Julian Zubek, Marcin Tatjewski, Adam Boniecki, Maciej Mnich, Subhadip Basu, Dariusz Plewczynski∗

• Corresponding author: d.plewczynski@cent.uw.edu.pl

Aim of this project is to predict PPI using multiscale features. Accurate predictors could be then used for recovering protein interaction network.

The starting point for this experiment is extracting complexes of proteins from PDB. Their sequences are sliced into segments using sliding window technique. Pairs of interacting segments are used as a dataset for training level-I classifier. Level-II classifier uses the predictions of level-I classifier as an input and decides whether two proteins with known sequences interact or not.

Experimental pipeline is implemented in Python 2.

External software:

• DSSP

Binary DSSP packages can be downloaded from ftp://ftp.cmbi.ru.nl/pub/software/dssp/ and placed in any directory listed in the PATH variable of the environment.

Python 2 packages:

• prody
• ruffus
• pandas
• numpy
• scipy
• scikit-learn
• affinity
• sqlite3
• urllib
• lxml

The src directory contains various scripts used in experimental pipeline. The scripts are generally independent, working in batch mode: they feed on input files and produce the output. Ruffus pipelining library is used to glue all the parts together. experiments directory should contain subdirectories corresponding to different experiments. Each experimental directory contains experiment.py file with Ruffus pipeline configuration, and all necessary input files, intermediate files and results files.

If all the dependencies are satisfied, you can run full experimental pipeline with:

> python experiment.py

If you are interested in repeating just one specific experimental step, analyze the structure of experiment.py to learn the API.

You can find training and testing data sets as well as trained classifiers in the experiments/yeast_experiment directory:

• pdb_yeast.id_list -- list of PDB complexes used in the experiment.
• pdb_yeast.ppi.sqlite -- sqlite3 database with protein-protein interactions.
• yeast_train.csv -- training set for level-I predictor.
• yeast_train.uid -- Uniprot ids of protein pairs in the train set.
• yeast_test.csv -- testing set for level-I predictor.
• yeast_test.uid -- Uniprot ids of protein pairs in the test set.
• uid2secstr_psipred.csv -- sequences and PSIPRED-predicted secondary structures for the proteins used in the experiment.

The following files can be easily recreated by running experiment.py script:

• pdb_yeast.segment.sqlite -- sqlite3 database with segment interactions.
• ppi_yeast_pos_neg_psipred_train.pkl -- pickled list of positive and negative pairs for training level-II predictor.
• ppi_yeast_xy_psipred_train.pkl -- pickled list of positive and negative predicted contact maps for training level-II predictor.

If you are just interested in running the trained classifiers to predict interactions for supplied protein pairs, check the script ppi_predict.py in src directory.

> usage: ppi_predict.py [-h] levelI_predictor levelII_predictor data_file
>
> Full PPI prediction for new proteins.
>
> positional arguments:
>  levelI_predictor   Trained level I predictor to use.
>  levelII_predictor  Trained level II predictor to use.
>  data_file          CSV file with protein pairs for prediction.
>
> optional arguments:
>  -h, --help         show this help message and exit

levelI_predictor and levelII_predictor should be pickled classifiers trained during full experimental pipeline. data_file should be a CSV (comma separated values) file with no header containing the following columns:

1. Protein A ID
2. Protein B ID
3. Protein A sequence
4. Protein B sequence
5. Protein A secondary structure
6. Protein B secondary structure

Zubek, J., Tatjewski, M., Boniecki, A., Mnich, M., Basu, S., & Plewczynski, D. (2015). Multi-level machine learning prediction of protein–protein interactions in Saccharomyces cerevisiae. PeerJ, 3, e1041.

Zubek, J., Tatjewski, M., Basu, S., & Plewczynski, D. (2015). Evaluating Multi-level Machine Learning Prediction of Protein-protein Interactions. In Computational Methods in Data Analysis ITRIA 2015 (pp. 199–211). Institute of Computer Science, Polish Academy of Sciences.