MAGE_ab_generation - GitLite

KO | EN

gitlite — search

#python #llm #golang #android #machine-learning #deep-learning #linux #nodejs #macos #kubernetes #vue #mcp #openai #open-source

MAGE_ab_generation

(★ 79)

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

파일 탐색기

최종 버전 다운로드 (.zip)

- - covabdab_gen_curation_allBinders_allVariants_v2_24-03-12.ipynb
  - NTD_SA.npy
  - RBD_SA.npy
- - crowe_ebola_processing_24-03-07.ipynb
- - denovo_trastuzamab_24-02-29.ipynb
- - detagging_script_24-03-04.ipynb
  - final_concat_v2_24-03-12.ipynb
- - wang_preprint_flu_cleaning_24-01-03.ipynb
- - atlas_antigen_alignment_v1_24-03-04.ipynb
  - published_LIBRA-seq_concatenation_24-05-31.ipynb
- - ebola_Bornholdt_24-03-07.ipynb
  - ebola_EhrHardt_24-01-03.ipynb
  - gilman_RSV_24-01-04.ipynb
  - hcv_24-01-03.ipynb
- - zurbuchen_data_23-03-08.ipynb
- ASAb_database_metadata_v2_24-03-12.xlsx
- example_training_data.csv
- full_model_training_24-03-12.py
- selected_abs_Figure_2-3_24-06-10.ipynb
- figure1_plots_24-06-10.ipynb
- output_filtering_annotation.ipynb
- RBD_1K_sequences_24-03-16.csv
- selected_abs_Figure_2-3_24-06-10.ipynb
- selected_abs_for_testing.csv
- selection_pipeline_24-03-16.ipynb
generate_antibodies.py
LICENSE
MAGE_annotated_training_data.zip
README.md

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 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