GraphPI Protein Inference

Folder Structure

└── code: contains the source code for the project.
    └── main_train.py: main file to train the model.
    └── main_eval.py: main file to make predictions.
    └── configs.py : contains the configuration for the project.
└── data: contains our test data, which includes:
    └── PSM features.
    └── Fasta database (these two constitute the input to our algorithm). 
    └── Results from other algorithms.
└── results: our trained model is saved here, also includes:
    └── Protein groups.
    └── Prediction made by our algorithm.
└── knime_workflows: contains the knime workflows.
    └── generate_decoy: generate decoy database based on fasta file.
    └── comet_search: execute peptide search algorithm
    └── epifany_pipeline: run epifany on Percolator results, also includes Percolator to generate PSM features.
    └── fido_pipeline, pia_pipeline: run fido and pia.

Usage

Environment Setup

For environment preparing, please use conda:

conda env create -f environment.yml

Train

Pretrain data

We use data from the public repository (promeXchange) for training purpose, the link to download the raw files are included in Supporting Table S2. Please run the following procedure to acquire training data.

1. Download the raw files from proteomeXchange, the fasta database from Uniprot.
2. Run the generate_decoy knime workflow to generate a decoy fasta database.
3. Convert raw file into mzML format using ProteoWizard (with peaks picking).
4. Run the comet_search knime workflow with mzml and decoy fasta files for peptide database search, please make sure to use correct search parameters from the dataset provider.
5. Run the epifany_pipeline knime workflow with idxml and decoy fasta files to generate the psm features and epifany scores.
6. Collect the psm and epifany scores for all training data.

To run this program for demonstration, we provide the psm and epifany scores for one single dataset which is stored in data/PXD005388/.

Please make sure to use KNIME version 4.7.7 for our workflows.

Run

Change the DATA_LIST in the configs file to the updated pretraining PXD datasets. (e.g., if you want to only use PXD005388 for training, then set DATA_LIST=["PXD005388"].)

Run python main_train.py --save_result_dir=test_run in the code folder to train the model.

Inference

Test data

For our comparison study, our test data (and download links) are listed in Supporting Material S1. The processed PSM features and decoy databases are included in the data folder.

Run

Please set the TEST_DATA variable in configs.py to our test datasets, for instance,

TEST_DATA = ['ups2', '18mix', 'yeast', 'iPRG2016_A', 'iPRG2016_AB', 'iPRG2016_B', 'hela_3t3']

To make predictions on our test datasets using your trained model, please run python main_eval.py --save_result_dir=test_run, ROC plots will be generated for every test data under test_run/predictions/.

To generate the predictions on the test datasets using our trained model, please run python main_eval.py directly. The results will be stored in results/predictions/.

Evaluate on any dataset with trained model

To evaluate any dataset with just raw and fasta files, please follow the steps below:

1. Run the generate_decoy knime workflow to generate a decoy fasta database.
2. Convert raw file into mzml format using ProteoWizard (with peaks picking).
3. Run the comet_search knime workflow with mzml and decoy fasta files for peptide database search, please make sure to use correct search parameters from the dataset provider.
4. Run the epifany_pipeline knime workflow with idxml and decoy fasta files to generate the psm features.
5. Create a folder under data, the datafolder must have precix TestDataset, put the psm feature and decoy fasta files to appropriate positions in the folder. The psm feature file should be stored as data/TestDataset{your_data_name}/psm_features/result.csv, and the decoy fasta file should be stored as data/TestDataset{your_data_name}/database/decoy.fasta
6. Modify the config file to include the new test dataset, the name of the test dataset should be the same as the data folder.
7. Run main_eval.py under the code folder to make predictions.
8. Prediction result will be stored in results/predictions/TestDataset{your_data_name}/
9. Please execute code/read_prediction_result.ipynb notebook to check out prediction result, and apply decoy FDR filtering.

Note that epifany_pipeline will run both Percolator and Epifany, and here we only need the result from Percolator.

For example, we included TestDataset004789 as an example of how to evaluate on a new dataset, it is stored under data/, and is derived from the PXD004789 ProteomeXchange dataset..