Molecular Psychiatry https://doi.org/10.1038/s41380-020-01011-y

ARTICLE

Drug repurposing for use disorders: integration of computational prediction, clinical corroboration, and mechanism of action analyses

1,2 1 1 3,4,5 6 1 Mengshi Zhou ● QuanQiu Wang ● Chunlei Zheng ● A. John Rush ● Nora D. Volkow ● Rong Xu

Received: 10 September 2020 / Revised: 11 December 2020 / Accepted: 17 December 2020 © The Author(s), under exclusive licence to Springer Nature Limited 2021

Abstract Morbidity and mortality from opioid use disorders (OUD) and other substance use disorders (SUD) is a major public health crisis, yet there are few medications to treat them. There is an urgency to accelerate SUD medication development. We present an integrated drug repurposing strategy that combines computational prediction, clinical corroboration using electronic health records (EHRs) of over 72.9 million patients and mechanisms of action analysis. Among top-ranked repurposed candidate drugs, , , , , and atomoxetine were associated with increased odds of OUD remission (adjusted odds ratio: 1.51 [1.38–1.66], 1.90 [1.66–2.18], 1.38 [1.31–1.46], 1.37 [1.29–1.46], 1.48 [1.25–1.76], p value < 0.001, 1234567890();,: 1234567890();,: respectively). Genetic and functional analyses showed these five candidate drugs directly target multiple OUD-associated genes including BDNF, CYP2D6, OPRD1, OPRK1, OPRM1, HTR1B, POMC, SLC6A4 and OUD-associated pathways, including opioid signaling, G-protein activation, receptors, and GPCR signaling. In summary, we developed an integrated drug repurposing approach and identified five repurposed candidate drugs that might be of value for treating OUD patients, including those suffering from comorbid conditions.

Introduction diseases of the brain that in its most severe manifestation result in an escalating and an uncontrollable use of the drug Substance use disorders (SUD) when manifested in their despite its adverse consequences to the individual [1–3]. moderate or severe presentation are considered chronic SUD are estimated to affect 10.8% of the adult population in the US [4] and account for 1.5% of global disease burden [5]. In the US overdose death associated with were five times higher in 2016 than 1999, leading the US These authors contributed equally: Mengshi Zhou, QuanQiu Wang government to declare the opioid crisis a public health Supplementary information The online version of this article (https:// emergency [6]. As of now opioid overdose fatalities as well doi.org/10.1038/s41380-020-01011-y) contains supplementary as drug fatalities in general have continue to rise in the US material, which is available to authorized users. [7]. There are few medications for SUD and while effective they are limited by limited utilization and high relapse rates * Rong Xu [email protected] [8]. There are no approved medications to treat , marihuana, , benzodiazepine, or inhalant 1 Center for Artificial Intelligence in Drug Discovery, Case Western use disorders. The traditional drug discovery process for Reserve University, Cleveland, OH, USA medication development is lengthy and costly [8, 9]. In 2 Department of Mathematics & Statistics, Saint Cloud State addition, the very modest investment from the pharmaceutical University, Saint Cloud, MN, USA sector in SUDs has limited the discovery of new medications 3 Duke University School of Medicine, Durham, NC, USA to a greater extent than for other neuro-psychiatric disorders. 4 Duke-National University of Singapore, Singapore, Singapore Thus, novel strategies to evaluate the potential for repurposing 5 Texas-Tech Health Sciences Center, Permian Basin, Odessa, TX, existing drugs to treat SUD could accelerate access to addi- USA tional medications [8, 10]. 6 National Institute on Drug Abuse, National Institutes of Health, Drug repurposing is a strategy that can help identify Bethesda, MD, USA potential new therapies for complex diseases, including M. Zhou et al.

Fig. 1 Flowcharts of the integrate SUD repurposing system. a Computational drug prediction. b EHR-based clinical corroboration. c Mechanism of action analysis.

SUD [11]. For example, lofexidine was approved for the the phenome-driven drug discovery system identifies drug treatment of and was recently approved for the candidates that share similar drug phenotypes (i.e., side treatment of acute opioid withdrawal [12]. With the accu- effects) and/or common targets with drugs causing or mulation of relevant data in machine-actionable formats, treating SUD. Second, we then performed retrospective data-driven computational approaches have been developed case-control studies to evaluate the clinical efficacy of to automate the drug repurposing process [13–18]. promising repositioned candidate drugs using EHRs of 72.9 However, clinically validating a promising repurposed million patients (20% of the US population). Finally, we candidate drug generated by computational algorithms examined potential mechanisms of action of promising remains a challenge. repurposed candidate drugs in targeting SUD by developing We propose an integrated drug repurposing strategy that data-driven informatics approaches. The knowledge and combines computational-based drug prediction, patient data generated by our study (i.e., promising candidate drugs Electronic Health Records (EHRs)-based clinical corro- with both supporting clinical evidence and potential boration and mechanisms of action analysis. First, we mechanism of actions) can set the foundation of experi- developed phenome-driven network-based drug discovery mental testing in animal models for SUD or for pilot testing system that prioritized repurposed anti-SUD candidate in clinical trials. drugs. The phenotypic and genetic relationships among drugs, drug phenotypes (side effects or SEs), and genes were modeled using the novel context-sensitive network Material and methods (CSN)-based modeling techniques that we previously developed [16, 19, 20, 22, 23]. Then phenome-driven net- Our study entailed three steps (Fig. 1): (1) We constructed a work-based prioritization approaches, which we recently drug side effect-gene (DSEG) computational drug prediction developed both for understanding disease mechanisms and system to prioritize drugs to treat SUD. (2) We performed for drug discovery [14, 16, 18, 19, 21–23], were used to retrospectives case-control studies to evaluate top candidate prioritized repurposed candidate drugs based on their phe- drugs using patient EHR data. (3) We performed genetic and notypic and genetic relevance to the input disease (i.e., pathway enrichment analysis of top candidate drugs to SUD). In our study, we significantly leveraged the context- understand their potential mechanisms of action. sensitive drug side effect network that we constructed based on known drug side effects. Side effects are observable Computational drug prediction phenotypes of drugs manifested at the level of the whole- body system and are mediated by a drug interacting with its Constructing drug side effect-gene prediction system on- or off-targets through a cascade of downstream pathway perturbations. While mechanisms of action (on- and off- We constructed a DSEG drug prediction system that models targets) of many drugs and the underlying molecular the interconnections among drugs, side effects, and genes mechanisms of diseases remain largely unknown, we can using the CSN-based modeling techniques that we pre- infer novel connections between drugs and diseases (drug viously developed [19–23]. The DSEG system included two repurposing) based on the observed drug and disease phe- networks (Fig. 1a): (1) a drug phenotypic (drug side effects) notypes as well as known drug-targets and protein–protein network (DPN) and (2) a protein–protein interaction net- connections [22]. Instead of directly identifying drugs that work (PPIN). DPN was constructed using drug-SE pairs target SUD’s mechanisms, which remain largely unknown, from the Side Effect Resource (SIDER) database [24] and Drug repurposing for opioid use disorders: integration of computational prediction, clinical. . . consisted of 1430 drug nodes, 4251 SE nodes, and 145,321 using de-identified population-level EHR data collected by drug-SE edges. PPIN was directly constructed from the IBM Watson Health from 360 hospitals and 317,000 protein–protein interactions in STRING [25] and consisted providers across 50 states from 1999 up to August, 2020, of 17,906 gene nodes and 2,091,571 gene–gene edges. representing 20% of the US population [30]. The EHRs are Drug nodes (899 drugs) on DPN were connected to gene de-identified according to the Health Insurance Portability and nodes (1021 genes) on PPIN using drug-target associations Accountability Act and the Health Information Technology from the DrugBank database [26]. for Economic and Clinical Health Act standards. After the de- identification process, the curation process normalizes the data Prioritize anti-SUD drug candidates by mapping key elements to widely-accepted biomedical terminologies [31]. Specifically, disease terms are coded using We prioritized candidate drugs using the network-based the Systematized Nomenclature of Medicine-Clinical Terms ranking algorithms that we previously used for drug (SNOMED-CT), a global standard for health terms that pro- repurposing, gene discovery, and gut microbial metabolite vides the core general terminology for EHRs [32]. We have discovery for the disease [16–23, 27–29]. In brief, given an recently used this EHR database for drug repositioning for input or seeds (drug abuse, drug dependence, and drug Alzheimer’s disease [23, 33] and for studying the risk and withdrawal syndrome in our study), the ranking score for outcomes of COVID-19 in patients with SUD, mental dis- each drug on the entire network was iteratively updated by: orders, and cancers [34–37]. The EHR data are de-identified and aggregated (not patient-level) and institutional review ¼ð À αÞ T þ α ; ð Þ pk þ 1 1 T pk p0 1 board review was exempt. We used the odds of remission from OUD as the in which α (α = 0.15 in this study) denotes the probability outcome measure. The status of OUD was based on the of restarting from the seed nodes at each step. The algorithm diagnosis of “Opioid dependence (disorder)” (SNOMED- À < À8 was iterated until convergence ( pk þ 1 pk 10 ). CT concept code 75544000) and the outcome measure We used D and G to represent DPN and PPIN, respec- “remission of OUD” was based on the diagnosis of “Opioid tively. T denotes the transition matrix of DSEG network: dependence in remission (disorder)” (SNOMED-CT con-  cept code 191821007). At the time of the study, there were TDD TDG 326,620 patients with OUD and 28,940 of whom had a T ¼ ; ð2Þ TGD TGG record of OUD in remission. We investigated the associa- tions between top-ranked drugs and the odds of remission In Eq. (2), the diagonal sub-matrices Txx (x ∈ {D, G}) from OUD. were calculated through normalizing the adjacency matrix of For a candidate drug A for repurposing, we identified a D and G, the off-diagonal sub-matrices Txy (x, y ∈ {D, G}) study cohort of OUD patients diagnosed with drug A’s were calculated through normalizing the adjacency matrix of original indication. We obtained drug indication informa- the bipartite network connecting D and G. tion from the National Library of Medicine DailyMed, a comprehensive and up-to-date resource of medication con- Evaluation tent and labeling as found in medication package inserts [38] and other authoritative health information sites. For We evaluated how the drug prediction algorithm ranked the example, “depressive disorder” is the initial treatment dis- four drugs (, , , and ease of . In the study cohort, the exposure group ) approved for the treatment of OUD or opioid (drug group) was comprised of OUD patients with depres- overdose reversal. Lofexidine, an alpha 2 adrenergic ago- sive disorder who were prescribed with drug A (citalopram) nists that was recently approved for the treatment of acute before remission of their OUD. The comparison group (no opioid withdrawal [10], was not included since it was not in drug group) was comprised of OUD patients with depres- the SIDER database and on the DPN. The average ranking sive disorders but without drug A (citalopram). We com- of these four drugs among all FDA-approved drugs was pared the odds of remission from opioid dependence calculated. between the exposure group and the comparison group. The EHR data are de-identified population-level Electronic health records-based large-scale clinical (not patient-level) data, therefore we used odds ratios (OR) corroborations for predicted anti-SUD drug instead of regression analyses, as previously done by studies candidates that used the Explorys EHR database [23–25, 34–37]. For a given input set of patient characteristics (e.g., age, gender, We performed a retrospective case-control study to eval- race, diagnosis, comorbidities), the Explorys Explore uate top repurposed candidate drugs for treating OUD Cohort Discovery tool built a patient cohort by querying the M. Zhou et al.

EHR database for patients matching the inputs. Patients enrichment analysis as in our previous studies [45, 46]. To with missing values for the input queries were not included identify genetic pathways significantly enriched for OUD, in the returned cohorts. The associations between candidate genetic pathways were first obtained from MSigDB for each drugs and remission were estimated by an OR for remission OUD gene. For each pathway, we assessed its probability of comparing the exposure group versus the comparison group being associated with the given set of 19 OUD-associated in which both groups suffered OUD. The adjusted odds genes as compared to its probability of being associated with ratios (AORs), 95% confidence intervals (CIs), and p values the same number of randomly selected genes. The random were calculated using Cochran–Mantel–Haenszel (CMH) process was repeated 1000 times and a t-test was used to method [39], controlling for age (adult: 18–64 years, assess the statistical significance. We obtained a total of seniors: >64 years), gender (male, female), race (Caucasian 53 significantly enriched pathways for 19 OUD genes, which and non-Caucasian), original indication, and existing treat- include G-protein activation, opioid signaling, gap junction, ments for opioid dependence. A two-sided p value < 0.05 and serotonin receptors, among others. was considered as statistical significance. All analyses were Similarly, we identified genetic pathways significantly done using R, version 3.6.3. Drugs with insufficient data enriched for each candidate drug. Genetic targets of each (e.g., rarely used drugs in OUD patients) to perform the candidate drug were obtained from the STITCH database. For CMH analysis were excluded. Among the top 20 drugs, one each drug-targeted gene, its associated pathways were drug was excluded. obtained from the gene–pathway pairs form the MSigDB database. For each pathway, we assessed its probability of Analysis of top anti-SUD drug candidates in the being associated with the given set of drug-associated genes context of OUD at the genetic and functional levels as compared to its probability of being associated with the same number of randomly selected genes. For example, a Genetic-level analysis total of 317 pathways were significantly enriched for the set of 62 olanzapine-associated genes, which include 41 of the 53 We obtained 19 OUD-associated genes from the published OUD-associated pathways. For each repurposed candidate literature [40–42], including APBB2, BDNF, CNIH3, drug, we identified OUD pathways that are also significantly COMT, CYP2D6, DRD2, DRD4, HTR1B, KCNC1, enriched for the drug by intersecting drug-associated path- KCNG2, OPRD1, OPRK1, OPRM1, PDYN, PENK, ways with OUD-associated genes and pathways. POMC, RGMA, SLC6A4, and TPH2. For each repurposed candidate drug, we used the STITCH (Search Tool for Interactions of Chemicals) database to obtain its associated Results genes. STITCH contains data on the interactions between 500,000 small molecules and 9.6 million proteins from Our drug prediction system ranked approved SUD 2031 organisms [43]. In this study, we used chemical-gene treatments highly associations found in the human body. The scores of chemical-gene associations ranged from 100 to 999 and we Table 1 lists the ranks (in top percentage) for the four used the median score of 500 as the cutoff value. For medications used clinically to treat OUD and opioid over- example, at cutoff score of 500, olanzapine (a top repur- dose reversal (naloxone). Our system prioritized these four posed candidate) is associated with 62 genes. We identified drugs within the top 3.4% among 1430 FDA-approved OUD genes that are directly targeted by a candidate drug by drugs on the network. intersecting OUD-associated genes with drug-associated Table 2 shows the top 20 predicted drug candidates for genes. For example, among 62 gene targets of olanzapine, SUD. Overall, 17 out of 20 are implicated as SUD treat- seven are OUD-associated genes (BDNF, CYP2D6, DRD2, ments through different sources, including FDA drug DRD4, HTR1B, POMM, and SLC6A4).

Functional-level analysis Table 1 Ranks for approved OUD medications generated by our prediction system. We then investigated how the repurposed anti-OUD candidate Drug name Rank drugs are functionally related to OUD by directly targeting Methadone Top 3.2% OUD-associated pathways. We used rich pathway informa- Buprenorphine Top 2.7% tion from the Molecular Signatures Database (MSigDB), Naltrexone Top 0.91% currently the most comprehensive resource of 17,779 anno- Naloxone Top 2.0% tated pathways and gene sets [44], to perform pathway Drug repurposing for opioid use disorders: integration of computational prediction, clinical. . .

Table 2 Top 20-ranked Rank Drug Original Indication Evidence repurposed drug candidates. 1 Diabetic neuropathic pain NCT00142883 2 Pain/acute pain 3 Pain/acute pain 4 Pain/acute pain NCT00218374 5 Pain/acute pain NCT00218361 6 Citalopram NCT01535573 7 Schizophrenia PMID24628830 8 Atomoxetine ADHD NCT01498549 9 Tramadol Pain NCT00301210, NCT01188421, NCT00142896, NCT03365817 10 Bupropion Depression NCT02111798 11 Mirtazapine Depression NCT02541526, NCT00249444, NCT00322309, NCT00732901 12 Cortisol Rheumatoid arthritis NCT00759005, NCT01718964 13 Naltrexone OUD FDA-approved 14 Pain NCT00687713 15 ADHD NCT00421603 16 Methylprednisolone Rheumatoid arthritis 17 Triazolam Insomnia NCT00218166 18 Olanzapine Schizophrenia NCT02643355 19 Seizures NCT00396734, NCT00421603 20 Depression NCT00142779 NCT: SUD drugs from clinical trials. PMID: SUD drugs from biomedical literature.

Fig. 2 Odds ratios of remission from opioid dependence and the corresponding 95% CI of 10 of top 20-ranked drugs. Triazolam was excluded due to insufficient cases for patients with opioid dependence.

labels, published literature, and clinical trials. In addition, not surprising that opioid analgesics were ranked high by 17 of the top 20 drugs were approved for treating the DSEG system. central nervous system-related diseases particularly for depression and pain, which are frequently comorbid with EHR-based analysis provided evidence that top- SUD [47]. We noticed that top-ranked drugs include ranked drug candidates have clinical effectiveness some opioid analgesics with high addiction risks (e.g., in OUD patients fentanyl). Since two of the approved OUD medications (e.g., methadone and buprenorphine) are opioid analge- We performed case-control studies to evaluate top 20 sics but with slower than fentanyl, it is repdurposed candidates. Figure 2 shows the AORs and 95% M. Zhou et al.

Table 3 OUD genes targeted by Drug Total target genes Target OUD genes OUD genes candidate drugs. Atomoxetine 11 3 BDNF, CYP2D6, SLC6A4 Bupropion 22 2 CYP2D6, SLC6A4 Mirtazapine 30 5 CYP2D6, DRD2, DRD4, OPRK1, SLC6A4 Olanzapine 62 7 BDNF, CYP2D6, DRD2, DRD4, HTR1B, POMC, SLC6A4 Tramadol 16 5 CYP2D6, OPRD1, OPRK1, OPRM1, SLC6A4

CIs of remission from OUD for the 19 top candidate drugs. Analysis of repurposed candidate drugs in the Triazolam was excluded due to insufficient cases for context of OUD at the genetic and functional level patients with OUD. Naltrexone, an FDA-approved OUD treatment, has the highest effects on remission from OUD. We analyzed how each of the five promising drugs (redu- Specifically, we studied OUD patients who were not pre- cing OUD relapse) (Fig. 2) relates to OUD at the genetic scribed methadone or buprenorphine. In this cohort, those and functional level. At the genetic level, each candidate who have prescribed naltrexone had odds of remission drug directly targets multiple OUD-associated genes. For nearly 4.00 times as compared to those not prescribed example, olanzapine targets a total of 62 genes, among naltrexone (AOR: 3.54, 95% CI: [3.31,3.69], p value < which seven genes are implicated in OUD. Tramadol targets 0.0001), adjusting for age, gender, and race. This result 16 genes, of which five are OUD genes, including three supports the validity of the EHR-based evaluation of clin- opioid receptors (OPRD1, OPRK1, and OPRM1) (Table 3). ical efficacy of candidate drugs for treating OUD. While each candidate drug is associated with higher odds of For the remaining 18 drugs, eight had an AOR of remis- OUD remission based on EHR analysis, these drugs target sion significantly higher than 1.0. Olanzapine, a drug different aspects of OUD and OUD genes as shown in approved to treat schizophrenia, has the highest AOR of Table 3. The complete list of OUD- and drug-associated remission from OUD. Specifically, we studied patients with genes are in Supplementary file S1. OUD and schizophrenia but not prescribed methadone, We then analyzed how these candidate drugs are func- buprenorphine, or naltrexone. In this cohort, OUD patients tionally involved in OUD. We identified 53 genetic path- who were prescribed olanzapine had odds of OUD remission ways that are significantly enriched for the 19 OUD- nearly 2.00 times compared to patients not prescribed olan- associated genes. The top enriched OUD pathways are zapine (AOR: 1.90; 95% CI [1.66,2.18], p value < 0.0001), opioid signaling, G-protein activation, gap junction, potas- adjusting for age, gender, and race. Three sium channels (full list is in Supplementary file 2). We bupropion, mirtazapine and atomoxetine had odds of OUD identified significantly enriched pathways for each candi- remission nearly 50% higher compared to patients with date drug and the shared pathways between the drug and depression not prescribed these antidepressants (AOR = 1.38 OUD. As shown in Table 4, each candidate drug targets [1.31–1.46], p value < 0.0001; 1.37 [1.29–1.46], p value < multiple OUD-associated pathways. For example, 317 0.0001; 1.48 [1.25–1.76], p value < 0.0001, respectively). pathways are significantly enriched for olanzapine, among The prescription of opioid analgesics was negatively which 41 are also significantly enriched for OUD (77% of associated with OUD remission. For example, patients who 53 OUD pathways) including G-protein activation, Ser- were prescribed morphine had odds of remission 16% lower otonin receptors among others. A total of 42 pathways are compared to patients with no prescription of morphine significantly enriched for tramadol, among which 24 were (AOR: 0.84; 95% CI [0.76,0.92], p value < 0.0001). These OUD-associated pathways (45% of 53 OUD pathways) results show that opioid analgesics can aggravate OUD. including G-protein activation, Serotonin receptors, opioid Interestingly, we found that tramadol was the only opioid signaling among others. The full list of OUD-associated analgesic positively associated with OUD remission (AOR: pathways targeted by each candidate drug is in Supple- 1.51; 95% CI [1.38,1.66], p value < 0.0001). Tramadol has mentary file S2. a lower risk of addiction compared to other opioid analge- sics [48] and previous studies support benefits in the man- agement of opioid withdrawal symptoms and long-term Discussion OUD treatment [49, 50]. Our findings alongside the prior studies in the literature support further evaluation of tra- We developed an integrated drug repurposing strategy for madol for OUD treatment, including treatment of comorbid OUD that is applicable to other SUDs. The approach OUD in pain patients. combined computational drug candidate prediction, EHRs- Drug repurposing for opioid use disorders: integration of computational prediction, clinical. . .

Table 4 Number of significantly enriched pathways for each drug, number of significantly enriched pathways shared between the drug and OUD, and top 10 OUD pathways targeted by each candidate drug. Drug Total (n) Shared (n) Top 10 pathways

Atomoxetine 42 16 Xenobiotics Voltage gated potassium channels Cytochrome P450—arranged by substrate type Phase 1—functionalization of compounds Neuroactive ligand–receptor interaction Biological oxidations Posttranslational regulation of adherens junction stability and disassembly Potassium channels SHP2 signaling Neurotrophic factor-mediated Trk receptor signaling Bupropion 305 21 Serotonin receptors Xenobiotics Amine ligand-binding receptors Cytochrome P450—arranged by substrate type Opioid Signaling Phase 1—functionalization of compounds Class A/1 (Rhodopsin-like receptors) Neuroactive ligand–receptor interaction Biological oxidations Posttranslational regulation of adherens junction stability and dissassembly Mirtazapine 93 24 Serotonin receptors Repression of pain sensation by the transcriptional regulator DREAM Xenobiotics Regulation of ck1/cdk5 by type 1 glutamate receptors Amine ligand-binding receptors G alpha (1) signaling events Cytochrome P450—arranged by substrate type G alpha (1) pathway Phase 1—functionalization of compounds Class A/1 (Rhodopsin-like receptors) Olanzapine 317 41 G-protein activation Androgen biosynthesis Serotonin receptors Repression of pain sensation by the transcriptional regulator DREAM Peptide hormone biosynthesis Endogenous sterols Xenobiotics Regulation of ck1/cdk5 by type 1 glutamate receptors Syndecan-3-mediated signaling events Amine ligand-binding receptors Tramadol 42 24 G-protein activation Serotonin receptors Repression of pain sensation by the transcriptional regulator DREAM Xenobiotics Amine ligand-binding receptors G alpha (1) signaling events Cytochrome P450—arranged by substrate type Opioid Signaling Peptide ligand-binding receptors Phase 1—functionalization of compounds based clinical corroboration, and mechanisms of action one (olanzapine), two antidepressants (mirta- analysis using genetic, genomic, phenotypic, and patient zapine and bupropion), and one selective EHR data. We identified one opioid analgesic (tramadol), (atomoxetine), that might be of value in M. Zhou et al. improving outcomes in OUD patients, including those associated with the rewarding effects of nicotine [56], since suffering from comorbid conditions. it is an approved treatment for smoking cessation. Fur- Our study shows that patients with schizophrenia and thermore, preclinical studies have reported that bupropion OUD who were prescribed olanzapine had higher odds of prevented the development of tolerance to opioids and the OUD remission than individuals without olanzapine (AOR = emergence of withdrawal symptoms [57]. Our analysis 1.90). Olanzapine is an antipsychotic drug that acts as an shows that bupropion directly targets multiple OUD- antagonist at D1 and D2 receptors, 5H2A and associated pathways including serotonin receptors, opioid 5H6 serotonin receptors, H1 histamine receptors, adrenergic and GPCR signaling, neuroactive ligand–receptor interac- alpha one receptor, and muscarinic M1–M5 receptors [51]. In tion, which suggests that bupropion may also be beneficial our study, we showed that olanzapine directly targets 7 of 19 in OUD treatment. OUD-associated genes, including BDNF, CYP2D6, DRD2, Mirtazapine enhances noradrenergic and presumably DRD4, HTR1B, POMC, and SLC6A4. Based on pathway dopaminergic signaling through its antagonistic effects at enrichment analysis, olanzapine significantly targets 41 of the serotonin receptors (5HT2 family) and alpha 2 adrenergic 53 OUD-associated pathways (77%). Some of the pathways, receptors. Preliminary evidence has shown a therapeutic such as G-protein activation, serotonin receptors, neuroactive benefit of mirtazapine in the treatment of patients with ligand–receptor interaction, and GPCR signaling are known methamphetamine use disorder alone or when combined to be involved in opioid addiction [52, 53]. with naltrexone [58]. Our genetic and pathway analyses In our analysis, patients with pain and OUD who were showed that mirtazapine directly target five OUD- prescribed with tramadol had higher odds of OUD remission associated genes (CYP2D6, DRD2, DRD4, OPRK1, and than individuals without tramadol (AOR = 1.51). Tramadol is SLC6A4) and 24 OUD-associated pathways including ser- a prodrug whose metabolite (O-desmethyltramadol) acts as an otonin receptors, G alpha signaling, GPCR signaling, and opioid agonist and a serotonin-norepinephrine reuptake inhi- GAP junctions, among others. bitor [54]. Tramadol is widely used as an analgesic and is Atomoxetine is a selective norepinephrine (noradrena- believed to be less addictive than other equipotent opioid line) reuptake inhibitor for the treatment of attention-deficit analgesics. The increased odds of OUD remission, we hyperactivity disorder (ADHD) [59, 60]. It is a highly observed when tramadol was prescribed to patients with selective and potent inhibitor of the presynaptic nora- comorbid pain identifies it as another opioid agonist poten- drenalin transporter and could be of use in the treatment of tially beneficial for OUD treatment. Consistent with this, depression associated with OUD as a monotherapy or as an studies have reported the benefit of Tramadol use for opioid augmentation agent [61]. Preclinical studies have shown detoxification [55]. Intriguingly whereas the prescription of that atomoxetine attenuated self-administration opioid analgesics was associated with decreased risk for OUD perhaps by modifying the reinforcing properties of stimu- remission, the prescription of tramadol showed the opposite lants [62]. Our EHR-based analysis showed that atomox- association. In this respect, tramadol might be particularly etine is associated with significant OUD remission in useful in the management of pain patients suffering from patients with ADHD and OUD. In addition, atomoxetine OUD. Our analysis of genes and genetic pathways targeted by directly targets three OUD genes (BDNF, CYP2D6, and tramadol show that tramadol directly target OUD-associated SLC6A4) and 16 of the 53 OUD pathways including genes and opioid receptors (CYP2D6, OPRD1, OPRK1, potassium channels, neuroactive ligand–receptor interac- OPRM1, and SLC6A4). Among 42 pathways significantly tion, and chemical synapsis. enriched for tramadol, 24 were OUD-associated pathways One important finding is that all five candidate drugs (45% of 53 OUD pathways) including G-protein activation, directly target multiple OUD genes and pathways, sug- serotonin receptors, opioid signaling, GPCR signaling among gesting that these drugs may treat OUD in the general others. These potential mechanisms of action of tramadol population and might be particularly beneficial in patients suggest tramadol may help treat OUD in general and not only with comorbid conditions (pain, depression, schizophrenia, in OUD patients with pain. ADHD). In addition, these drugs target different OUD Two antidepressants (bupropion and mirtazapine) and genes and pathways, suggesting that combinations of these one selective norepinephrine reuptake inhibitor (atomox- drugs may have synergistic efficacy in treating OUD. etine) that also has properties were identified Our study has several limitations. First, we used patient as anti-OUD candidates. Inasmuch as depression is fre- EHR data to corroborate top-ranked anti-OUD candidate quently comorbid with OUD and contributes to relapse this drugs generated by the network-based prediction algorithm. could be a mechanism for improving outcomes in OUD. Patient EHR data were collected for clinical convenience and Most notable is bupropion, a blocker of dopamine and billing, not for research purposes. Though EHR data have norepinephrine transporters and an antagonist at various been widely used for research purposes, they have inherent nicotine receptors including the alpha 4 beta 2 receptor limitations including under-diagnosis, over-diagnosis, or mis- Drug repurposing for opioid use disorders: integration of computational prediction, clinical. . . diagnosis, limited information on time-series, timing, and of particular values for ADHD patients with OUD. These adherence of medications and treatments, among others [63– medications might also be of value in OUD patients in 66]. Second, disease diagnoses in the patient EHR data are general by themselves or in combination with approved coded using SNOMED-CT terminology and for SUD these medication for OUD. While our approach that seamlessly differ from categories used by DSM-5. A previous study integrates computational drug prediction, patient EHR- showed EHRs alone were inadequate in capturing mental based evaluation and bioinformatics-based mechanisms of health diagnoses, visits, specialty care, hospitalizations, and action analysis is at the in-silico level it provides the medications [67]. In our recent study, we showed that 10.3% foundation for future hypothesis-driven preclinical and of the study population in the Explorys EHR database had a clinical studies of these candidate drugs in treating OUD. diagnosis of SUD [34], which is similar to the reported pre- valence of 10.8% among people aged 18 or older in the US Acknowledgements We acknowledge support from Eunice Kennedy according to the 2018 National Survey on Drug Use and Shriver National Institute of Child Health & Human Development of the National Institutes of Health under the NIH Director’s New Health [68]. However, the under-diagnosis of other mental Innovator Award number DP2HD084068, NIH National Institute on disorders including depression, schizophrenia, and ADHD, Aging R01 AG057557, R01 AG061388, R56 AG062272, National some of which are the original indications of identified can- Institute on Drug Addiction UG1DA049435, American Cancer didate drugs, may have impacts on the EHR-based corro- Society Research Scholar Grant RSG-16-049-01—MPC, and The Clinical and Translational Science Collaborative (CTSC) of Cleveland boration in our study. Third, though patients in this EHR 1UL1TR002548-01. database represent 20% of US population, they represent patients who had encounters with healthcare systems and are Author contributions RX conceived, designed, and supervised the not necessarily representative of the general US population. In study. MZ performed drug prediction and clinical evaluation. QW per- addition, EHRs in the Explorys database are population-level formed the mechanism of action study. RX, QW, MZ, and NDV drafted (not patient-level) data, on which more advanced statistical the manuscript. NDV, CLZ, and AR critically contributed to data interpretation and result discussion. All authors approved the manuscript. testing could not be performed. The candidates identified by the computational prediction algorithm need to be evaluated in other databases and in other populations or to be investi- Compliance with ethical standards gated in a prospective manner to control for confounding Conflict of interest The authors declare that they have no conflict of factors. Because of the inherent limitations in EHR-based interest. drug discovery, our drug discovery scheme comprised of three complementary components (computational drug pre- Publisher’s note Springer Nature remains neutral with regard to diction, hypothesis-driven EHR-based testing, and mechan- jurisdictional claims in published maps and institutional affiliations. isms of action analyses), in which EHR-based testing was not used for drug discovery but for gathering clinical evidence to References further corroborate candidates output from the computational prediction system. Finally, another concern is that our current 1. Volkow ND, McLellan AT. Opioid abuse in chronic pain- misconceptions and mitigation strategies. N. Engl J Med. knowledge regarding genes and pathways involved in OUD is 2016;374:1253–63. https://doi.org/10.1056/NEJMra1507771. likely to be incomplete. In the future as new knowledge 2. Volkow ND, Koob GF, McLellan AT. Neurobiologic advances emerges one can investigate how newly identified OUD- from the brain disease model of addiction. N. Engl J Med. related genes and pathways interact with the previously 2016;374:363–371. 3. Volkow ND, Morales M. The brain on drugs: from reward to known OUD genes and pathways and determine how addiction. Cell. 2015;162:712–725. including them affects the prediction for recognizing ther- 4. National Survey on Drug Use and Health—SAMHSA. Key sub- apeutic drugs for OUD or other SUD. stance use and mental health indicators in the United States: In summary, we developed an integrated drug repur- results from the 2018 National Survey on Drug Use and Health. National Survey on Drug Use and Health. 2018. https://www.sa posing system for OUD by combining computational pre- mhsa.gov/data/sites/default/files/cbhsq-reports/NSDUHNationa diction, clinical corroboration, and mechanisms of action lFindingsReport2018/NSDUHNationalFindingsReport2018.pdf. studies. We identified one opioid analgesic (tramadol), that 5. GBD 2019 Diseases and Injuries Collaborators. Global burden of might be of particular value for the management of pain 369 diseases and injuries in 204 countries and territories, 1990- 2019: a systematic analysis for the Global Burden of Disease patients suffering from OUD; one antipsychotic (olanza- Study 2019. Lancet. 2020;396:1204–22. https://doi.org/10.1016/ pine), that might be of particular value for the management S0140-6736(20)30925-9. of bipolar or schizophrenia patients with OUD, two anti- 6. Department of Health and Human Services. HHS acting secretary depressant (mirtazapine and bupropion) that might be of declares public health emergency to address national opioid crisis. Department of Health and Human Services; 2017. p. 2016–7. value in improving OUD outcomes in OUD patients in https://www.hhs.gov/about/news/2017/10/26/hhs-acting-secreta whom depression contributes to relapse and one nor- ry-declares-public-health-emergency-address-national-opioid- epinephrine transporter blocker (tomoxetine) that might be crisis.html. M. Zhou et al.

7. Centers for Disease Control and Prevention (CDC) National 2018. Nucleic Acids Res. 2018;46:D1074–82. https://doi.org/10. Center for Health Statistics. Mortality. Centers for Disease Control 1093/nar/gkx1037. and Prevention. 2020. https://www.cdc.gov/nchs/data/health_ 27. Xu R, Wang Q, Li L. A genome-wide systems analysis reveals policy/Provisional-Drug-Overdose-Deaths-by-Quarter-Demogra strong link between colorectal cancer and trimethylamine N-oxide phic-Characteristics-Q1-2020.pdf. (TMAO), a gut microbial metabolite of dietary meat and fat. BMC 8. Volkow ND, Skolnick P. New medications for substance use Genom. 2015;16:S4 https://doi.org/10.1186/1471-2164-16-S7-S4. disorders: challenges and opportunities. Neuropsychopharmacology. 28. Xu R, Wang Q. Towards understanding brain-gut-microbiome 2012;37:290–292. connections in Alzheimer’s disease. BMC Syst Biol 2016;10:63 9. Mullard A. New drugs cost US $2.6 billion to develop. Nat Rev https://doi.org/10.1186/s12918-016-0307-y. Drug Discov. 2014;13:877. https://doi.org/10.1038/nrd4507. 29. Wang Q, Xu R. CoMNRank: an integrated approach to extract 10. National Academies of Sciences, Engineering, and Medicine; and prioritize human microbial metabolites from MEDLINE Health and Medicine Division; Board on Health Sciences Policy; records [published online ahead of print, 2020 Aug 11]. J Biomed Forum on Neuroscience and Nervous System Disorders. Advan- Inf. 2020;109:103524 https://doi.org/10.1016/j.jbi.2020.103524. cing Therapeutic Development for Pain and Opioid Use Disorders 30. IBM Explorys. IBM Watson Health. IBM Explorys. https://www. Through Public-Private Partnerships: Proceedings of a Workshop. ibm.com/watson/health/explorys/. Accessed June–August 2020. Washington (DC): National Academies Press (US); 2018. Accessed June–Aug 2020. 11. Ashburn TT, Thor KB. Drug repurposing: identifying and developing 31. Bodenreider O. The Unified Medical Language System (UMLS): new uses for existing drugs. Nat Rev Drug Discov. 2004;3:673–683. integrating biomedical terminology. Nucleic Acids Res. 2004;32: 12. Srivastava AB, Mariani JJ, Levin FR. New directions in the D267–D270. https://doi.org/10.1093/nar/gkh061. treatment of opioid withdrawal. Lancet. 2020;395:1938–48. 32. SNOMED International. The systematized nomenclature of https://doi.org/10.1016/S0140-6736(20)30852-7. medicine—Clinical Terms (SNOMED CT). SNOMED Interna- 13. Li J, Zheng S, Chen B, Butte AJ, Swamidass SJ, Lu Z. A survey tional. 2020. http://www.snomed.org/snomed-ct/why-snomed-ct. of current trends in computational drug repurposing. Brief 33. Zhou M, Xu R, Kaelber DC, Gurney ME. Tumor necrosis factor Bioinform. 2016;17:2–12. (TNF) blocking agents are associated with lower risk for Alz- 14. Wang Q, Xu R. Disease comorbidity-guided drug repurposing: a case heimer’s disease in patients with rheumatoid arthritis and psor- study in schizophrenia. AMIA Annu Symp Proc. 2018;2018:1300–9. iasis. PLoS ONE. 2020;15:e0229819. https://doi.org/10.1371/ 15. Nagaraj AB, Wang Q, Joseph P, Zheng CL, Chen Y, Kovalenko journal.pone.0229819. O, et al. Using a novel computational drug repurposing approach 34. Wang QQ, Kaelber DC, Xu R, Volkow ND. COVID-19 risk and (DrugPredict) to rapidly identify potent drug candidates for cancer outcomes in patients with substance use disorders: analyses from treatment. Oncogene. 2017. https://doi.org/10.1038/onc.2017.328. electronic health records in the United States. Mol Psychiatry. 16. Chen Y, Xu R. Context-sensitive network-based disease genetics 2020:1–10. https://doi.org/10.1038/s41380-020-00880-7. prediction and its implications in drug discovery. Bioinformatics 35. Wang QQ, Xu R, Volkow ND. Increased risk of COVID-19 2017;33:1031–9. https://doi.org/10.1093/bioinformatics/btw737. infection and mortality in people with mental disorders: analysis 17. Xu R, Wang Q. A genomics-based systems approach towards drug from electronic health records in the United States. World Psy- repurposing for rheumatoid arthritis. BMC Genom. 2016;17:518 chiatry. 2020. https://doi.org/10.1002/wps.20806. https://doi.org/10.1186/s12864-016-2910-0. 36. Wang QQ, Berger NA, Xu R. Analyses of risk, racial disparity, 18. Xu R, Wang Q. PhenoPredict: a disease phenome-wide drug repur- and outcomes of cancers and coronavirus disease 2019 from posing approach towards schizophrenia drug discovery. J Biomed electronic health 4 records in the US. JAMA Oncol. 2020. https:// Inf. 2015;56:348–55. https://doi.org/10.1016/j.jbi.2015.06.027. doi.org/10.1001/jamaoncol.2020.6178. 19. Chen Y, Xu R. Context-sensitive network-based disease genetics 37. Wang QQ, Berger NA, Xu R. When hematologic malignancies prediction and its implications in drug discovery. Bioinformatics. meet COVID-19 in the United States: infections, death and dis- 2017;33:1031–9. https://doi.org/10.1093/bioinformatics/btw737. parities. Blood Rev. https://doi.org/10.1016/j.blre.2020.100775. 20. Chen Y, Xu R. Context-sensitive network analysis identifies food 38. The National Library of Medicine (NLM). DailyMed. The metabolites associated with Alzheimer’s disease: an exploratory National Library of Medicine. https://healthdata.gov/dataset/da study. BMC Med Genom. 2019;12:17 https://doi.org/10.1186/ ilymed. Accessed June–Aug 2020. s12920-018-0459-2. 39. Kuritz SJ, Landis JR, Koch GG. A general overview of Mantel- 21. Zheng C, Xu R. Large-scale mining disease comorbidity rela- Haenszel methods: applications and recent developments. Annu tionships from post-market drug adverse events surveillance data. Rev Public Health. 1988;9:123–160. https://doi.org/10.1146/a BMC Bioinforma. 2018;19:500. nnurev.pu.09.050188.001011. 22. Zhou M, Chen Y, Xu R. A drug-side effect context-sensitive 40. Wang SC, Chen YC, Lee CH, Cheng CM. Opioid addiction, network approach for drug target prediction. Bioinformatics. genetic susceptibility, and medical treatments: a review. Int J Mol 2019;35:2100–7. https://doi.org/10.1093/bioinformatics/bty906. Sci 2019;20:4294 https://doi.org/10.3390/ijms20174294. 23. Zhou M, Zheng C, Xu R. Combining phenome-driven drug-target 41. Darcq E, Kieffer BL. Opioid receptors: drivers to addiction? Nat interaction prediction with patients’ electronic health records-based Rev Neurosci. 2018;19:499–514. https://doi.org/10.1038/s41583- clinical corroboration toward drug discovery. Bioinformatics. 018-0028-x. 2020;36:i436–i444. https://doi.org/10.1093/bioinformatics/btaa451. 42. Crist RC, Reiner BC, Berrettini WH. A review of opioid addiction 24. Kuhn M, Letunic I, Jensen LJ, Bork P. The SIDER database of genetics. Curr Opin Psychol. 2019;27:31–35. drugs and side effects. Nucleic Acids Res. 2016;44:D1075–9. 43. Szklarczyk D, Santos A, von Mering C, Jensen LJ, Bork P, Kuhn https://doi.org/10.1093/nar/gkv1075. M. STITCH 5: augmenting protein-chemical interaction networks 25. Franceschini A, Szklarczyk D, Frankild S, Kuhn M, Simonovic M, with tissue and affinity data. Nucleic Acids Res. 2016;44: Roth A, et al. STRING v9.1: protein-protein interaction networks, D380–D384. https://doi.org/10.1093/nar/gkv1277. with increased coverage and integration. Nucleic Acids Res. 44. Liberzon A, Subramanian A, Pinchback R, Thorvaldsdóttir H, 2013;41:808–15. https://doi.org/10.1093/nar/gks1094. Tamayo P, Mesirov JP. Molecular signatures database (MSigDB) 26. Wishart DS, Feunang YD, Guo AC, Lo EJ, Marcu A, Grant JR, 3.0. Bioinformatics. 2011;27:1739–40. https://doi.org/10.1093/ et al. DrugBank 5.0: a major update to the DrugBank database for bioinformatics/btr260. Drug repurposing for opioid use disorders: integration of computational prediction, clinical. . .

45. Wang Q, Xu R. Data-driven multiple-level analysis of gut- 58. Brackins T, Brahm NC, Kissack JC. Treatments for methamphe- microbiome-immune-joint interactions in rheumatoid arthritis. BMC tamine abuse: a literature review for the clinician. J Pharm Pr. Genom. 2019;20:124 https://doi.org/10.1186/s12864-019-5510-y. 2011;24:541–50. https://doi.org/10.1177/0897190011426557. 46. Wang Q, Li L, Xu R. A systems biology approach to predict and 59. Wong DT, Threlkeld PG, Best KL, Bymaster FP. A new inhibitor characterize human gut microbial metabolites in colorectal cancer. of norepinephrine uptake devoid of affinity for receptors in rat Sci Rep. 2018;8:6225 https://doi.org/10.1038/s41598-018-24315-0. brain. J Pharm Exp Ther. 1982;222:61–5. 47. National Institute on Drug Abuse. Common comorbidities with 60. Garnock-Jones KP, Keating GM. Atomoxetine: a review of its use substance use disorders research report. National Institute on Drug in attention-deficit hyperactivity disorder in children and adoles- Abuse. 2018. https://www.drugabuse.gov/publications/drugfacts/ cents. Paed Drugs. 2009;11:203–26. comorbidity-substance-use-disorders-other-mental-illnesses. 61. Carpenter LL, Milosavljevic N, Schecter JM, Tyrka AR, Price LH. 48. Kaye AD. Tramadol, pharmacology, side effects, and serotonin Augmentation with open-label atomoxetine for partial or non- syndrome: a review. Pain Physician. 2015;18:395–400. response to antidepressants. J Clin Psychiatry. 2005;66:1234–8. 49. Lofwall MR, Babalonis S, Nuzzo PA, Siegel A, Campbell C, 62. Lofwall MR, Walsh SL, Bigelow GE, Strain EC. Modest opioid Walsh SL. Efficacy of extended-release tramadol for treatment of withdrawal suppression efficacy of oral tramadol in humans. prescription opioid withdrawal: a two-phase randomized con- Psychopharmacology. 2007;194:381–393. trolled trial. Drug Depend. 2013;133:188–97. 63. Cowie MR, Blomster JI, Curtis LH, Duclaux S, Ford I, et al. 50. Sarkar S, Varshney M, Patil V, Lal R. Maintainence treatment Electronic health records to facilitate clinical research. Clin Res of opioid dependence with tramadol. J Neurosci Rural Pr. Cardiol. 2017;106:1–9. https://doi.org/10.1007/s00392-016-1025-6. 2017;8:S98. 64. Coorevits P, Sundgren M, Klein GO, Bahr A, Claerhout B, Daniel C, 51. Bymaster FP, Rasmussen K, Calligaro DO, Nelson DL, DeLapp NW, et al. Electronic health records: new opportunities for clinical research. Wong DT, et al. In vitro and in vive biochemistry of olanzapine: a J Intern Med. 2013;274:547–560. https://doi.org/10.1111/joim.12119. novel, drug. J Clin Psychiatry. 1997;58:28–36. 65. Ahmad FS, Chan C, Rosenman MB, Post WS, Fort DG, Greenland P, 52. Al-Hasani R, Bruchas MR. Molecular mechanisms of opioid et al. Validity of cardiovascular data from electronic sources: the receptor-dependent signaling and behavior. Anesthesiology. 2011; Multi-Ethnic Study of Atherosclerosis and HealthLNK. Circulation. 115:1363–81. https://doi.org/10.1097/ALN.0b013e318238bba6. 2017;136:1207–1216. https://doi.org/10.1161/CIRCULATIONAHA. 53. Koob GF, Volkow ND. Neurobiology of addiction: a neuro- 117.027436. circuitry analysis. Lancet Psychiatry. 2016;3:760–773. 66. Safran C, Bloomrosen M, Hammond WE, Labkoff S, Markel-Fox 54. Grond S, Sablotzki A. Clinical pharmacology of tramadol. Clin S, Tang PC, et al. Toward a national framework for the secondary Pharmacokinet. 2004;43:879–923. use of health data: an American Medical Informatics Association 55. Radcliff JA, Vellanki SC, Moore PS, Weisensell BJ, Cather JK. A White Paper. J Am Med Inf Assoc. 2007;14:1–9. https://doi.org/ retrospective chart review of outcomes resulting from a three-day 10.1197/jamia.M2273. tramadol taper for acute opioid withdrawal. J Addict Dis. 2018; 67. Madden JM, Lakoma MD, Rusinak D, Lu CY, Soumerai SB. 37:252–8. Missing clinical and behavioral health data in a large electronic 56. Stahl SM, Pradko JF, Haight BR, Modell JG, Rockett CB, health record (EHR) system. J Am Med Inform Assoc. 2016; Learned-Coughlin S. A review of the neuropharmacology of https://doi.org/10.1093/jamia/ocw021. bupropion, a dual norepinephrine and dopamine reuptake inhi- 68. The Substance Abuse and Mental Health Services Administration bitor. Prim Care Companion J Clin Psychiatry. 2004;6:159–66. (SAMHSA). Key substance use and mental health indicators in https://doi.org/10.4088/pcc.v06n0403. the United States: results from the 2018 National Survey on Drug 57. Hamdy MM, Elbadr MM, Barakat A. Bupropion attenuates Use and Health. Substance Abuse and Mental Health Services morphine tolerance and dependence: Possible role of glutamate, Administration. https://www.samhsa.gov/data/sites/default/files/ norepinephrine, inflammation, and oxidative stress. Pharm Rep. cbhsq-reports/NSDUHNationalFindingsReport2018/NSDUHNa 2018;70:955–62. https://doi.org/10.1016/j.pharep.2018.04.003. tionalFindingsReport2018.pdf. Accessed 12 Nov 2020.