Computed Biological Relations among Five Select Treatment-Related

Apr 10, 2016 - School of Mechanical Engineering, National Technical University of Athens, Athens, .... phenotypes and then curated them by referencing...
0 downloads 0 Views 2MB Size
Article pubs.acs.org/crt

Computed Biological Relations among Five Select Treatment-Related Organ/Tissue Toxicities Theodore Sakellaropoulos,†,§ Timothy J. Herod,†,⊥ Leonidas G Alexopoulos,‡ and Jane P.F. Bai*,† †

Office of Clinical Pharmacology, Center for Drug Evaluation and Research, U.S. Food and Drug Administration, Silver Spring, Maryland 20993, United States ‡ School of Mechanical Engineering, National Technical University of Athens, Athens, Greece S Supporting Information *

ABSTRACT: Drug toxicity presents a major challenge in drug development and patient care. We set to build upon previous works regarding select drug-induced toxicities to find common patterns in the mode of action of the drugs associated with these toxicities. In particular, we focused on five disparate organ toxicities, peripheral neuropathy (PN), rhabdomyolysis (RM), Stevens−Johnson syndrome/toxic epidermal necrosis (SJS/TEN), lung injury (LI), and heart contraction-related cardiotoxicity (CT), and identified biological commonalities between and among the toxicities in terms of pharmacological targets and nearest neighbors (indirect effects) using the hyper-geometric test and a distance metric of Spearman correlation. There were 20 significant protein targets associated with two toxicities and 0 protein targets associated with three or more toxicities. Per Spearman distance, PN was closest to SJS/TEN compared to other pairs, whereas the pairs involving RM were more different than others excluding RM. The significant targets associated with RM outnumbered those associated with every one of the other four toxicities. Enrichment analysis of drug targets that are expressed in corresponding organ/tissues determined proteins that should be avoided in drug discovery. The identified biological patterns emerging from the mode of action of these drugs are statistically associated with these serious toxicities and could potentially be used as predictors for new drug candidates. The predictive power and usefulness of these biological patterns will increase with the database of these five toxicities. Furthermore, extension of our approach to all severe adverse reactions will produce useful biological commonalities for reference in drug discovery and development.



INTRODUCTION Adverse reactions (ARs) to drugs present health risks to patients1 and are responsible for drug attrition during late stage drug development.2 Improving the ability to predict potentially serious adverse effects would enable early elimination of problematic drug candidates, reduce drug attrition, and increase the cost effectiveness of pharmaceutical R&D. A drug could cause extended pharmacological effects through its mode of action, which is reflected by a drug-perturbed toxico-pharmacological network rippling from its pharmacological targets through biological protein−protein interactions. The term of “class effect” is often used to describe specific AR(s) for the same class of drugs in their approved labels (Drugs@FDA). Previous works had explored how the intended pharmacological targets and unintended targets of a list of drugs that share a specific treatment-related adverse reaction (hereafter, a toxicity) interacted with one another, how such a network of interactions could be used as a predictor of drug toxicity, and how specific drug toxicity was related to the biological pathways perturbed by these drugs.3−7 Furthermore, a network analysis of integrating the transcriptomic profiles of drugs causing lung injury, a prior knowledge network of proteins, and the pharmacological targets of these drugs was shown to be useful for repurposing approved drugs to treat lung injury.6 These previous successes have demonstrated the power of network biology. With the biological © 2016 American Chemical Society

network/pathways accurately defined in detail, one could be able to use the pharmacological targets of all approved drugs and their safety profiles of concern to create a curated toxicopharmacological network; this network would be expected to have a high predictive power of drug safety for future drug candidates. In this study, we extended the results of previous work addressing five specific groups of conditions: drug-induced peripheral neuropathy (PN),3 drug-induced rhabdomyolysis (RM),5 drug-induced severe cutaneous reactions of Stevens− Johnson syndrome and toxic epidermal necrosis (SJS/TEN),4 drug-induced lung injury (LI),6 and drug-induced contractionrelated cardiotoxicity (CT).7 These toxicities were selected because the drugs causing these potentially fatal or disabling toxicities are approved for different indications, and there is a need to understand the underlying biology between and among toxicities to guide the development of safe and effective drugs with fewer toxicities. Our goal was to determine the interrelationship among these five organ/tissue toxicities using computational methods analyzing the known protein target drugs with either pharmacological or indirect effects (hereafter, targets). We compiled the targets of the drugs that individually Received: February 22, 2016 Published: April 10, 2016 914

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

Figure 1. Drug toxicity association. (A) Bar graph of the number of drugs per toxicity; drugs unique to each toxicity are in red. Each toxicity was caused by an average of 178.2 drugs with a standard deviation of 16.84 drugs. (B) Histogram showing the number of drugs per number of drug-induced toxicities. Approximately half of all drugs caused just one toxicity, whereas the other half caused multiple toxicities with decreasing frequency. shown in Table S2 (Drug_info). The original drug list of the five toxicities contained 519 drugs, but 10 of them could not be mapped to a DrugBank ID so they were removed (see Drugs with No DrugBank-ID in Table S2, Drug_Info). When combined, all of the drugs studied were directly associated with 574 pharmacological targets from DrugBank and directly or indirectly with 4497 proteins based on STITCH (search tools for interactions of chemicals; http://stitch.embl.de/), yielding a total of 5071 proteins (targets). Another 15 drugs that could not be linked to any human protein target were removed from subsequent analysis (see Drug with No Interactions in Table S2, Drug_Info). The corresponding targets for each drug were collected using both STICH 4.0 and DrugBank 4.4 (Table S3, Targets_info). In particular, targets from DrugBank were pharmacological targets that were considered already curated and were included without any filtering. On the other hand, targets from STITCH are derived using automated text-mining and thus also capture indirect effects of a drug. Our protein list consisted of 5,071 pharmacological as well as indirect drug targets, all of which are referred to as “targets” (Table S3, Targets_info) hereafter. Association of Drug Targets with Toxicities. From STITCH, we collected all the drug−protein interactions that were characterized as “high confidence” (internal score greater than 700/1000) (see Table S4, Target to Drugs). We then used each drug as a mediator to associate its protein targets with its treatment-related toxicities and did so across all five toxicities and all drugs at hand to obtain statistical associations. In particular, we calculated the percentage of drugs that were associated with a toxicity with respect to a specific protein, and a toxicity was considered to be associated with a protein if this percentage was significantly greater than the overall percentage of drugs that targeted this protein based on the hypergeometric test. All of the protein− toxicity pairs with p-values less than 0.05 were considered significantly linked. To further examine commonalities among toxicities, we used the computed p-values as a measure of importance of a protein to a toxicity and then compared the five different toxicities based on how they ranked the 5,071 proteins. Because we were interested in the ranking of the different proteins and wanted to moderate the effect of extreme values, we used, as a comparison metric, Spearman’s correlation coefficient. Relating Highly-Targeted Proteins to Highly-Expressed Proteins in a Corresponding Organ. To further support our results, we tested whether the proposed proteins were expressed in the organs/

caused at least one of the five organ/tissue toxicities and then performed computations for biological closeness and commonalities between toxicity pairs across all five toxicities, and then related the results to human biology. We believe that, with a clear understanding of the biological interrelationship among clinical drug toxicities, one would be able to conduct a robust common biology-driven prediction of the potential clinical adverse reactions of a drug candidate at its late stage of development and beyond based on the toxicities observed during its early phase of development. With the goal of enabling prediction of drug toxicity, this study is a case study of relating one drug toxicity phenotype to another a biological basis. Our results would provide a useful reference for predicting specific potential toxicities of concern by utilizing the biological relations among drug toxicities.



EXPERIMENTAL PROCEDURES

Compilation of Drug and Target Lists. The lists of drugs associated with individual toxicities were pooled together from previously published lists3−5,7 except for the drug-induced lung injury (LI) list.6 The previously published LI list of 200 drugs only included those that are in both Pneumox (http://www.pneumotox.com/) and Connectivity Map (https://www.broadinstitute.org/cmap/). To be consistent with how we previously compiled individual CT, SJS/TEN, PN, and RM lists by referencing DRUGS@FDA along with the side effect resource (SIDER) database, we searched the SIDER2 database to collect an initial list of drugs with treatment-related lung-toxic phenotypes and then curated them by referencing DRUGS@FDA. The number of LI phenotypes turned out to be much larger than those of the other four toxicities. The longer list of LI phenotypes compared to those of the other four toxicities could be attributed to the adverse reaction terms reported relative to the anatomical feature of lungs (Table S1, five toxicities). Though bronchitis and pneumonia are two common drug-induced injury phenotypes, drugs on the LI list tend to cause multiple adverse pulmonary phenotypes. The drug list associated with each toxicity was further reduced by eliminating those approved to treat infectious diseases because their pharmacological targets (direct) are nonhuman proteins. The list of drugs associated with each toxicity is 915

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

Figure 2. Histograms of drug target interaction frequency. (A) The number of drugs is plotted against the number of targets interacting with drug. All five toxicity drugs interacted with a median number of 6 targets and a mean number of 36.1 targets. (B) The number of targets is plotted against the number of drugs interacting with target. All five toxicity targets interacted with a median number of 2 drugs and a mean number of 3.6 drugs.

Table 1. List of 27 Drugs that All Interacted with at Least 150 Proteins name

DrugBank ID

PubChem ID

CT

LI

PN

RM

SC

# toxicities

# targets

Carbenoxolone Sufentanil Montelukast Mifepristone Hydroxyzine Tocainide Mitoxantrone Carboplatin Imatinib Nifedipine Meprobamate Mefenamic Acid Entacapone Oxaprozin Miglustat Ondansetron Oxycodone Troglitazone Cerivastatin Doxylamine Phenelzine Letrozole Simvastatin Cabergoline Ketoprofen Bexarotene Triazolam

DB02329 DB00708 DB00471 DB00834 DB00557 DB01056 DB01204 DB00958 DB00619 DB01115 DB00371 DB00784 DB00494 DB00991 DB00419 DB00904 DB00497 DB00197 DB00439 DB00366 DB00780 DB01006 DB00641 DB00248 DB01009 DB00307 DB00897

238 197 271 312 158 175 206 174 283 224 191 187 274 232 264 231 280 317 222 119 160 177 54454 304 169 389 314

0 0 0 0 0 1 1 0 1 1 0 1 0 1 0 0 1 1 0 0 0 1 0 1 1 1 0

0 1 1 0 0 0 1 0 1 0 0 1 1 1 0 1 1 0 0 0 0 1 1 0 1 1 0

0 0 1 0 0 0 1 1 1 0 0 0 0 0 1 0 0 0 1 0 0 0 1 0 0 1 0

1 0 0 1 1 0 0 0 1 0 1 0 1 0 0 0 0 0 1 1 1 0 1 0 0 0 1

0 0 1 0 0 0 0 0 0 1 0 1 0 1 0 0 0 0 1 0 0 1 1 0 1 0 0

1 1 3 1 1 1 3 1 4 2 1 3 2 3 1 1 2 1 3 1 1 3 4 1 3 3 1

2097 1844 1165 709 539 492 480 398 358 350 331 325 318 277 263 250 241 241 215 203 188 181 170 168 165 157 151

gene expression across multiple tissues/organs from diseased but otherwise healthy human donors.8 In particular, we were interested in samples from heart, lung, nerve, muscle, and skin tissues, which were matched to CT, LI, PN, RM, and SCAR accordingly. For each protein target in our data, we computed the enrichment score for each tissue by counting how many times it was differentially expressed in the samples

tissues where individual corresponding drug-induced toxicities occurred. For that purpose, we used the Enrichr database (http:// amp.pharm.mssm.edu/Enrichr/) to collect differentially expressed genes, either up- or downregulated, in different human tissues based on the genotype−tissue expression database (GTEx database, http:// www.gtexportal.org/home/). GTEx is a rich database of differential 916

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

Table 2. List of 31 Protein Targets (Direct and Indirect) Interacting with at Least 25 Drugs and Their Associated KEGG Pathwaysa gene (protein) CHRM1 (cholinergic receptor, muscarinic 1) CHRM2 (cholinergic receptor, muscarinic 2) CHRM3 (cholinergic receptor, muscarinic 3) CHRM4 (olinergic receptor, muscarinic 4) ARA2A (adrenoceptor alpha 2A) ARA2B (adrenoceptor alpha 2B) ARA2C (adrenoceptor alpha 2C) DRD1 (dopamine receptor D1) DRD2 (dopamine receptor D2) DRD3 (dopamine receptor D3) DRD4 (dopamine receptor D4) ADRA1D (adrenoceptor alpha 1D) ADRA1B (adrenoceptor alpha 1B) ADRA1A (adrenoceptor alpha 1A) ADRB1 (adrenoceptor beta 1) ADRB2 (adrenoceptor beta 2) SLC6A2 (solute carrier family 6 (neurotransmitter transporter), member 2) SLC6A4 (solute carrier family 6 (neurotransmitter transporter), member 4) HTR1A (5-hydroxytryptamine (serotonin) receptor 1A, G proteincoupled) HTR1B (5-hydroxytryptamine (serotonin) receptor 1B, G proteincoupled) HTR2A (5-hydroxytryptamine (serotonin) receptor 2A, G proteincoupled) HTR2B (5-hydroxytryptamine (serotonin) receptor 2B, G proteincoupled) HTR2C (5-hydroxytryptamine (serotonin) receptor 2C, G proteincoupled) HRH1 (histamine receptor H1)

PTGS1 (prostaglandin-endoperoxide synthase 1) PTGS2 (prostaglandin-endoperoxide synthase 2)

INS-IGF2 (INS-IGF2 readthrough) AKT1 (v-akt murine thymoma viral oncogene homologue 1) GABRA1 (gamma-aminobutyric acid (GABA) A receptor, alpha 1) GABRA2 (gamma-aminobutyric acid (GABA) A receptor, alpha 2) NOS1 (nitric oxide synthase 1) a

associated KEGG pathways shared KEGG pathways among CHRM1, CHRM2, and CHRM3 include cholinergic synapse, calcium signaling pathway, neuroactive ligand−receptor interaction, and regulation of actin cytoskeleton; the pathways unique to individual genes are PI3K-Akt signaling pathway for CHRM1, cAMP signaling and PI3K-Akt signaling pathways for CHRM2, insulin secretion, salivary secretion, gastric acid secretion, and pancreatic secretion for CHRM3; CHRM4 does not share the calcium signaling pathway with the other three CHRM proteins and has no unique pathway these proteins share KEGG pathways of cGMP-PKG signaling pathway and neuroactive ligand−receptor interaction shared KEGG pathways among DRD1, DRD2, DRD3, and DRD4 include dopaminergic synapse and neuroactive ligand−receptor interaction; shared KEGG pathways by DRD1 and DRD2 include cocaine addiction, gap junction, cAMP signaling pathway, Parkinson’s disease, and alcoholism; KEGG pathways unique to DRD1 protein include amphetamine addition and morphine addition; a KEGG pathway unique to DRD2 is the Rap1 signaling pathway; no KEGG pathways are unique to DRD3 or DRD4 shared pathways include the calcium signaling pathway, cGMP-PKG signaling pathway, neuroactive ligand− receptor interaction, adrenergic signaling in cardiomyocytes, vascular smooth muscle contraction, and salivary secretion; a pathway unique to ADRA1A protein is the AMPK signaling pathway shared KEGG pathways include the calcium signaling pathway, cGMP-PKG signaling pathway, cAMP signaling pathway, neuroactive ligand−receptor interaction, endocytosis, adrenergic signaling in cardiomyocytes, regulation of lipolysis in adipocytes, renin secretion, and salivary secretion; KEGG pathways unique to ADRB1 include gap junction and dilated cardiomyopathy no known KEGG pathway associated with this gene in norepinephrine homeostasis (http://www.ncbi.nlm.nih. gov/gene/6530) a KEGG pathway is serotonergic synapse shared KEGG pathways include serotonergic synapse and neuroactive ligand−receptor interaction; a KEGG pathway unique to both HTR1A and HTR2B is the cAMP signaling pathway; KEGG pathways unique to HTR2A, HTR2B, and HTR2C include the calcium signaling pathway, gap junction, and inflammatory mediator regulation of TRP channels

contraction of smooth muscles, increase in capillary permeability, the release of catecholamine from adrenal medulla, and neurotransmission in the central nervous system; associated with memory and learning, circadian rhythm, and thermoregulation; KEGG pathways include the calcium signaling pathway, neuroactive ligand−receptor interaction, and inflammatory mediator regulation of TRP channels shared KEGG pathways include arachidonic acid metabolism, metabolic pathways, platelet activation, serotonergic synapse, regulation of lipolysis in adipocytes; KEGG pathway unique to PTGS1 is platelet activation; KEGG pathways unique to PTGS2 inlude the oxytocin signaling pathway, TNF signaling pathway, chemical carcinogenesis, NF-kappa B signaling pathway, VEGF signaling pathway, retrograde endocannabinoid signaling, ovarian steroidogenesis, Leishmaniasis, pathways in cancer, microRNAs in cancer, and small cell lung cancer unknown function a critical mediator of growth factor-induced neuronal survival; associated with 140 KEGG pathways shared KEGG pathways inlude neuroactive ligand−receptor interaction, retrograde endocannabinoid signaling, GABAergic synapse, taste transduction, morphine addiction, and nicotine addiction synthesizing nitric oxide (a biologic mediator in several processes); involved in 20 KEGG pathways

References: KEGG Orthology (KO): http://www.genome.jp/kegg/ko.html and NCBI Gene (http://www.ncbi.nlm.nih.gov/gene).

of a specific tissue and performing a hyper-geometric test for this ratio (Table S5, Enriched Targets per GTEx).

associated with only one toxicity, 8 with three toxicities, and only 2 drugs were associated with four toxicities. Most of protein targets were targeted by less than 5 drugs, but there were proteins targeted by 50 or more drugs (Figure 2B). Statistically, the whole list of drugs interacted with a median number of 6 targets and a mean number of 36.1 targets, and all the targets associated with the five toxicities interacted with a median number of 2 drugs and a mean number of 3.59 drugs. Table 2 contains the list of most highly targeted proteins (0.25%) with each of them being targeted by 25 or more drugs (directly or indirectly), and 31 proteins met this criterion. This list of proteins is mainly associated with three pathways, including neuroactive ligand receptor, calcium signaling, and gap junction; the Kegg Pathways associated with these 31 proteins are involved in cell functions, differentiation and growth, as well as a broad range of biological activities across multiple organs.



RESULTS The final drug list consisted of 504 drugs, 240 of which were shared among the five toxicities (Figure 1A and B). The majority of the drugs (∼400) were associated with no more than 2 toxicities (Figure 1B). On average, each toxicity was associated with 178.2 drugs with a standard deviation of 16.84 drugs. The number of protein targets interacting with a drug varied from drug to drug, as did the number of drugs associated with a target protein (Figure 2A and B). In particular, most drugs individually targeted less than 30 different proteins, but there were drugs that targeted more than 1,000 proteins (Figure 2A). Listed in Table 1 are 27 drugs, each of which interacted with more than 150 proteins (top 5%). More than half of these 27 drugs were 917

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

Figure 3. Number of targets significantly linked to each toxicity and to the number of diseases. (A) The number of targets was plotted for each toxicity. Each toxicity was significantly linked with an average of 43.2 targets with a standard deviation of 29.54 targets. Lung injury was linked to the largest number of targets. (B)The number of targets was plotted against the number of toxicities. The majority of targets were not significantly linked to any disease. No targets were linked to more than three diseases and only 20 targets were linked to two toxicities.

The number of proteins significantly linked to a toxicity, via the hyper-geometric test, also varied widely among the toxicities. Rhabdomyolysis was linked to the highest number of proteins, 94, whereas lung injury and cardiotoxicity were linked to the fewest, 21 and 25, respectively (Figure 3A). This is in agreement with our observation that rhabdomyolysis shared the fewest number of drugs with other toxicities. In total, 176 proteins were linked to at least one toxicity, but no protein was linked to more than two different toxicities (Figure 3B). A network of the five toxicities connected via their commonly enriched proteins is shown in Figure 4. Per the hyper-geometric test, the proteins linked to each toxicity were significantly targeted by the drugs that were associated with this toxicity. To determine the biological commonalities among the five toxicities, we used the enrichment score of the targets (based on the hypergeometric test) as a measure of their significance for each toxicity and computed the distance between each toxicity pair using Spearman’s correlation coefficient (Table 3). Spearman distance was selected partly because we were interested in how toxicities rank different proteins and partly because we wanted to avoid the effects of extreme cases and imposing extra assumptions on our data. Figure 5 shows the interrelationship among the five toxicities using a heatmap with a dendrogram of toxicity clustering. We found that PN, SJS/TEN, and LI were closely associated with one another whereas rhabdomyolysis differed significantly from the rest. When comparing the proposed targets (proteins) expressed in the organs/tissues where individual corresponding drug-induced toxicities occurred, we found 55 proteins (Table 4) to be both highly targeted by drugs causing a specific toxicity as well as highly expressed in the the corresponding tissue/organ.

Figure 4. Toxicity Network. Every drug-induced toxicity is linked to proteins significantly targeted by its drug list according to the hypergeometric test. Proteins are represented as small circles, and their color depends on the number of toxicities associated with them (green = 1, orange = 2).

whereas the other half caused multiple toxicities. The number of targets was inversely proportional to the number of toxicities. Per Spearman distance, PN was closest to SJS/TEN, and the pairs involving RM were more different than those pairs excluding RM. The significant protein targets associated with RM outnumbered those associated with each of the other four toxicities. There were 20 significant targets associated with two



DISCUSSION The majority of drugs on the combined list were linked to a single toxicity; approximately half of them caused a single toxicity, 918

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology Table 3. Pairwise Distance between Targets of Each Toxicitya toxicity 1

toxicity 2

Spearman Distance

peripheral neuropathy peripheral neuropathy SJS/TEN cardiotoxicity cardiotoxicity cardiotoxicity rhabdomyolysis rhabdomyolysis rhabdomyolysis rhabdomyolysis

SJS/TEN lung injury lung injury lung injury peripheral neuropathy SJS/TEN cardiotoxicity lung injury peripheral neuropathy SJS/TEN

0.479 0.629 0.731 0.791 0.912 0.951 1.159 1.381 1.590 1.658

toxicity pair, and the rank order was CT/LI (87 drugs) > PN/LI = SC/LI (74 drugs) > PN/SC (66 drugs) > PN/CT (60 drugs) > SC/CT (57 drugs) > RM/LI (46 drugs) > SC/RM (42 drugs) > RM/PN (38 drugs) > RM/CT (36 drugs), which differed from that suggested by pairwise Spearman distances. The rank order by Spearman distance analysis was PN/SC > PN/LI > SC/LI > CT/LI > CT/PN > CT/SC > RM/CT > RM/LI > RM/PN > RM/SC. SJS/TEN had the lowest number of unique drugs; however, it was only close to PN or LI per Spearman correlation. Clearly, delineating biological relationships among drug organ toxicities using Spearman distance analysis of targets provides insight into biological commonalities among drug toxicities. Conceivably, as the number and diversity of drugs associated with each of the five toxicities increase, different biological commonalities and interrelations will emerge, and the power of predicting specific drug toxicities using biological interrelationships among toxicities will increase. The 27 drugs that all interacted with 150 proteins or more (Table 1) are approved for treating diverse diseases, including cancers/tumors (cabergoline, imatinib, bexarotene, letrozole, ondansetron, and carboplatin), anti-inflammatory, pain and analgesic (ketoprofen, oxycodone, sufentanil, mefenamic acid, and oxaprozin), endocrine and metabolic diseases (troglitazone, cerivastatin, mifepristone, and simvastatin), CNS diseases (meprobamate, hydroxyzine, phenelzine, and triazolam), cardiovascular diseases (nifedipine and tocainide), antihistamine (doxylamine), GI diseases (carbenoxolone), asthma, allergic rhinitis, bronchoconstriction (montelukast), rare disease (miglustat as a substrate reduction therapy), Parkinson’s disease (entacapone), and multiple sclerosis (mitoxantrone). Surprisingly, the majority of these drugs were each associated with a

a

Pairwise distance between toxicities was calculated using Spearman’s rank correlation method for the 196 targets associated with one or more toxicities, as in Figure 4. Toxicity similarities were ranked accordingly. A lower pairwise distance value indicates greater similarity among the targets of given toxicities.

toxicities and 0 targets associated with three or more toxicities. Only three drugs (venlafaxine, sorafenib, and paroxetine) were associated with all five toxicities. Among the five toxicities, rhabdomyolysis had the highest number of unique drugs not shared by other toxicities (Figure 1), as reflected by Spearman distance analysis where rhabdomyolysis differed from other toxicities to the greatest extent. One might argue that comparison of common drugs shared by individual toxicity pairs would have also provided insight into interrelations between/among toxicities without utilizing targets (direct plus indirect) and computation. To address this question, we ranked toxicity pairs based on the number of drugs common to each

Figure 5. Heatmap with a dendrogram of toxicity clustering. Each of the 196 targets associated with one or more toxicities is represented by one row on the heatmap. The shading is based on the negative logarithm of probability that it is associated significantly with the given toxicity based on the hypergeometric test. Scores close to or above 1 (yellow) represent a high probability of association. Toxicities with similar ranking of protein targets are clustered together. The clustering does not change by including all of the protein targets. 919

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

Table 4. Proteins Significantly Enriched in Both Targets of Drug-Associated Specific Toxicity and in Gene Expression in Corresponding Tissue/Organ toxicity/tissue cardiotoxicity/ heart

lung injury/lung peripheral neuropathy/ nerve

rhabdomyolysis/ muscle

SJS/TEN/skin

matching proteins ADRB1 (adrenoceptor beta 1) KCNH2 (potassium channel, voltage-gated eag-related subfamily H, member 2) ACO1 (aconitase 1) SLC27A2 (solute carrier family 27 (fatty acid transporter), member 2) ACSS1 (acyl-CoA synthetase short-chain family member 1) ACAT2 (acetyl-CoA acetyltransferase 2) MAP2 (microtubule-associated protein 2) SMOX (spermine oxidase); PPARG (peroxisome proliferator-activated receptor gamma) PLA2G10 (phospholipase A2 group X) PRKCE (protein kinase C, epsilon) GALC (galactosylceramidase) MASP2 (mannan-binding lectin serine peptidase 2) MMP3 (matrix metallopeptidase) PNLIP (pancreatic lipase) APOB (apolipoprotein B) ATP5D (ATP synthase, H+ transporting, mitochondrial F1 complex, delta subunit) BMP2 (bone morphogenetic protein 2) CASP1 (caspase 1); CASP3 (caspase 3) CBS (cystathionine-beta-synthase) CHRM3 (cholinergic receptor, muscarinic 3); CHRM4 (cholinergic receptor, muscarinic 4); CHRM5 (cholinergic receptor, muscarinic 5) CTSS (cathepsin S) GABRB3 (gamma-aminobutyric acid (GABA) A receptor, beta 3) HRH2 (histamine receptor H2) HTR1B (5-hydroxytryptamine (serotonin) receptor 1B, G protein-coupled provided); IFNAR1 (interferon (alpha, beta, and omega) receptor 1) IFNAR2 (interferon (alpha, beta, and omega) receptor 2) IL10 (interleukin 10); IL6 (interleukin 6); MTHFD1 (methylenetetrahydrofolate dehydrogenase (NADP+ dependent) 1, methenyltetrahydrofolate cyclohydrolase, formyltetrahydrofolate synthetase); SCNN1A (sodium channel, non voltage gated 1 alpha subunit provided); SDS (serine dehydratase) SLC18A2 (solute carrier family 18 (vesicular monoamine transporter), member 2); SLC26A6 (solute carrier family 26 (anion exchanger), member 6); SLC6A4 (solute carrier family 6 (neurotransmitter transporter), member 4) SOCS3 (suppressor of cytokine signaling 3) ACE (angiotensin I converting enzyme) CA1 (carbonic anhydrase I); CA2 (carbonic anhydrase II); CA4 (carbonic anhydrase IV) CACNA1I (calcium channel, voltage-dependent, T type, alpha 1I subunit); CACNA2D2 (calcium channel, voltage-dependent, alpha 2/delta subunit 2) COX1 (cytochrome c oxidase subunit 1); COX5B (cytochrome c oxidase subunit Vb); COX6A2 (cytochrome c oxidase subunit VIa polypeptide 2) DHFR (dihydrofolate reductase) HMGCR (3-hydroxy-3-methylglutaryl-CoA reductase) IMPA2 (inositol(myo)-1(or 4)-monophosphatase 2) KLK5 (kallikrein-related peptidase 5) MDM4 (MDM4, p53 regulator) MMP7 (matrix metallopeptidase 7)

are known to be involved in cellular oxidation of fatty acids.10 SLC27A2 encodes the solute carrier family 27 (fatty acid transporter), member 2, ACSS1 encodes acyl-CoA synthetase short-chain family member 1; ACAT2 encodes acetyl-CoA acetyltransferase 2; and PPARG encodes peroxisome proliferator-activated receptor gamma.10 For LI and lung tissue, PRKCE encodes protein kinase C epsilon, which is reportedly involved in protecting human alveolar epithelial cells.11 For PN and nerves, there were only four targets identified in our enrichment analysis, among which GALC encoding galactosylceramidase is involved in Krabbe disease,12 a degenerative disease affecting the neural system; MMP3 encoding matrix metalloproteinase 3 has been suggested to be a therapeutic target following traumatic nerve injury.13 For RM and muscle, ATP5D encodes mitochondrial ATP synthase, which catalyzes ATP synthesis;10 ATP deficiency caused by genetic mutations or drugs could cause rhabdomyol-

single toxicity even though every one of them interacted with a larger number of targets. The range of biological functions associated with the 31 proteins in Table 2 is broad, as evidenced by the associated KEGG pathways. Common toxicological networks are expected to be shared among some ARs of a drug or among specific ARs shared by different drugs. Drugs of different classes could have similar adverse reactions if interacting with the same targets. Enrichment analysis of targets of drugs associated with a specific toxicity and gene expression in corresponding organ/ tissues enlisted 55 targets (Table 4), providing an important list of proteins to be avoided to reduce drug attrition. Normally, the main energy source for heart contraction comes from oxidation of fatty acids but is switched to glucose metabolism when the heart begins to fail.9 Among the proteins significantly enriched in cardiac toxic drug targets and in heart cells (Table 4), the proteins encoded by SLC27A2, ACSS1, ACT2, and PPARG genes 920

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

Figure 6. Integrative pharmacologically based predictions of drug toxicity. Drugs associated with specific categories of serious adverse reactions (organ toxicities) are curated and compiled along with targets (direct and indirect) for analysis of the common mode of action or biological commonalities shared among them. Such analysis across key toxicities of concern in combination with omics data and biological/medical evidence would enable robust predictions of toxicity profiles for new drug candidates.

ysis-related side effects.5,14 The proteins encoded by IFNAR1, IFNAR2, IL10, and IL6 are involved in regulating muscle inflammation and regeneration. Pro-inflammatory cytokines play a role in diseases involving rhabdomyolysis,15 and cytokines are regulators of muscle regeneration.16 Elevation of SOCS3 protein (suppressor of cytokine signaling 3 encoded by SOCS3) and IL6 have been associated with aging-related dysfunction of human muscle stem cells,17 and SOCS3 is also proposed to be a therapeutic target for aging-related muscle wasting.18 For SJS/ TEN, expression of carbonic anhydrase II encoded by CA2 is elevated in the skin of atopic dermatitis patients.19 This enzyme belongs to a family of different isoforms and maintains intracellular pH and thereby cellular homeostasis.20 Increased activity of kallikrein-related peptidase 5, encoded by KLK5, due to genetic deficiency of LEKTI serine protease inhibitor is linked to skin inflammation in Netherton syndrome.21 Matrix metallopeptidase 7 encoded by MMP7 is a member of metrix metallopeptidases that play a vital role in wound healing and skin health.22 In brief, our enrichment analysis of targets associated with specific toxicities and their gene expression profiles in corresponding organ/tissue provide insight into drug toxicity in the context of organ biology. This study enables us to understand drug toxicity in a different context than what we published previously3−7 using drug-induced gene perturbation data in cancer cell lines (Connectivity Map, https://www. broadinstitute.org/cmap/). Among the five drug-induced toxicities, human leukocyte antigen (HLA) alleles have been identified to be risk markers of drug-induced SJS/TEN,23 which is the only toxicity without a natural disease counterpart. Drugs causing SJS/TEN were previously shown to perturb major histocompatibility complex (MHC) genes,4 also known as HLA genes in humans. On the other hand, immune-mediated peripheral neuropathy constitutes a subset of peripheral neuropathy.24 In patients with chronic obstructive pulmonary disease (COPD), development of peripheral neuropathy25 or muscle weakness26 has been reported. Drug-induced SJS/TEN could cause complications of damaging internal organs, including heart, lung, liver, and

kidney.27 Cardiac complications including congestive heart failure resulting from drug-induced SJS cases have been reported.28 Cases of bronchiolitis obliterans following druginduced SJS/TEN have been reported.29,30 Specific biological commonalities are shared by individual groups of drug toxicities or by drug toxicities and natural diseases. These reports support learning the relations among and between natural diseases for our understanding and prediction of drug-induced toxicities across organs/tissues. Biological factors involved in the pathological progression of a natural disease are important references for understanding drug toxicity. In addition to the target organ or tissue (and tumor), a drug is in fact differentially distributed throughout the body by diffusion and convection following administration, depending on its physicochemical characteristics. The intensity of interaction between a drug and individual off-target proteins would vary depending on its local concentration, drug−protein interaction affinity, and the expression levels of individual targets. The major limitation of our approach is that we do not have numerical values to reflect differential affinity of each drug interacting with a target across the board for all the drugs and targets used. Another limitation in our analysis is that the temporal relationship between the duration of exposure to a drug treatment and differential onsets of a specific adverse reaction in individual patients cannot be captured by pathway analysis of the expression data at hand. Nor are temporal associations in relation to drug treatment among the five toxicities studied. Such temporal associations require additional benchtop-to-bedside quality data and are influenced by host factors including comorbidities and the indications to which individual drug regimens are prescribed; some drugs are only used for a short duration, such as antibiotics, whereas others are used chronically, such as drugs for treating diabetes, hypertension, and hyperlipidemia. Host factors due to genetic polymorphisms and epigenetic influences may cause differential expressions of proteins and thereby contribute to rare but serious drug-induced toxicities. Current bioinformatics approaches have not been able to take host factors into 921

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology

(2) Bowes, J., Brown, A. J., Hamon, J., Jarolimek, W., Sridhar, A., Waldron, G., and Whitebread, S. (2012) Reducing safety-related drug attrition: the use of in vitro pharmacological profiling. Nat. Rev. Drug Discovery 11, 909−922. (3) Hur, J., Guo, A. Y., Loh, W. Y., Feldman, E. L., and Bai, J. P. (2014) Integrated systems pharmacology analysis of clinical drug-induced peripheral neuropathy. CPT: Pharmacometrics Syst. Pharmacol. 3, e114. (4) Hur, J., Zhao, C., and Bai, J. P. (2015) Systems pharmacological analysis of drugs inducing stevens-johnson syndrome and toxic epidermal necrolysis. Chem. Res. Toxicol. 28, 927−934. (5) Hur, J., Liu, Z., Tong, W., Laaksonen, R., and Bai, J. P. (2014) Druginduced rhabdomyolysis: from systems pharmacology analysis to biochemical flux. Chem. Res. Toxicol. 27, 421−432. (6) Melas, I. N., Sakellaropoulos, T., Iorio, F., Alexopoulos, L. G., Loh, W. Y., Lauffenburger, D. A., Saez-Rodriguez, J., and Bai, J. P. F. (2015) Identification of drug-specific pathways based on gene expression data: application to drug induced lung injury. Integr. Biol. 7, 904−920. (7) Melas, I. N., Hur, J., and Bai, J. P .F. Pharmacological Networkbased Assessment of contraction-related drug-induced cardiotoxicity, unpublished data. (8) GTEx Consortium (2015) Human genomics. The GenotypeTissue Expression (GTEx) pilot analysis: multitissue gene regulation in humans. Science 348, 648−660. (9) Ingwall, J. S. (2009) Energy metabolism in heart failure and remodelling. Cardiovasc. Res. 81, 412−419. (10) NCBI Gene. http://www.ncbi.nlm.nih.gov/gene/ (accessed January 2016). (11) Wang, W., Jia, L., Wang, T., Sun, W., Wu, S., and Wang, X. (2005) Endogenous calcitonin gene-related peptide protects human alveolar epithelial cells through protein kinase Cepsilon and heat shock protein. J. Biol. Chem. 280, 20325−20330. (12) Genetics Home Reference. Krabbe disease. http://ghr.nlm.nih. gov/condition/krabbe-disease (accessed January 2016). (13) Chao, T., Frump, D., Lin, M., Caiozzo, V. J., Mozaffar, T., Steward, O., and Gupta, R. (2013) Matrix metalloproteinase 3 deletion preserves denervated motor endplates after traumatic nerve injury. Ann. Neurol. 73, 210−223. (14) Marciante, K. D., Durda, J. P., Heckbert, S. R., Lumley, T., Rice, K., McKnight, B., Totah, R. A., Tamraz, B., Kroetz, D. L., Fukushima, H., Kaspera, R., Bis, J. C., Glazer, N. L., Li, G., Austin, T. R., Taylor, K. D., Rotter, J. I., Jaquish, C. E., Kwok, P. Y., Tracy, R. P., and Psaty, B. M. (2011) Cerivastatin, genetic variants, and the risk of rhabdomyolysis. Pharmacogenet. Genomics 21, 280−288. (15) Hamel, Y., Mamoune, A., Mauvais, F. X., Habarou, F., Lallement, L., Romero, N. B., Ottolenghi, C., and de Lonlay, P. (2015) Acute rhabdomyolysis and inflammation. J. Inherited Metab. Dis. 38, 621−628. (16) De Paepe, B., and De Bleecker, J. L. (2013) Cytokines and chemokines as regulators of skeletal muscle inflammation: presenting the case of Duchenne muscular dystrophy. Mediators Inflammation 2013, 540370. (17) McKay, B. R., Ogborn, D. I., Baker, J. M., Toth, K. G., Tarnopolsky, M. A., and Parise, G. (2013) Elevated SOCS3 and altered IL-6 signaling is associated with age-related human muscle stem cell dysfunction. Am. J. Physiol. Cell. Physiol. 304, C717−728. (18) Palus, S., von Haehling, S., and Springer, J. (2014) Muscle wasting: an overview of recent developments in basic research. Int. J. Cardiol. 176, 640−644. (19) Kamsteeg, M., Zeeuwen, P. L., de Jongh, G. J., Rodijk-Olthuis, D., Zeeuwen-Franssen, M. E., van Erp, P. E., and Schalkwijk, J. (2007) Increased expression of carbonic anhydrase II (CA II) in lesional skin of atopic dermatitis: regulation by Th2 cytokines. J. Invest. Dermatol. 127, 1786−1789. (20) Boron, W. F. (2004) Regulation of intracellular pH. Adv. Physiol. Educ. 28, 160−179. (21) Furio, L., Pampalakis, G., Michael, I. P., Nagy, A., Sotiropoulou, G., and Hovnanian, A. (2015) KLK5 Inactivation Reverses Cutaneous Hallmarks of Netherton Syndrome. PLoS Genet. 11, e1005389.

consideration on a genome-wide scale to relate drug toxicities among one another. Understanding the relations among drug toxicities and/or natural diseases through the lens of common biology will enable evidence-based prediction of clinical drug toxicity along with ever-advancing omics technologies (transcriptomics, phosphoproteomics, and genomics) and imaging technologies (Figure 6) and pave the path for us to move forward the goal of depicting and identifying common biology among drug toxicities and natural diseases. In summary, predictive computation of druginduced toxicities in the context of human biology could potentially guide drug development most cost effectively.



ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.chemrestox.6b00060. Legends for the tables in the Supporting Information Excel file (PDF) (Table S1) The five lists of toxicities (SIDER (http:// sideeffects.embl.de/) side effects (name and internal id) and Drugs@FDA) that, upon manual curation, were deemed indicative of each toxicity; (Table S2) summary of the drugs examined in this study; (Table S3) information regarding the protein targets; (Table S4) the adjacency matrix of the protein−drug interaction network; and (Table S5) a summary of the tissue enrichment analysis (XLSX)



AUTHOR INFORMATION

Corresponding Author

*Tel: 301-796-2473. E-mail: [email protected]. Present Addresses §

T.S.: School of Mechanical Engineering, National Technical University of Athens, Athens, Greece ⊥ T.J.H.: Biomedical Engineering, University of Virginia, Charlottesville, VA 22904, United States Funding

This work was supported by the U.S. Food and Drug Administration ORISE Fellowships to T.S. and T.J.H. through a 2012 FDA Critical Path grant to J.P.F.B. The views expressed in this article do not represent the views of the U.S. Food and Drug Administration. Notes

The authors declare no competing financial interest.



ABBREVIATIONS PN, peripheral neuropathy; RM, rhabdomyolysis; SJS/TEN, Stevens−Johnson syndrome/toxic epidermal necrosis; LI, lung injury; CT, heart contraction-related cardiotoxicity; ARs, adverse reactions; SIDER, side effect resource; GTEx, genotype−tissue expression database; STITCH, search tools for interactions of chemicals; CNS, central nervous system; KEGG, Kyoto encyclopedia of genes and genomes; HLA, human leukocyte antigen; MHC, major histocompatibility complex; COPD, chronic obstructive pulmonary disease



REFERENCES

(1) Lee, A. Adverse Drug Reactions. Pharmaceutical Press, 2nd ed. Royal Pharmaceutical Society: London, UK, 2005. 922

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923

Article

Chemical Research in Toxicology (22) Philips, N., Auler, S., Hugo, R., and Gonzalez, S. (2011) Beneficial regulation of matrix metalloproteinases for skin health. Enzyme Res. 2011, 427285. (23) Leckband, S. G., Kelsoe, J. R., Dunnenberger, H. M., George, A. L., Jr., Tran, E., Berger, R., Muller, D. J., Whirl-Carrillo, M., Caudle, K. E., and Pirmohamed, M. (2013) Clinical Pharmacogenetics Implementation Consortium guidelines for HLA-B genotype and carbamazepine dosing. Clin. Pharmacol. Ther. 94, 324−328. (24) Leger, J. M., Guimaraes-Costa, R., and Muntean, C. (2016) Immunotherapy in Peripheral Neuropathies. Neurotherapeutics 13, 96− 107. (25) Oncel, C., Baser, S., Cam, M., Akdag, B., Taspinar, B., and Evyapan, F. (2011) Peripheral neuropathy in chronic obstructive pulmonary disease. COPD: Journal of Chronic Obstructive Pulmonary Disease 7, 11−16. (26) Breunung, L., and Roberts, M. (2011) Peripheral muscle dysfunction and chronic obstructive pulmonary disease. Br. J. Hosp. Med. 72 (17−18), 17−21. (27) Mayo Clinic Staff (accessed 2015). Stevens-Johnson Syndrome. http://www.mayoclinic.org/diseases-conditions/stevens-johnsonsyndrome/basics/definition/con-20029623. (28) Ahmed, Y. I., Azeem, S., Khan, O., Majid, T. H., Ahmed, D., Amin, A., Mahmood, A., and Hameed, K. (2004) Stevens Johnson syndrome in Pakistan: a ten-year survey. J. Pak. Med. Assoc. 54, 312−315. (29) Bakirtas, A., Harmanci, K., Toyran, M., Razi, C. H., and Turktas, I. (2007) Bronchiolitis obliterans: a rare chronic pulmonary complication associated with Stevens-Johnson syndrome. Pediatr. Dermatol. 24, E22− 25. (30) Tsunoda, N., Iwanaga, T., Saito, T., Kitamura, S., and Saito, K. (1990) Rapidly progressive bronchiolitis obliterans associated with Stevens-Johnson syndrome. Chest 98, 243−245.

923

DOI: 10.1021/acs.chemrestox.6b00060 Chem. Res. Toxicol. 2016, 29, 914−923