PeptDeep-HLA

A deep learning-model that predicts if a HLA peptide is present or not.

This is a sub-package of AlphaPeptDeep, and see our publication for details.

Quick start

Use Colab to train the models and predict HLA peptides, see:

Training from scratch: nbs/HLA1_Classifier.ipynb
Transfer learning: nbs/HLA1_transfer.ipynb

Installation

After installing anaconda, please clone and install this package using commands below:

cd path/to/place/this/package
git clone https://github.com/MannLabs/PeptDeep-HLA.git
cd PeptDeep-HLA
pip install .

Or install directly via pip:

pip install git+https://github.com/MannLabs/PeptDeep-HLA

CLI

After installation, we can use command line interface (CLI) to train sample-specific HLA models and predict HLA peptides either from fasta files or from peptide tables. Type the command below will show usage messages.

peptdeep_hla class1 -h

Here are the details of the CLI parameters/options:

--prediction-save-as TEXT: File to save the predicted HLA peptides [required]
--fasta TEXT: The input fasta files for training and prediction, multiple fasta files are supported, such as: --fasta 1.fasta --fasta 2.fasta .... If --peptide-file-to-predict is provided, these fasta files will be ignored in prediction.
--peptide-file-to-predict TEXT: Peptide file for prediction. It is an txt/tsv/csv file which contains peptide sequences in sequence column to be predicted. If not provided, this program will predict peptides from fasta files. Multiple files are supported. Optional, default is empty.
--pretrained-model TEXT: The input model for transfer learning or prediction. Optional, default is the built-in pretrained model.
--prob-threshold FLOAT: Predicted probability threshold to discriminate HLA peptides. Optional, default=0.7.
--peptide-file-to-train TEXT: Peptide file for transfer learning. It is an txt/tsv/csv file which contains true HLA peptide sequences in sequence column for training. Multiple files are supported. Optional, default is empty.
--model-save-as TEXT: File to save the transfer learned model. Optional, applicable if --peptide-file-to-train is provided.
--predicting-batch-size INTEGER: The larger the better, but it depends on the GPU/CPU RAM. Optional, default=4096.
--training-batch-size INTEGER: Optional, default=1024.
--training-epoch INTEGER: Optional, default=40.
--training-warmup-epoch: INTEGER Optional, default=10.
--min-peptide-length INTEGER: Optional, default=8.
--max-peptide-length INTEGER: Optional, default=14.
-h, --help Show this message and exit.

For example, use the following command to predict from fasta without trainfer learning:

peptdeep_hla class1 --fasta /Users/zengwenfeng/Workspace/Data/fasta/irtfusion.fasta --prediction-save-as /Users/zengwenfeng/Workspace/Data/fasta/irt_hla.tsv

Notebook

Using Jupyter notebooks might be easier if users are not familiar with CLI.

HLA1_Classifier.ipynb. We used this notebook to train the pretrained models:

HLA1_IEDB.pt: the LSTM model trained with HLA1 sequeces from IEDB. This is the default pretrained model in peptdeep_hla.
HLA1_94.pt: the LSTM model trained with 94 allele types.

HLA1_transfer.ipynb. A simple example of transfer learning to train the sample-specific model.

Spectral libraries

After HLA peptides are predicted, we can then use these peptides to predict spectral libraries with AlphaPeptDeep for HLA DIA analysis.

Citations

Wen-Feng Zeng, Xie-Xuan Zhou, Sander Willems, Constantin Ammar, Maria Wahle, Isabell Bludau, Eugenia Voytik, Maximillian T. Strauss & Matthias Mann. AlphaPeptDeep: a modular deep learning framework to predict peptide properties for proteomics. Nat Commun 13, 7238 (2022). https://doi.org/10.1038/s41467-022-34904-3