Exploring Landscape of Drug-Target-Pathway-Side Effect Associations

Abstract

Side effects are the second and the fourth leading causes of drug attrition and death in the US. Thus, accurate prediction of side effects and understanding their mechanism of action will significantly impact drug discovery and clinical practice. Here, we show REMAP, a neighborhood-regularized weighted and imputed one-class collaborative filtering algorithm, is effective in predicting drug-side effect associations from a drug-side effect association network, and significantly outperforms the state-of-the-art multi-target learning algorithm for predicting rare side effects. We also apply FASCINATE, an extension of REMAP for multi-layered networks, to infer associations among side effects and drug targets from drug-target-side effect networks. Then, using random permutation analysis and gene overrepresentation tests, we infer statistically significant side effect-pathway associations. The predicted drug-side effect associations and side effect-causing pathways are consistent with clinical evidences. We expect more novel drug-side effect associations and side effect-causing pathways to be identified when applying REMAP and FASCINATE to large-scale chemical-gene-side effect networks.

Introduction

Severe side effects are the second leading cause for drug attrition, and the fourth leading cause of death in the US. Severe side effects limit the use of the drugs, decrease their value, and negatively affect patients^1,2 Despite the importance of identifying potential side effects of a drug molecule in advance, it is daunting and prohibitive to test them experimentally. This results in biased, sparse and noisy knowledge about the biological and biochemical associations of side effect. To tackle the difficulty in studying drug side effects, systematic, large-scale methods have been developed to computationally predict drug-induced side effects^3,4,5,6. Although these approaches show acceptable accuracy for predicting common side effects of existing drugs, challenges remain to predict rare side effects as well as to systematically infer missing multi-scale drug-target-pathway-side effect associations. It is important to model drug actions on a multi-scale, since the drug response phenotypes result from complex interplay among biological pathways that are modulated by drug-target interactions.

It is not a trivial task for a machine learning method to infer novel drug-target-pathway-side effect associations based on incomplete, biased, and noisy data. Recently, we have developed a neighborhood-regularized weighted and imputed one-class collaborative filtering method REMAP to address this challenge⁷. REMAP has several unique features, making it particularly suitable to infer missing relations from incomplete and noisy data sets such as drug side effects. First, REMAP does not require negative data for model training by utilizing the imputation. The drug-side effect associations in the existing database are mainly positive. The known negative associations are extremely sparse. These limitations impose hurdles for most classification methods. Second, REMAP can handle mislabeling problem by assigning a confidence score to each label. Mislabeling is common in biological and clinical data sets due to systematic and random errors in experiments. Finally, by applying neighborhood regularization on drug, target, and side effect information, REMAP alleviates the cold start problem, where predicting new targets or side effects is difficult for chemicals without any known targets or side effects. In our earlier study, we have showed that REMAP can be successfully applied to predict unknown drug-target associations⁷. In this paper, we extend its application to drug side effect prediction.

While REMAP shows high prediction accuracy and potential in understanding drug actions, it has limitations. One of the most important issues is that REMAP can take only two types of biological entities (e.g. drugs and targets) and their relationship, and model them as nodes and edges in a bipartite graph. As mentioned above, however, drug actions involve multiple biological entities that are linked with each other on a multi-scale. Thus, integrating information from more than two types of biological entities may be crucial for predicting drug action. For example, a drug interacts with an off-target. The off-target is involved in a biological pathway. The pathway is associated with a side effect. These biological entities (e.g. drug, target, pathway, and side effect) and their relationships can be modeled as a multi-layered network (Figure 1). To infer missing relations from the multi-layered network, most of conventional methods model multiple pairwise relations independently, and integrate these binary relations subsequently. Such an approach ignores the inter-dependency among binary relations. FASCINATE has been developed to infer novel missing relationships from multi-layered networks by jointly optimizing multiple bipartite graphs⁸. In the benchmark studies, FASCINATE outperforms other state-of-the-art methods in inferring multiple relations⁸.

Figure 1.

Multi-layered network view of drugs causing side effects. Drugs may bind targets that are associated with side effects or relevant biological pathways. Thus, drugs may cause side effects through the interplay of biological networks. Solid lines: known associations used as training sets in this study. Dashed lines: no known associations used.

Here, we apply REMAP and FASCIANTE to the prediction of drug side effects and identification of pathways associated with side effects, respectively. We first show that our network-regularized weighted and imputed one-class collaborative filtering method, REMAP, outperforms state-of-the-art multi-target learning methods for drug-side effect associations. Then, we perform random permutation analysis using the FASCINATE algorithm to predict statistically significant gene-side effect associations based on known associations among drug-gene, drug-side effect, and gene-side effect associations. The gene overrepresentation analysis is performed to infer associations among biological pathways and side effects. The predicted results suggest biological pathways associated with cataracts, glaucoma, depression, and other side effects. These putative associations are consistent with existing clinical evidence. Thus, FASCINATE has demonstrated great potential for identifying novel associations between side effects and biological pathways systematically from drug-gene-side effect networks. Our method relies on the known associations and node similarity information, such as drug-drug similarity by chemical substructure comparison. It is expected that a more accurate chemical similarity and side effect similarity measurement as well as more complete chemical-gene-side effect association datasets can discover more novel associations that may be directly applied to clinical studies.

Contribution

The contributions of this paper are both methodological and translational. Methodologically, we demonstrate that a network-regularized weighted and imputed one-class collaborative filtering method when applied to bi-layered and multi-layered drug-target-side effect network is a potentially powerful tool to predict drug side effects and elucidate the molecular and cellular mechanisms of side effects. The translational application is that we predict side effects for several drugs in the database as well as biological pathways associated with these side effects. Majority of them are consistent with existing clinical evidences.

Related work

This section is a review of the existing methods for side effect-pathway prediction, and the core statistical and machine learning methods that have been applied to drug side effect predictions, followed by applications of nonnegative matrix factorization methods for association prediction problems.

To the best of our knowledge, only a few studies have been done to predict the associations between side effects and biological pathways. Fukuzaki et al. proposed a method, Cooperative Pathway Enumerator (CREPE) to predict associations between biological conditions (e.g. heat shock condition) and pathways using the frequent subgraph discovery method⁹. Wallach et al. presented a method to predict side effect-pathway associations by a combination of virtual docking and logistic regression methods, assuming that drugs affecting the same pathway may induce the same side effects¹⁰. More recently, Shaked et al. developed a prediction method, called Array of Model-based Phenotype Predictors (AMPP) to predict drug-target associations using Support Vector Machine (SVM), and extended it to connect key metabolic reactions and biomarkers to side effects, by applying Random Forest-based feature selection on the biomarker features^11,12.

Similarity Ensemble Approach (SEA) is a method developed to predict whether a drug will bind to a target protein based on the chemical structure similarity^13,14,15. Lounkine et al. applied SEA to drug-induced side effect prediction by connecting the predicted drug-target associations to the known target-side effect associations³. The strength of SEA is that it can capture information from sets of weak similarities.

Multi-label Learning (ML) is a class of machine learning methods suitable for predicting multiple labels that are not necessarily mutually exclusive, and ML applications on several domains have been developed¹⁶. Zhang and Zhou developed k-nearest neighbor-based ML (MLKNN)¹⁷ and it refined by applying mutual information-based feature selection on top of MLKNN (FS-MLKNN) to predict drug-induced side effects⁶. FS-MLKNN can be classified an ensemble learning method, which combines weak base learners to improve prediction.

Canonical Correlation Analysis (CCA) is used to find the weights that maximize the correlation between the feature vectors¹⁸. Pauwels et al. developed a chemical substructure-based prediction method that applies CCA and refined it to Sparse CCA (SCCA) for easier interpretation^5,19 In addition to the predicted outcome, SCCA can also return to the optimized weights that represent the importance of each feature. Liu and Altman applied CCA and found correlations between binding affinities of drugs to essential genes and side effects of drugs. They combined CCA with singular value decomposition to extract features representing biological signals, excluding the frequency of information in the dataset²⁰.

While these sophisticated methods are reasonably accurate, each class of methods has drawbacks. First, SEA, FS-MLKNN and SCCA only use chemical features. Therefore, these methods cannot utilize other side information, such as target-target similarity without integrating other methods. Second, these methods may not be able to handle the activity cliff, where a small change in the chemical structure results in a dramatic change in the biological outcome²¹. Third, the methods mentioned above generate models based on the assumption that the unobserved associations are negative. However, the unobserved associations may be positive, the idea which the weighted imputation-based method tries to adopt.

Nonnegative matrix factorization (NMF), first introduced by Paatero and Tapper in 1994, became a popular choice for recommendation problems after the development of the multiplicative update algorithm by Lee and Seung^22,23. One of the most successful NMF applications is the Netflix competition, where the user-video preferences are predicted by NMF²⁴. To better predict the user-item preferences, advanced NMF algorithms take side information, such as user-user and item-item similarities to reflect the idea that similar users tend to choose similar items. Advanced NMF methods such as wiZAN-dual can also take weights and imputation values to account the fact that some of the unknown associations may be positive²⁵. Based on the mathematical framework of wiZAN-dual, REMAP and its variation COSINE were developed to make predictions on unobserved chemical-protein associations^7,26. FASCINATE is an extension of REMAP on a multi-layered network, and has been applied to study drug-gene-disease networks⁸. In this paper, for the first time, we utilize REMAP and FASCINATE to infer novel drug-gene-pathway-side effect associations.

Method

A. Overview of REMAP and FASCIANTE algorithms

As illustrated in Figure 2, the hypothesis of REMAP and its extension FASCINATE algorithms is that the observed sparse and noisy drug-gene profiles and drug-side effect profiles rise from the latent features (hidden variables) of Methodchemicals, genes, and side effects. Similar chemicals, genes, or side effects will have similar latent features. The operation between chemical and gene latent features (e.g. inner product) will restore the complete chemical-gene associations; similarly for chemical-side effect and gene-side effect associations.

B. Dataset description

Our method has three types of nodes: drugs, genes, and side effects. We collected known associations from publicly available databases. Drug-side effect associations for small molecule drugs were obtained from SIDER4.1 database²⁷. Drug-target associations are obtained from the combination of three databases: ZINC, ChEMBL, and DrugBank^{28,29, 30} Drug-target associations were excluded if the drug was not included in the SIDER database. Gene-side effect associations were obtained from Lounkine et al³. The known associations were treated as binary (i.e. 1 for observed associations, and 0 otherwise). The drug-drug similarity scores were calculated based on chemical 2D structure similarity as described in REMAP paper⁷. Protein-protein interactions were obtained from BioGRID database³¹ and used as gene-gene similarity matrix. We have tested semantic similarity-based side effect-side effect similarity scores using UMLS-Similarity software³². However, the semantic similarity scores did not help gene-side effect or drug-side effect prediction accuracy. The best performance with semantic similarity was within one standard deviation of the performance without using it (data not shown). Thus, an identity matrix was used as side effect-side effect similarity matrix. The data statistics are listed in Table 1.

Table 1.

Data statistics of drug-gene-side effect network.

	Type	Count
Nodes	Drugs	1549
	Genes	19116
	Side effects (diseases)	4251
Associations	Drug-gene	4727
	Drug-side effect	127595
	Gene-side effect	2209
Similarities	Drug-drug	5303
	Gene-gene	458135
	Side effect-side effect	4251

C. Description of prediction method

REMAP is a variation of nonnegative matrix factorization algorithm for inferring novel binary relations. FASCINATE is an extension of REMAP on multi-layered network⁸. The three layers used in this study are drugs, genes, and side effects, and the aim is to find low-rank matrices of latent features for each layer that minimizes the following equation:

$$\underset{{}_{\text{Fi³0 (i=1,}\mathrm{\dots }\text{,g)}}}{\text{min}}\text{J}={\displaystyle \sum \left|\right|}{\text{W}}_{\text{i,j}}\text{e}\left({\text{D}}_{\text{i,j}}{\text{+P}}_{\text{i,j}}{\text{-F}}_{\text{i}}{\text{F}}_{\text{j}}{}^{\prime }\right){\text{||}}_{\text{F}}^{\text{2}}+\alpha {\displaystyle \sum _{i-1}^g{\text{tr(F}}_{\text{i}}^{\prime }{\text{(T}}_{\text{i}}{\text{-A}}_{\text{i}}{\text{)F}}_{\text{i}}\text{)+β}{\displaystyle \sum _{i=1}^g{\text{||F}}_{\text{i}}{\text{||}}_{\text{F}}^{\text{2}}}}$$

Here, g is the number of total layers (3 in this study), and F_i (e.g. F₁, F₂, or F₃) is the low-rank matrix for the i^th layer, D_i,j is the cross-layer association matrix between the i^th and j^th layers (e.g. drug-gene), A_i is the within-layer similarity matrix (e.g. drug-drug), T_i is the diagonal degree matrix of A_i, W_i,j is the weight matrix, P_i,j is the imputation matrix, ⊙ is an element-wise product operator (Hadamard product), and ||M||_F is the Frobenius norm of matrix M. For the algorithmic efficiency, the weight matrix was replaced with a scalar, global weight for all layers, and the imputation value was uniformly set to a constant for all unobserved associations. Once the low-rank matrices for each layer are calculated by FASCINATE, the prediction matrix between the i^th and j^th layer can be obtained by the matrix product of F_i and F_j. If there are only two layers, the objective function in Eq. (1) is identical to that of REMAP. More details about the algorithm can be found in the reference⁸.

We evaluated the effectiveness of REMAP using multiple metrics. For varying degree of density in the gene-side effect network, we performed 10-fold cross-validation to measure four different performance metrics that are commonly used to evaluate a recommender system. The four metrics are the standard area under the receiver operating characteristic curve (AUC), Half-Life Utility (HLU)³³, Mean Average Precision (MAP)³⁴, and Mean Percentage Ranking (MPR)³⁵. HLU estimates the likelihood that a user accepts the recommended items under the assumption that it exponentially decays with the item’s ranking³⁶. MAP measures the average precision over all users in the test dataset at different recall levels³⁴. MPR measures the average of percentile-ranks of positive associations in the test dataset³⁵. For example, an MPR of 0% is obtained when all positive associations are top-ranked. On the other hand, 100% MPR is obtained when they are ranked to the least. The higher AUC, HLU, MAP and the lower MPR, the better performance. During the cross-validation, we categorized side effects by the number of known drug-side effect associations and measured performances from rare side effects to frequently observed ones. For comparison, we tested feature selection-based multi-label k-nearest neighbor method (MLKNN)⁶ under the same cross-validation and rarity conditions as those for REMAP.

Since FASCINATE takes a few tuning parameters, we found the optimal parameters by using a grid search and 10-fold cross-validation. For each fold, we randomly selected approximately 10% of the associations from both drug-gene and drug-side effect layers as test sets, and measured the AUC for both layers separately. Then, we evaluated the performance by the average of the test AUC for 10 folds weighted by the number of associations in each layer. In this way, we found the optimal low-rank and maximum iteration number to be 770 and 100, respectively, while fixing the other parameters to be the default values. Using the optimal rank and iteration number, we found the optimal parameters to be α=0.5, β=1.0, and weight=0.25.

D. Random permutation and overrepresentation analysis

Next, we performed random permutation and overrepresentation analysis on the gene-side effect layer using the optimal parameters above. Briefly, we estimated the distribution of prediction scores from random permutations, and using the distribution, we filtered predicted pairs from FASCINATE without permutation. The filtered pairs were then used to search for significant pathway-side effect associations by an overrepresentation test. For each permutation test, we randomly permutated the drug-gene and drug-side effect associations while keeping the total number of associations in each layer. Using the permuted layers and the intact gene-side effect layer as inputs, we ran FASCINATE to obtain the prediction scores for gene-side effect pairs. We performed 200 permutation iterations and recorded the prediction scores for each gene-side effect pair. Then, we collected raw prediction scores from randomly selected 1,000 gene-side effect pairs from unknown associations. With the 200,000 raw scores (1,000 pairs with 200 scores each), Epanechnikov kernel was used to fit the distribution (Figure 3). Except those for known gene-side effect associations, the raw prediction scores were converted to p-values by the Epanechnikov kernel probability density function obtained above, followed by the False Discovery Rate adjustment. To infer associations among biological pathways and the side effects, the gene sets for each side effect were used for KEGG overrepresentation test provided in the cluster Profiler package³⁷. The enriched pathway-side effect associations were filtered out if the pathway is associated with more than 5 side effects, or the side effect is associated with more than 5 pathways. Also, predicted pairs with q-value (from hypergeometric test) higher than 0.05 were filtered out.

Results

A. Significantly improved performance of REMAP in predicting rare side effects

REMAP performs better than MLKNN for all four metrics in all tested rarity categories (Table 2). The performance gain of REMAP over MLKNN is more significant when the side effect is rare. The AUC of REMAP drops from 0.95 for all side effects to 0.84 for the rarest side effect class, while that of MLKNN drops from 0.92 to 0.68. This suggests that REMAP can be effectively applied for drug-side effect predictions from rare to common side effects. It is noted that while REMAP shows better AUC and MPR for more common side effects, REMAP’s performance is not necessarily improving as the rarity decreases, based on HLU and MAP. The same trend is observed for MLKNN, suggesting that presenting the four metrics provides more objective evaluations. In summary, our extensive benchmark studies show that REMAP has a great potential to predict novel associations.

B. Literature supports of top ranked predictions

Furthermore, the top ranked predictions from REMAP are supported by existing clinical evidences, as shown in Table 3. Newton et al. reported that patient groups receiving diazepam showed decreased libido compared to other groups including placebo in reviewing total of 984 patients with generalized anxiety disorder³⁸. An ovarian carcinoma patient treated with paclitaxel and carboplatin was reported to develop acute myeloid leukemia (AML) in 2006 while the main culprit was unclear³⁹. In 2016, another group reported a case of a patient who developed AML after endometrial chemotherapy including paclitaxel, where paclitaxel was the likely cause of AML⁴⁰. Although indirectly related to ichthyosis and less severe, patients treated with lenalidomide showed significant dermatologic side effects⁴¹. It was reported that one of the most common side effect of erlotinib, when treated with emibetuzumab for non-small cell lung cancer (NSCLC), is hypocalcemia⁴². In addition, erlotinib alone was reported to cause hypocalcemia⁴³. Two clinical case reports and human exposure studies suggest that propranolol may cause cardiac arrest if overdosed^44,45. Although rare and the cases were limited to the elderly population, ranolazine treatment was reported to cause visual hallucinations, which resolved after discontinuation of treatment⁴⁶. In a study of 5,194 males, approximately 35% of patients experienced erectile dysfunction, and it was associated with cumulative exposure to zalcitabine or enfuvirtide⁴⁷. Ejaculatory failure was reported from oxcarbazepine treatment⁴⁸. In three cases of long-term corticosteroid treatment, including two cases with prednisolone, the patients suffered from mild to severe impairment of vision, which could have resulted in permanent loss of vision⁴⁹. In a study analyzing more than a million cases from September 1999 to April 2012, levofloxacin was found to be the greatest risk for patients. Patients who took levofloxacin had higher hazard ratios for serious cardiac arrhythmia and death compared to amoxicillin and azithromycin⁵⁰. In a case report, Mahendran and Liew suggested that when prescribing alprazolam, symptoms of depression need to be carefully monitored. They described a healthy woman in her 20s who developed severe depression with suicidal thoughts after taking the prescribed doses of alprazolam three times⁵¹.

C. Identification of pathway-side effect associations

A total of 7,572 gene-side effect pairs for 409 unique genes and 1,150 unique side effects were found with q-values lower than 0.001 by FASCINATE and subsequent random permutation analysis. Then these genes are associated with biological pathways using gene set overrepresentation analysis. The final predicted pathway-side effect associations with their q-value are listed in Table 4. The literature support for predicted top-ranked side-effect pathways suggest that our method can infer missing one type of binary relations by jointly learning other types of binary relations from a multi-layered network (Table 4). Wirtz and Keller identified mutations and significant up or down regulation in different kinds of an interleukin family of cytokines in mice with glaucoma, suggesting that disruption of interleukin-20 signaling in the eye may be associated with glaucoma⁵². More recently, Gupta et al. showed that the concentration of tear film cytokines were significantly lower in patients with primary open-angle glaucoma⁵³. Liu and Neufeld showed that the epidermal growth factor receptor (EGFR) is significantly upregulated and tyrosine phosphorylated in the human glaucomatous optic nerve head in vivo. They demonstrated that it induces nitric oxide synthase to generate excessive nitric oxide, leading to increased intraocular pressure and glaucoma⁵⁴. In another study, the authors suggested that EGFR is a common regulatory pathway for neural injuries in the optic nerve, including glaucoma⁵⁵. Kwon and Tomarev showed that several components in integrin-focal adhesion kinase – serine/threonine kinase signaling pathway are activated in the eyes of mice expressing high level of myocilin, a protein known to be associated with glaucoma⁵⁶. Piotrowska et al. surveyed patients treated with a selective EGFR tyrosine kinase inhibitor, where they found a significant proportion of patients developed drug-induced cataracts, including those who had to undergo surgical repair⁵⁷. Although the association between cataract and focal adhesion pathway is subtle, Kokkinos et al. showed that the expression and phosphorylation pattern of focal adhesion kinase changes as the lens develops in mice⁵⁸. As one may simply guess, sex steroid hormone deficiency is known to be associated with erectile dysfunction⁵⁹. Although we could not find pertinent references, it may be trivial to validate the predicted association between steroid hormone synthesis and decreased libido. A more indirect association was described between erectile dysfunction and gastric acid secretion, which is clinically regulated by histamine receptor antagonists^60,61. In a study involving 55 patients with stage IV NSCLC, patients having EGFR mutations showed less severe depression compared to those without EGFR mutation, suggesting the association between depression and EGFR⁶². In a review paper, evidence was found for omega-6 polyunsaturated fatty acid (e.g. linoleic acid) affecting the function of leukocytes⁶³. Several interesting points rise from the predicted side effect-pathway associations. First, glaucoma and cataract share two common pathways: EGFR tyrosine kinase inhibitor resistance and focal adhesion pathways. Second, the results suggest that erectile dysfunction and decreased libido are relevant to the steroid hormone biosynthesis pathway. Further studies with more details and experimental validations are necessary to refine the FASCINATE method for gene-side effect-biological pathway association prediction.

Conclusion

In this study, REMAP outperformed state-of-the-art MLKNN on drug-side effect prediction as measured by the four metrics (AUC, HLU, MAP, and MPR), especially for rare side effects. Furthermore, FASCINATE, a multi-layer extension of REMAP, was applied to jointly predict the drug-target-side effect network and to infer missing pathway-side effect associations in combination with random permutation analysis and the overrepresentation test. Many of the predicted drug-side effect and pathway-side effect associations are supported by existing experimental and clinical evidences (Table 3 and Table 4). These results suggest that network-regularized weighted imputed one-class collaborative filtering is a powerful tool for inferring missing biological relations on a multi-scale. There are many ways to improve the performance of REMAP and FASCINATE. Their performances are dependent on the similarity measure of drug-drug, target-target, and side effect-side effect pairs as well as the coverage of reconstructed drug-gene-side effect networks. The side effect-side effect semantic similarity used in this study does not help much with the prediction accuracy. Thus, new side effect-side effect similarity measurement is needed. In the earlier studies, it was shown that the structure-based ligand binding site similarity improves the prediction accuracy of side effect⁵⁶. In addition, the datasets used in this study are not comprehensive. Only 1,549 unique FDA approved drugs and 4,727 drug-gene interactions are included in the current network. It is expected that more novel relations can be identified if the reliable chemical-gene network which include approximately 500,000 chemicals (including drugs) and thousands of targets genes⁷ and improved drug-side effect database are applied⁶⁴.

Figure 2.

Illustration of the rationale of REMAP and FASCINATE algorithms. These algorithms calculate the latent features of chemicals, genes, and side effects from the observed associations. The latent features are used to reconstruct the complete association matrices.

Figure 3.

(A) Epanechnikov kernel fitting on FASCINATE prediction scores from random permutation analysis on gene-side effect associations. (B) Inset of (A) for smaller ranges of both axes.

Table 2.

Performances of REMAP and MLKNN predicting drug-side effect associations.

	Avg. AUC		Avg. HLU		Avg. MAP		Avg. MPR
Number of drugs per SE*	REMAP	MLKNN	REMAP	MLKNN	REMAP	MLKNN	REMAP	MLKNN
< 12	0.84	0.68	5.6	3.6	.048	.031	0.26	0.33
< 25	0.88	0.76	4.9	2.4	.045	.026	0.20	0.24
< 50	0.90	0.82	4.1	1.6	.044	.026	0.16	0.19
< 100	0.92	0.85	3.9	1.1	.043	.024	0.13	0.16
< 200	0.93	0.88	3.5	0.7	.040	.023	0.11	0.13
< 400	0.94	0.90	3.4	0.4	.042	.024	0.10	0.11
< 800	0.95	0.91	3.6	0.2	.047	.027	0.08	0.10
All	0.95	0.92	4.0	0.3	.052	.031	0.08	0.09

^*Side effects having a certain number of associated drugs per side effect.

Table 3.

Predicted top ranked drug-side effect associations with literature support.

Drug	Predicted Side Effect	Reference
Diazepam	Libido disorder	38
Paclitaxel	Acute myeloid leukemia	39, 40
Lenalidomide	Ichthyosis	41
Erlotinib	Hypocalcemia	42, 43
Propranolol	Cardiac failure	44, 45
Ranolazine	Visual impairment	46
Zalcitabine	Erectile dysfunction	47
Oxcarbazepine	Erectile dysfunction	48
Prednisone	Blindness	49
Levofloxacin	Cardiac failure	50
Alprazolam	Suicide attempt	51

Table 4.

Predicted pathway-side effect associations by random permutation and overrepresentation analysis.

Side Effect (disease)	KEGG Pathway	Q-value	Reference
Glaucoma	Cytokine-cytokine receptor interaction	1.1E-2	52, 53
Glaucoma	EGFR tyrosine kinase inhibitor resistance	9.2E-4	54, 55
Glaucoma	Focal adhesion	6.9E-3	56
Cataract	EGFR tyrosine kinase inhibitor resistance	7.3E-3	57
Cataract	Focal adhesion	2.2E-2	58
Erectile dysfunction	Steroid hormone biosynthesis	5.6E-3	59
Libido decreased	Steroid hormone biosynthesis	9.3E-6	*
Erectile dysfunction	Gastric acid secretion	8.1E-3	60, 61
Depression	EGFR tyrosine kinase inhibitor resistance	1.2E-2	62
Leukopenia	Linoleic acid metabolism	1.4E-2	63

^*No pertinent references found.