Prediction of adverse drug reactions associated with drug-drug interactions using hierarchical classification


0 
 

 
 
 
 
 
 

Prediction of adverse drug reactions associated with drug-drug interactions  

using hierarchical classification 

 

 

Catherine Kim1* and Nicholas Tatonetti2 

 

 

1. Jericho Senior High School, 99 Cedar Swamp Rd, Jericho, NY 11753 

2. Department of Biomedical Informatics, Department of Systems Biology, & 

Department of Medicine, Columbia University, 622 West 168th St. PH20 

New York, NY 10032 

 

*Corresponding author: cathy.kim@jerichoapps.org 

 

 

 

 

 

 

 

 

 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


1 
 

ABSTRACT 

Adverse drug reactions (ADRs) associated with drug-drug interactions (DDIs) represent a 

significant threat to public health. Unfortunately, most conventional methods for prediction of 

DDI-associated ADRs suffer from limited applicability and/or provide no mechanistic insight 

into DDIs. In this study, a hierarchical machine learning model was created to predict DDI-

associated ADRs and pharmacological insight thereof for any drug pair. Briefly, the model takes 

drugs’ chemical structures as inputs to predict their target, enzyme, and transporter (TET) 

profiles, which are subsequently utilized to assess occurrences of ADRs, with an overall 

accuracy of ~91%. The robustness of the model for ADR classification was validated with DDIs 

involving three widely prescribed drugs. The model was then applied for interstitial lung disease 

(ILD) associated with DDIs involving atorvastatin, identifying the involvement of multiple 

targets, enzymes, and transporters in ILD. The model presented here is anticipated to serve as a 

versatile tool for enhancing drug safety.  

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


2 
 

INTRODUCTION 

Adverse drug reactions (ADRs) represent a significant threat to public health worldwide, 

accounting for considerable morbidity and mortality with estimated costs of ~$500 billion 

annually [1, 2]. As ADRs continue to present a growing concern in modern health care systems, 

their identification and prevention are quintessential for improved drug safety and patient care. 

While drugs are subjected to preclinical in vitro safety profiling and clinical drug safety trials to 

assess drug safety, many ADRs occur in small subsets of the human population, making ADRs 

not readily detectable in advance [3]. Moreover, ADRs are more difficult to analyze when 

multiple, rather than single, drugs are administered, which has become common amongst a 

growing elderly population [4]. Drug-drug interactions (DDIs) between co-administered drugs 

appear in various forms of ADRs by different mechanisms, adding additional complexity [3].   

To better address DDI-associated ADRs, an understanding of their pharmacological 

mechanisms is strongly required. DDIs can occur when drugs compete for the same target [5]. 

DDIs also involve drug metabolizing enzymes (e.g. cytochrome P450 (CYP) enzymes) and 

influx and efflux drug transporters — all of which determine the adsorption, distribution, 

metabolism, and excretion (ADME) of drugs [6]. Thus, interference with target binding, enzyme-

mediated metabolization, and/or uptake and excretion of drugs may cause DDIs [5, 7-9].  

Moreover,  the comprehensive evaluation of entire TET profiles — many of which are dependent 

on the chemical structures of drugs — and their interplay between the drugs is critical for an 

enhanced understanding of DDI-associated ADRs  [10].   

With the current inability to reliably assess DDIs in preclinical testing and clinical trials 

and the complex nature of DDI-associated ADRs, a data-driven computational approach is well- 

suited for predicting such ADRs. This approach may benefit from extensive ADR databases, 

such as the FDA Adverse Event Reporting System, where data representative of a large 

population are collected from patients, clinicians, and pharmaceutical companies [11]. While 

various machine learning models have been previously developed for predicting DDI-associated 

ADRs with considerable accuracy, they suffer from major limitations. Most currently available 

models are based on drug similarity, providing accurate prediction only when the drug in 

question is similar to existing drugs with known TET profiles and/or ADR information [12-15]. 

This requirement makes these models not readily applicable when such information is 

unavailable, for example, when a drug is still under development.  Moreover, conventional 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


3 
 

models provide no pharmacological insight into DDI-associated ADRs. The availability of such 

a priori mechanistic understandings can lay out theoretical foundations on which a novel, 

effective pharmacological strategy can be developed. Overall, a novel computational approach to 

evaluate associations between DDIs and ADRs and to determine their molecular basis is urgently 

needed for better drug design and enhanced drug safety.  

This study reports the development of a hierarchical machine learning model to predict 

risks of various DDI-associated ADRs and their underlying pharmacological mechanisms. This 

model consists of two layers of classifiers for the prediction of TET profiles and occurrences of 

ADRs from chemical structures of a drug pair, requiring no drug similarity. The model was 

tested for its robustness with three case studies and then employed to elucidate the origin of an 

ADR of a rare disease, interstitial lung disease (ILD), associated with DDIs. 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


4 
 

METHODS 

All computations were performed with Python 3.8.3 on Jupyter Notebook 6.0.1 

Anaconda Navigator 1.9.12, unless otherwise noted. 

 

Statistical Analyses of ADRs 

Proportional reporting ratios (PRRs) were calculated for each drug pair and 

corresponding adverse drug reactions (ADRs) from the TWOSIDES v0.1 database [3] using the 

equation ��� �
�/�����

�/�����
 as described previously [3, 16]: where a = the number of patients who 

were administered the drug pair and were reported for the ADR, b = the number of patients who 

were administered the drug pair and were not reported for the ADR, c = the number of patients 

were not administered the drug pair and were reported for the ADR, and d = the number of 

patients who were not administered the drug pair and were not reported for the ADR. The 

TWOSIDES v0.1 database was created by application of propensity score matching to the FDA 

Adverse Event Reporting System  [11] in order to account for covariates in the dataset and 

eliminate potential bias [3] and used directly for the PRR calculations in this study. The numbers 

of unique drug pairs and ADRs used in this study were 211,990 and 12,726, respectively. 

For drug pairs containing one of three widely prescribed drugs — levothyroxine, 

omeprazole, and atorvastatin — all of their reported ADRs were extracted from the TWOSIDES 

v0.1 database using the Python pandas 1.1.2 library [17].  

 

Determination of Chemical Fingerprints of Drugs 

For all the drugs listed in the DrugBank 5.1.7 database [18], their chemical structures in 

the format of the simplified molecular-input line-entry system (SMILES) were obtained directly 

from the database or PubChem v1.6.3.b [18, 19]. The SMILES were stored in a 2D 

representation with the Python RDKit 2020.03.1 library and used to produce a chemical 

fingerprint for each drug by calculating its Molecular Access System (MACCS) keys [20]. 

Binary string representations of the MACCS keys were stored in a Python pandas 1.1.2 

dataframe [17].  

 

Construction of Target, Enzyme, and Transporter Profiles of Drugs 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


5 
 

Annotations about 4,263 unique targets, 316 unique enzymes, and 286 unique 

transporters were collected from DrugBank 5.1.7 [18] to create TET vectors of drugs using the 

Python NumPy 1.19.4 library [17]. Each of all unique TETs was assigned a position in a TET 

vector. For each drug, a TET vector was created with the Python NumPy 1.19.4 library to 

represent its pharmacological profile. Briefly, in each position of a drug’s TET vector, the value 

of “1” was assigned if any action of the drug (e.g., as a ligand, substrate, inhibitor, activator, 

agonist, antagonist) on each target, enzyme, and transporter was noted in DrugBank 5.1.7, 

whereas the value of “0” otherwise. 

 

Development of RFCs for Prediction of Target, Enzyme, and Transporter Profiles of Drugs  

Random forest classifiers (RFCs) were constructed for prediction of targets, enzymes, 

and transporters from the chemical structure of a drug using the Python sci-kit learn 0.23.2 

library [17]. These models formed the first layer of the hierarchical model. The dataset of the 

drugs’ MACCS keys and TET vectors were split into training (75% of dataset) and testing (25% 

of dataset) sets. RFCs were trained and tested to predict TET profiles from MACCS key 

representations (i.e., chemical fingerprints) of the drugs. During training and testing of the RFCs, 

model accuracies were measured and averaged.  

 

Development of a Model for Prediction of DDI-associated ADRs from TET Profiles of Drugs  

TET vectors of a drug pair were combined to form its TET matrix, which was then 

matched to the drug pair’s PRRs for various ADRs reported in the TWOSIDES v0.1 database. 

From the TWOSIDES v0.1 database, the calculated PRRs for different ADRs of each drug pair 

were categorically encoded with the value of “0” when 0 ≤ PRRs < 1, “1” when PRRs = 1, and 

“2” when PRRs > 1. The processed PRR dataset with the matched TET matrices for the drug 

pairs were split into training (75% of dataset) and testing (25% of dataset) sets. The machine 

learning algorithms, Random Forest Classifiers (RFC) [21], Logistic Regression (LR) [22], and 

Support Vector Machines (SVM) [23], were constructed as classifiers for ADR prediction using 

the Python sci-kit learn 0.23.2 library. The models were fit with default tuning parameters in the 

Python sci-kit learn 0.23.0 library. Model accuracies were measured using a 10-fold cross-

validation, as described elsewhere [24]. The SVM model was chosen as a second layer of the 

hierarchical model.  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


6 
 

 

Pathway Analysis of ADRs 

 The key genes/proteins involved in ILD associated with DDIs involving atorvastatin were 

determined from the pathway database Reactome [25]. Various repositories on gene/protein 

interactions and pathways, such as BioGRID [26], Proteomics DB [27], STRING [28], and 

CORUM [29], were applied to identify interactions between drug targets and genes/proteins 

involved in ADRs. The NCBI Gene database [30] was used to determine tissue-specific gene 

expression levels.  

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


7 
 

RESULTS AND DISCUSSION 

Model Overview 

 A hierarchical model to predict ADRs from the chemical structures of a drug pair was 

developed. In this model, the input variables were the chemical structures of drugs (Fig. 1A) and 

the output variables were the PRRs for various ADRs (Fig. 1B). A value of PRR > 1 indicates a 

high risk of an ADR for a given drug pair, whereas a PRR<1 suggests that a given ADR is less 

commonly reported for a drug pair, relative to other drug pairs [3, 16]. The PRR of 1 indicates a 

statistically neutral association between a drug pair and an ADR. Instead of attempting to 

correlate chemical structures of drugs directly with PRRs for ADRs, an intermediate tier of a 

pharmacological profile, namely a target, enzyme, and transporter (TET) profile, of drugs was 

introduced to connect chemical fingerprints of drugs with various ADRs (Fig. 1).  

The TET profiles depend on the drugs’ chemical structures [24, 31]. On the other end, the 

TET profiles determine the drugs’ ADME processes and their ultimate pharmacological actions 

through on- and off-targeting, all of which play a dominant role in ADRs [5-9]. Thus, the 

intermediate tier of the TET profiles serves as an essential component in the hierarchical 

machine learning model that connects the input variables (i.e., chemical structures of drugs) and 

the output variables (i.e., PRRs for various ADRs), while allowing for a deeper mechanistic 

understanding of ADRs (Fig. 1). 

  

Prediction of Target, Enzyme, and Transporter Profiles from Chemical Fingerprints of Drugs 

To predict the pharmacological profiles (i.e. TET profiles) from the chemical fingerprints 

(i.e. MACCS keys) of drugs, Random Forest Classifiers (RFCs) were constructed. The RFCs 

achieved high (>95%) accuracy across TET profile prediction (Table 1). The accuracies of these 

models were higher than other machine learning algorithms previously developed for the 

classification of drugs inhibiting a specific transporter [24]. For prediction of the entire TET 

profile, a testing accuracy of the RFC models is estimated to be 94.63% (=99.54% × 96.82% × 

98.19 %; Table 1).  

The TET prediction method presented in this study may address limitations of costly and 

often time-inefficient preclinical in vitro experiments to determine TET profiles [32]. Moreover, 

in vitro methods to assess drug’s action on transporters are not well established, presenting 

another limitation [33]. Other computational approaches, such as molecular docking, require the 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


8 
 

3D chemical structures of TETs [34], which are often lacking in newly identified drug targets 

and many transporters [24]. Requiring no such 3D structural information, the RFCs presented 

here allow for accurate, thorough, and inexpensive evaluations of TET profiles of drugs — even 

those under the development stage.  

ADR prediction from Target, Enzyme, and Transporter Profiles of Drug Pairs 

To predict ADRs of a drug pair from its TET profiles, Random Forest Classifier (RFC), 

Logistic Regression (LR), and Support Vector Machine (SVM) models were developed and 

evaluated using a 10-fold cross-validation. Compared to the RFC and LR models performing at 

mean classification accuracies (i.e. a fraction of a correctly classified ADR from a drug pair’s 

TET matrix) of 91.96% (Fig. 2A) and 86.63% (Fig. 2B), respectively, the SVM model 

outperformed with a greater mean classification accuracy of 95.73% (Fig. 2C).  

 

Application of the SVM model for DDI-associated ADRs Involving Three Major Drugs  

The SVM model was further tested for its robustness with DDIs involving three 

commonly prescribed drugs: levothyroxine, a synthetic hormone to treat hypothyroidism [35], 

omeprazole, a proton pump inhibitor for gastric acid-related disorders [36], and atorvastatin, an 

inhibitor of 3-hydroxy-3-methyl-glutaryl-CoA (HMG-CoA) reductase used for lowering lipids 

concentrations to treat hypercholesterolemia [37]. 

 

(1) Case study 1: Levothyroxine 

To apply the SVM model for DDIs associated with levothyroxine, eptifibatide was 

chosen as the concomitant drug, since the co-administration of levothyroxine and a blood thinner 

(e.g., eptifibatide) was previously found to cause a  bleeding-related ADR [38, 39] via inhibition 

of platelet aggregation [38, 40]. Levothyroxine has four major targets (integrin subunit αV 

(ITGAV), integrin subunit βIII (ITGB3), thyroid hormone receptor α (THRA), and thyroid 

hormone receptor β (THRB)), two metabolizing enzymes (cytochrome P450 (CYP) 2C8 

(CYP2C8) and UDP-glucuronosyltransferase 1A1 (UGT1A1)), and nine transporters (ATP-

binding cassette sub-family B member 1 (ABCB1), solute carrier (SLC) family 7 member 5 

(SLC7A5), SLC16A2, solute carrier organic anion transporter (SLCO) 1A2 (SLCO1A2), 

SLCO1B1, SLCO1B3, SLCO1C1, SLCO2B1, and SLCO4A1; Fig. 3A). THRA and THRB are 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


9 
 

nuclear receptors for levothyroxine, regulating transcription of hormone-responsive genes 

(referred to as genomic actions) [41, 42]. A heterodimeric complex, ITGAV-ITGB3, consisting 

of ITGAV and ITGB3 is another receptor for levothyroxine [41, 42], mediating the drug’s 

nongenomic actions, such as the proliferation of endothelial cells [41, 42]. Eptifibatide’s TET 

profile contains only one target, ITGB3 (Fig. 3A), which is complexed with integrin subunit αIIb 

[43] to form a heterodimeric complex, ITGB3-ITGA2B, mediating platelet aggregation [40, 43]. 

The co-administration of eptifibatide, which inhibits ITGB3-ITGA2B’s binding to fibrinogen, 

can reduce platelet aggregation [38, 40]. Thus, the sharing of ITGB3 by levothyroxine and 

eptifibatide (Fig. 3A) may be responsible for ADRs, as is the case with other combinations of 

drugs binding to the same pharmacological targets [44].  

The predictive power of the SVM model, particularly in the role of the shared target (i.e. 

ITGB3) in DDI-associated ADRs, was evaluated through comparisons with statistical results. 

Briefly, the PRRs for various ADRs associated with the co-administration of levothyroxine and 

eptifibatide were calculated from statistical analyses of TWOSIDES v0.1. The PRRs for 

levothyroxine alone and eptifibatide alone were calculated similarly using OFFSIDES v0.1 [3]. 

Then, for a given ADR, its PRR for the co-administration of levothyroxine and eptifibatide was 

subtracted from the average PRR for the single administrations of levothyroxine and eptifibatide 

(i.e. Δ PRR = average PRR for single administrations of levothyroxine and eptifibatide – PRR 

for their co-administration). Highly negative values of this difference (e.g., Δ PRR < the 99% 

confidence interval of Δ PRRs for the “No Change” group) are indicative of strong DDIs. The 

calculated PRR differences were then compared with prediction results, which were obtained 

using the SVM model upon removal of the ITGB3 as a target from the TET profile of 

eptifibatide. The comparison result suggests that the risks of most ADRs associated with strong 

DDIs between levothyroxine and eptifibatide are predicted to decrease if the TET profile of 

eptifibatide lacks ITGB3 as a target, suggesting the critical role of shared ITGB3 in the DDIs 

(Fig. 3B). 

 

(2) Case study 2: Omeprazole 

For a subsequent analysis, clopidogrel, an antiplatelet drug for the treatment of 

cardiovascular diseases [45], was chosen as the concomitant drug with omeprazole. Omeprazole 

has two major targets (aryl hydrocarbon receptor (AHR) and potassium-transporting ATPase α 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


10 
 

chain 1 (ATP4A)), nine metabolizing enzymes (CYP1A1, CYP1A2, CYP1B1, CYP2C8, 

CYP2C9, CYP2C18, CYP2C19, CYP2D6, and CYP3A4), and three transporters (ABCB1, ABC 

subfamily C member 3 (ABCC3), and ABC subfamily G member 2 (ABCG2); Fig. 4A) 

While omeprazole and clopidogrel share no same pharmacological targets, they have 

multiple enzymes and a single transporter in common (Fig. 4A). The concomitant use of 

omeprazole was found to lower the platelet inhibitory effects of clopidogrel [46, 47], increasing a 

risk of reinfarction [48, 49] and major cardiovascular events [50], compared to those receiving 

clopidogrel alone. Accordingly, the FDA has recommended not using omeprazole together with 

clopidogrel unless absolutely required [47]. 

To identify the major pharmacological determinants responsible for DDIs between 

omeprazole and clopidogrel, each of the targets, enzymes and transporters was removed one at a 

time from the TET profile of omeprazole and the effects of each removal on the DDI-associated 

ADRs were predicted using the SVM model developed in this study. The result from this 

analysis showed that CYP1A2, CYP2C8, CYP2C9, CYP2C19, CYP3A4, ABCB1, and ABCC3 

may play key roles in DDIs between omeprazole and clopidogrel (Fig. 4B). 

Consistent with this result, CYP2C19 was previously identified as a key enzyme to 

mediate DDIs between omeprazole and clopidogrel [47]. For its anti-platelet aggregation effect, 

clopidogrel needs to be converted by CYP2C19 to an active metabolite [51, 52], which prevents 

activation of P2RY12 required for platelet activation and aggregation [53]. Thus, omeprazole, an 

inhibitor of CYP2C19 [47, 54], can prevent the biotransformation of clopidogrel required for 

efficacy, causing DDI-associated ADRs [46-49]. The analysis also suggests the possible 

involvement of other enzymes and transporters in DDIs between omeprazole and clopidogrel 

(Fig. 4B), as supported by previous reports. For example, the metabolic activation of clopidogrel 

was also found to be mediated by CYP3A4 [55, 56]. In addition, the high likelihood of CYP1A2 

mediating DDIs involving omeprazole was previously proposed based on omeprazole’s ability to 

induce CYP1A2 activity [57, 58], though still under debate [47, 59, 60]. Omeprazole is a weak 

inhibitor of CYP2D6 relative to CYP2C19 and CYP3A4 [47], making CYP2D6-mediated DDIs 

less likely. An efflux transporter, ABCB1, may be an active player in these DDIs, as omeprazole 

interferes with the efflux of other drugs (e.g., digoxin [61] and nifedipine [62]) by ABCB1 [47]. 

Out of these enzymes and transporters, CYP2C19 was chosen as a key enzyme for 

subsequent comparative analyses. For this examination, Δ PRR (= an average PRR of 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


11 
 

omeprazole alone and clopidogrel alone − a PRR for the co-administration of omeprazole and 

clopidogrel) was calculated as a DDI index for each ADR, as described above. Δ PRR values 

were then compared with the PRR changes predicted by the SVM model when CYP2C19 was 

removed from the TET profile of omeprazole. Overall, the predictions and the calculations were 

in good agreement for ADRs significantly associated with DDIs (either negatively or positively, 

as judged by Δ PRR relative to the 99% interval of Δ PRRs for the “No change” group) between 

omeprazole and clopidogrel (Fig. 4C), supporting the role of CYP2C19 in their DDIs, as 

described elsewhere [47]. The comparative result also indicates that correct predictions of drug 

pairs with little to no DDIs (that is, Δ PRR ~ 0) are difficult with this model.  

 

(3) Case study 3: Atorvastatin 

To further validate the SVM model, a similar computational approach was applied to 

atorvastatin for its well-known DDI-associated ADR, myopathy [63-66]. Atorvastatin has five 

major targets (AHR, dipeptidyl peptidase 4 (DPP4), histone deacetylase 2 (HDAC2), 3-hydroxy-

3-methylglutaryl-coenzyme A reductase (HMGCR), and nuclear receptor subfamily 1 group I 

member 3 (NR1I3)), ten metabolizing enzymes (CYP2B6, CYP2C8, CYP2C9, CYP2C19, 

CYP2D6, CYP3A4, CYP3A5, CYP3A7, UGT1A1, and UDP-glucuronosyltransferase 1A3 

(UGT1A3)), and ten transporters (ABCB1, ATP-binding cassette sub-family B member 11 

(ABCB11), ATP-binding cassette sub-family C member (ABCC) 1 (ABCC1), ABCC2, ABCC4, 

ABCC5, SLCO1A2, SLCO1B1, SLCO1B3, and SLCO2B1; Fig. 5A). Atorvastatin’s actions on 

multiple TETs indicate its pharmacological complexity.   

The SVM model was used to derive TETs important in atorvastatin-induced myopathy 

through predicted PRR changes of drug pairs upon removal of each of the targets, enzymes, and 

transporters from atorvastatin’s TET profile. The SVM model predicted the importance of 

CYP2C9, CYP2C19, CYP3A4, UGT1A1, ABCB1, ABCB11, ABCC1, ABCC2, ABCC4, 

SLCO1A2, and SLCO1B1 in atorvastatin DDI-associated myopathy (Fig. 5B). Consistent with 

this result, the co-administration of drugs that are either inhibitors or substrates of CYP3A4 was 

found to decrease the metabolism of atorvastatin [67]. As a result, the plasma concentration of 

atorvastatin increases, leading to the onset of ADRs [67], including myopathy [66]. In addition, 

polymorphisms in the CYP2C19, UGT1A1, ABCB1 and SLCO1B1 genes are associated with 

systemic exposure of atorvastatin, an important risk factor for myopathy [68, 69]. Drugs 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


12 
 

inhibiting SLCO1B1 and ABCB1, most of which are CYP3A4 inhibitors [70], can cause DDIs 

with atorvastatin [70]. While the crucial role of CYP2C8 in DDIs involving other statins (e.g., 

simvastatin and lovastatin) has been documented [70, 71], the involvement of this enzyme in 

atorvastatin-mediated DDIs remains unclear.  

The model was then tested through comparisons between the predicted and calculated 

PRR changes of drug combinations for myopathy. Out of the identified key molecules, CYP3A4 

was chosen for further analyses due to its direct involvement in the onset of myopathy associated 

with DDIs involving atorvastatin, as reported previously [72, 73]. PRRs of drugs with higher 

degrees of DDIs (as judged by Δ PRR relative to the 99% confidence interval of Δ PRRs for the 

“No Change” group) with atorvastatin were predicted to decrease upon removal of atorvastatin’s 

CYP3A4 interaction (Fig. 5C), consistent with the literature reports on the importance of this 

CYP enzyme in atorvastatin-mediated myopathy [72, 73]. Similar to the results with omeprazole 

(case study 2), the accurate prediction of PRR changes for drug combinations with Δ PRR~ 0 

was difficult (Fig. 5C). 

  

Overall, the results obtained with levothyroxine, omeprazole and atorvastatin in this 

study demonstrate the high applicability of the machine learning model for predicting DDI-

associated ADRs and providing underlying pharmacological insight. 

 

Model Application for Interstitial Lung Disease Involving DDIs with Atorvastatin 

 Motivated by its high prediction power, the model developed in this study was applied to 

a rare yet life-threatening ADR, interstitial lung disease (ILD), associated with DDIs involving 

atorvastatin. The PRR of the single administration of atorvastatin for ILD was calculated from 

OFFSIDES v0.1 to analyze its statistical associations to ILD. A similar statistical analysis was 

extended to drug pairs containing atorvastatin for ILD.  Δ PRRs (i.e. DDI indices) were 

calculated and plotted with PRRs of the co-administration for ILD (Fig. 6). 

The PRR of atorvastatin alone for ILD was 1.01, a value indicative of a statistically 

neutral association between atorvastatin and ILD. Between ΔPRRs and PRRs for the co-

administration of atorvastatin, a strong negative linear relationship was detected (Fig. 6). The 

implication is that drug pairs of atorvastatin and concomitant drugs reported with high risks of 

ILD are due to DDIs between the drugs. When calculated from the linear regression line, Δ PRR 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


13 
 

is 0.0508 (~ 0, no DDI) for drug pairs showing no associations with ILD (i.e., PRR = 1; Fig. 6), 

as expected, validating this analysis.  

To identify important TETs in ILD associated with DDIs involving atorvastatin, the PRR 

changes of drug pairs were predicted by the model upon removal of each of the targets, enzymes, 

and transporters from atorvastatin’s TET profile (Fig. 7A). Among atorvastatin’s five targets, 

AHR, DPP4, HDAC2, HMGCR and NR1I3, the importance of DPP4 was minimal (Fig. 7A). In 

this analysis, different metabolizing enzymes seemed equally important, suggesting that DDI-

associated ILD involving atorvastatin may be mediated by a set of multiple enzymes, which may 

be responsible for previous contradictory findings on the role of metabolizing enzymes in this 

type of ADR [74, 75]. Among transporters, ABCB11, SLCO1B1 and SLCO1B3, all of which are 

primarily expressed in liver [30], were found to be important. ABCB11, a primary transporter of 

bile salts [76], was found to be involved in the biliary excretion of statins [77]. SLCO1B1 and 

SLCO1B3 are responsible for the uptake of atorvastatin into hepatocytes [78, 79]. Thus, the 

removal of these three transporters from atorvastatin’s TET profile may increase its plasma 

concentration, increasing risks of ILD [78, 79].  

To further distinguish among the five targets, a similar procedure was conducted to 

calculate the number of drug pairs with predicted increases in PRRs for ILD when a target was 

removed from atorvastatin’s TET profile (Fig. 7B). Interestingly, when atorvastatin’s action on 

HMGCR and DPP4 became nullified, PRRs for ILD further increased (Fig. 7B). The implication 

is that when its binding to HMGCR and DPP4 becomes ineffective, atorvastatin may bind to the 

other three targets more strongly, increasing risks of DDIs. No such PRR increases were 

observed with the removal of the other three targets (Fig. 7B). Thus, AHR, HDAC2, and NR1I3 

were identified as important targets for DDI-associated ILD involving atorvastatin.  

To validate these computational results, two key molecules, NR1I3 and ABCB11, which 

were identified by the SVM model in DDI-associated ILD with atorvastatin, were used for 

further analyses. For this examination, Δ PRRs (the average PRR for single administrations of 

atorvastatin and a concomitant drug – the PRR for their co-administration) were calculated and 

compared with PRR predictions by the SVM model for the drug pairs upon the removal of 

NR1I3 (Fig. 7C) and ABCB11 (Fig. 7D) from the TET profile of atorvastatin. In these analyses, 

PRRs for ILD were predicted to decrease with most drug pairs involving significant DDIs, when 

judged by Δ PRR < the 99% confidence interval of Δ PRRs for the “No change” group (Fig. 7C-

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


14 
 

D), supporting the critical roles of NR1I3 and ABCB11 in DDI-associated ILD involving 

atorvastatin. 

 A few potential pathological pathways underlying DDI-associated ILD involving a high 

plasma concentration of atorvastatin were determined around the three important targets — 

AHR, NR1I3, and HDAC2 — identified in this study. In this analysis, only genes/proteins 

significantly expressed in lung, as recorded in the NCBI Gene database were [30] considered. 

The analyses revealed the high likelihood that atorvastatin binding to AHR, NR1I3, and HDAC2 

may cause ILD through a major ILD mechanism — the dysregulation of surfactant production 

and homeostasis [80, 81]. Many interactors, such as SP1 transcription factor and estrogen 

receptor 1 (ESR1), are highly interconnected in the pathways around AHR, NR1I3, and HDAC2 

(Fig. 8A-B). Genes/proteins important in surfactant metabolism also create a strong network 

(Fig. 8A-B). Thus, any impact from AHR, NR1I3 and HDAC2 can be amplified, influencing one 

another in these pathways.  

Literature survey identified a few plausible routes atorvastatin can take to cause ILD. 

AHR is a transcription factor inducible by aromatic hydrocarbon-based xenobiotics, such as 

atorvastatin [82, 83]. Upon binding to a ligand, AHR is complexed with aryl hydrocarbon 

receptor nuclear translocator (ARNT; Fig. 8A) [82, 83]. The AHR/ARNT complex can then 

induce expression of AHR’s target genes, which code for enzymes and transporters required for 

xenobiotic metabolism [82, 84]. Activated AHR inhibits estrogen receptor (ESR1) activity [83], 

by redirecting ESR1 away from ESR1 target genes [85], such as ATP-binding cassette sub-

family A member 3 (ABCA3; Fig. 8A) [86]. ABCA3 plays a critical role in the formation of 

pulmonary surfactant by transporting phospholipids from the endoplasmic reticulum to a 

surfactant storage organelle in type II epithelial cells [87, 88]. Thus, the binding of atorvastatin to 

AHR may cause pulmonary surfactant metabolism dysfunction by downregulating the ABCA3 

gene via inhibition of ESR1 activity [89]. Different interactors (e.g., histone acetyltransferase 

p300 (EP300)) may be involved in ILD, amplifying the effects of AHR through networks of 

other interactors and ILD genes/proteins.  

In addition, NR1I3 is a nuclear receptor that mediates transcriptional activation of target 

genes required for the metabolism and elimination of xenobiotics [82, 90, 91], such as CYP2B6 

and CYP3A4 [90, 92]. Upon the binding of xenobiotics, NR1I3 is dephosphorylated for nuclear 

translocation and transactivation [93], which requires reduced SRC kinase activity [94]. On the 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


15 
 

other hand, P2Y purinoceptor 2 (P2RY2), a G protein-coupled receptor, activates SRC [95, 96] 

in order to promote surfactant secretion from alveolar type II cells [97]. Thus, the binding of 

atorvastatin to NR1I3 [98] can dysregulate normal surfactant secretion via interference with SRC 

kinase activity. 

Atorvastatin inhibits HDAC2 [99]. The connections between HDAC2 and ILD 

genes/proteins are highly interconnected, also sharing many interactors with AHR and NR1I3, 

suggesting the existence of many different paths that cause ILD from HDAC2 inhibition (Fig. 

8B). Interestingly, the binding of atorvastatin to HDAC2 may be related to atorvastatin’s 

beneficial effect against cancer [100]. HDAC2 plays a key role in the epigenetic regulation of 

gene expression in cancer [101] and HDAC2 inhibitors (e.g., atorvastatin [99]) can display anti-

cancer activities [101]. The anticancer effect of atorvastatin was also statistically analyzed. In 

this analysis, combinations of atorvastatin and a drug that have PRRs > 1 for ILD were identified 

and their PRRs for lung cancer were also calculated. None of combinations of atorvastatin and 

concomitant drugs that had PRRs > 1 for ILD had PRRs > 1 for lung cancer, supporting 

atorvastatin’s anti-cancer effects. Overall, the model-based computational examinations and 

pathway analyses revealed key molecules important in DDI-associated ILD involving 

atorvastatin and proposed underlying pathological pathways.  

 

CONCLUSIONS 

 This study presented a novel computational approach to accurately predict the 

occurrences of ADRs using a machine learning model consisting of hierarchically structured 

classifiers. The hierarchical model presented here addresses the limitations of conventional 

models relying on drug similarity for the prediction of ADRs. The method developed here is 

based on TET profile-dependencies of ADRs derived from drugs’ chemical structures, requiring 

no high chemical similarity of drugs. Given basic structural characteristics of drugs, this 

hierarchical model integrating the RFCs for TET profile prediction and the SVM for specific 

DDI-associated ADRs can accurately predict ADRs with an overall ~91% (=94.63% for TET 

prediction × 95.73% for ADR prediction) accuracy. As DDIs typically appearing as various 

forms of ADRs have been another primary issue in past predictions of DDI-associated ADRs 

[102], the presented model deconvolutes this complexity, as judged by its accurate prediction of 

various ADRs for any drug pair. In addition, pharmacological insight offered by the hierarchical 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


16 
 

model was successfully connected to pathway analyses underlying ADRs, making the described 

computational approach powerful for not only predicting the occurrence of DDI-associated 

ADRs but also enhancing mechanistic understandings.  

Notably, the constructed model can accurately predict TET profiles and DDI-associated 

ADRs from most basic information of drugs - chemical structures – for any pair. Thus, the model 

presented in this study can also be used for drug design. For example, the MACCS keys 

descriptors can be manipulated and inputted into the hierarchical model to identify a drug’s key 

structural characteristics that increase a risk of ADRs. Once the structural hot spots are 

identified, an array of drug variants with different chemical moieties at the locations can be 

designed and evaluated for DDIs and ADRs prior to synthesis. As a result, many drugs can 

readily be evaluated for their potential DDIs in advance, avoiding costly preclinical and clinical 

tests. Thus, the hierarchical model developed is anticipated to pave new way to enhance drug 

safety and reduce drug development costs.   

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


17 
 

REFERENCES 

1. Kovacevic, M., Vezmar Kovacevic, S., Radovanovic, S., Stevanovic, P. & Miljkovic, B. 

(2019). Adverse drug reactions caused by drug-drug interactions in cardiovascular 

disease patients: introduction of a simple prediction tool using electronic screening 

database items. Curr Med Res Opin, 35(11), 1873-1883. 

2. Luo, J., Eldredge, C., Cho, C. C. & Cisler, R. A. (2016). Population Analysis of Adverse 

Events in Different Age Groups Using Big Clinical Trials Data. JMIR Med Inform, 4(4), 

e30. 

3. Tatonetti, N. P., Ye, P. P., Daneshjou, R. & Altman, R. B. (2012). Data-driven prediction 

of drug effects and interactions. Sci Transl Med, 4(125), 125ra131. 

4. Nguyen, T., Wong, E. & Ciummo, F. (2020). Polypharmacy in older adults: practical 

applications alongside a patient case. The Journal for Nurse Practitioners, 16(3), 205-

209. 

5. Kaneko, S. & Nagashima, T. (2020). Drug Repositioning and Target Finding Based on 

Clinical Evidence. Biol Pharm Bull, 43(3), 362-365. 

6. Neve, E. P., Artursson, P., Ingelman-Sundberg, M. & Karlgren, M. (2013). An integrated 

in vitro model for simultaneous assessment of drug uptake, metabolism, and efflux. Mol 

Pharm, 10(8), 3152-3163. 

7. Benet, L. Z., Cummins, C. L. & Wu, C. Y. (2003). Transporter-enzyme interactions: 

implications for predicting drug-drug interactions from in vitro data. Curr Drug Metab, 

4(5), 393-398. 

8. Kalliokoski, A. & Niemi, M. (2009). Impact of OATP transporters on pharmacokinetics. 

Br J Pharmacol, 158(3), 693-705. 

9. Poirier, A., Funk, C., Lave, T. & Noe, J. (2007). New strategies to address drug-drug 

interactions involving OATPs. Curr Opin Drug Discov Devel, 10(1), 74-83. 

10. Jamal, S., Goyal, S., Shanker, A. & Grover, A. (2017). Predicting neurological Adverse 

Drug Reactions based on biological, chemical and phenotypic properties of drugs using 

machine learning models. Sci Rep, 7(1), 872. 

11. Sakaeda, T., Tamon, A., Kadoyama, K. & Okuno, Y. (2013). Data mining of the public 

version of the FDA Adverse Event Reporting System. Int J Med Sci, 10(7), 796-803. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


18 
 

12. Brouwers, L., Iskar, M., Zeller, G., van Noort, V. & Bork, P. (2011). Network neighbors 

of drug targets contribute to drug side-effect similarity. PLoS One, 6(7), e22187. 

13. Munoz, E., Novacek, V. & Vandenbussche, P. Y. (2016). Using Drug Similarities for 

Discovery of Possible Adverse Reactions. AMIA Annu Symp Proc, 2016, 924-933. 

14. Seo, S., Lee, T., Kim, M. H. & Yoon, Y. (2020). Prediction of Side Effects Using 

Comprehensive Similarity Measures. Biomed Res Int, 2020, 1357630. 

15. Vilar, S., Uriarte, E., Santana, L., Lorberbaum, T., Hripcsak, G., Friedman, C. & 

Tatonetti, N. P. (2014). Similarity-based modeling in large-scale prediction of drug-drug 

interactions. Nat Protoc, 9(9), 2147-2163. 

16. Noguchi, Y., Ueno, A., Otsubo, M., Katsuno, H., Sugita, I., Kanematsu, Y., Yoshida, A., 

Esaki, H., Tachi, T. & Teramachi, H. (2018). A simple method for exploring adverse 

drug events in patients with different primary diseases using spontaneous reporting 

system. BMC Bioinformatics, 19(1), 124. 

17. Stancin, I. & Jovic, A. (2019). An overview and comparison of free Python libraries for 

data mining and big data analysis. 2019 42nd International Convention on Information 

and Communication Technology, Electronics and Microelectronics (MIPRO). Anal 

Chem, 88(7), 3539-3547. 

18. Wishart, D. S., Feunang, Y. D., Guo, A. C., Lo, E. J., Marcu, A., Grant, J. R., Sajed, T., 

Johnson, D., Li, C., Sayeeda, Z., Assempour, N., Iynkkaran, I., Liu, Y., Maciejewski, A., 

Gale, N., Wilson, A., Chin, L., Cummings, R., Le, D., Pon, A., Knox, C. & Wilson, M. 

(2018). DrugBank 5.0: a major update to the DrugBank database for 2018. Nucleic Acids 

Res, 46(D1), D1074-D1082. 

19. Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., 

Thiessen, P. A., Yu, B., Zaslavsky, L., Zhang, J. & Bolton, E. E. (2019). PubChem 2019 

update: improved access to chemical data. Nucleic Acids Res, 47(D1), D1102-D1109. 

20. Du, H., Cai, Y., Yang, H., Zhang, H., Xue, Y., Liu, G., Tang, Y. & Li, W. (2017). In 

Silico Prediction of Chemicals Binding to Aromatase with Machine Learning Methods. 

Chem Res Toxicol, 30(5), 1209-1218. 

21. Breiman, L. (2001). Random Forests. Machine Learning, 45, 5-32. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


19 
 

22. Dreiseitl, S. & Ohno-Machado, L. (2002). Logistic regression and artificial neural 

network classification models: a methodology review. J Biomed Inform, 35(5-6), 352-

359. 

23. Cortes, C. & Vapnik, V. (1995). Support-Vector Networks. Machine Learning, 20, 

273−297. 

24. Khuri, N., Zur, A. A., Wittwer, M. B., Lin, L., Yee, S. W., Sali, A. & Giacomini, K. M. 

(2017). Computational Discovery and Experimental Validation of Inhibitors of the 

Human Intestinal Transporter OATP2B1. J Chem Inf Model, 57(6), 1402-1413. 

25. Jassal, B., Matthews, L., Viteri, G., Gong, C., Lorente, P., Fabregat, A., Sidiropoulos, K., 

Cook, J., Gillespie, M., Haw, R., Loney, F., May, B., Milacic, M., Rothfels, K., Sevilla, 

C., Shamovsky, V., Shorser, S., Varusai, T., Weiser, J., Wu, G., Stein, L., Hermjakob, H. 

& D'Eustachio, P. (2020). The reactome pathway knowledgebase. Nucleic Acids Res, 

48(D1), D498-D503. 

26. Oughtred, R., Stark, C., Breitkreutz, B. J., Rust, J., Boucher, L., Chang, C., Kolas, N., 

O'Donnell, L., Leung, G., McAdam, R., Zhang, F., Dolma, S., Willems, A., Coulombe-

Huntington, J., Chatr-Aryamontri, A., Dolinski, K. & Tyers, M. (2019). The BioGRID 

interaction database: 2019 update. Nucleic Acids Res, 47(D1), D529-D541. 

27. Schmidt, T., Samaras, P., Frejno, M., Gessulat, S., Barnert, M., Kienegger, H., Krcmar, 

H., Schlegl, J., Ehrlich, H. C., Aiche, S., Kuster, B. & Wilhelm, M. (2018). 

ProteomicsDB. Nucleic Acids Res, 46(D1), D1271-D1281. 

28. Szklarczyk, D., Gable, A. L., Lyon, D., Junge, A., Wyder, S., Huerta-Cepas, J., 

Simonovic, M., Doncheva, N. T., Morris, J. H., Bork, P., Jensen, L. J. & Mering, C. V. 

(2019). STRING v11: protein-protein association networks with increased coverage, 

supporting functional discovery in genome-wide experimental datasets. Nucleic Acids 

Res, 47(D1), D607-D613. 

29. Giurgiu, M., Reinhard, J., Brauner, B., Dunger-Kaltenbach, I., Fobo, G., Frishman, G., 

Montrone, C. & Ruepp, A. (2019). CORUM: the comprehensive resource of mammalian 

protein complexes-2019. Nucleic Acids Res, 47(D1), D559-D563. 

30. Brown, G. R., Hem, V., Katz, K. S., Ovetsky, M., Wallin, C., Ermolaeva, O., Tolstoy, I., 

Tatusova, T., Pruitt, K. D., Maglott, D. R. & Murphy, T. D. (2015). Gene: a gene-

centered information resource at NCBI. Nucleic Acids Res, 43(Database issue), D36-42. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


20 
 

31. Yu, W. & MacKerell, A. D., Jr. (2017). Computer-Aided Drug Design Methods. Methods 

Mol Biol, 1520, 85-106. 

32. Meibohm, B. & Derendorf, H. (2002). Pharmacokinetic/pharmacodynamic studies in 

drug product development. J Pharm Sci, 91(1), 18-31. 

33. Sudsakorn, S., Bahadduri, P., Fretland, J. & Lu, C. (2020). 2020 FDA Drug-drug 

Interaction Guidance: A Comparison Analysis and Action Plan by Pharmaceutical 

Industrial Scientists. Curr Drug Metab, 21(6), 403-426. 

34. El-Hachem, N., Haibe-Kains, B., Khalil, A., Kobeissy, F. H. & Nemer, G. (2017). 

AutoDock and AutoDockTools for Protein-Ligand Docking: Beta-Site Amyloid 

Precursor Protein Cleaving Enzyme 1(BACE1) as a Case Study. Methods Mol Biol, 1598, 

391-403. 

35. Jonklaas, J., Bianco, A. C., Bauer, A. J., Burman, K. D., Cappola, A. R., Celi, F. S., 

Cooper, D. S., Kim, B. W., Peeters, R. P., Rosenthal, M. S., Sawka, A. M. & American 

Thyroid Association Task Force on Thyroid Hormone, R. (2014). Guidelines for the 

treatment of hypothyroidism: prepared by the american thyroid association task force on 

thyroid hormone replacement. Thyroid, 24(12), 1670-1751. 

36. Shi, S. & Klotz, U. (2008). Proton pump inhibitors: an update of their clinical use and 

pharmacokinetics. Eur J Clin Pharmacol, 64(10), 935-951. 

37. Mohammadkhani, N., Gharbi, S., Rajani, H. F., Farzaneh, A., Mahjoob, G., Hoseinsalari, 

A. & Korsching, E. (2019). Statins: Complex outcomes but increasingly helpful treatment 

options for patients. Eur J Pharmacol, 863, 172704. 

38. Amoroso, G., van Boven, A. J., van Veldhuisen, D. J., Tio, R. A., Balje-Volkers, C. P., 

Petronio, A. S. & van Oeveren, W. (2001). Eptifibatide and abciximab exhibit equivalent 

antiplatelet efficacy in an experimental model of stenting in both healthy volunteers and 

patients with coronary artery disease. J Cardiovasc Pharmacol, 38(4), 633-641. 

39. Coutinho, J., Field, J. B. & Sule, A. A. (2018). Armour(R) Thyroid Rage - A Dangerous 

Mixture. Cureus, 10(4), e2523. 

40. Schror, K. & Weber, A. A. (2003). Comparative pharmacology of GP IIb/IIIa 

antagonists. J Thromb Thrombolysis, 15(2), 71-80. 

41. Davis, P. J., Leonard, J. L. & Davis, F. B. (2008). Mechanisms of nongenomic actions of 

thyroid hormone. Front Neuroendocrinol, 29(2), 211-218. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


21 
 

42. Hammes, S. R. & Davis, P. J. (2015). Overlapping nongenomic and genomic actions of 

thyroid hormone and steroids. Best Pract Res Clin Endocrinol Metab, 29(4), 581-593. 

43. Phillips, D. R., Charo, I. F. & Scarborough, R. M. (1991). GPIIb-IIIa: the responsive 

integrin. Cell, 65(3), 359-362. 

44. Torres, N. B. & Altafini, C. (2016). Drug combinatorics and side effect estimation on the 

signed human drug-target network. BMC Syst Biol, 10(1), 74. 

45. Lee, C. H., Franchi, F. & Angiolillo, D. J. (2020). Clopidogrel drug interactions: a review 

of the evidence and clinical implications. Expert Opin Drug Metab Toxicol, 16(11), 

1079-1096. 

46. Ho, P. M., Maddox, T. M., Wang, L., Fihn, S. D., Jesse, R. L., Peterson, E. D. & 

Rumsfeld, J. S. (2009). Risk of adverse outcomes associated with concomitant use of 

clopidogrel and proton pump inhibitors following acute coronary syndrome. JAMA, 

301(9), 937-944. 

47. Ogawa, R. & Echizen, H. (2010). Drug-drug interaction profiles of proton pump 

inhibitors. Clin Pharmacokinet, 49(8), 509-533. 

48. Evanchan, J., Donnally, M. R., Binkley, P. & Mazzaferri, E. (2010). Recurrence of acute 

myocardial infarction in patients discharged on clopidogrel and a proton pump inhibitor 

after stent placement for acute myocardial infarction. Clin Cardiol, 33(3), 168-171. 

49. Stockl, K. M., Le, L., Zakharyan, A., Harada, A. S., Solow, B. K., Addiego, J. E. & 

Ramsey, S. (2010). Risk of rehospitalization for patients using clopidogrel with a proton 

pump inhibitor. Arch Intern Med, 170(8), 704-710. 

50. Gaglia, M. A., Jr., Torguson, R., Hanna, N., Gonzalez, M. A., Collins, S. D., Syed, A. I., 

Ben-Dor, I., Maluenda, G., Delhaye, C., Wakabayashi, K., Xue, Z., Suddath, W. O., 

Kent, K. M., Satler, L. F., Pichard, A. D. & Waksman, R. (2010). Relation of proton 

pump inhibitor use after percutaneous coronary intervention with drug-eluting stents to 

outcomes. Am J Cardiol, 105(6), 833-838. 

51. Savi, P., Pereillo, J. M., Uzabiaga, M. F., Combalbert, J., Picard, C., Maffrand, J. P., 

Pascal, M. & Herbert, J. M. (2000). Identification and biological activity of the active 

metabolite of clopidogrel. Thromb Haemost, 84(5), 891-896. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


22 
 

52. Umemura, K., Furuta, T. & Kondo, K. (2008). The common gene variants of CYP2C19 

affect pharmacokinetics and pharmacodynamics in an active metabolite of clopidogrel in 

healthy subjects. J Thromb Haemost, 6(8), 1439-1441. 

53. Herbert, J. M. & Savi, P. (2003). P2Y12, a new platelet ADP receptor, target of 

clopidogrel. Semin Vasc Med, 3(2), 113-122. 

54. Li, X. Q., Andersson, T. B., Ahlstrom, M. & Weidolf, L. (2004). Comparison of 

inhibitory effects of the proton pump-inhibiting drugs omeprazole, esomeprazole, 

lansoprazole, pantoprazole, and rabeprazole on human cytochrome P450 activities. Drug 

Metab Dispos, 32(8), 821-827. 

55. Clarke, T. A. & Waskell, L. A. (2003). The metabolism of clopidogrel is catalyzed by 

human cytochrome P450 3A and is inhibited by atorvastatin. Drug Metab Dispos, 31(1), 

53-59. 

56. Farid, N. A., Payne, C. D., Small, D. S., Winters, K. J., Ernest, C. S., 2nd, Brandt, J. T., 

Darstein, C., Jakubowski, J. A. & Salazar, D. E. (2007). Cytochrome P450 3A inhibition 

by ketoconazole affects prasugrel and clopidogrel pharmacokinetics and 

pharmacodynamics differently. Clin Pharmacol Ther, 81(5), 735-741. 

57. Diaz, D., Fabre, I., Daujat, M., Saint Aubert, B., Bories, P., Michel, H. & Maurel, P. 

(1990). Omeprazole is an aryl hydrocarbon-like inducer of human hepatic cytochrome 

P450. Gastroenterology, 99(3), 737-747. 

58. Rost, K. L., Brosicke, H., Brockmoller, J., Scheffler, M., Helge, H. & Roots, I. (1992). 

Increase of cytochrome P450IA2 activity by omeprazole: evidence by the 13C-[N-3-

methyl]-caffeine breath test in poor and extensive metabolizers of S-mephenytoin. Clin 

Pharmacol Ther, 52(2), 170-180. 

59. Rizzo, N., Padoin, C., Palombo, S., Scherrmann, J. M. & Girre, C. (1996). Omeprazole 

and lansoprazole are not inducers of cytochrome P4501A2 under conventional 

therapeutic conditions. Eur J Clin Pharmacol, 49(6), 491-495. 

60. Xiaodong, S., Gatti, G., Bartoli, A., Cipolla, G., Crema, F. & Perucca, E. (1994). 

Omeprazole does not enhance the metabolism of phenacetin, a marker of CYP1A2 

activity, in healthy volunteers. Ther Drug Monit, 16(3), 248-250. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


23 
 

61. Oosterhuis, B., Jonkman, J. H., Andersson, T., Zuiderwijk, P. B. & Jedema, J. N. (1991). 

Minor effect of multiple dose omeprazole on the pharmacokinetics of digoxin after a 

single oral dose. Br J Clin Pharmacol, 32(5), 569-572. 

62. Soons, P. A., van den Berg, G., Danhof, M., van Brummelen, P., Jansen, J. B., Lamers, 

C. B. & Breimer, D. D. (1992). Influence of single- and multiple-dose omeprazole 

treatment on nifedipine pharmacokinetics and effects in healthy subjects. Eur J Clin 

Pharmacol, 42(3), 319-324. 

63. Bataillard, M., Beyens, M. N., Mounier, G., Vergnon-Miszczycha, D., Bagheri, H. & 

Cathebras, P. (2019). Muscle Damage Due to Fusidic Acid-Statin Interaction: Review of 

75 Cases From the French Pharmacovigilance Database and Literature Reports. Am J 

Ther, 26(3), e375-e379. 

64. Boonmuang, P., Nathisuwan, S., Chaiyakunapruk, N., Suwankesawong, W., Pokhagul, 

P., Teerawattanapong, N. & Supsongserm, P. (2013). Characterization of Statin-

Associated Myopathy Case Reports in Thailand Using the Health Product Vigilance 

Center Database. Drug Saf, 36(9), 779-787. 

65. Brahmachari, B. & Chatterjee, S. (2015). Myopathy induced by statin-ezetimibe 

combination: Evaluation of potential risk factors. Indian J Pharmacol, 47(5), 563-564. 

66. du Souich, P., Roederer, G. & Dufour, R. (2017). Myotoxicity of statins: Mechanism of 

action. Pharmacol Ther, 175, 1-16. 

67. Hirota, T. & Ieiri, I. (2015). Drug-drug interactions that interfere with statin metabolism. 

Expert Opin Drug Metab Toxicol, 11(9), 1435-1447. 

68. Marusic, S., Lisicic, A., Horvatic, I., Bacic-Vrca, V. & Bozina, N. (2012). Atorvastatin-

related rhabdomyolysis and acute renal failure in a genetically predisposed patient with 

potential drug-drug interaction. Int J Clin Pharm, 34(6), 825-827. 

69. Stormo, C., Bogsrud, M. P., Hermann, M., Asberg, A., Piehler, A. P., Retterstol, K. & 

Kringen, M. K. (2013). UGT1A1*28 is associated with decreased systemic exposure of 

atorvastatin lactone. Mol Diagn Ther, 17(4), 233-237. 

70. Neuvonen, P. J., Niemi, M. & Backman, J. T. (2006). Drug interactions with lipid-

lowering drugs: mechanisms and clinical relevance. Clin Pharmacol Ther, 80(6), 565-

581. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


24 
 

71. Ogilvie, B. W., Zhang, D., Li, W., Rodrigues, A. D., Gipson, A. E., Holsapple, J., Toren, 

P. & Parkinson, A. (2006). Glucuronidation converts gemfibrozil to a potent, 

metabolism-dependent inhibitor of CYP2C8: implications for drug-drug interactions. 

Drug Metab Dispos, 34(1), 191-197. 

72. Canestaro, W. J., Austin, M. A. & Thummel, K. E. (2014). Genetic factors affecting 

statin concentrations and subsequent myopathy: a HuGENet systematic review. Genet 

Med, 16(11), 810-819. 

73. Gluba-Brzozka, A., Franczyk, B., Toth, P. P., Rysz, J. & Banach, M. (2016). Molecular 

mechanisms of statin intolerance. Arch Med Sci, 12(3), 645-658. 

74. Fernandez, A. B., Karas, R. H., Alsheikh-Ali, A. A. & Thompson, P. D. (2008). Statins 

and interstitial lung disease: a systematic review of the literature and of food and drug 

administration adverse event reports. Chest, 134(4), 824-830. 

75. Zanger, U. M. & Schwab, M. (2013). Cytochrome P450 enzymes in drug metabolism: 

regulation of gene expression, enzyme activities, and impact of genetic variation. 

Pharmacol Ther, 138(1), 103-141. 

76. Pedersen, J. M., Matsson, P., Bergstrom, C. A., Hoogstraate, J., Noren, A., LeCluyse, E. 

L. & Artursson, P. (2013). Early identification of clinically relevant drug interactions 

with the human bile salt export pump (BSEP/ABCB11). Toxicol Sci, 136(2), 328-343. 

77. Hirano, M., Maeda, K., Hayashi, H., Kusuhara, H. & Sugiyama, Y. (2005). Bile salt 

export pump (BSEP/ABCB11) can transport a nonbile acid substrate, pravastatin. J 

Pharmacol Exp Ther, 314(2), 876-882. 

78. Hirota, T., Fujita, Y. & Ieiri, I. (2020). An updated review of pharmacokinetic drug 

interactions and pharmacogenetics of statins. Expert Opin Drug Metab Toxicol, 16(9), 

809-822. 

79. Zhang, L., Lv, H., Zhang, Q., Wang, D., Kang, X., Zhang, G. & Li, X. (2019). 

Association of SLCO1B1 and ABCB1 Genetic Variants with Atorvastatin-induced 

Myopathy in Patients with Acute Ischemic Stroke. Curr Pharm Des, 25(14), 1663-1670. 

80. Akella, A. & Deshpande, S. B. (2013). Pulmonary surfactants and their role in 

pathophysiology of lung disorders. Indian J Exp Biol, 51(1), 5-22. 

81. Whitsett, J. A., Wert, S. E. & Weaver, T. E. (2010). Alveolar surfactant homeostasis and 

the pathogenesis of pulmonary disease. Annu Rev Med, 61, 105-119. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


25 
 

82. Larigot, L., Juricek, L., Dairou, J. & Coumoul, X. (2018). AhR signaling pathways and 

regulatory functions. Biochim Open, 7, 1-9. 

83. Matthews, J. & Gustafsson, J. A. (2006). Estrogen receptor and aryl hydrocarbon receptor 

signaling pathways. Nucl Recept Signal, 4, e016. 

84. Hankinson, O. (1995). The aryl hydrocarbon receptor complex. Annu Rev Pharmacol 

Toxicol, 35, 307-340. 

85. Matthews, J., Wihlen, B., Thomsen, J. & Gustafsson, J. A. (2005). Aryl hydrocarbon 

receptor-mediated transcription: ligand-dependent recruitment of estrogen receptor alpha 

to 2,3,7,8-tetrachlorodibenzo-p-dioxin-responsive promoters. Mol Cell Biol, 25(13), 

5317-5328. 

86. Lin, C. Y., Strom, A., Vega, V. B., Kong, S. L., Yeo, A. L., Thomsen, J. S., Chan, W. C., 

Doray, B., Bangarusamy, D. K., Ramasamy, A., Vergara, L. A., Tang, S., Chong, A., 

Bajic, V. B., Miller, L. D., Gustafsson, J. A. & Liu, E. T. (2004). Discovery of estrogen 

receptor alpha target genes and response elements in breast tumor cells. Genome Biol, 

5(9), R66. 

87. Klugbauer, N. & Hofmann, F. (1996). Primary structure of a novel ABC transporter with 

a chromosomal localization on the band encoding the multidrug resistance-associated 

protein. FEBS Lett, 391(1-2), 61-65. 

88. Yamano, G., Funahashi, H., Kawanami, O., Zhao, L. X., Ban, N., Uchida, Y., Morohoshi, 

T., Ogawa, J., Shioda, S. & Inagaki, N. (2001). ABCA3 is a lamellar body membrane 

protein in human lung alveolar type II cells. FEBS Lett, 508(2), 221-225. 

89. Shulenin, S., Nogee, L. M., Annilo, T., Wert, S. E., Whitsett, J. A. & Dean, M. (2004). 

ABCA3 gene mutations in newborns with fatal surfactant deficiency. N Engl J Med, 

350(13), 1296-1303. 

90. Auerbach, S. S., Dekeyser, J. G., Stoner, M. A. & Omiecinski, C. J. (2007). CAR2 

displays unique ligand binding and RXRalpha heterodimerization characteristics. Drug 

Metab Dispos, 35(3), 428-439. 

91. Qatanani, M. & Moore, D. D. (2005). CAR, the continuously advancing receptor, in drug 

metabolism and disease. Curr Drug Metab, 6(4), 329-339. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


26 
 

92. Goodwin, B., Hodgson, E., D'Costa, D. J., Robertson, G. R. & Liddle, C. (2002). 

Transcriptional regulation of the human CYP3A4 gene by the constitutive androstane 

receptor. Mol Pharmacol, 62(2), 359-365. 

93. Mutoh, S., Osabe, M., Inoue, K., Moore, R., Pedersen, L., Perera, L., Rebolloso, Y., 

Sueyoshi, T. & Negishi, M. (2009). Dephosphorylation of threonine 38 is required for 

nuclear translocation and activation of human xenobiotic receptor CAR (NR1I3). J Biol 

Chem, 284(50), 34785-34792. 

94. Groll, N., Petrikat, T., Vetter, S., Wenz, C., Dengjel, J., Gretzmeier, C., Weiss, F., Poetz, 

O., Joos, T. O., Schwarz, M. & Braeuning, A. (2016). Inhibition of beta-catenin signaling 

by phenobarbital in hepatoma cells in vitro. Toxicology, 370, 94-105. 

95. Liu, J., Liao, Z., Camden, J., Griffin, K. D., Garrad, R. C., Santiago-Perez, L. I., 

Gonzalez, F. A., Seye, C. I., Weisman, G. A. & Erb, L. (2004). Src homology 3 binding 

sites in the P2Y2 nucleotide receptor interact with Src and regulate activities of Src, 

proline-rich tyrosine kinase 2, and growth factor receptors. J Biol Chem, 279(9), 8212-

8218. 

96. Woods, L. T., Jasmer, K. J., Munoz Forti, K., Shanbhag, V. C., Camden, J. M., Erb, L., 

Petris, M. J. & Weisman, G. A. (2020). P2Y2 receptors mediate nucleotide-induced 

EGFR phosphorylation and stimulate proliferation and tumorigenesis of head and neck 

squamous cell carcinoma cell lines. Oral Oncol, 109, 104808. 

97. Rice, W. R. & Singleton, F. M. (1986). P2-purinoceptors regulate surfactant secretion 

from rat isolated alveolar type II cells. Br J Pharmacol, 89(3), 485-491. 

98. Rezen, T., Hafner, M., Kortagere, S., Ekins, S., Hodnik, V. & Rozman, D. (2017). 

Rosuvastatin and Atorvastatin Are Ligands of the Human Constitutive Androstane 

Receptor/Retinoid X Receptor alpha Complex. Drug Metab Dispos, 45(8), 974-976. 

99. Lin, Y. C., Lin, J. H., Chou, C. W., Chang, Y. F., Yeh, S. H. & Chen, C. C. (2008). 

Statins increase p21 through inhibition of histone deacetylase activity and release of 

promoter-associated HDAC1/2. Cancer Res, 68(7), 2375-2383. 

100. Archibugi, L., Arcidiacono, P. G. & Capurso, G. (2019). Statin use is associated to a 

reduced risk of pancreatic cancer: A meta-analysis. Dig Liver Dis, 51(1), 28-37. 

101. Bolden, J. E., Peart, M. J. & Johnstone, R. W. (2006). Anticancer activities of histone 

deacetylase inhibitors. Nat Rev Drug Discov, 5(9), 769-784. 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


27 
 

102. Noguchi, Y., Tachi, T. & Teramachi, H. (2020). Subset Analysis for Screening Drug-

Drug Interaction Signal Using Pharmacovigilance Database. Pharmaceutics, 12(8),  

 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


28 
 

 Training Accuracy Testing Accuracy 

Target 99.69% 99.54% 

Enzyme 98.74% 96.82% 

Transporter 99.39% 98.19% 

 

Table 1. Average training and testing accuracies for target, enzyme, and transporter 

prediction by the RFC models.  

 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


29 
 

FIGURE CAPTIONS 

Fig. 1. Hierarchical classification model overview for the prediction of DDI-associated 

ADRs from drugs’ chemical structures through predictions of (A) TET profiles from 

chemical fingerprints and (B) ADRs from TET matrices of drug pairs. (A) Drugs’ chemical 

structures were represented with MACCS keys and used as features to predict TET profiles of 

drugs using a Random Forest Classifier (RFC). (B) TET profiles of a drug pair were combined 

into a TET matrix, which was then used as a feature to predict encoded PRRs for all ADRs in 

RFC, Logistic Regression (LR), and Support Vector Machine (SVM) models. 

 

Fig. 2. Repeated 10-fold cross-validation for (A) Random Forest Classifier (RFC), (B) 

Logistic Regression (LR), and (C) Support Vector Machine (SVM) models.  

 

Fig. 3. Adverse drug reactions associated with drug-drug interactions (DDIs) between 

levothyroxine and eptifibatide. (A) Target, enzyme, and transporter (TET) profiles of 

levothyroxine and eptifibatide. Y: the presence of a drug’s action on TETs. N: the absence of a 

drug’s action on TETs. (B) Comparisons between ΔPRRs (as DDI indices) and PRR 

prediction upon removal of ITGB3 from eptifibatide’s TET profile for ADRs associated 

with the co-administration of levothyroxine and eptifibatide. For a given ADR, the average 

PRR of single administrations of levothyroxine and eptifibatide – the PRR of their co-

administration was calculated, and the PRR change for co-administration of levothyroxine and 

eptifibatide was predicted upon alteration of TET profiles of eptifibatide for integrin β-3 

(ITGB3) from Y to N. Y: Inclusion of ITGB3 in eptifibatide’s TET profile. N: Removal of 

ITGB3 from eptifibatide’s TET profile. **: Outside of the 99% confidence interval of the “No 

change” group.  

 

Fig. 4. Adverse drug reactions associated with drug-drug interactions (DDIs) between 

omeprazole and clopidogrel. (A) Target, enzyme, and transporter (TET) profiles of 

omeprazole and clopidogrel. Y: the presence of a drug’s action on TETs. N: the absence of a 

drug’s action on TETs. (B) The impacts of omeprazole’s TET profile on its DDI-associated 

ADRs with clopidogrel. The PRR changes for ADRs associated with co-administration of 

omeprazole and clopidogrel were calculated using the SVM model when each of omeprazole’s 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


30 
 

TETs was removed. (C) Comparisons between ΔPRRs (as DDI indices) and PRR predictions 

upon removal of CYP2C19 from omeprazole’s TET profile for ADRs associated with the 

co-administration of omeprazole and clopidogrel. For a given ADR, the average PRR of 

single administrations of omeprazole and clopidogrel – the PRR of their co-administration was 

calculated, and the PRR change for co-administration of omeprazole and clopidogrel was 

predicted using the SVM model upon alteration of the TET profile of omeprazole for CYP2C19 

from Y to N. Y: Inclusion of CYP2C19 in omeprazole’s TET profile. N: Removal of CYP2C19 

from omeprazole’s TET profile. **: Outside of the 99% confidence interval of the “No change” 

group.  

 

Fig. 5. Myopathy associated with drug-drug interactions (DDIs) involving atorvastatin. (A) 

Target, enzyme, and transporter (TET) profiles of atorvastatin and concomitant drugs, 

such as ramipril and warfarin. Y: the presence of a drug’s action on TETs. N: the absence of a 

drug’s action on TETs. (B) The impacts of atorvastatin’s TET profile on its DDI-associated 

ADR of myopathy with various concomitant drugs. The PRR changes for myopathy 

associated with co-administration of atorvastatin and other drugs were calculated using the SVM 

model when each of atorvastatin’s TETs was removed. (C) Comparisons between ΔPRRs (as 

DDI indices) and PRR predictions upon removal of CYP3A4 from atorvastatin’s TET 

profile for myopathy associated with the co-administration of atorvastatin and a 

concomitant drug. For myopathy, the average PRRs of single administrations of atorvastatin 

and a concomitant drug – the PRRs of their co-administration were calculated, and PRR changes 

for the co-administration of atorvastatin and the drug were predicted using the SVM model upon 

alteration of the TET profile of atorvastatin for cytochrome P450 3A4 (CYP3A4) from Y to N. 

Y: Inclusion of CYP3A4 in atorvastatin’s TET profile. N: Removal of CYP3A4 from 

atorvastatin’s TET profile. **: Outside of the 99% confidence interval of the “No change” group.  

 

Fig. 6. Drug-drug interactions between atorvastatin and concomitant drugs for interstitial 

lung disease (ILD). For ILD, Δ PRR (= the average PRR of single administrations of 

atorvastatin and the concomitant drug – the PRR of their co-administration) was calculated, and 

plotted with the PRR for their co-administration. 

 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


31 
 

Fig. 7. Interstitial lung disease (ILD) associated with drug-drug interactions (DDIs) 

involving atorvastatin. (A) The impacts of atorvastatin’s TET profile on its DDI-associated 

ILD with various concomitant drugs. The PRR changes for ILD associated with the co-

administration of atorvastatin and other drugs were calculated using the SVM model when each 

of atorvastatin’s TETs was removed. (B) The number of drug pairs containing atorvastatin 

with a predicted increase in PRRs for ILDs when each of atorvastatin’s targets was 

removed. (C) Comparisons between ΔPRRs (as DDI indices) and PRR predictions upon 

removal of (A) NR1I3 and (B) ABCB11 from atorvastatin’s TET profile for ILD associated 

with the co-administration of atorvastatin and a concomitant drug. For ILD, the average 

PRR of single administrations of atorvastatin and a concomitant drug– the PRR of their co-

administration was calculated, and the PRR changes for the co-administration of atorvastatin and 

the drug were predicted using the SVM model upon alteration of the TET profile of atorvastatin 

for (A) NR1I3 and (B) ABCB11 from Y to N. Y: Inclusion of (A) NR1I3 and (B) ABCB11 in 

atorvastatin’s TET profile. N: Removal of (A) NR1I3 and (B) ABCB11 from atorvastatin’s TET 

profile. **: Outside of the 99% confidence interval of the “No change” group. 

 

Fig. 8. Pathway analyses for the enhanced risk of ILD, associated with DDIs involving 

atorvastatin created around (A) AHR and NR1I3 and (B) HDAC2. Interactions among 

genes/proteins were determined using an array of bioinformatics databases, including BioGRID , 

Proteomics DB, STRING, CORUM and Reactome.  

 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


32 
 

 

Fig. 1 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


33 
 

 

 

Fig. 2 

 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


34 
 

 

Fig. 3  

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


35 
 

 

Fig. 4  

 

 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


36 
 

 

Fig. 5  

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


37 
 

 

Fig. 6 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


38 
 

 

Fig. 7 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


39 
 

 

Fig. 8 

  

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/


40 
 

KEYWORDS 

adverse drug reactions; drug-drug interaction; drug safety; hierarchical classification; machine 

learning; prediction, metabolizing enzyme, target; transporter. 

 

 

.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made 

The copyright holder for this preprintthis version posted February 11, 2021. ; https://doi.org/10.1101/2021.02.10.430512doi: bioRxiv preprint 

https://doi.org/10.1101/2021.02.10.430512
http://creativecommons.org/licenses/by-nc-nd/4.0/