PLTNUM is a protein language model designed to predict protein half-life from amino acid sequences. It is built upon the SaProt model and is trained on the dataset from Schwanhäusser, B et al. Nature. 2011, which was derived using mass spectrometry and SILAC.
PLTNUM achieves not only highly accurate protein half-life prediction but also identification of amino acid residues that significantly influence these predictions utilizing SHAP analysis.
The models and demo are available on the Hugging Face.
To set up PLTNUM, you can either clone the repository and create a conda environment or directly install the dependencies via conda:
Clone and Setup Environment:
git clone https://github.com/sagawatatsuya/PLTNUM.git
cd PLTNUM
conda env create -f environment.ymlDirect installation:
conda create --name pltnum python=3.11.8
conda activate pltnum
conda install anaconda::requests
conda install conda-forge::biopython
conda install anaconda::pandas
conda install anaconda::numpy
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
conda install anaconda::scikit-learn
conda install conda-forge::transformers
conda install -c conda-forge shap1. Create Structure-Aware Sequences
To use the SaProt-based PLTNUM model, generate structure-aware amino acid sequences with Foldseek:
python scripts/apply_foldseek_to_pdb.py \
--pdb_dir="pdb_files" \
--num_processes=4 \
--output_dir="./data"
2. Half-life prediction using PLTNUM
Predict protein half-life using the PLTNUM model:
python scripts/predict_with_PreTrainedModel.py \
--data_path="data/demo_input.csv" \
--model_path="sagawa/PLTNUM-SaProt-NIH3T3" \
--architecture="SaProt" \
--batch_size=32 \
--use_amp \
--num_workers=4 \
--output_dir="output" \
--sequence_col="aa_foldseek"
3. SHAP analysis
Interpret predictions with SHAP analysis and search for the sequence that significantly influences the prediction:
python scripts/calculate_shap.py \
--data_path="data/demo_input.csv" \
--model="facebook/esm2_t33_650M_UR50D" \
--model_path="sagawa/PLTNUM-SaProt-NIH3T3" \
--architecture="SaProt" \
--batch_size=32 \
--output_dir="output" \
--sequence_col="aa_foldseek" \
--max_evals=100
You can train PLTNUM with your own dataset.
python scripts/train.py \
--data_path="data/demo_input.csv" \
--model="westlake-repl/SaProt_650M_AF2" \
--architecture="SaProt" \
--lr=2e-5 \
--epochs=10 \
--batch_size=4 \
--use_amp \
--num_workers=4 \
--max_length=512 \
--used_sequence="left" \
--mask_ratio=0.05 \
--mask_prob=0.2 \
--n_folds=10 \
--output_dir="./model_output/" \
--task="classification" \
--target_col="T1/2 [h]" \
--sequence_col="aa_foldseek"
PLTNUM/
├── bin/ # Foldseek's binary file
├── data/ # Datasets
├── pdb_files/ # PDB files
├── scripts/ # Scripts for training and prediction, etc.
├── environment.yml # Conda environment file
└── README.md # This README file
Tatsuya Sagawa
@article {Sagawa2024.09.10.612367,
author = {Sagawa, Tatsuya and Kanao, Eisuke and Ogata, Kosuke and Imami, Koshi and Ishihama, Yasushi},
title = {Prediction of Protein Half-lives from Amino Acid Sequences by Protein Language Models},
elocation-id = {2024.09.10.612367},
year = {2024},
doi = {10.1101/2024.09.10.612367},
publisher = {Cold Spring Harbor Laboratory},
abstract = {We developed a protein half-life prediction model, PLTNUM, based on a protein language model using an extensive dataset of protein sequences and protein half-lives from the NIH3T3 mouse embryo fibroblast cell line as a training set. PLTNUM achieved an accuracy of 71\% on validation data and showed robust performance with an ROC of 0.73 when applied to a human cell line dataset. By incorporating Shapley Additive Explanations (SHAP) into PLTNUM, we identified key factors contributing to shorter protein half-lives, such as cysteine-containing domains and intrinsically disordered regions. Using SHAP values, PLTNUM can also predict potential degron sequences that shorten protein half-lives. This model provides a platform for elucidating the sequence dependency of protein half-lives, while the uncertainty in predictions underscores the importance of biological context in influencing protein half-lives.Competing Interest StatementThe authors have declared no competing interest.},
URL = {https://www.biorxiv.org/content/early/2024/09/14/2024.09.10.612367},
eprint = {https://www.biorxiv.org/content/early/2024/09/14/2024.09.10.612367.full.pdf},
journal = {bioRxiv}
}