Code for paper: On Pre-trained Language Models For Antibody.

The AnTibody Understanding Evaluation (ATUE) is released on zenodo.

Our implementation is based on ParaGen. To install ParaGen from source:

cd ParaGen
pip install -e .

More information can be found in the official repo.

We put the source code under the antibody directory. It contains the modified AbLang and ESM to fit this framework. Some of our code was borrowed from them. Thanks for their great work!

The file organization is as follows:

antibody
├── ablang                              # AbLang-H / AbLang-L
│   ├── configs     
│   ├── src      
│   └── vocab
├── esm                                 # ESM / MSA-Transformer 
│   ├── configs
│   └── src  
├── src                                 # EATLM
│   ├── models
│   │   ├── axial_transformer_layer.py
│   │   ├── axial_attention.py
|   │   ├── transformer_germline_encoder.py
|   │   └── germline_model.py
│   └── tasks      
│       ├── germline_masked_lm_task.py
|       ├── germline_pretrained_task.py
|       └── germline_task.py 
├── configs                             # EATLM config files     
├── utils                              
│   ├── analyze.py
│   ├── preprocess.py
│   ├── draw.ipynb
│   └── metric.py      
├── scripts                             
│   ├── cross_valid.sh
│   ├── disease_diagnosis.sh
│   └── individual.sh    
└── README.md                           # Readme 

The ParaGen has many options for training and inference. Usually we use a *.yaml file to organize these arguments. These arguments can also be fed to the framework via the command line.

[Attention] If an option is set more than once, the last value will overwrite the previous one. And the command arguments have a higher priority than the *.yaml configuration files.

We provide several sample configuration files for each model and task with default hyperparameters. These files can be found under the configs/ directory of each model. If you want to use these configuration files, please change the options as you need. For example, change the default path.

The antibody tasks are defined under antibody/src/tasks/. Therefore, the antibody/src/ should be added as user-specific library.

To run with our EATLM, the command looks like this:

$ paragen-run --config antibody/configs/downstream/cell_germ.yaml --lib antibody/src

For esm and ablang, the specific library should be added:

# esm
$ paragen-run --config antibody/esm/configs/cell.yaml --lib antibody/src,antibody/esm/src

# ablang
$ paragen-run --config antibody/ablang/configs/cell.yaml --lib antibody/src,antibody/ablang/src

We provide some scripts for repeative experiments under the antibody/scripts directory. We can use these scripts for k-cross validation.

$ bash antibody/scripts/cross_valid.sh <config> <model> <data> k [save/drop]

The details explanations of each option can be found in the cross_valid.sh.

Besides, we also provide the postprocessing script for disease diagnosis. The entry is disease_diagnosis.sh. It can download sequence-level results from HDFS and calculate the individual-level results:

$ bash antibody/scripts/disease_diagnosis.sh [download/seq/ind] <disease> <model> <name>

[Attention] Please modify the default HDFS path before using these scripts.