Recent work in identifying a drug’s marketed disease, or indication, from protein interactions found that this identification was maximized by considering protein interactions that are in close proximity to a drug’s target [22,23,24]. Thus, we hypothesized that protein-protein interactions that are proximal to a drug’s target(s) could provide insight into mechanisms of drug safety and efficacy.

To create drug interaction pathways, we pulled interaction data from iRefWeb[15], Reactome[25], PharmGKB[6], and a curated set of predicted drug-protein binding data (see Methods section). We merged and scored (explained in methods) these data to yield an interaction network of 25,604 nodes and 318,644 edges. The number of interactions and interaction score distributions are in S1 Fig.

Our algorithm, PathFX, selects a drug target’s most relevant interaction edges (local interaction neighborhoods), merges neighborhood networks from all drug targets, and then identifies enriched phenotypes–which could represent either safety or efficacy phenotypes–in the interaction neighborhood (Fig 1, and usage summary in S4 Fig). In this context, safety phenotypes included associations such as adverse events or side effects (e.g. “pancreatitis”, “adverse weight gain”), and efficacy phenotypes included disease associations (e.g. “diabetes”, “major depressive disorder”); some phenotypes (e.g. “hypertension”) could belong to both of these groups. To identify which phenotypes are associated with the drug target network, we merged data from multiple sources: DisGeNet[9,10], Phenotype Genotype Integrator (PheGenI)[26], ClinVar [27], OMIM [8], and PheWas [11,12]. In this process we controlled for multiple biases as follows:

PathFX uses a threshold parameter for selecting proteins included in each drug network. We derived this threshold based on the available data and did not tune the parameter to improve identification accuracy (explained in An optimal threshold parameter for the tissue non-specific network in Methods).

We first applied the PathFX method to the diabetic medication, metformin. DrugBank[29] listed five protein targets for metformin—SLC22A2, SLC22A3, PRKAB1, SLC47A1, and SLC29A4 –for metformin that were in our interactome (note that some of these are transport proteins that may be included because metformin inhibits them, not as a pharmacological effect). We created a drug interaction network pathway based on all listed proteins using PathFX. This yielded a 25-protein final neighborhood (20 proteins + 5 drug targets) significantly associated with 18 phenotypes (Fig 2A, S1 Table). Diabetes mellitus type 2 and diabetes mellitus type 1 are both associated with the metformin pathway via interactions with 12 genes (Fig 2B, S1 Table). Metformin’s protein targets were not sufficient to describe the association to diabetes mellitus type 1 and diabetes mellitus type 2 when we analyzed phenotypic associations with these targets. However, with the full 25 protein network identified by PathFX, we recovered the association to diabetes mellitus type 1 and diabetes mellitus type 2.

We collected a benchmarking set of approved drugs to test our algorithm’s utility in accurately identifying diseases that the drug is known to effectively treat. This set included marketed drugs with approved disease indications. We first started with marketed, non-palliative drugs analyzed in [22] to compare our performance with this seminal work. This data set included 238 drugs associated with one or more disease indications, yielding a total of 403 drug-indication pairs. We augmented this dataset by using repoDB[30] to add additional approved indications for the original drug set (we excluded data from repoDB where the trial was terminated or ongoing); using repoDB, we added 1353 drug-indication pairs, yielding a total of 1756 drug-indication pairs for testing. The full list of drugs and approved indications are included in S1 File and consists of a list of drugs associated with one or more disease indications. Most drugs were approved for fewer than 10 indications, though, prednisolone and hydrocortisone are used to treat 105 and 98 indications respectively (S1 File, Fig 3A & 3B).

The dataset included drugs used to treat 594 indications. We binned these 594 indications into 62 clusters based on the semantic similarity of the disease indications. We used the interactive mmlite interface to metamap[31], a highly configurable program developed to map biomedical text to the UMLS Metathesaurus, to map diseases to the nearest Unified Medical Language System (UMLS) CUI (Concept Unique Identifier) identifier. We selected the UMLS terminology because this system had the greatest coverage of phenotypes in our dataset and contained mappings from many popular languages (such as MedDRA). We then clustered these diseases based on ontological, semantic similarity using the UMLS::Similarity package in Perl[32] (cluster membership in S1 File, Fig 3C). For instance, cluster four contained two CUI terms–C0497327, C0002395 –that mapped to 24 Alzheimer’s and dementia phenotypes (S1 File). Cluster five contained five CUI terms–C0042842, C0042875, C0030783, C0016412, C0936215 –that mapped to nine diseases associated with vitamin deficiency. When testing PathFX, we analyzed and reported whether the algorithm identified the drug’s original, un-clustered indication, and also reported results based on the indications’ cluster to assess trends in the types of diseases where we had better identification capacity.

We used the UMLS::Similarity tools for determining if PathFX identified phenotypes that matched the drug’s marketed indication. In this case, we regarded a match as any phenotype significantly associated to the network; for most drugs, PathFX identified multiple phenotypes as statistically significantly associated to the drug’s network. For each drug, we pulled these significant phenotypes from our PathFX analysis as above, and matched these identified phenotypes to CUI identifiers, also using mmlite[31]. We measured semantic similarity using Lin distance in the UMLS::Similarity package[32]. For example, the drug enoxaparin is indicated for deep vein thrombosis and myocardial infarction; our algorithm identified that enoxaparin’s drug pathway was significantly associated with “deep venous thrombosis” (direct match, semantic similarity = 1.0), and other similar diseases such as “thrombosis” (semantic similarity = 0.8615), “venous thromboembolism” (semantic similarity = 0.6853), and “myocardial infarction” (semantic similarity = 0.9377) (full results in S1 File, Fig 4A). Of the 1756 drug-indication pairs, PathFX could not create an interaction network for two pairs (tolazamide + diabetes type 1, and tolazamide + diabetes type 2) because this drug’s target was not mapped to a gene symbol in our interactome. For 389 drug-indication pairs, metamap was unable to map the marketed indication to a CUI term so we could not assess whether the PathFX identified indications matched the marketed indications (cluster number 58 in Fig 4C and in S1 File). This left 1366 pairs for further analysis.

For this analysis, PathFX found pathway information for 171 of the 403 drug-indication pairs (42.4%) from Guney et al[22] and 558 of the 1364 (40.9%) eligible drug-disease pairs in our expanded set of benchmarking drugs; this is our best estimate of PathFX sensitivity. At the drug level, 141 out of 236 drugs (there were two of the original 238 drugs without sufficient binding information in DrugBank) had at least one identified phenotype that was similar to one of the drug’s marketed indication(s) (59.8%). Guney et al[22] reported a sensitivity of 15.4% (they matched 62 of 403 drugs-indication pairs), demonstrating improved sensitivity from our drug-target-centric approach.

For comparison with PathFX, we analyzed the disease associations of the drug targets alone without the local interaction information. Using only drug targets, we identified statistically significant associations between 751 drug and disease pairs (54.98% sensitivity). Of these pairs, 409 were also identified using PathFX pathway information, leaving 342 drug-disease pairs that were identified by targets alone and 147 drug-disease pairs that were only identified when pathway information was included from PathFX (S1 File). After merging gene-to-phenotype associations from multiple data sources, each phenotype had a set of associated genes and the size of this gene set distinguished which diseases each method detected (Kruskal-Wallis statistics = 33.6, p-value = 5.04x10^-8) (Fig 4B). PathFX was biased towards selecting phenotypes with fewer genes (median gene set size of 90 genes); targets-only analysis was biased to selecting phenotypes with more associated genes (median gene set size of 342.5 genes) (Mann-Whitney-U statistic: 33089, p-value 1.43x10^-8) (Fig 4B). The median gene set size where both methods detected the phenotype was 257 genes.

For completeness, we calculated positive and negative predictive values (PPV, NPV) (S2 File). To calculate PPV and NPV, we made a conservative assumption that any phenotype associated with a drug that was not a marketed disease indication was a false positive. Because PathFX was designed to search broadly for drug-associated phenotypes, the PPV and NPV were deflated and inflated respectively (S3 Fig).

We analyzed PathFX identifications in the context of the 62 disease clusters (Table 1, Fig 4C). For 58 of the 62 clusters, PathFX found pathway evidence supporting the drug’s marketed indication for at least one of the drug-disease pairs assigned to that cluster (Table 1, S1 File). For instance, the top cluster contained three CUI terms which mapped to five disease phenotypes diseases: inappropriate adh syndrome, acromegaly somatic, hyperprolactinemia, acromegaly, and prolactin excess (CUI terms C0021141, C0001206, and C0020514) (Table 1). There were five drug-indication pairs for these diseases: tolvaptan (inappropriate adh syndrome), octreotide (acromegaly somatic), bromocriptine (hyperprolactinemia), bromocriptine (acromegaly somatic), and cabergoline (hyperprolactinemia). PathFX identified phenotypes for all five drug-disease pairs. The cluster with the second highest identification rate contained four CUI terms that mapped to seven disease terms, which are pathophysiologically distinct: alcohol withdrawal delirium, restless legs syndrome, premenstrual dysphoric disorder, insomnia, nicotine dependence, late insomnia, sleeplessness. There were seven drug-indication pairs included in this cluster: gabapentin (restless legs syndrome), ropinirole (restless legs syndrome), rotigotine (restless legs syndrome), diphenhydramine (late insomnia), estradiol (premenstrual dysphoric disorder), nicotine (nicotine dependence), and diazepam (alcohol withdrawal syndrome). PathFX identified the original, un-clustered phenotype for six of the seven pairs (PathFX did not identify estradiol’s association to premenstrual dysphoric disorder) (Table 1).

For the remaining four clusters, PathFX did not identify the drugs’ marketed indication for any of the drug-indication pairs assigned to these clusters (S1 File). These clusters are numbered 7, 13, 35, and 1. Additionally, cluster 58 contained 216 disease indications, of which 123 diseases were not mapped to a CUI term.

Understanding and prioritizing drug safety signals are important regulatory concerns[33,34]. The FDA Adverse Event Reporting System (FAERS) is a repository of voluntarily submitted case reports of adverse events that occur when a patient is on a particular medication. Multiple confounding variables, including comorbidities, incomplete reports, and polypharmacy[35], make it difficult to determine when a drug is causative for an AE. This makes signal detection and triaging reports difficult. Further, when associations are not directly explained by drug targets, pathway models may provide additional justification for why a drug is associated with an AE. Anticipating and identifying drug-induced AEs are critical for novel therapeutics in development as well; and thus, further characterizing which drug targets and interacting proteins may cause AEs is an important goal of this work. Designated medical events (DMEs) are serious, significant adverse drug events of special concern to regulators. As such, DMEs, selected based on expert medical review, were evaluated in this study.

To assess the utility of PathFX for signal detection, we extracted drug-adverse event pairs from FAERs for 1906 drugs across 35 DMEs and ran PathFX on these 1906 drugs. Case reports were associated with one of the 35 DMEs if the adverse event mapped to the preferred MedDRA DME term. Close synonyms were used for some DMEs to better capture reporting (e.g., “pancreatitis” was captured by “pancreatitis” and “pancreatitis acute”). PathFX identified the adverse event phenotype for the input drugs between 0.24% - 39.57%(Table 2 and S3 File), depending on the DME. For instance, in this noisy data set, 1045 drugs were reported as having an adverse association with pancreatitis. PathFX identified that 282 of these drugs were associated with pancreatitis (26.99%). Similarly, 1150 drugs were reported to have an association with myocardial infarction and PathFX identified 391 (34.00%) of these drug-DME associations. For eight of the DME phenotypes, we found no pathways associations to the reported drugs (S3 File).

We used this FAERS dataset to estimate a lower bound on the specificity of PathFX. Because FAERS contains many more drug-DME associations than are real, we treated any drugs without a reported DME association as silver-standard negatives; reasoning that if a noisy sampling of the FAERS system contained no association between the drug and the DME, that these pairs were sufficient negatives. We asked how often PathFX associated a drug with a DME when no case report existed to calculate the specificity rate for the 35 DMEs in this analysis; the rate varied from 80.84%-99.91% (Table 2, S3 File). For instance, 1261 drugs were reported to have an association with hypertension, leaving 645 of our original 1906 drugs without an association to hypertension. Of these 645 silver-standard negatives, PathFX identified an association with hypertension for 112 (17.4%) of the drugs (82.64% specificity). For cardiac arrest, 1211 drugs were reported to have an association, leaving 695 drugs without an association. Of these drugs, PathFX identified seven drugs to have an association, estimating a specificity of 98.99% (Table 2).

PathFX identified multiple phenotypes for each drug even if the drug only has a single approved indication. We sought support for the additional identified phenotypes from two data sets: (1) a list of off-label drug uses extracted from the electronic medical record[36] and (2) drugs currently in clinical trials. In the EMR data, we found support for six drugs applied in 11 off-label indications (Table 3, listed as ‘Jung CUI’ and ‘Jung Disease’). For instance, telmisartan is indicated for hypertension, though PathFX and the EMR dataset supported the use of this drug as an anti-diabetic; this conclusion is further supported in the literature[37]. PathFX identified that thiothixene would be broadly applicable to depressive disorders beyond the indicated use for schizophrenia. PathFX also identified that etanercept, an anti-TNF-alpha drug used in auto-immune disorders, would be applicable to colitis and this identification was supported by the Jung dataset (Table 3). PathFX identified associations for 4 drug-indication pairs supported by additional clinical trials (Table 4): sunitinib for viral infections[38], erlotinib for viral infections[38], ketoprofen for lymphoedema[39], and sirolimus for dystrophic bullosa [40].

Given evidence that EMR and clinical trial data supported PathFX predictions, we further scrutinized PathFX identifications to identify drug repurposing opportunities (Fig 5A); we inferred that a non-marketed indication could be a repurposing opportunity if the interaction path was found in the network of a drug marketed for this indication. We demonstrated an example with leuprolide and triptorelin, both gonadotropin-releasing hormone (GnRH) agonists (Fig 5B). PathFX also identified that the triptorelin pathway is enriched for associations with endometriosis, and these interaction pathways are supported by leuprolide’s pathway. PathFX also identified that the antispasmodic, flavoxate, could be indicated for urticaria based on interaction paths shared with cyproheptadine and promethazine, two anti-histamines already approved for urticaria. In total, we identified 2,043 new drug-disease associations for 215 drugs (S4 File). We ranked these predictions based on the number of diseases identified for a drug (top 20 in Table 5, remainder in S4 File), and the number of interaction paths supporting a drug-disease association (top 20 in Table 6, remainder in S4 File).

We downloaded data from iRefWeb version 13.0 human, Reactome, and PharmGKB. We chose iRefWeb because the source contains interactions from BIND, BioGRID, CORUM, DIP, IntAct, HPRD, MINT, MPact, MPPI, and OPHID. We extracted protein-protein interactions from http://irefindex.org/wiki/index.php?title=iRefIndex, drug-variant interactions from PharmGKB, and protein-protein interactions from Reactome. We scored iRefWeb protein-protein interactions using the MIScore framework[45] and removed low-scoring interactions below the median score value, 0.244, to save memory in later computations. We kept the relative weighting of each score component equal (i.e. K_m = K_p = K_t = 1). The scoring framework represents the amount of evidence supporting the interaction of two proteins: the S_m represents the method used to detect the interaction and is higher for more dedicated experimental techniques. S_p reflects the number of publications supporting an interaction. This score increases with the number of publication and plateaus. S_t reflects the interaction type. Because we only used ‘direct’ interactions, this score is always 1.

We adapted the MIScore framework for PharmGKB data and used publication, and ‘clinical evidence’ to score drug-variant relationships. Whenever an interaction with a variant was added to our network, we also added an interaction edge from the variant to the gene and scored this interaction as 0.99, the maximum possible score in the interactome. We used the following equation where K_p = K_e = 1, S_p was the same as published in [45]. S_e reflects the clinical level evidence available from PharmGKB and we crafted a scoring framework similar to [45].

Where scv_’1A’ = 0.99, scv_’1B’ = 0.86625, scv_’2A’ = 0.7425, scv_’2b’ = 0.61875, scv_’3’ = 0.495, and scv_’4’ = 0.2475. Because interactions in PharmGKB only receive one level of clinical evidence, a and b collapse to: a=scvi;b=a+∑scvi

We adapted this scoring framework for Reactome pathways using the following equation: SMI=KmSm(cv)+KpSp(n)+KtSt(cv)Km+Kp+Kt

S_p was the kept the same as in [45]. We derived a scheme for the method component, S_m, where scv_{’direct_complex’} = 1.00, scv_{’neighboring_reaction’} = 0.66, scv_{’indirect_complex’} = 0.8, scv_{’reaction’} = 1.0, and scv_{’unknown’} = 0.05.

Because the two maximum scoring categories were ‘direct_complex’ and ‘reaction’, the Max(Gscv_i) term = 2.0. Because Reactome interactions did not include a method of detection but are curated interactions, we gave these interactions a cv_i = 0.8. Because they all received the same ‘method’ score, calculating a and b yields a S_m = 0.615.

We lastly incorporated predicted drug to protein binding data based on PocketFEATURE [46] where drug-protein pairs were scored based on the similarity between the drug’s known targets and other protein targets from the Protein Data Bank[47] (See methods below). PocketFEATURE has been extensively validated on predicting drug protein interactions in multiple applications [46,48,49]. In all cases, we estimated interaction scores based on the quality of evidence available; these edge scores were fixed before applying PathFX and we did not alter these parameters to improve prediction accuracy.

We downloaded variant and phenotype association data from PheWAS[11,12], disease to gene associations from DisGeNet[10,11], Phenotype-Genotype Integrator (PheGenI) [26], ClinVar[27], and OMIM[8], and eQTL data from the GWAS catalogue[7]. We collapsed all phenotype names to CUI identifiers using MetaMap lite and took the union of all data sources to create our source of gene to phenotype annotations. This yielded a database associating 29785 genes to 20524 phenotypes.

To look for enriched phenotypes, we used a Fisher’s exact test and Benjamini-Hochberg multiple hypothesis correction to assess whether a disease or phenotype had more associations to the network genes relative to the total number of associations in the interactome. We filtered out disease or phenotypes associated with fewer than 25 genes. Recent work suggested that current interactomes are insufficient for analyzing disease pathways with fewer than 25 genes[28]. More specifically, interactomes are incomplete and these missing interactions reduce the ability to find pathways between smaller disease modules. This recent estimate discovered that missing interactions disproportionally affect phenotype pathways with fewer than 25 genes and thus, we eliminated these phenotypes knowing that our interaction network was insufficient for studying these phenotypes.

We further filtered phenotypic predictions by deriving a p-value threshold from networks created with randomly-selected, druggable proteins. We realized that if any random set of input drug-targeting proteins could discover a statistically-significant association to a given phenotype, that associating a real set of drug proteins with this phenotype would reflect bias in our data instead of a biologically-meaningful result. To assess this bias in our data, we created 100 networks using randomly selected drug targets from the intersection of all drug-targeting proteins in DrugBank and ran PathFX using these targets as inputs (i.e. using the same statistical approaches to assess phenotypes that are significantly associated to networks created with random inputs). Because the distribution of p-values from the 100 randomizations was not normally distributed, we used the median value as the threshold p-value for determining if a drug network was associated to a phenotype. The number of randomly selected input proteins matched the number of targets of the drug of interest. PathFX retained a phenotype if the association is more significant than the p-value threshold for that phenotype run with the same number of random, input protein targets.

When analyzing drug targets alone, we again used the Fisher’s exact text, the Benjamini-Hochberg correction, and filtering relative to the expected p-value threshold.

We created a depth-first search tool which ‘walked’ away from a drug’s protein target through the interactome and evaluated these paths using the multiplicative sum of all interaction edges between a gene and the target. Genes with a path score above the threshold were retained in the drug pathway. We expedited the search with “fast-tracking”. This process reflects the fact that molecules exist in highly interconnected pathways and assumed that we could reduce the searchable interaction spaces by looking for molecular redundancies. As the search explored an interaction path, fast-tracking searched the remaining que of interaction paths for genes that had already been added to the network and added these interaction paths to the network. Interaction edge scores were used from the scoring system above. We used specificity analysis to determine an optimal threshold for the interactome (S2 Fig, explained below). In the case of multiple protein targets, we created a pathway around each protein target, and merged these neighborhoods to create the full protein network.

We evaluated the interactome specificity by comparing a gene’s path score to all possible path scores for that gene. To measure all path scores, we created pathways for all genes in the interactome network, treating each molecular entity as a drug target and creating a pathway as described above. We used these empirically-derived scores to calculate an enrichment score for an entity in the pathway of a real drug target by subtracting the average path score to that gene from the gene’s score in the drug pathway (Fig 1, ‘Interaction Specificity Analysis’).

We selected an optimal threshold by evaluating gene specificity at threshold values from 0.7 to 0.9. At each of these values we created a drug pathway around the drug’s protein target(s), calculated the gene specificity, and then tabulated the fraction of genes that are specific to a drug target (i.e. have a specificity score > 0). We plot the normalized histograms of specificity values in S2A Fig and a distribution of the proportion of specific paths at each threshold value in S2B Fig.

The PathFX code is available at: https://github.com/jenwilson521/PathFX. Using the algorithm requires minimal inputs and creates a network and several association files as depicted in S4 Fig. The user provides three inputs: 1. an analysis name. 2.the name of the drug. 3. an optional list of proteins (if the drug-binding proteins are not in DrugBank or the user wishes to complete a more specific analysis).The algorithm creates a set of output files: 1. networks for individual target proteins and a merged interaction network combining networks from each target proteins. These files are tab-delimited files with one interaction per line and the score for that interaction. 2. An association table containing one significantly-associated network phenotype, a p-value for that association, and network genes associated with that phenotype. 3. A table listing the database source for individual phenotype-gene associations.

For all diseases, these phenotypes were mapped to CUI terms using Metamap lite[31]. This was the same process used in assembling the phenotype dataset.

We downloaded the UMLS Metathesaurus, version 2017AA and used the Perl packages UMLS::Interface[32] and UMLS::Similarity[32] to measure the lin distance between diseases in a set. For the gold-standard drug set, we calculated a matrix of similarity values for all approved indications and we used SciPy in Python to perform hierarchical clustering. We identified 62 as the optimal number of clusters using the elbow method. For visualization of the dendrogram, we counted the top five disease-associated words in each cluster. To determine how well PathFX identified a drug’s approved indication, we again used the umls-similarity.pl scripts to calculate similarity between the approved indication and the PathFX identified phenotypes.

Because the number of true positives and true negatives varied for each drug and for each phenotype, we calculated the PPV and NPV separately for each drug and for each phenotype. To calculate PPV, we assumed that false positives were any PathFX identified phenotype that was not a marketed indication. The PPV was the ratio of correctly identified marketed indications to the total of marketed indications and additional PathFX phenotypes. To calculate NPV, we considered any phenotype from our dataset that was not a marketed disease indication to be a true negative. The NPV was the ratio of these true negatives to the sum of the true negatives and the unidentified, marketed indications. When calculating PPV for each phenotype, we assumed that a false positive was any drug identified by PathFX to be associated to the phenotype but was not marketed for that phenotype. When calculating NPV for each phenotype, we considered as true negatives any drugs not marketed for or identified by PathFX to be associated with the phenotype.

We calculated the number of genes associated with the original marketed indication for drug-disease pairs identified by PathFX only, targets only, or identified by both methods. We first used a Kruskal-Wallis test implemented in the Python package, SciPy, to determine that these populations were not from the same distribution. We tested the hypothesis that the targets-only analysis was biased towards diseases with more genes using the Mann-Whitney-U statistic implemented in the Python package, SciPy.

The FDA Adverse Event Reporting System (FAERS) data was extracted using the Oracle Health Sciences Empirica Signal software. Although, the software is not publicly available, we used only publicly-available data at the time of analysis (2004-2017Q3). All drugs that had at least one case reported for the 35 MedDRA Preferred Terms identified as Designated Medical Events by review and medical experts at the FDA were included in this analysis.

From the successfully matched drug-indications pairs, we created a catalogue of interaction paths that supported the association between these drugs and their approved indications. We searched through remaining drug pathways and asked if any associations with the drugs’ non-marketed indications were supported by the catalogue of interaction paths. These non-marketed associations became our cohort of repurposing predictions.

The Drug-binding Dataset collects 984 high-quality 3D structures (x-ray resolution higher than 2.5 Å) that co-crystalized with FDA approved small molecule drugs (non-nutraceuticals), representing binding environments of 284 distinct drugs [50]. The Human Off-target Dataset comprises 2271 proteins representing a non-redundant representative set (90% percent identity) of human proteins and their close homologs that have high quality 3D structures (x-ray resolution higher than 2.5 Å) in PDB[51]. We have applied PocketFEATURE[46] to predict the probability of binding between the 284 drugs and the 2271 potential off-targets. PocketFEATURE uses the FEATURE representation to calculate site similarities by aligning microenvironments between two sites. A more negative score suggests binding site similarity and thus a higher probability of drug binding to a site similar (off-target) to its known binding site. Given a pair of drug and off-target, we used an average score of similarity between the binding sites and the off-target. For each drug, we generated a profile of its binding probability to each of the 2271 potential off-targets.