This repository contains the code and resources for a protein abundance imputation framework using a graph neural network approach. The framework is designed to predict missing protein abundances by leveraging relationships between gene-derived molecules such as proteins, peptides, and mRNAs. It offers a flexible and high-performing solution for imputing protein abundance values in mass spectrometry-driven proteomics data.
- Inductive, attention-based graph neural network framework
- Ability to utilize all available abundances for protein-peptide-mRNA tuples
- Flexible graph-building and model training scheme
- Benchmarking against different imputation methods on diverse human datasets
To use this framework, you'll need to set up a Python environment. We recommend using a virtual environment to manage your dependencies. Here are the basic installation steps:
-
Clone this repository to your local machine:
git clone https://github.com/DILiS-lab/GNN_imputer.git
-
Navigate to the project directory:
cd GNN_imputer -
Install the required packages:
pip install -r requirements.txt
To run the GNN model, follow these steps:
- Navigate to the "train" folder:
cd train - Run the GNN model with the following command:
python3 train_gnn_model.py <FOLD_NUM>
The FOLD_NUM is the fold that you wish to train on (Should be between 0 to 4). For running training for all folds use 'all' for FOLD_NUM. The 'data' folder contains the DGL networks for the Blood Plasma dataset on which the GNN is to be trained. These networks are derived from real world dataset as described in [1] and processed by methods described in [2].
The training generates results in the form of:
- The foldwise performance over epochs in the file train/performance_GNN_regression.csv.
- The pytorch saved model in the train/saved_models directory with file name as best_model_fold_<FOLD_NUM>.pt
- The outputs predictions for the known test molecules in the train/saved_models directory with file name as mapping_abundances_prediction_vs_original_fold_<FOLD_NUM>.json.
- The outputs predictions for the unknown test molecules in the train/saved_models directory with file name as mapping_abundances_prediction_vs_missing_original_fold_<FOLD_NUM>.json.
This project is licensed under the MIT License.
[1] Ahmed, Saima, et al. "Using plasma proteomics to investigate viral infections of the central nervous system including patients with HIV-associated neurocognitive disorders." Journal of NeuroVirology 28.3 (2022): 341-354.
[2] Viode, Arthur, et al. "A simple, time-and cost-effective, high-throughput depletion strategy for deep plasma proteomics." Science advances 9.13 (2023): eadf9717.