Peptide Encoder

An encoder for peptides (short amino acid sequences) based on BLOSUM similarity.

In particular, this package includes a model for learning peptide embeddings such that the embedding of two peptides in the vector space is proportional to their BLOSUM62 similarity.

Installation

This project is written in python3 and can be installed with pip. It is available on PyPI.

pip3 install --find-links https://download.pytorch.org/whl/cu113/torch_stable.html peptide-encoder

Alternatively, the package can be installed from source.

git clone https://github.com/bmmalone/peptide-encoder.git
cd piptide-encoder
pip3 install -r requirements.txt .

(The "period" at the end is required.)

Prerequisites: This project relies on quite a few prerequisites, such as pytorch, ray, cudnn, and others. Both the requirements.txt and setup.py files aim to install these dependencies correctly; nevertheless, it may still be preferable to install these dependencies before installing this package.

In particular, the find-links argument to pip may need to be adjusted depending on the available version of CUDA.

Usage

After installation, models can be trained using a command similar to the following:

train-pepenc-models /prj/peptide-encoder/conf/base/config.yaml --num-hyperparameter-configurations 500 --max-training-iterations 30 --name my-pepenc-tune-exp --out-dir /tmp

The --help flag can be used to see a description of all arguments to the script.

For adjusting the hyperparameter search space, algorithms, or schedulers, the pepenc/models/train_pepenc_models.py script can be adjusted. If the package was not installed in pip "editable" mode, then make sure to re-run pip install so that the changes take effect for the next run of ray.

Documentation

Unfortunately, there is no sphinx, etc., documentation at this time. The file conf/base/config.yaml shows examples of all hyperparameters, data files, etc., for training models.

Data format

The models in this project require an input csv file that has one row which is a header and remaining rows which are the peptides for the various datasets. The column in the csv file with the peptide sequences must be named sequence. (This can be adjusted if calling the pepenc library from python code.)

Tensorboard visualization

The <out_dir>/<name> directory (based on the arguments to train-pepenc-models) will contain output suitable for visualization with Tensorboard. The following command uses Docker to expose the results on port 6007.

docker run --rm --mount type=bind,source=/tmp/my-pepenc-tune-exp,target=/tensorboard --publish 6007:6006 nvcr.io/nvidia/tensorflow:21.12-tf2-py3 tensorboard --logdir /tensorboard

The tensorflow image can be updated as necessary.

N.B. The source of the bind mount must be the <out_dir>/<name> directory (based on the arguments to train-pepenc-models).

Training the model

The model consistently experiences vanishing (or, improbably, exploding) gradient issues when using a single LSTM layer. It is not clear why this happens, and it is currently suggested to avoid allowing lstm_layers == 1 in the hyperparameter search space (or directly setting it that way in the config).

Testing the code

The project uses pytest for unit testing. The testing prerequisites (though not other dependencies, as described above) can be installed as follows.

pip3 install .[test]

After installing pytest and other testing dependencies, tests can be performed as follows.

cd /path/to/peptide-encoder
pytest .