Conceived and designed the experiments: PB VN. Performed the experiments: LB MI GZ. Analyzed the data: LB MI GZ. Wrote the paper: LB MI GZ VN PB.
The authors have declared that no competing interests exist.
In pharmacology, it is essential to identify the molecular mechanisms of drug action in order to understand adverse side effects. These adverse side effects have been used to infer whether two drugs share a target protein. However, side-effect similarity of drugs could also be caused by their target proteins being close in a molecular network, which as such could cause similar downstream effects. In this study, we investigated the proportion of side-effect similarities that is due to targets that are close in the network compared to shared drug targets. We found that only a minor fraction of side-effect similarities (5.8 %) are caused by drugs targeting proteins close in the network, compared to side-effect similarities caused by overlapping drug targets (64%). Moreover, these targets that cause similar side effects are more often in a linear part of the network, having two or less interactions, than drug targets in general. Based on the examples, we gained novel insight into the molecular mechanisms of side effects associated with several drug targets. Looking forward, such analyses will be extremely useful in the process of drug development to better understand adverse side effects.
As almost 30% of drug candidates fail in clinical stages of drug discovery due to toxicity or concerns about clinical safety
Since most drugs have in addition to their primary target many off-targets
Previous integrative studies of human disease states, protein-protein interaction networks and expression data have uncovered common pathways and cellular processes that are dysregulated in human disease or upon drug treatment
In this work, we aim to quantify the contribution of protein network neighborhood on the observed side-effect similarity of drugs. We developed a pathway neighborhood measure that assesses the closest distance of drug pairs based on their target proteins in the human protein-protein interaction network. We show that this measure is predictive of the side-effect similarity of drugs. By investigating the unique overlap between pathway neighborhood and side-effect similarity of drugs, we find known and unexpected associations between drugs and provide novel mechanistic insights in drug action and the phenotypic effects they cause.
Our network neighborhood measure is based on the protein associations in the database STRING
To investigate whether drug targets that are close to each other in the network tend to have similar side effects, both the normalized pathway neighborhood scores and the direct confidence scores in STRING were used to predict drug pairs with significant side-effect similarity (
This performance is estimated with a ROC curve (
Since both the Recall (7.9 %) of side-effect similarity by the top 500 normalized scores and the Precision (29.8%) are higher than the Recall and Precision by the top 500 direct neighborhood scores (1.8%; 5.7%) (see Methods/
We conclude that drug pairs targeting proteins that are network neighbors indeed have higher side-effect similarity. However, while many drug pairs that have similar side effects target the same network neighborhood, protein network neighborhood doesn't appear to be a good predictor for novel, so far undetected side-effect similarities of drugs.
Previous work has shown that sharing of drug targets is often reflected by similarity in side effects and now we find that also drugs targeting the same network neighborhood show similarity in side effects. We aim to quantify the percentage of side-effect similarities that arise from drugs that target a similar part of the protein-protein network as opposed to drugs that share a target. To this end, we define the drug pairs that target neighboring proteins as those that have a normalized neighborhood score ≥1, i.e. those protein pairs that have STRING confidence which are more than twice the average confidence of the proteins. At this cutoff, 25,263 drug pairs are classified as targeting the same protein network neighborhood.
Of all drug pairs with significant side-effect similarity (N = 1,534), we observe that both drugs are targeting a similar protein network neighborhood in 47.3% of the cases (N = 726) (
(A) Venn diagram of drug pairs with side-effect similarity, shared targets and targeting network neighborhood. We define drug pairs that have side effect p-values ≤0.10 as pairs having significant side-effect similarity. Pairs that target neighboring proteins are defined as having normalized neighborhood score ≥1. Drug pairs that share one or more drug targets are based on data from DrugBank, Matador and PDSP Ki. Only drug-pairs are taken into consideration where at least one drug target is known for both drugs and the side-effect similarity is also available. After removing 12 drug pairs (from 101) where we might expect target-sharing based on chemical or protein similarity, 89 drug pairs are left that target neighboring proteins and have similar side-effects. This is 5.8% of drug pairs with side-effect similarity where we have both target and network information. A minimum of 986 (64%) of side-effect similarities can be explained by sharing drug-targets in the set where at least one drug target is known. (B) Degree distribution of drug pairs with side-effect similarity that target the same network neighborhood. The drugs have been divided in two categories, drugs that target proteins with two or less interaction partners and more than two interaction partners. The drugs in drug pairs that have side-effect similarity target significantly more target proteins with fewer interaction partners than when we consider all drug pairs that target the same network neighborhood. Drug pairs with high chemical similarity or with high sequence similarity of protein binding partners have been removed from the overlapping set, to avoid possible undetected shared targets between drug pairs.
Since it is known that drugs that are chemically similar or have targets that are similar in sequence and/or structure are likely to share a target
To investigate if the local network topology is markedly different for the proteins that are targeted by these drugs, we investigated the degree (number of interaction partners) of drug targets (
We visualized the drug-drug relationships of the 89 remaining cases in a network (
Drugs are drawn as yellow circles, grey lines between them indicate drug targets that are network neighbors.
The network analysis also reveals novel mechanistic insights, illustrated, for example by the association of the alcohol sensitivity drug disulfiram with isoniazid, which is both an antitubercular agent and antidepressant. Common adverse effects of both drugs include liver related pathologies (“jaundice”, “hepatitis”), but also neural and neuronal conditions (“encephalopathy”, “neuritis”, “psychosis”, “eye pain”). Both drugs have long been suspected to interact with each other when taken concurrently
Another example for revealing mechanistic insights of drug actions can be derived from the association between tegaserod and phenylephrine, both GPCR agonists. Tegaserod is an agonist of the serotonin receptor 5-hydroxytryptamine 4 (5-HT4) and has been used for treating chronic constipation in patients with irritable bowel syndrome and chronic idiopathic constipation
A final example is the association between the drugs tolcapone and pergolide, which are both used in the treatment of Parkinson's disease
In this study we have shown that the similarity of adverse effects for a number of drugs can uniquely be explained by the common protein subnetwork that they target. While network neighborhood on its own is not predictive for side-effect similarity, it does lead to novel mechanistic insights into the molecular basis of side effects. It must be noted that the percentage of drug pairs with significant side-effect similarity sharing a common target is much larger than the percentage of drug pairs targeting non-overlapping proteins that are neighbors in a pathway (64% compared to 5.8%). Previous studies relied on the assumption that common adverse effects between drugs generally arise due to the binding of the same (off-)targets
The figure of 5.8% should be treated with caution and is likely to be an underestimate of the role of the protein interactions play in causing adverse drug effects. Since our pathway neighborhood measure only accounts for direct neighbors in the network, further relations between protein network neighborhood and phenotypic effects might be found if larger parts of the network are considered. The number is even more likely to increase if the limited knowledge of the human protein-protein interaction network, even after transferring information from other species, will be extended by more experimental data. The integration of protein network data with other molecular and cellular readouts (e.g., gene expression) should also provide a more sensitive and comprehensive understanding of the role that pathway perturbations play in establishing adverse drug reactions. On the other hand, more complete knowledge of the drug target profiles of small molecules could increase the number of side-effect similarities that are associated with a shared drug target, making our figure an overestimation.
In the drug-drug network that is presented here, we observe multiple drug pairs where both drugs are known to negatively interact (such as disulfiram and isoniazid) or are used in combination therapies (amiloride and thiazide, for example). Most
Drugs and their protein targets were extracted from the drug target databases; DrugBank, Matador
In order to investigate the role of pathways in the side-effect similarity of drugs, we created two datasets: one for pairwise comparisons between drugs in terms of the adverse effects that they cause and another one that contains a measure for the closeness of proteins in the human protein-protein network. The side-effect similarity of drug pairs is calculated as previously described
To obtain a measure for the relatedness of proteins in the human protein-protein network, we use the confidence scores between proteins in the STRING functional protein association database
For every possible pair of drugs in our dataset of 827 drugs, we go through the list of their associated targets and retrieve the confidence scores for every target pair where there is an edge present in the STRING network. We normalize these confidence scores by dividing them by the sum of the average confidence scores of all edges both targets have in the network. The idea behind this normalization is that an interaction with high confidence between two proteins is more significant if it has a higher confidence score than would be expected from the average confidence score of the edges of both proteins. The overlap between the datasets on side-effect similarity and pathway neighborhood consists of 129,975 unique drug pairs.
The chemical similarity of drugs is calculated using the commonly used Tanimoto/Jaccard 2D chemical similarity scores
For every possible pair of drugs in our dataset of 827 drugs, we go through the list of their associated targets and retrieve the confidence scores for every target pair where there is an edge present in the STRING network. We normalize these confidence scores by dividing them by the sum of the average confidence scores of all edges
The idea behind this normalization is that an interaction with high confidence between two proteins is more significant if it has a higher confidence score than would be expected from the average confidence score of the edges of both proteins.
(TIF)
(XLSX)
We thank members of the Bork group for helpful discussions.