Understanding and predicting the drug-drug interactions is an important in both drug development and clinical application, especially for co-administered medications. We propose a new information model for drug-drug interaction analysis, based on the common biological targets and therapeutic similarity. Based on the data from DrugBank, our model calculates target similarity and therapeutic similarity features. To predict the possible drugdrug interactions it uses a semi-supervised approach, defined in two steps: adding the missing labels using the clustering algorithm K-Means, and then, executing a classification with a supervised learning model Support Vector Machine. Proposed model is tested for the known data set and had shown the high classification rate, with the AUC=98.5+-0.05.
Identification and prediction of drug-drug interactions (DDI) is a widespread topic of the research
in the healthcare, and studying such aspect is a big part of the drug development process [
A number of computational methods have been employed for the prediction of DDIs based on drugs
structures and/or functions: physiologically based pharmacokinetic modeling, molecular structural
similarity analysis, ontology and annotation-based analysis, network modeling, QSAR modeling [
We can divide the machine learning-based methods used for the prediction of DDIs according to the approach used: unsupervised, supervised, and semi-supervised machine learning-based algorithms [21].
In one study [
The aforementioned studies proceeded a vast amount of data [
Several studies focused on predicting the DDI though protein-protein-interaction networks with
chemical features [
Only few of the studies proceeded the ‘in vitro’ experiments to evaluate their models [
We have observed that the supervised learning models analyzed significantly smaller amount of
possible DDIs [15, 16]. It was caused by the relatively small number of the known DDIs in the literature,
and therefore, picking the corresponding amount of the drug pairs without indication of possible
interaction in the literature. Besides, we assume that there should be more complex procedure of data
labeling, because it improves the performance of the next predictions [
So, predicting DDIs is a complex problem [
In this study, we propose a new information model for drug–drug interaction analysis, based on the common biological targets and therapeutic similarity. The information model is able to proceed the data extraction from the source, execute the calculation of the features, execute the data labeling process and make predictions of the possible DDIs.
Regarding that many researchers obtained data for their investigations of DDI from a database DrugBank [17], we used DrugBank as the data source for our work as well. Also we used the calculation methods of target and therapeutic similarities features proposed by Cheng et al. [l5, 16] – such approach, combined with the supervised learning algorithm, Support Vector Machine has shown a significant accuracy in predicting DDI on the sample.
We examined the hypothesis of predicting the based on the common biological targets and therapeutic use, instead of including chemical and physical descriptors of the drugs.
We addressed the following problem: the researchers added labels meaning the absence of DDI when the drug pair didn’t have the DDI indication in the data source, however there might be the unreported, and used only 3% of the original input data. In this study we want to improve the data labeling process, and that would enable us to use the whole dataset as well for the next predictions. So, we followed a semi-supervised approach, which consists of first, clustering algorithm for obtaining the missing data labels, and then, classification to predict the possible drug-drug interactions.
We used a database DrugBank as a source of data. Drug bank is freely accessible, online database containing information on drugs and drug targets [17]. It contains both bioinformatics, cheminformatics details about drugs such as resource, chemical, pharmacological and pharmaceutical data with comprehensive drug target (sequence, structure, and pathway) information.
Obtaining data for the study was performed by parsing the xml document. We extracted the following drug information: drug name, DrugBank ID of the drug, targets, ATC code (anatomicaltherapeutic-chemical classification), known drug-drug interactions. From the obtained set of drugs and characteristics, we removed drugs that did not contain ATC code (experimental drugs, homeopathic and herbal traditional medicinal products) as well as drugs of antibodies and inorganic salts. 2.2.
The information model was designed to proceed all the steps required for the drug-related data analysis and prediction of the possible drug-drug interactions. element of the vector contains a value of 1, if the drug has no effect on the target - the value is 0. After that, for each combination of drugs, we constructed the target similarity by calculating the Tanimoto coefficient for binary vectors of drugs: combination; where represents the quantity of the biological targets, which are common for both drugs of the represents quantity of the biological targets, which the drug a affects; represents quantity of the biological targets, which the drug b affects.
For the therapeutic similarity feature (ST) we used the method proposed by Cheng et al. [15]. We created five sets with unique ATC codes representing each of the five ATC classification levels for each drug pair. Next, for each drug pair for each ATC classification level the therapeutic similarity feature was calculated: ( , ) =
, ( , ) =
+ − , where k represents an ATC classification level (from 1 to 5); ( ) represents ATC codes of the k-level for the drug a; ( ) represents ATC codes of the k-level for the drug b. levels: ( , ) =
∑ =1 ( , )
After that, the general therapeutic similarity was calculated considering all five ATC classification (1) (2) (3) where n represents the overall number of ATC classification levels (=5); ( , )represents the previously calculated therapeutic similarity for each ATC classification level.
The drug pairs that were indicated in the DrugBank as known, received the label 1. For the remaining drug combinations there is not enough information in DrugBank to assert or deny the drug-drug interaction, so no labels were added.
In order to predict the labels for the drug pairs that contained no label, the clustering method
KMeans was applied. We have chosen K-Means method because it has been used by the researchers in
the healthcare fields as the first step of the semi-supervised machine learning approach to define the
missing data labels [
In K-Means method the centroids are randomly initialized from the dataset. Then, from each centroid the Euclidean distance is calculated to each data point, and depending upon the minimum distance between the centroids and data points, that data point is assigned to that centroid. This is repeated until there is no change of the centroids. In this way, the clusters are formed.
The accuracy of the clustering was calculated according to the percentage of how much of the clustered labels 1 match the original labels 1 for the drug pairs.
For the predicting of the possible drug-drug interactions, we used the supervised machine learning model Support Vector Machine (SVM), namely Linear Support Vector Classification. We made our choice based on the literature review: in drug-related research this method is used to solve classification problems. In the studies we investigated, such method has shown significantly good performance [14, 16].
The AUC value was calculated, and the confusion matrix was composed to evaluate the performance of the model.
The research was performed using the programming language Python3. The xml parsing was proceeded using the library by using library xml.etree.ElementTree in Python3. The data analysis was performed using libraries numpy, pandas, sklearn. Data visualization was executed using libraries matplotlib, seaborn.
We used open-source software which is freely available and contributed by the global community of developers.
From DrugBank we obtained information about 721 drugs, which has been used as an input for the information model.
For the feature construction, 266085 unique drug-drug combinations were created. The target and therapeutic similarities were calculated and assigned to the corresponding drug pairs. 6946 drug pairs were actually indicated in DrugBank as having the drug-drug interaction, so they received the label 1.
Whole dataset was used as an input for the clustering algorithm, with k=2 clusters. The accuracy calculated with our method is 54%.
After that, for all known drug combinations that contained the label 1 before clustering, we left the original labels, and for the drug combinations with the missing ones, we assigned the labels obtained as a result of clustering.
We investigated the distribution of each feature according to their labels.
The distribution is binomial, the values are contained in range [
To execute the classification, the whole dataset was splitted into the train set (70%) and test set (30%). Such splitting ratio (70/30) has been widely used by the scientific community for data analysis. In our research we didn’t notice the significant difference of performance with various splitting, but with the 70/30 ration the AUC value was the highest – 98.53 (Comparing to AUC=98.39 for 90/10, AUC=98.41 for 80/20, AUC=98.47 for 60/40).
We applied the Linear Support Vector Classification method of the Support Vector Machine algorithm with the linear kernel.
Based on the prediction, we received the Area Under the Curve (AUC) value = 98.5+-0.05 illustrate the absolute values of prediction of the training set, we composed the following confusion matrix:
The count of predicted drug-drug interactions is 35 186. The number of correctly predicted drug pairs that do not have drug-drug interactions is 43 469. There are 217 drug pairs with label 1 that were predicted as such not having the drug-drug interaction. 954 drug pairs were wrongly predicted as having the drug-drug interaction, although they were labeled as 0.
The information system for the drug-related data analysis and the prediction of the possible drugdrug combinations based on their calculated target and therapeutic similarities was created. It uses a semi-supervised learning approach, in order to firstly, define the missing labels using the clustering algorithm, and then, execute a classification using a supervised learning model.
Our examined hypothesis to use data about biological targets and therapeutic use has received reinforcement in the form of high predictive performance on the dataset from DrugBank, verified with the test set.
By executing the data labeling process, we were able to use for the further predictions all amount of drug combinations, including the 97.39% that didn’t have the labels at the beginning.
In the similar studies that use same dataset, included biological targets or therapeutic use into their examined hypothesis, the results were AUC=0.968 [14] and AUC=0.912 [15]. So, our implementation of the proposed information system has shown accuracy of classification about 98.5+0.05 (AUC) for the DrugBank dataset and it outperforms other similar systems.
This information system can be enhanced with the functionality to calculate more features such as enzyme similarity and transporter similarity.
We thank Olga Boretska (Danylo Halytskyi Lviv National Medical University) for providing insight and expertise that greatly assisted the research. 6. References [14] Song, Dalong, et al. "Similarity‐based machine learning support vector machine predictor of drug‐ drug interactions with improved accuracies." Journal of clinical pharmacy and therapeutics 44.2 (2019): 268-275. [15] Cheng, F., Li, W. et al. Prediction of polypharmacological profiles of drugs by the integration of chemical, side effect, and therapeutic space. Journal of chemical information and modeling, 53(4), (2013): 753-762. [16] Cheng, F., & Zhao, Z. Machine learning-based prediction of drug–drug interactions by integrating drug phenotypic, therapeutic, chemical, and genomic properties. Journal of the American Medical Informatics Association, 21(e2), (2014): e278-e286. [17] Wishart, David S., et al. "DrugBank: a knowledgebase for drugs, drug actions and drug targets." Nucleic acids research 36.suppl_1 (2008): D901-D906. [18] Singh, Reetu, and E. Rajesh. "Prediction of Heart Disease by Clustering and Classification
Techniques." International Journal of Computer Sciences and Engineering 7 (2019): 861-866. [19] Li, Xiangyi, et al. «Biomolecular network-based synergistic drug combination discovery» BioMed research international2016 (2016). [20] Wu, Zikai, Xing-Ming Zhao, and Luonan Chen. «A systems biology approach to identify effective cocktail drugs» BMC systems biology. Vol. 4. No. 2. BioMed Central, 2010. [21] Li, Xiangyi, et al. "Biomolecular network-based synergistic drug combination discovery." BioMed research international 2016 (2016).