=Paper=
{{Paper
|id=Vol-1748/paper-13
|storemode=property
|title=Application of Machine Learning in Knowledge Discovery for Pharmaceutical Drug-drug
Interactions
|pdfUrl=https://ceur-ws.org/Vol-1748/paper-13.pdf
|volume=Vol-1748
|dblpUrl=https://dblp.org/rec/conf/kdweb/Herrero-ZazoLB16
}}
==Application of Machine Learning in Knowledge Discovery for Pharmaceutical Drug-drug
Interactions==
https://ceur-ws.org/Vol-1748/paper-13.pdf
Application of machine learning in knowledge discovery
for pharmaceutical drug-drug interactions
Maria Herrero-Zazo1,*, Maxime Lille2, David J. Barlow1
1 Institute of Pharmaceutical Science, King’s College London, London
{maria.herrero, dave.barlow}@kcl.ac.uk
2 Faculty of Pharmacy, University of Auvergne, BP 38-63001, Clermont-Ferrand Cedex 1, France
maxime.lille@etu.udamail.fr
Abstract. Artificial neural networks (ANNs) have been developed to predict
the clinical significance of drug-drug interactions (DDIs) for a set of 35 phar-
maceutical drugs using data compiled from the Web-based resources, Lexi-
comp® and Vidal®, with inputs furnished by various drug pharmacokinetic
(PK) and/or pharmacodynamic (PD) properties, and/or drug-enzyme interaction
data. Success in prediction of DDI significance was found to vary according to
the drug properties used as ANN input, and also varied with the DDI dataset
used in training. The Lexicomp® dataset is found to give predictions marginal-
ly better than those obtained using the Vidal® dataset, with the best prediction
of minor DDIs achieved using a multi-layer perceptron (MLP) model trained
using enzyme variables alone (F1 82%) and the best prediction of major DDIs
achieved using a MLP model trained on PK/PD properties alone (F1 54%).
Given a more comprehensive and more consistent dataset of DDI data, we con-
clude that machine learning tools could be used to acquire new knowledge on
DDIs, and could thus facilitate the regulatory agencies’ pharmocovigilance of
newly licensed drugs.
Keywords: drug-drug interactions, pharmacovigilance, machine learning, arti-
ficial neural networks
1 INTRODUCTION
Recent years have witnessed an increasing number of publically availa-
ble databases for drug knowledge including chemical and pharmacolog-
ical information [1], drug-protein relationships and drug mechanisms of
action [2,3] and adverse effects [4,5].
The creation of these various resources has afforded new opportunities
in drug discovery and development, wherein data mining techniques are
employed, for example, to deduce combinations of chemical and bio-
logical characteristics of use in drug repurposing [6,7] – that is, the
identification of new therapeutic indications for approved drugs – and
also to allow the prediction of drug-protein relationships [8], and likely
drug side effects [9]. Such predictions as these are of great relevance in
pharmacovigilance, providing for the detection, assessment, under-
standing, and prevention of the adverse effects of drugs (and any other
drug-related problems) [10]. Licensed drugs are frequently seen to
cause adverse effects that are not observed in the clinical trials con-
ducted prior to approval when used in large populations comprising
individuals with very different physiological and pathological charac-
teristics. The early detection of these adverse effects is crucial to ensure
patient safety and is the primary purpose of pharmacovigilance.
Among the various adverse effects of interest, those that arise as a con-
sequence of drug-drug interactions (DDIs) are arguably the most com-
mon. These adverse effects are seen in patients that are prescribed two
or more drugs that interact in some way. In these cases, one of the
drugs taken affects the blood levels and/or efficacy of a co-
administered drug, thereby giving rise to unexpected toxicological
problems or therapeutic failure [11]. Problems of this nature have seen
increased frequency of late, partly because of the increased numbers of
elderly patients that suffer multiple co-morbidities [12] or the common
use of cocktails of drugs to treat complex pathologies [13].
Historically, information on DDIs was collated in manually curated
compendia [11], [14], and many of these resources are nowadays acces-
sible online. There are many such pharmacological databases and semi-
structured resources that are available to assist healthcare professionals
in the prevention of DDIs (e.g., Vidal® and Lexicomp®) but their qual-
ity is variable and the consistency of their contents rather limited.
One of the most relevant discrepancies among the different DDI infor-
mation sources is their assessment of the significance of the recorded
interactions. Here, significance refers to the clinical relevance of the
DDI, and describes the risk that the DDI might pose for a patient’s
health [14]. For any given DDI, the clinical significance will vary ac-
cording to the nature of the patient – their age, ethnicity, and genetic
profile – and also on the drugs’ pharmacological characteristics – in-
cluding their target(s), metabolism, and side effects. The grading of
DDI significance is generally assessed subjectively, through a pre-
established set of evaluation criteria, and thus discrepancies among dif-
ferent DDI compendia and information sources are very common. In-
deed, different researchers have identified important discrepancies be-
tween different information compendia [15] and between these and cri-
teria laid down by clinicians [16]. These studies highlight the extreme
difficulty of assessing the severity of DDIs. One might expect, howev-
er, that most of the information sources would show a high degree of
overlap, at least for those DDIs that would have severe health conse-
quences (interactions that we may thus consider as major DDIs), and a
similar degree of overlap for those DDIs that would not be expected to
do so (interactions which we might thus call minor DDIs). The devel-
opment of an in-silico system that could automatically identify DDIs
and provide an initial assessment of their likely clinical significance
(classifying each as major or minor) would likely be of great benefit,
therefore, in the field of signal detection in pharmacovigilance.
The prediction of DDIs through the application of machine learning
methods is an active research field. Cheng et al. [17] used phenotypic,
therapeutic, chemical structure and genomic similarity between drugs
as input properties to train and evaluate different machine learning
methods: naive Bayes (NB), decision tree (DT), k-nearest neighbors (k-
NN), logistic regression (LR), and support vector machines (SVM).
The last of these proved to lead to the highest performance for the pre-
diction of DDIs between 721 drugs. Recently, Sridhar et al. [18] de-
scribed a probabilistic approach using Probabilistic Soft Logic (PSL)
and different drug-drug and protein-protein similarity measures as input
variables. Focusing on enzyme-related DDIs only, Hunta et al. [19]
evaluated the performance of ANN, SVM and k-NN for the prediction
of DDIs, while Polak et al. [20] constructed several ANN models using
drug physicochemical and metabolic properties of drugs as input data.
One of the challenges in all these projects is the identification of relia-
ble negative examples of non-interacting drug pairs. So far, these ap-
proaches have used pairs of drugs not included in the selected infor-
mation source as examples of non-DDIs. Because of the aforemen-
tioned limitations of manually created DDI compendia and databases
[15,16], it is impossible to know if one pair of drugs is not described in
the selected source because it is not known yet or because there is no
DDI between the drugs [21]. In contrast, prediction of significance of
DDIs based on examples of graded DDIs could overcome this issue,
enabling the distinction between minor (or potentially non-harmful)
and major (or potentially serious) DDIs. However, to the best of our
knowledge there has been no attempt made to date to develop a more
sophisticated reasoning engine to furnish systematic predictions of DDI
severity.
In the work reported here, we describe a novel approach for the identi-
fication of clinically relevant DDIs using machine learning techniques;
we use input provided by relevant chemical and pharmacological drug
characteristics extracted from online information sources, together with
a bespoke DDI dataset containing information extracted from known
DDI compendia.
2 MATERIALS AND METHODS
2.1 Drug data
A total of 35 drugs were selected for study, according to the criteria that
each of those selected exhibit a high potential for interaction with other
drugs and/or are representative of a major therapeutic class and had no
missing data for any of the input variables. Many of the selected drugs
were previously used in related work by Vitry et al. [15].
Through a review of the general literature on DDIs [11], [14], those
drug characteristics considered relevant to their interaction profiles
were identified. The 20 drug characteristics selected included both
pharmacokinetic (PK) properties − those descriptive of drug disposition
in the body − and pharmacodynamic (PD) properties − those descrip-
tive of the drugs’ effects in patients. The drug properties data were tak-
en from the manually-curated database DrugBank [1], the online ver-
sions of Martindale [22] and Clarke's Analysis of Drugs and Poisons
[23] and the Database of Intravenous Pharmacokinetic Parameters in
Humans [24].
Given the frequency with which DDIs result as a consequence of drug-
induced changes in the activity of metabolic enzymes [25], data were
also collected on drug-enzyme relationships, using information extract-
ed from the SuperCyp database [26]. We collected drug relationships
with different isoenzymes and – for each drug in a DDI drug pair – we
represented these as a set of 26 binary variables (with 1 signifying an
effect of a given drug on a given enzyme, and 0 signifying no such rela-
tionship).
2.2 DDI data
Information on DDIs was collated from the Lexicomp® and Vidal®
compendia. In the former case, the online Lexi-Interact™ Online Inter-
action Service, was used to acquire the DDI information [27]. DDI in-
formation from both compendia were graded according to their respec-
tive five and four point scales. In the case of Lexicomp®, the DDIs
were graded from ‘no known interaction’ through to ‘avoid combina-
tion’. In the case of Vidal®, the DDIs were graded from ‘none’ to ‘con-
traindication’ [28] (Table 1). The two sets of DDI data were compiled
separately, and are referred to below as the DDI-L and DDI-V datasets.
2.3 Construction of in-silico models for DDI prediction
Artificial neural network (ANN) models to predict DDIs were devel-
oped using the data mining tools provided in Statistica®. This applica-
tion provides a wide selection of network types and the training algo-
rithms BFGS (Broyden-Fletcher-Goldfarb-Shanno) and Scaled Conju-
gate Gradient algorithms [29].
Separate models were trained and tested using the datasets described
above, which included a total of 142 variables, and using different
combinations of drug properties. Experiments were initially performed
to predict interacting vs. non-interacting drug pairs and these then re-
peated to predict the significance/grade of the DDIs.
In an initial analysis, we explored two types of ANN architecture: Mul-
tilayer Perceptron Neural Networks (MLP) and Radial Basis Function
Neural Networks (RBF). Only MLP networks with a number of hidden
units in the range of 8-20 were retrieved as the best ones.
For the construction of the final models, therefore, the methodology
adopted was as follows: 200 MLP networks were trained with a range
of 8-20 hidden units retrieving the best five networks. Activation func-
tions were not restricted, so we explored the set of neuron activation
functions for the hidden and output neurons available in Statistica:
identity, logistic sigmoid, tanh and exponential. The error function used
was either sum of squares (SOS) or cross entropy. Overfitting was pre-
vented by manually dividing the input data into training (70% of the
dataset), test (10%) and validation (20%) datasets, ensuring a balanced
representation in each of these for all classes.
From the five best networks per analysis, we selected the one with best
performance for training, test and validation sets. The generalization
ability of the models was quantified by means of precision (P), recall
(R) and F1 in the validation dataset.
3 Results
The two DDI datasets differ considerably in terms of both their cover-
age and their significance gradings. The total number of drug pairs is
561, of which 210 (37%) are labelled as interacting in Lexicomp®,
while in Vidal® they are only 124 (22%). The overlap between them is
small (421 coincidences), with only 97 DDIs and 324 non-DDIs in
common. Regarding significance grading, the number of coincidences
is also limited (Table 1). Because of the different scales used in the two
DDI datasets and the small number of examples for some types (such
as contraindication/avoid combination), we combined the examples
into two gradings: minor and major DDIs.
Table 1. Occurrences of DDIs (n) by significance grade in Vidal® and Lexicomp® datasets
and common occurrences.
Significance grade
n Lexicomp n Vidal Matches
Vidal/Lexicomp
none/not known 351 437 324
to take into account 10 - -
Minor
precaution for use/monitor therapy 139 42 30
avoid/consider therapy modification 52 56 14
Major
contraindication/avoid combination 9 26 3
As shown in Table 2, results vary for the different datasets and the dif-
ferent input variables. In the case of the DDI-L dataset, the best results
are achieved using enzyme properties alone with a MLP network with
204-10-2 input, hidden and output neurons, respectively. The hidden
activation function is exponential and the output activation function is
logistic. The training algorithm is BFGS and the error function SOS.
Although the results show higher relevance of enzyme properties alone
compared to a combination of all variables, this difference is very small
(F1 64% vs 60%). In contrast, the use of PK/PD variables alone outper-
formed the other models in the DDI-V dataset, mainly because of a de-
crease in recall. As with the previous dataset, the performance of the
best model does not differ greatly from that achieved through use of all
combined properties (F1 58% vs 50%). This model is a MLP with 101-
14-2 input, hidden and output units trained using a BFGS algorithm.
The hidden and output activation functions are exponential and logistic
respectively, while the error function is SOS.
Table 2. Evaluation metrics for DDI prediction models in the DDI-L and DDI-V datasets for
validation instances.
DDI-Lexicomp dataset DDI-Vidal dataset
Variables Precision Recall F1 Precision Recall F1
PK/PD 0.75 0.429 0.545 0.577 0.60 0.588
Enzyme 0.649 0.632 0.64 0.571 0.32 0.41
PK/PD + enzyme 0.71 0.524 0.603 0.521 0.48 0.5
Regarding the prediction of the significance of DDIs, we created anoth-
er six different models using the same datasets DDI-L and DDI-V but
excluding the non-interacting pairs. As with the results presented
above, it is not possible to establish a relationship between a set of var-
iables and better models’ performance (Table 3). The two datasets are
unbalanced and both showed better performance for the majoritarian
class (minor in DDI-L dataset and major in DDI-V dataset). A larger
DDI dataset would solve this issue and would allow us to establish
more significance classes.
Table 3. Evaluation metrics for significance prediction models in the DDI-L and DDI-V
datasets for validation instances classified as minor and major DDIs.
DDI-Lexicomp dataset
Variables P minor P major R minor R major F1 minor F1 major
PK/PD 0.821 0.5 0.767 0.583 0.793 0.538
Enzyme 0.806 0.545 0. 833 0.5 0.82 0.522
PK/PD + enzyme 0.722 0. 333 0. 867 0.167 0.789 0.222
DDI-Vidal dataset
Variables P minor P major R minor R major F1 minor F1 major
PK/PD 0.857 0.882 0.75 0.938 0.8 0.909
Enzyme 0.25 0.833 0.625 0.938 0.357 0.882
PK/PD + enzyme 0. 667 0.867 0.75 0.813 0.71 0.839
Although the best results correspond to the model based on PK/PD var-
iables in the DDI-V dataset, the small number of instances (194 DDIs)
and the high results suggest that the model might be over-fitted. There-
fore, we believe that the most reliable results correspond to the model
trained using enzyme variables alone for minor DDIs (F1 82%) and the
model based on PK/PD properties alone for major DDIs (F1 54%) in
the DDI-L dataset. The first model is a MLP with 204-16-2 input, hid-
den and output units and hidden and output activation function than.
The second one is a MLP with 100-13-2 input, hidden and output neu-
rons with exponential and identity hidden and output activation func-
tions. Both models are trained with a BFGS algorithm and use SOS as
error function.
4 DISCUSSION AND CONCLUSIONS
Here, we have described a preliminary analysis for the prediction of
DDIs and their clinical significance through the creation of machine
learning models that exploit drug information collected from different
information sources available on the web.
Different research groups have applied machine learning for the predic-
tion of DDIs. These projects differ considerably in the original datasets,
the properties used as input variables, the machine learning methods
studied and the evaluation of their performance. Thus, a straight com-
parison with our results is difficult and beyond the scope of this project.
The closest work in terms of evaluation metrics is the prediction of
DDIs based on a probabilistic approach using PSL, which reported an
F1 of 67% on a dataset of 4,293 known interactions between 315 drugs.
This approach outperformed state-of-the-art works for DDI prediction
that obtained F1 values of 51% and 60% [18].
In our case, the ANN models led to F1 of 64% and 59% for the valida-
tion instances in the DDI-L and DDI-V datasets, respectively. We be-
lieve that there is still room for improvement, in part because the en-
zyme properties included in our approach represent only a relatively
small selection of those likely to lead to DDIs. However, there are
many different DDI mechanisms not related to metabolic processes
[11]. Therefore, in our future work we plan to include other drug-
protein relationships – including targets, transporters and carriers – that
will enable the identification of DDIs occurring by other mechanisms.
Also, we believe that representation of adverse effects profiles will be
very useful to identify DDIs due to the addition of side effects [30].
The previous approaches rely on unknown DDIs as examples of non-
interacting pairs, which may lead to incorrect predictions and hinder the
identification of new DDIs. In contrast, we have proposed a new strate-
gy based on the prediction of the significance of known DDIs. Our
model yielded interesting results, with an F1 of 82% for the best model.
To the best of our knowledge, this is the first work attempting the pre-
diction of DDI significance. The main limitations are, however, the size
of the current dataset, the inconsistent information between different
DDI sources and the frequent missing data for some input variables.
Automatic methods for knowledge extraction from the web is crucial
for the creation of a larger dataset of graded DDIs combining consistent
information from different sources, which will lead to more sophisti-
cated prediction models.
We believe that further improvements in this area could represent an
important tool in pharmacovigilance, for example as an initial signal
detection tool for the editorial boards maintaining and updating current
DDI compendia.
Acknowledgements
María Herrero-Zazo holds a C. W. Maplethorpe Postdoctoral Fellow-
ship for Pharmaceutical Education and Research at King’s College
London. Maxime Lille was supported by an internship provided from
the Fondation de l'Universite d'Auvergne, France.
References
[1] D. S. Wishart, C. Knox, A. C. Guo, S. Shrivastava, M.
Hassanali, P. Stothard, Z. Chang, and J. Woolsey, “DrugBank: a
comprehensive resource for in silico drug discovery and
exploration.,” Nucleic Acids Res., vol. 34, no. Database issue,
pp. D668–D672, 2006.
[2] T. Liu, Y. Lin, X. Wen, R. N. Jorissen, and M. K. Gilson,
“BindingDB: a web-accessible database of experimentally
determined protein-ligand binding affinities.,” Nucleic Acids
Res., vol. 35, no. Database issue, pp. D198–201, Jan. 2007.
[3] A. Gaulton, L. J. Bellis, a P. Bento, J. Chambers, M. Davies, A.
Hersey, Y. Light, S. McGlinchey, D. Michalovich, B. Al-
Lazikani, and J. P. Overington, “ChEMBL: a large-scale
bioactivity database for drug discovery.,” Nucleic Acids Res.,
vol. 40, no. Database issue, pp. D1100–7, Jan. 2012.
[4] M. Kuhn, M. Campillos, I. Letunic, L. J. Jensen, and P. Bork, “A
side effect resource to capture phenotypic effects of drugs,” Mol.
Syst. Biol., vol. 6, no. 343, pp. 1–6, 2010.
[5] R. Xu and Q. Wang, “Large-scale combining signals from both
biomedical literature and the FDA Adverse Event Reporting
System (FAERS) to improve post-marketing drug safety signal
detection,” BMC Bioinformatics, vol. 15, no. 1, p. 17, 2014.
[6] T. I. Oprea, S. Kim Nielsen, O. Ursu, J. J. Yang, O. Taboureau,
S. L. Mathias, I. Kouskoumvekaki, L. a. Sklar, and C. G.
Bologa, “Associating drugs, targets and clinical outcomes into
an integrated network affords a new platform for computer-aided
drug repurposing,” Mol. Inform., vol. 30, no. 2–3, pp. 100–111,
2011.
[7] P. Sanseau and J. Koehler, “Editorial: Computational methods
for drug repurposing,” Brief. Bioinform., vol. 12, no. 4, pp. 301–
302, 2011.
[8] M. Campillos, M. Kuhn, A. Gavin, L. J. Jensen, and P. Bork,
“Drug Target Identification Using Side-Effect Similarity,”
Science (80-. )., vol. 321, no. 5886, pp. 263–266, 2008.
[9] Z. Li, M. Huang, W. Zhong, Z. Liu, Y. Xie, Z. Dai, and X. Zou,
“Identification of drug–target interaction from interactome
network with ‘guilt-by-association’ principle and topology
features,” Bioinformatics, vol. 32, no. 7, pp. 1057–1064, Apr.
2016.
[10] WHO, “The Importance of Pharmacovigilance: Safety
Monitoring of Medicinal Products,” 2002.
[11] K. Baxter, Stockley’s Drug Interactions, 10th ed. London:
Pharmaceutical Press, 2013.
[12] M. Lea, S. E. Rognan, R. Koristovic, T. B. Wyller, and E.
Molden, “Severity and Management of Drug–Drug Interactions
in Acute Geriatric Patients,” Drugs Aging, vol. 30, no. 9, pp.
721–727, Sep. 2013.
[13] T. O’Hare, A. S. Corbin, and B. J. Druker, “Targeted CML
therapy: controlling drug resistance, seeking cure,” Curr. Opin.
Genet. Dev., vol. 16, no. 1, pp. 92–99, Feb. 2006.
[14] D. Tatro, Drug interaction facts 2010: The Authority on Drug
Interactions. St. Louis, MO: Wolters Kluwer Health, 2010.
[15] A. I. Vitry, “Comparative assessment of four drug interaction
compendia,” Br. J. Clin. Pharmacol., vol. 63, no. 6, pp. 709–
714, Jun. 2007.
[16] P. L. Smithburger, S. L. Kane-Gill, N. J. Benedict, B. A.
Falcione, and A. L. Seybert, “Grading the Severity of Drug-Drug
Interactions in the Intensive Care Unit: A Comparison Between
Clinician Assessment and Proprietary Database Severity
Rankings,” Ann. Pharmacother., vol. 44, no. 11, pp. 1718–1724,
Nov. 2010.
[17] F. Cheng and Z. Zhao, “Machine learning-based prediction of
drug-drug interactions by integrating drug phenotypic,
therapeutic, chemical, and genomic properties,” J. Am. Med.
Informatics Assoc., vol. 21, no. e2, pp. e278–e286, Oct. 2014.
[18] D. Sridhar, S. Fakhraei, and L. Getoor, “A probabilistic approach
for collective similarity-based drug–drug interaction prediction,”
Bioinformatics, vol. 450, no. June, p. btw342, Jun. 2016.
[19] S. Hunta, N. Aunsri, and T. Yooyativong, “Drug-Drug
Interactions prediction from enzyme action crossing through
machine learning approaches,” in 12th International Conference
on Electrical Engineering/Electronics, Computer,
Telecommunications and Information Technology (ECTI-CON),
2015, pp. 1–4.
[20] S. Polak, J. Brandys, and a. Mendyk, “Neural System for in
silico Drug-Drug Interaction Screening,” Int. Conf. Comput.
Intell. Model. Control Autom. Int. Conf. Intell. Agents, Web
Technol. Internet Commer., vol. 2, 2005.
[21] M. Herrero-Zazo, I. Segura-Bedmar, J. Hastings, and P.
Martínez, “DINTO: Using OWL Ontologies and SWRL Rules to
Infer Drug–Drug Interactions and Their Mechanisms,” J. Chem.
Inf. Model., vol. 55, no. 8, pp. 1698–1707, 2015.
[22] E. Sweetman and S. Martindale, “Martindale: the complete drug
reference.,” 2016. .
[23] “Clarke’s Analysis of Drugs and Poisons,” 2016. .
[24] R. S. Obach, F. Lombardo, and N. J. Waters, “Trend Analysis of
a Database of Intravenous Pharmacokinetic,” Pharmacology,
vol. 36, no. 7, pp. 1385–1405, 2008.
[25] S. Zhou, S. Yung Chan, B. Cher Goh, E. Chan, W. Duan, M.
Huang, and H. L. McLeod, “Mechanism-based inhibition of
cytochrome P450 3A4 by therapeutic drugs.,” Clin.
Pharmacokinet., vol. 44, no. 3, pp. 279–304, Jan. 2005.
[26] S. Preissner, K. Kroll, M. Dunkel, C. Senger, G. Goldsobel, D.
Kuzman, S. Guenther, R. Winnenburg, M. Schroeder, and R.
Preissner, “SuperCYP: a comprehensive database on
Cytochrome P450 enzymes including a tool for analysis of CYP-
drug interactions,” Nucleic Acids Res., vol. 38, no. Database, pp.
D237–D243, Jan. 2010.
[27] Lexicomp, “Lexi-InteractTM Online.” .
[28] Vidal, “Interctions medicamenteuses.” .
[29] C. M. Bishop, Neural networks for pattern recognition. Oxford :
Clarendon Press, 1995.
[30] P. Zhang, F. Wang, J. Hu, and R. Sorrentino, “Label Propagation
Prediction of Drug-Drug Interactions Based on Clinical Side
Effects,” Sci. Rep., vol. 5, p. 12339, Jul. 2015.