p-IgGen

p-IgGen is a paired antibody auto-regressive langauge model. This package provides utlity functions for generating and scoring antibody sequences using p-IgGen, with model weights stored on HuggingFace.

Features

• Generate full-length antibody sequences.
• Generate a heavy chain given a light chain and vice versa.
• Generate full-length antibody sequences given an initial sequence.
• Calculate log likelihoods of sequences.
• VH and VL chains of generated sequences can be optionally seperated using ANARCI.

Installation

We advise installing using a conda environment.

Prerequisites

• Conda

Step-by-Step Setup

1. Create a new conda environemnt:

   conda create -n my_env
   conda activate my_env
   conda install python=3.11 pip -y

2. Install this repository:

   pip install git+https://github.com/OliverT1/p-IgGen.git

3. Install optional ANARCI dependency (for --separate_chains option):

   conda install -c bioconda anarci

Usage

Command Line Interface

Generate Sequences

To generate new antibody sequences, use the piggen_generate command:

piggen_generate --output_file output_sequences.txt --n_sequences 100 

Sequences are generated by default in direction VH->VL, from C-term to N-term. Alternatively, they can be genreated in reverse from VL->VH, from N-term to C-term using the --backwards flag. This allows generation given an VH or VL sequence of any length.

Note:

• If --backwards is used, the --initial_sequence should be provided in reverse, starting from the N-term of the VL chain.
• If heavy_chain_file or light_chain_file this inversion is handled autmoamtically, and the VH and VL chains should be provide in the standard direction.

Options:

• --developable: Use the developable model.
• --heavy_chain_file FILE: File containing heavy chain sequences to generate light chains from.
• --light_chain_file FILE: File containing light chain sequences to generate heavy chains from.
• --initial_sequence TEXT: Initial sequence to generate from.
• --n_sequences INTEGER: Number of sequences to generate, per input sequence if applicable.
• --top_p FLOAT: Top-p sampling value (default: 0.95).
• --temp FLOAT: Temperature for generation (default: 1.2).
• --bottom_n_percent FLOAT: Bottom n percent of sequences to discard based on likelihood (default: 5).
• --backwards: Generate sequences in reverse.
• --output_file FILE: File to save the generated sequences (required).
• --separate_chains: Output VH and VL sequences separately, requires ANARCI.

Using bottom_n_percent requires n_sequences to be at least 100, otherwise this option is ignored.

Calculate Log Likelihoods

To calculate the log likelihoods of sequences, use the piggen_likelihood command:

sh
python cli.py likelihood --sequence_file input_sequences.txt --output_file log_likelihoods.txt

Options:

• --developable: Use the developable model.
• --sequence_file FILE: The file containing sequences to calculate log likelihoods.
• --batch_size INTEGER: Batch size for processing sequences.
• --output_file FILE: File to save the log likelihoods.
• --local: Load model from local path.

Examples

Generate Light Chains for Provided Heavy Chain :

piggen_generate --heavy_chain_file heavy_chains.txt --n_seqeunces 5 --top_p 0.95 --temp 1.25 --output_file generated_sequences.txt

Heavy chains should be seperate by new lines. Here, five light chains will be generated for each heavy chain.

Calculate Log Likelihoods for Sequences:

piggen_likelihood --sequence_file sequences.txt --batch_size 2 --output_file log_likelihoods.txt