kokrah/IgLM

Generative Language Modeling for Antibody Design

IgLM

Official repository for IgLM: Generative Language Modeling for Antibody Design.

The code and pre-trained models from this work are made available for non-commercial use under the terms of the JHU Academic Software License Agreement. For commercial inquiries, please contact Johns Hopkins Tech Ventures at awichma2@jhu.edu. The training and test datasets for IgLM are available for download here.

Try antibody sequence generation in Google Colab.

Setup

To use IgLM, install via pip:

pip install iglm

Alternatively, you can clone this repository and install the package locally:

$ git clone git@github.com:Graylab/IgLM.git 
$ pip install IgLM

Command line usage

IgLM supports sequence infilling, sequence generation (with prompting), and sequence evaluation from the command line.

Re-design spans of an antibody sequence

To use IgLM to re-design spans of an antibody sequence, supply the fasta file, the fasta record ID corresponding to the sequence to design, the start index of the span (0-indexed), and the end index of the span (0-indexed, exclusive).

To generate 100 unique sequences of the anti-tissue factor antibody (1JPT) heavy chain with an IgLM-designed CDR3:

iglm_infill data/antibodies/1jpt/1jpt.fasta :H 98 106 --chain_token [HEAVY] --species_token [HUMAN] --num_seqs 100 

Full antibody sequence generation

IgLM can be used to generate full antibody sequences while conditioning on the chain type and species-of-origin.

To generate 100 unique human heavy chain sequences starting with EVQ:

iglm_generate --prompt_sequence EVQ --chain_token [HEAVY] --species_token [HUMAN] --num_seqs 100 

To generate 100 unique nanobody sequences starting with QVQ:

iglm_generate --prompt_sequence QVQ --chain_token [HEAVY] --species_token [CAMEL] --num_seqs 100 

Sequence evaluation

IgLM can be used to calculate the log likelihood of a sequence given a chain type and species-of-origin.

Full sequence log likelihood calculation:

iglm_evaluate data/antibodies/1jpt/1jpt.fasta :H --chain_token [HEAVY] --species_token [HUMAN]

Infilled sequence log likelihood calculation:

iglm_evaluate data/antibodies/1jpt/1jpt.fasta :H --start 98 --end 106 --chain_token [HEAVY] --species_token [HUMAN]

Package usage

IgLM may also be used as a Python package, enabling the above use cases and more flexible usage.

Re-design spans of an antibody sequence

To use IgLM to re-design spans of an antibody sequence, supply the sequence to design, the start index of the span (0-indexed), and the end index of the span (0-indexed, exclusive).

To generate 100 unique sequences of the anti-tissue factor antibody (1JPT) heavy chain with an IgLM-designed CDR3:

from iglm import IgLM

iglm = IgLM()

parent_sequence = "EVQLVESGGGLVQPGGSLRLSCAASGFNIKEYYMHWVRQAPGKGLEWVGLIDPEQGNTIYDPKFQDRATISADNSKNTAYLQMNSLRAEDTAVYYCARDTAAYFDYWGQGTLVTVS"
chain_token = "[HEAVY]"
species_token = "[HUMAN]"
infill_range = (98, 106)
num_seqs = 100

generated_seqs = iglm.infill(
    parent_sequence,
    chain_token,
    species_token,
    infill_range=infill_range,
    num_to_generate=num_seqs,
)

Full antibody sequence generation

IgLM can be used to generate full antibody sequences while conditioning on the chain type and species-of-origin.

To generate 100 unique human heavy chain sequences starting with EVQ:

from iglm import IgLM

iglm = IgLM()

prompt_sequence = "EVQ"
chain_token = "[HEAVY]"
species_token = "[HUMAN]"
num_seqs = 100

generated_seqs = iglm.generate(
    chain_token,
    species_token,
    prompt_sequence=prompt_sequence,
    num_to_generate=num_seqs,
)

To generate 100 unique nanobody sequences starting with QVQ:

from iglm import IgLM

iglm = IgLM()

prompt_sequence = "QVQ"
chain_token = "[HEAVY]"
species_token = "[CAMEL]"
num_seqs = 100

generated_seqs = iglm.generate(
    chain_token,
    species_token,
    prompt_sequence=prompt_sequence,
    num_to_generate=num_seqs,
)

Sequence evaluation

IgLM can be used to calculate the log likelihood of a sequence given a chain type and species-of-origin.

Full sequence log likelihood calculation:

import math
from iglm import IgLM

iglm = IgLM()

sequence = "EVQLVESGGGLVQPGGSLRLSCAASGFNIKEYYMHWVRQAPGKGLEWVGLIDPEQGNTIYDPKFQDRATISADNSKNTAYLQMNSLRAEDTAVYYCARDTAAYFDYWGQGTLVTVS"
chain_token = "[HEAVY]"
species_token = "[HUMAN]"

log_likelihood = iglm.log_likelihood(
    sequence,
    chain_token,
    species_token,
)
perplexity = math.exp(-log_likelihood)

Infilled sequence log likelihood calculation:

import math
from iglm import IgLM

iglm = IgLM()

sequence = "EVQLVESGGGLVQPGGSLRLSCAASGFNIKEYYMHWVRQAPGKGLEWVGLIDPEQGNTIYDPKFQDRATISADNSKNTAYLQMNSLRAEDTAVYYCARDTAAYFDYWGQGTLVTVS"
chain_token = "[HEAVY]"
species_token = "[HUMAN]"
infill_range = (98, 106)

log_likelihood = iglm.log_likelihood(
    sequence,
    chain_token,
    species_token,
    infill_range=infill_range,
)
perplexity = math.exp(-log_likelihood)

Citing this work

@article{shuai2021generative,
  title={Generative language modeling for antibody design},
  author={Shuai, Richard W and Ruffolo, Jeffrey A and Gray, Jeffrey J},
  journal={bioRxiv},
  pages={2021--12},
  year={2021},
  publisher={Cold Spring Harbor Laboratory}
}