/SecrifyDL

Primary LanguagePython

Massively parallel interrogation of protein fragment secretability using SECRiFY reveals features influencing secretory system transit

This is the python code for training, evaluating and visualizing a convolutional neural network for secretability prediction. The corresponding paper (preprint) can be found at https://www.biorxiv.org/content/10.1101/241349v4

Packages used:

tensorflow v1.10.0
numpy v1.14.5
sklearn v0.23.1

(i) Create data files for cross-validation

From the results tables found online, and put in a directory initial_data, you can prepare the datafiles used in these scripts. This can be done by simply running

Uses files:

  • initial_data/Pp_resultstable_enriched.txt
  • initial_data/Pp_resultstable_depleted.txt
  • initial_data/Sc_resultstable_enriched.txt
  • initial_data/Sc_resultstable_depleted.txt
python create_data_files.py

The files containing the different folds for cross-validation are saved in the directory data.

(ii) Train a convolutional neural network model

The neural network model is constructed and trained by executing the main.py script.

python main.py <dataset> <varlen_reduction_strategy>

Args:

  • dataset one of pp, sc
  • varlen_reduction_strategy one of global_maxp, k_maxp, gru, zero_padding

Predictions over all folds are stored in a single file in the predictions directory. Trained model parameters for all folds are stored in the parameters directory, with one file per fold. Finally, the AUROC and AUPRC are calculated for all predictions.

It is also possible to calculate the AUROC and AUPRC afterwards. To this end, execute

python eval.py <prediction_file>

(iii) Visualize important input features using integrated gradients

Using the model parameters (using the stored files with their original names), feature contribution can be calculated by running the following command. Note that output is written to stdout.

python vis_calc_ig.py <parameter_file> <dataset_name>

Args:

  • parameter_file one of the parameter directories (ending in foldX) in the parameters directory
  • dataset one of pp, sc

Then, the calculated contributions can be normalized using vis_normalize_ig. Note that output is written to stdout.

python vis_normalize_ig.py <old_file_name> > <normalized_file_name>

And finally, the normalized values can be aggregated by running:

python vis_quantify_ig.py <normalized_file_name>

We recommend to then copy this to an excel spreadsheet to start analyzing the result.