On the agenda today is a fascinating paper titled: predicting new molecular targets for known drugs, Keiser et al., Nature 2009 [PMC]. This is one of my favorite papers of the past two years. I have to admit that it took my quite a while to really understand the potential of the methods and findings of the paper and I was torn as to whether to include the paper in the class because it is likely a bit much to digest in a single class period; however, the findings are so cool and the implications of the method are so far-reaching that I ultimately decided I couldn’t pass up the opportunity to do this paper for the class. Let’s dive in…
The idea is simple. Drugs work because they have an action at a molecular target. However, most drugs have side effects and we don’t always know why. Moreover, while we often presume that drugs work because they hit a particular target, this may not actually be true. There are generally several drugs of a particular class but one particular drug may be better for treating a disease. This could be explained by a number of factors and one potential possibility is that a given drug has a mix of activities that is a perfect fit for a particular disorder. The problem is that we don’t necessarily know what that polypharmacological mix is. If there was a computation model that could help us predict appropriate polypharmacologies and off-target hits at the same time we could potentially make large leaps forward in terms of drug development.
This is, more or less, what the author’s set out to do:
The creation of target-specific ‘magic bullets’ has been a therapeutic goal since Ehrlich, and a pragmatic criterion in drug design for 30 years. Still, several lines of evidence suggest that drugs may have many physiological targets. Psychiatric medications, for instance, notoriously act through multiple molecular targets, and this ‘polypharmacology’ is probably therapeutically essential. Recent kinase drugs, such as Gleevec and Sutent, although perhaps designed for specificity, modulate several targets, and these ‘off-target’ activities may also be essential for efficacy. Conversely, anti-Parkinsonian drugs such as Permax and Dostinex activate not only dopamine receptors but also 5-HT2B serotonin receptors, thereby causing valvular heart disease and severely restricting their use.
Drug polypharmacology has inspired efforts to predict and characterize drug–target associations. Several groups have used phenotypic and chemical similarities among molecules to identify those with multiple targets, and early drug candidates are screened against molecular target panels. To predict new targets for established drugs, a previous group looked for side-effects shared between two molecules, whereas another group linked targets by drugs that bind to more than one of them. Indeed, using easily accessible associations, one can map 332 targets by the 290 drugs that bind to at least two of them, resulting in a network with 972 connections (Fig. 1a, below). It seemed interesting to calculate a related map that predicts new off-target effects.
So what does this network look like?
Panel A is a network of known drugs with known targets linked together. It shows, basically, some pharmacological promiscuity between receptors that would be expected by any pharmacologist. For instance, 5HT receptor drugs tend to hit other 5HT receptors and these drugs also tend to have effects at other GPCRs within the amine ligand class. Nuclear receptors and their ligands also group together and on and on. Where this really gets interesting is in panel B:
Accordingly, we used a statistics-based chemoinformatics approach to predict new off-targets for 878 purchasable FDA-approved small-molecule drugs and 2,787 pharmaceutical compounds. Unlike bioinformatics methods, which might use the sequence or structural similarity among targets, this similarity ensemble approach (SEA) compares targets by the similarity of the ligands that bind to them, expressed as expectation values, adapting the BLAST algorithms (other methods such as naive Bayesian classifiers may also be used, see Supplementary Table 1). The approach thus captures ligand-based similarities among what would otherwise be considered disparate proteins. The 3,665 drugs were compared against 65,241 ligands organized into 246 targets drawn from the MDL Drug Data Report (MDDR) database, yielding 901,590 drug–target comparisons.
Most drugs had no significant similarities to most ligand sets. However, 6,928 pairs of drugs and ligand sets were similar, with expectation values (E-values) better than 1 × 10-10. We analysed these predictions retrospectively against known associations and prospectively for unreported drug polypharmacology.
So, they essentially used ligand similarities to predict new pharmacological targets for drugs based on known pharmacological interactions. The rest of the paper includes a number of experiments aimed at figuring out if these predictions have any validity (lots of binding experiments are on the way — if you read the paper).
There is a ton of interesting stuff here and I’m not going to cover it all. Instead I will focus on two predictions that led to some really interesting findings and then delve into how one might use this dataset to move ahead in some specific areas.
The new targets can improve our understanding of drug action. DMT is an endogenous metabolite and a notorious hallucinogen. Recently, the molecule was characterized as a σ1-receptor regulator at micromolar concentrations, an association implicated in its hallucinogenic properties. This surprised us because many drugs, including non-hallucinogens, bind promiscuously to the σ1 receptor with higher affinity than DMT. Also, the hallucinogenic characteristics of DMT are consistent with other hallucinogens thought to act through serotonergic receptors, some of which the molecule is known to bind. We therefore screened DMT against the 1,133 WOMBAT targets. SEA predicted it to be similar against multiple serotonergic (5-HT) ligand sets, with E-values ranging from 9.2 × 10-81 to 7.4 × 10-6. Upon testing, we find DMT binds 5-HT1A, 5-HT1B, 5-HT1D, 5-HT2A, 5-HT2B, 5-HT2C, 5-HT5A, 5-HT6 and 5-HT7 receptors with affinities ranging from 39 nM to 2.1 μM (Supplementary Table 4 and Supplementary Fig. 2). Of these, three were previously unknown (Table 1), and all had substantially greater affinities for DMT than that represented by its 14.75 μM dissociation constant (Kd) for σ1. To investigate further the role of serotonin receptors in DMT-induced hallucination, we turned to a cell-based assay and an animal model that are predictive of hallucinatory actions. Consistent with SEA prediction, we find that DMT is not only a potent partial agonist at 5-HT2A (Fig. 2g) as has been reported31, but it also induces head twitch response in wild-type but not in 5-HT2A knockout mice (Fig. 2h), which is new to this study. The half-maximum effective concentration (EC50) of DMT at 5-HT2A is 100-fold lower (better) than that observed for σ1. These observations support 5-HT2A as the primary target for DMT’s hallucinogenic effects.
This is pretty remarkable. We start with a σ1 receptor ligand that is a known hallucinogen, DMT, but σ1 receptor activity doesn’t really make sense for a hallucinogen. Most hallucinogens are 5HT receptor modulators, more specifically, 5HT2A receptor agonists. Their model predicts that DMT should have activity at 5HT receptors (this was not known) and they demonstrate through binding experiments that DMT actually has more affinity at several 5HT receptors than it does at σ1. They then demonstrate, in mice, that the hallucinogenic effects of DMT are likely mediated by 5HT2A receptors. Problem solved (well, maybe not) but this is certainly strong support for the hypothesis.
Let’s have another example of the predictive power of the network:
Whereas many of the predicted off-targets occur among aminergic GPCRs, a target class for which cross-activity is well-known, four of the drugs bound to targets unrelated by sequence or structure to their canonical targets (Table 2). For instance, the reverse transcriptase (enzyme) inhibitor Rescriptor was predicted and shown to bind to the histamine H4 receptor, a GPCR. These two targets share no evolutionary history, functional role, or structural similarity whatsoever. Intriguingly, although the Ki value of Rescriptor for the H4 receptor is high at 5.3 μM (Table 2 and Supplementary Fig. 1), this is within its steady-state plasma concentration (minimum plasma concentration averages 15 μM) and is consistent with the painful rashes associated with Rescriptor use; likewise, H4 dysregulation has been associated with atopic dermatitis.
I think this is absolutely amazing. Here you have a reverse transcriptase (RT) inhibitor that is predicted to bind histamine receptors (they’re right, who ever would have guessed this?). They find that the prediction is accurate and that this receptor/drug interaction may explain skin rashes caused by the drug. If you are making new RT inhibitors you may want to check for histamine receptor activity. Moreover, if you are the maker of an RT inhibitor that causes skin rashes you now have a pretty good idea of how to avoid a bad side-effect, include an antihistamine.
There are some caveats to all this and they state them all pretty clearly:
Certain caveats merit mention. Not all of the new off-targets predicted here would surprise specialists. For instance, Dimetholizine has antihypertensive activity and so its affinity for adrenergic receptors is not wholly unanticipated. Similarly, Kalgut is classified as a selective β1 agonist, thought to have little activity on other adrenergic receptors. Whereas the observation that it does bind to the β3 receptor goes against this classification, structurally this seems easy to credit (Table 1 and Supplementary Fig. 1). Indeed, 10 of the 14 drugs reported here are active against aminergic GPCRs (Fig. 3), and so their cross-activities against other aminergic GPCRs have some precedent. Finally, although most of the drugs were active at their predicted off-targets, one-third were not; these are examples of the false-positives to which this method is susceptible (Supplementary Table 7). Thus, the anxiolytics Valium and Centrax scored well against cholecystokinin B ligands, the antipsychotic Emilace was predicted to bind 5-HT4, the anaesthetic Duocaine the κ-opioid receptor, the antihypertensive Doralese neurokinin receptors, and the narcotic Dromoran and the bradycardic Zatebradine scored well against the D2 and D1 receptors. None of these bound their predicted off-targets with affinities better than 10 μM. SEA ignores pharmacophores in its predictions, comparing drugs to ligand sets based on all shared chemical patterns. This is at once a strength, in that it is model-free, and a weakness, in that it may predict activity for drugs that share many features with the ligands of a target, and yet miss a critical chemotype.
Where to go from here. Surely you know that my main research interest is pain and there is a wealth of info here for a pain researcher interested in new drugs. Two major drug classes that are excellent for pain are NMDA receptor antagonists and opioids. Both of these receptor types are included in their receptor network and a little look at Fig 1B shows that their is extensive connectivity between these drug classes. Presumably one could delve into this data a bit deeper and find compounds that have NMDA antagonist activity with predicted mu opioid activity (or vice-versa). You could then do some simple screening experiments to find if that prediction is true and you may just have discovered a drug that would hit two major targets at the same time. In this case, you might be able to prevent some central sensitization (NMDA antagonist) and provide effective analgesia (mu opioid) at the same time. I’ll stop there because I’m not in the mood for giving away all my ideas on this paper for free but needless to say, its a long list.
Keiser, M., Setola, V., Irwin, J., Laggner, C., Abbas, A., Hufeisen, S., Jensen, N., Kuijer, M., Matos, R., Tran, T., Whaley, R., Glennon, R., Hert, J., Thomas, K., Edwards, D., Shoichet, B., & Roth, B. (2009). Predicting new molecular targets for known drugs Nature, 462 (7270), 175-181 DOI: 10.1038/nature08506