLHAPATVCGPKKSTNLVKNKCVNF
  • This is provided as a string named antigen_prompt within the python script
• When running the script from the terminal, specify n number of sequences to generate, and an output csv name:
  • Example: python generate_antibodies.py --n=1 --output=MAGE_antibodies.csv
• This was tested on an Nvidia A6000 and took ~15 seconds to generate one antibody sequence against RBD.
• To annotate and analyze output sequences without installing the follow-up analyses libraries listed above, generated sequences can be uploaded to the IMGT Domain Gap Align webserver: https://www.imgt.org/3Dstructure-DB/cgi/DomainGapAlign.cgi