By providing methods for integration of multiple type of information about drugs and their effects, systems pharmacology studies facilitate the prediction and identification of drug off-targets.17
Prediction of drug off-targets can utilize many different sources and types of information about drugs to make their predictions (). For example, there exist multiple databases containing known drug target information or ligand affinities.18
They have been used to construct ligand-based descriptions of drug targets which can be used to predict additional ligands that may bind. Keiser et al.14
used known ligand protein relationships and drug chemical structure information to make predictions using their similarity ensemble approach to predict drug targets. Drug targets were represented as the set of all ligands known to bind the protein. The structures of all drugs were compared to these ligand sets. As such, they were able to augment a network connecting drugs to their targets with new predictions of drug and drug target associations. They then analyzed how these predicted off-targets could explain the various side effects of different drugs ().
FIGURE 2 Different strategies can be used to predict drug–target relationships. (a) Drug chemical structure can be compared to known ligands of a potential target. (b) Drug chemical structure and known target protein structure can be combined through computational (more ...)
Xie et al.19
studied the cholesteryl ester transfer protein (CETP) inhibitor, Torcetrapib. This medication was developed to treat high cholesterol, but has the side effect of inducing dangerous hypertension through an unknown mechanism. They used computational methods to dock the torcetrapib to its target CETP. After identification of putative binding sites, they searched for other proteins with known or predicted structures that were structurally similar to the drug-binding site (). The drug was then docked back to these predicted off-targets. This allowed the creation of predicted protein–ligand networks that demonstrated differences between CETP inhibitors which induce hypertension and those not known to have this side effect. Incorporating these off-target binding networks with known biological pathways allows further explanation of the pathophysiology of torcetrapib-induced hypertension by the binding of various nuclear receptors to dys-regulate the renin-angiotensin-aldosterone system, a major component of blood pressure regulation in humans.
Yang et al.20
computationally docked sets of drugs into binding pockets of 845 proteins to predict binding affinities. They applied this to sets of drugs known to cause two dangerous dermatological adverse events, Stevens–Johnson syndrome (SJS) and toxic epidermal necrolysis (TEN), as well as a set of controls. They reported that the drugs with the side effects had a distinct binding profile compared to the controls. From this analysis, combined with a gene co-citation network, which connects genes to genes discussed together in papers about the side effects, they were able to propose that MHC I protein heavy chain Cw*4 (1QQD) is the mediator of adverse event: drug-induced TEN. They report that using their docking-based approach, they can distinguish between different alleles of HLA-B for risk of abacavir-induced SJS (). Using a similar approach, they predicted candidate drug targets related to other severe drug adverse events including deafness, rhabdomyolysis, and cholestasis. Their methods have been implemented in SePreSA, a Server for the Prediction of Population Susceptible to Serious Adverse Drug Reactions, and SADR Gengle, which are internet-accessible programs available online.21,22
Campillos et al.23
also utilized the knowledge of drug chemical structures for the prediction of drug off-targets. However, they used additional knowledge provided by the known adverse reactions associated with each drug. Using drug package inserts, they assigned each drug a profile of adverse events associated with it. Combining drug structural similarity with effect similarity, measured by comparing the side effect profiles of drugs, they created a network that connected drugs to each other if they had a sufficiently similar side effect profile and chemical structure. In several cases, the drugs connected in the network were not known to share a protein target. They were then able to validate their approach by demonstrating that one of the drugs could in fact bind at least one of the targets of the other drug. This approach was thus able to predict previously unknown off-targets of drugs.