Over the past years many methods for similarity-based (a.k.a. knowledge-based, guilt-by-association-based) drug repurposing, yet most of these studies do not provide the code or the model used in the study. To improve reproducibility, we present a Python-platform offering

• drug feature data parsing and similarity calculation
• data balancing
• (disjoint) cross validation
• classifier building

Using this platform we investigate the effect using unseen data in the test set in similarity-based classification.

See Jupyter (IPython) Repurpose Notebook for reproducing the analysis presented in the manuscript and example runs.

See DDI Notebook for the analysis of drug-drug interaction prediction using drug similarity.

The Python platform has the following dependencies:

• Numpy
• Scikit-learn

Just download (i.e. clone) the files to your computer, no additional install is required. Several test cases for the methods in utilities.py are provided in tests.py. To run these, type

python tests.py

It should give an output similar to below

......

Ran 6 tests in 0.002s

OK

The data sets used in the analysis are freely available online

We have modified these data sets slightly for parsing in Python by

• converting all drug, disease and side effect terms to lowercase
• removing the quotations and making the text tab delimited
• we also added the 'Drug' text to the header

These modified files are available under data/ folder.

We have also retrieve pharmocokinetic drug-drug interaction (DDI) information from DrugBank database (v4.5.0) and mapped the drugs on the data set above.

For running the code with the default parameters defined in the default.ini in src/ directory, type

config_file = "default.ini"
config_section = "DEFAULT"
python main.py -c config_file -s config_section 

Alternatively, for using the check_ml method that builds a machine learning classifier to predict drug-disease associations using a cross-validation scheme, include the following in the python code

import ml
ml.check_ml(data, n_run, knn, n_fold, n_proportion, n_subset, model_type, prediction_type, features, recalculate_similarity, disjoint_cv, split_both = False, output_file = None, model_fun = None, verbose = False, n_seed = None)

data can be loaded using the following function

import utilities
data = utilities.get_data(drug_disease_file, drug_side_effect_file, drug_structure_file, drug_target_file, drug_interaction_file=None)

See the Repurpose Notebook for several use cases on repurposing drugs using chemical, target profile and side effect similarity. For drug-drug interaction prediction using drug similarity, see the DDI Notebook.

The configuration information for the experiments are in default.ini. The path of the data file has to be defined based on your local file structure.

Parameters in default.ini:

• drug_disease_file: File containing drug-disease associations (a binary matrix where rows are drugs, columns are diseases)
• drug_side_effect_file = File containing drug-side effect associations (a binary matrix where rows are drugs, columns are side effects)
• drug_structure_file = File containing drug-chemical sub structure mapping (a binary matrix where rows are drugs, columns are substructures)
• drug_target_file: File containing drug-target mapping (a binary matrix where rows are drugs, columns are targets)
• output_file: File in which the output AUC and AUPRC values are going to be stored
• random_seed: A number to assign use as seed to random package functions (set it an integer for reproducibility, if -1 the output would vary depending on the random selection)
• model_type: Machine learning model to be used to build the classifier, either svm | logistic | knn | tree | rf | gbc
• prediction_type = Whether the classifier will be build to predict drug-disease ('disease') or drug-side effect ('side effect') associations
• features = Features to be used to build the classifier, a combination of chemical | target | phenotype
• disjoint: Whether the cross-validation folds contain overlapping drugs (True) or not (False)
• pairwise_disjoint : Whether the cross-validation folds should group both of the pairs within the same group
• recalculate_similarity = Whether to recalculate k-NN based drug-disease and drug-side effect association score within training and test sets (True: recalculate, default, False: do not recalculate)
• knn = Number of most similar drugs to consider while calculating drug-disease and drug-side effect association score
• n_fold: Number of cross-validation folds
• n_proportion: Proportion of negative instances compared to positives (e.g., 2 means for each positive instance there are 2 negative instances)
• n_subset: If not -1, it uses a random subset of size n_subset of the positive instances (to reduce the computational time for large data sets)
• n_run = Number of repetitions of cross-validation analysis

• Data balancing and cross validation (in utilities.py)

balance_data_and_get_cv(pairs, classes, n_fold, n_proportion, n_subset=-1, disjoint=False, split_both=False, n_seed=None)

Input parameters: pairs: all possible drug-disease pairs classes: labels of these drug-disease associations (1: known, 0: unknown) n_fold: number of cross-validation folds n_proportion: proportion of negative instances compared to positives (e.g., 2 means for each positive instance there are 2 negative instances) n_subset: if not -1, it uses a random subset of size n_subset of the positive instances (to reduce the computational time for large data sets) disjoint: whether the cross-validation folds contain overlapping drugs (True) or not (False) split_both: whether the cross-validation folds should group both of the pairs within the same group n_seed: number to feed to the random generator to for reproducibility (of the cross-validation folds)

Output: This function returns (pairs, classes, cv) after balancing the data and creating the cross-validation folds. cv is the cross validation iterator containing train and test splits defined by the indices corresponding to elements in the pairs and classes lists.

• Classifier model (in utilities.py)

get_classification_model(model_type, model_fun = None)

Input parameters: model_type: custom | svm | logistic | knn | tree | rf | gbc model_fun: the function implementing classifier when the model_type is custom

The allowed values for model_type are custom, svm, logistic, knn, tree, rf, gbc
corresponding to custom model provided in model_fun by the user or the default 
models in Scikit-learn for support vector machine, k-nearest-neighbor, 
decision tree, random forest and gradient boosting classifiers, respectively. 

Output: Returns the classifier object that provides fit and predict_proba methods.

Guney E., REPRODUCIBLE DRUG REPURPOSING: WHEN SIMILARITY DOES NOT SUFFICE. Pac Symp Biocomput. 2016;22:132-143. Pubmed