ML tools like Weka and XGBoost are used to train the binary classifier to predict wether a amino acids attached to prolines have cis (0 degree) omega-angle (ω) or trans (180 deree).

To do this, a protein dataset is built using a file listing all PDB chains used:

# see protein dataset 
less pdb/cullpdb_pc25.0_res0.0-3.0_noBrks_noDsdr_len40-10000_R0.25_Xray_d2021_06_05_chains3940

Using the script scripts/pdb_database_build.sh, the correct protein chains are extracted and the torsion angles are extracted into a text file by the pdbtorsions which is part of the BiopTools (extras torsion all torsion angles from a PDB file):

# example of how pdbtorsions is used
pdbtorsions example_pdb_file.ent > torsion_file.txt

Once the torsion files are created for each pdb file the proline dataset csv file is made using the scripts/getproline_py.py program:

# extracting the proline datasets using 
# a directory that contains all torsion files

scripts/./getpro_py.py -d <directory with torsion>  -w <an intiger for window size> > proline_dataset.csv

Once the proline datasets are exracted to a single csv file its encoded using the encoder program scripts/encoder.py with the choice of an encoder e.g. BLOSUM90

The before building and training the Machine Learning models, the dataset needs to be split into training and testing samples using the program scripts/build_train_test_sets.py

Once the training and testing samples are created the XGBoost models are trained using XGBoost python wrapper and its performance is evaluated on MCC (Matthews correlation coefficient) score.

Random Forest model is built using weka (a Machine Learning software) and the parameters used are shown in

a bash script: scripts/models/weka_RF/random_forest_model.sh

Both the XGBoost models and Random Forest models are located in the models directory under scripts.

Another Machine Learning predictor is built using both the predictions from Random Forest and XGBoost, where the predicted

class is the model with the higer power (higher probalility).

This is done using the program scripts/models/prediction_ensemble_method.py

© Ayub Hareed, Birkbeck: University of London