|Home | About | Journals | Submit | Contact Us | Français|
Molecular biology now dominates pharmacology so thoroughly that it is difficult to recall that only a generation ago the field was very different. To understand drug action today, we characterize the targets through which they act and new drug leads are discovered on the basis of target structure and function. Until the mid-1980s the information often flowed in reverse: investigators began with organic molecules and sought targets, relating receptors not by sequence or structure but by their ligands. Recently, investigators have returned to this chemical view of biology, bringing to it systematic and quantitative methods of relating targets by their ligands. This has allowed the discovery of new targets for established drugs, suggested the bases for their side effects, and predicted the molecular targets underlying phenotypic screens. The bases for these new methods, some of their successes and liabilities, and new opportunities for their use are described.
So dominant has the molecular biology view of pharmacology become that it is difficult to remember that even 25 years ago it was little more than an aspiration. Today we understand the activity of drugs and reagents first through the specific, clonable receptor molecules with which they interact. To understand biological mechanism we begin with a specific molecular receptor, and to discover new leads for pharmacological intervention, we screen a library of compounds against a particular isolated target. Even when we screen for a phenotype against a cell or organism we subsequently seek to isolate the receptor responsible for that phenotype. Two targets are similar when their sequences and structures are similar, and when we think about side effects our first thoughts are of those proteins that are most related by sequence and structure to the particular targets in which we are interested.
A generation ago this view was inverted: investigators more often began with small molecules and sought targets, and receptors were related not by sequence or structure but by their ligands (Figure (Figure1).1). Except for some soluble enzymes, these receptors were rarely used in isolation. Most were characterized by the patterns of agonists and antagonists that modulated their activity, often in experiments conducted on entire organs such as the guinea pig ileum or atrium perfused in baths of reagent. Thus, Ahlquist first divided the adrenergic responding receptors into α and β subclasses based on differing dose responses to norepinephrine, epinephrine, and isoproterenol in organ systems such as the uterus, the cat nictitating membrane, pupil dilation, and gut contraction, in 1948 (1). Twenty years later, Lands further divided the β-adrenergic family into β1 and β2 receptors based on differing specificities for these same agonists on fatty acid mobilization, and bronchiodilation and vasodepression, also in whole organs (2). The distinction between the α and β adrenergic receptors was strengthened by the appearance of the first β-blockers, such as propanolol. Subtype selective agonists, such as salbutamol for β2, and antagonists, such as atenolol for β1, further solidified the classification of the β-AR family. Meanwhile, the α-adrenergics were subdivided into α1 and α2 classes based on postsynaptic and presynaptic sites of action and the differing specificities of related antagonists. The histamine receptor family was subdivided into the H1 and H2 classes based on the ability to distinguish receptors that responded to histamine but not to mepyramine yet could be antagonized by Burimamide and molecules related to it, initially based on organ level experiments on guinea pig atrium and ilium (3). Gaddum first differentiated receptors responding to serotonin into two subtypes in the 1950s based on the contraction of smooth muscle or on the depolarization of the cholinergic nerves. These targets were subsequently classified in the 5-HT1, 5-HT2, and 5-HT3 families using specific and distinguishing agonism and especially antagonism by drugs like Bemestron and Tropisetron (for the 5-HT3 family).
Classifying targets by small molecules often led directly to new therapeutics. Thus, the subdivision of the β-adrenergic family into the β1 and β2 subtypes both allowed and was itself confirmed by the development of the β-blockers and β2 agonists. Similarly, the specific antagonist Burimamide defined Black’s elucidation of the histamine H2 family, and this led directly to the first histamine-acting anti-ulcer drug, Cimetidine, cited in Black’s receipt of the Nobel Prize in Physiology or Medicine in 1988. ICS 205-930 not only was the molecule that defined 5-HT3 as a unique and specific receptor (4) but also became the anti-nausea drug Tropisetron (Figure (Figure2).2). The classification of receptors by small molecules remains with us to this day, and we still talk of the α1, α2, β1, β2, H1, H2, and 5-HT1−3 receptors.
This antique chemical taxonomy leaves us with, from the molecular biology perspective, some odd bestiaries. All of the serotonin receptors are G-protein-coupled receptors (GPCRs)1 except for 5-HT3, which is an ion channel. By sequence and structure, the 5-HT3 receptor has no similarity whatsoever to the GPCRs whose name it shares. Conversely, 5-HT3 responds to serotonin and its close analogues, as do all of the other members of the 5-HT family, and drugs and reagents that classify this ion channel as 5-HT3 also bind to 5-HT2 at low micromolar concentrations (5) and to the 5-HT4 receptors at midnanomolar concentrations (6); however, these latter are GPCRs. Similarly, the metabotropic glutamate receptors are GPCRs, while the ionotropic glutamate receptors are ion channels; both respond to glutamate and related molecules. The same is true for the nicotinic and muscarinic acetylcholine receptors, both canonical drug targets. Many transporters, which are dissimilar in structure and sequence to both ion channels and GPCRs, are modulated by drugs and ligands that are characteristic of these latter receptors. Serotonergic receptor drugs modulate serotonin and norepinephrine transporters, and putatively “selective” serotonin reuptake inhibitors modulate adrenergic receptors (below). Conversely, at the molecular biology level, many receptors that are closely related share no ligand similarity. The μ-opioid receptor is by sequence and structure similar to the metabatropic serotonin receptors; both are seven-transmembrane GPCRs. There is, however, little similarity among the ligands that modulate them, and many GPCRs sharing high sequence identity have no ligands in common. From a small molecule perspective, saying that 5-HT3 is related to 5-HT4, or that a serotonin transporter is related to an adrenergic receptor, sensibly organizes pharmacology (Figure (Figure3),3), whereas from a molecular biology view, this is, at least superficially, baffling.
The organization of pharmacology is thus bicameral. On one hand, the molecular biology view can be quantified and reduced to specific, clonable targets and reflects a deep understanding of biology and evolution. On the other hand, the chemical view is the basis of our everyday taxonomy of receptors. Over the past five years, the formal basis of this classic, premolecular biology view has been reinvestigated, leading to new maps of pharmacology and the discovery of new drug activities. Some of these methods are available online and may be used by nonspecialists to frame chemistry-guided questions of biology (http://sea.bkslab.org). These might include the following. To what targets might my organic molecule(s) bind, and to what other receptors is my target linked by chemistry? Here, we consider just how common it is to find drugs and reagents that bind to apparently unrelated receptors, the informatics and databases that have allowed this investigation, and some of the new tools developed to exploit them. We consider applications of the chemical view to predict new activities for drugs, to understand their side effects, and to identify the targets for molecules active in phenotypic screens. We close by considering opportunities for a synthesis between the two views, neither of which alone seems fully complete.
Over most of the history of pharmacology, investigators made do with few characteristic ligands for any target and it is a testament to the thought and care that went into target characterization that so much was learned from so few molecules. Today, hundreds of thousands of ligands are characterized for thousands of targets; one of the challenges in pharmacology is organizing the weight of information under which the field unsteadily groans. Databases of target−ligand associations have begun to address this problem.
Two sources of target−ligand associations are patents and scientific publications, particularly the medicinal chemistry literature. The drug patent literature has been curated into electronic form by Prous Science (Barcelona, Spain) and formatted and marketed by MDL (now Symyx/Accelrys) as the MDL Drug Data Report (MDDR). Investigators often work with a subset of the MDDR that Schuffenhauer and colleagues associated with a specific biological target in their ontology, as opposed to, for instance, a more general category such as “anti-cancer” (7) (Table (Table1).1). The CMC database, also curated and sold by Symyx (8), contains compounds identified in the United States Approved Names (USAN) list. It describes chemical structures, biological activity, drug class, originating company, and literature references. It does not contain immediately useful drug−target associations but remains a valuable source of information for predicting drug−target associations and complements the MDDR. Similarly, Boyer and colleagues at IBM have created a fully automated patent parsing engine to create a database of ligand−target−disease associations; this database is available commercially (9).
Two databases of literature-based target−ligand associations are widely used. Among the first of these was the World of Biomolecular Activity database (WOMBAT) (10), which covers most of the past 20 years of the Journal of Medicinal Chemistry and nearly a decade of the next three most important medicinal chemistry journals and has partial coverage of several others. It annotates targets with SwissProt and Uniprot codes, where available, and differentiates agonists from antagonists, a level of detail helpful to target prediction and often not available in other databases. WOMBAT is a commercial product but is also accessible collaboratively from its authors at Sunset Molecular (http://www.sunsetmolecular.com/). Recently, the ChEMBL database has become freely available via the European Bioinformatics Institute (EBI) (11). This library is actively curated and annotates more than half a million ligands with more than 3000 targets; it is freely accessible at http://www.ebi.ac.uk/chembl/.
Interrogating the relationships among the hundreds of thousands of ligands and thousands of targets that are described in ligand−target annotation databases demands ligand representations that support rapid comparisons. This is often accomplished with molecular fingerprints, usually expressed as a bit string. Widely used examples include Daylight (12) (Figure (Figure4a)4a) and Scitegic extended connectivity fingerprints (13), which encode topological [two-dimensional (2D)] information, e.g., atom types and the bond connectivity among them, though there are many others that are also popular. To those trained in biochemistry and biophysics, the idea that a topological fingerprint of a small molecule can be informative seems hard to credit, and indeed, there has been considerable effort to develop more information-rich three-dimensional (3D) methods (14−17). Still, topological fingerprints have proven themselves to be surprisingly robust for many chemoinformatics approaches and are what we rely on for our own work.
The most common way to compare molecular fingerprints for similarity uses the Tanimoto coefficient (Tc) (18,19), which compares the number of “on” bits shared between two fingerprints to all the on bits that could have been matched between them (Figure (Figure4b).4b). Developed in 1957, this metric (20) extends the Jaccard coefficient, used in 1901 to compare similarity and diversity among alpine flowers (21). The Tc measures the overall level of similarity between two molecules and is symmetric, e.g., Tc(fpa,fpb) = Tc(fpb,fpa). Like the 2D fingerprints that it is asked to compare, the Tanimoto coefficient has substantial theoretical and practical limitations; it is not a true distance measurement as it violates the triangle inequality, nor is there any accepted demarcation in Tc that identifies ligands that are functionally related, notwithstanding much effort (19). This has limited the reliability of simple chemical similarity in predicting ligand-based associations and has inspired the weighting of chemical similarity using statistical models of significance. These models have improved our ability to assign confidence to measurements of ligand similarity and especially to the similarity of sets of ligands (13,22).
An ambition of the chemical approach is to reorganize pharmacological maps, associations among proteins, on the basis of ligand similarities rather than sequence, structural, or pathway similarities. Several approaches have been explored, most mining the rich veins of ligand−target associations available in the databases. Among the first to illuminate the unexpected relationships that emerge from such an analysis were Paolini, Hopkins, and colleagues (23), who found that many bioactive small molecules possessed extensive polypharmacology, often across target boundaries (Figure (Figure5b).5b). For instance, ligands active on aminergic GPCRs were often observed to have activity on protein kinases; protein kinase ligands in turn had unexpected activities on ion channels and on phosphodiesterases. Vidal and colleagues (24) analyzed graph connectivity of drug target networks (Figure (Figure5a),5a), and Mestres (25) combined data sources to build expanded target networks. The drug−target associations studied by Vidal et al. suggested that most new drugs acted on targets that had been previously drugged, not itself surprising, but more encouragingly, there was an evolution toward more diverse targets over time. Surprisingly, correlating the drug−target map with a disease−protein map suggested that many drugs were not acting on a protein most directly implicated in a disease but rather were acting at one or two degrees of separation, at proteins that themselves were linked to the disease genes.
Proteins may be associated on the basis of not only known but also predicted polypharmacology (Figure (Figure5c).5c). In our own work, we have linked receptors on the basis of similarities among the sets of ligands annotated to bind to them. For two proteins to be related, no single ligand need be shared between them, but overall, the patterns of chemistry among their ligand sets must be similar, hence a “Similarity Ensemble Approach” (SEA) (22,26). It is here where we found that a statistical model for relating similarities to those expected at random was critical. The model was motivated by empirical BLAST theory (13,27), where individual ligands now replaced the unordered “words” used in heuristic sequence alignment, with both scoring systems using extreme value distributions.
Common to these chemocentric networks is the reorganization of the target boundaries and associations to which we have become accustomed from molecular biology. To those trained in the molecular, reductionist paradigm, as we ourselves have been, it may seem peculiar that an ion channel will be associated with a transporter, a transporter with a GPCR, a human adrenergic GPCR with a parasitic ribosome, and an aminergic GPCR with peptide and chemokine GPCRs. From a chemical perspective, however, the similarities among the ligand sets are striking. They are also generative, predicting previously unknown associations and crosses. Because they are based on specific, organic molecules, these predictions may be directly tested by an experimental assay on the same molecules that articulate them. It is to such testing, and its relevance to drug biology and target identification, that we now turn.
In the past 4 years, more than 30 drugs have been tested against more than 40 novel off-targets based on chemocentric predictions [summarized in Table Table2;2; others have been proposed on the basis of target structure-based approaches (28−39), but these fall out of the remit of this paper]. Some of these new off-targets are consistent with drug side effects, whereas others may bolster a drug’s on-target action; we consider each case in turn. Such off-target binding may cross target structure and fold categories, such as when an ion channel inhibitor is found also to modulate GPCRs, and we consider examples of such molecules at the close of the section.
Unintended off targets are widely associated with adverse drug reactions and are widely feared in drug discovery. An innovative idea pioneered by Campillos, Bork, and colleagues was to exploit known side effects to organize drugs into networks by similarities among the profiles listed on their package inserts (40). From these networks, they predicted and experimentally confirmed 13 cases of novel drug off-target activity (selections in Table Table2).2). In one example, they identified a subnetwork in which the CNS drugs pergolide, paroxetine, fluoxetine, and zolmitriptan were clustered around the anti-ulcer drug rabeprazole, a proton pump inhibitor. This led them to predict and show that rabeprazole would bind two CNS targets known for these drugs, the dopamine D3 (1.6 μM) and 5-HT1D (7.6 μM) receptors (Table (Table2).2). As rabeprazole plasma concentrations reach these levels, this may suggest that it should also be investigated for the side effects already associated with these nervous system targets (40) (whether the fraction unbound, FU, reaches these levels is not clear).
Side effect targets also emerge from approaches based strictly on ligand chemistry [Table [Table22(22,41)]. The SEA method was used to predict that the amebicide emetine would modulate the α2-adrenergic receptor. Whereas these two targets have no obvious structural similarity, inspection of emetine’s structure reveals its striking similarity to adrenergic ligands. This prediction was subsequently tested experimentally and shown to occur at 1 μM (22). Consistent with adrenergic activity, the side effects of emetine include hypotension, tachycardia, and congestive heart failure. Similarly, the well-known μ-opioid agonist methadone was predicted to bind to the muscarinic M3 receptor (22); this is consistent with, though of course far from establishes the basis of, methadone’s unusual side effects for an opioid agonist, including the heavy sweating that patients report with it. Using the same approach, Motilium, used by nursing mothers to stimulate lactation, was predicted and found to bind to the α1A receptors, here at 71 nM (41). This activity is consistent with the cardiac effects observed with Motilium (though admittedly so is its known hERG activity, at 5 μM). Finally, the widely used SSRIs Prozac and Paxil were predicted to bind β1-adrenergic receptors, the blockade of which is consistent with changes in heart rate observed in SSRI discontinuation syndrome and the sexual dysfunction induced by these antidepressants (41). Prozac and Paxil’s β-binding was only at the threshold of their plasma concentrations, without considering the fraction unbound, but a pilot study has recently correlated a common β1-adrenergic gene single-nucleotide polymorphism (SNP) known to increase sensitivity to β-blockers with these Prozac- and Paxil-induced changes in heart rate and diastolic blood pressure (42). Efforts to predict adverse drug reactions (ADRs) are also well advanced in several pharmaceutical companies, though most reports in the open literature have been restricted to retrospective correlation. The extent of these studies nevertheless suggests that this is an active area of research (43−47).
Predicted targets can also illuminate the primary mechanism of action of drugs, or opportunities for repurposing drugs to treat new diseases. di Bernardo and colleagues have made a case for the use of Fasudil in cancer and in certain neurodegenerative diseases (48). Using their Mode of Action by Network Analysis (MANTRA) method, the authors leveraged the Connectivity Map (49,50) collection to group drugs into “communities” by similarities in their patterns of specific transcriptional responses. By asking which known drugs if any were most similar to 2-deoxy-d-glucose, a known inducer of autophagy, they predicted and demonstrated strong activation of autophagic degradation by Fasudil in both human fibroblasts and HeLa cells (48). In other examples of potentially useful new targets, Bork and colleagues noted that the acetylcholinesterase inhibitor donepezil may also find use in depression, consistent with its binding to the serotonin re-uptake transporter, and Distefano’s group demonstrated miconazole’s ability to disrupt H-ras oncogene localization in cells, consistent with its predicted inhibition of protein farnesyltransferase (51). While these two activities were found for weakly binding off targets, this is not always the case; the antihypertensive Doralese unexpectedly bound the dopamine D4 receptor a log order more tightly (18 nM) than it does its canonical α-adrenergic on target (200−600 nM) (41).
Where a drug’s mode of action is unknown, chemocentric approaches can narrow the field of inquiry. The hallucinogen N,N-dimethyltryptamine (DMT) was observed to have a Kd of 14.8 μM for the σ1 receptor, implicating this target in its hallucinogenic properties and potentially identifying an endogenous ligand for σ152,53. This was surprising given the promiscuous and sometimes potent activity on σ1 of many non-hallucinogens, and DMT binds targets already implicated in hallucination, the serotonin receptors (54−56). In a blind computational panel, SEA predicted both known and novel serotonin receptor subtype binding for DMT, and subsequent ligand displacement studies suggested that it does so with affinities in the range of 100 nM, 2 log more potent than its σ1 binding. More compellingly, whereas DMT shows strong activity in a mouse model for hallucination, the 5-HT2A knockout mouse, one of the predicted and observed targets, did not respond to DMT, consistent with its status as DMT’s primary target (41). Other efforts to “deorphanize” drugs and candidates more broadly are ongoing.
A new direction in drug discovery and chemical biology is phenotypic screening. Compound libraries are screened for phenotypic outcomes in a cell or whole organism. This returns to an older pharmacological modality, in which a “model system” might be a guinea pig ileum perfused in an solution of compound (57), except that now tens of thousands of compounds are screened. This can result in interesting whole animal phenotypes for several chemical series, whose targets nonetheless remain unclear. Identifying these targets is perhaps the key challenge in such “forward chemical genetic” screens. In a recent study, Peterson and colleagues quantified the response of zebrafish embryos to light and the modulation of this response by preincubation of the fish with small molecules (58,59). Similarities among the phenotypes organized compounds into broad activity classes (59). Where a compound’s activity was unknown, SEA was used to suggest specific molecular targets. Consistent with SEA prediction, one such phenotypic hit, MAG-1, was found to be a 1 nM inhibitor of MAO in direct kinetic inhibition studies (58). Work in this area is ongoing.
Other approaches to uncovering mechanisms of action incorporate an even greater number of data sources. Using factor analysis over phenotypic profiles, chemical similarity, and predicted protein binding, Feng and colleagues derived mechanism of action inferences from a high-content cellular screen (HCS, selections in Table Table2)2) (60). Fluorescent cell cycle markers were observed in microscopy to derive phenotypic profiles associated with particular compounds. Clustering these phenotypic profiles suggested structure−activity relationships among the small molecules consistent with their structural patterns and known activities. For instance, a subcluster associated with cell death contained several known cytotoxics such as Diperamycin and Kendomycin, whereas a subcluster associated with G1 arrest contained corticosteroids such as Dexamethasone and Triamcinolone (60). To illustrate the value of phenotypic profiles in predicting targets for small molecules, the authors then showed that from the phenotypes a common target, α-tubulin, could be inferred for three groups of phenotypically similar yet structurally distinct molecules (e.g., colchicine, quinoline, and pseudolarix acid B), a prediction that they confirmed via micrographs of stained cells (60).
It is almost perplexing that the chemical view of pharmacology, which has little basis in physical or biological theory, works as well as it does to relate targets and discover drugs. Conversely, the molecular biology view, representing our best understanding of biology, has curious gaps in pharmacological organization and a checkered career in drug discovery. Pharmaceutical research is by now dominated by the reductionist program, and even a new direction like chemical biology models itself on molecular biology and molecular genetics. Still, in the past 15 years, the pharmaceutical industry has struggled to produce enough new drugs to keep up with the expectations raised for it by those introduced in the late 1980s and 1990s, most discovered using the older, chemical approach to pharmacology. How might this discrepancy, between the successes of a theoretically impoverished chemical view and the failures of a rich molecular biology one, be reconciled?
The dilemma is partly resolved by the domain of questions that the two views are asked to address. As long as pharmacology involves the actions of drugs and reagents on biology, then a view that begins and ends with these will have an advantage. The chemical view does not pretend to characterize all of biology or its mechanisms, which is the molecular biology program, but restricts itself to those targets with which bioactive molecules interact. Thus, the observation that the μ-opioid and 5-HT2A receptors are related by sequence never occurs to the chemical view. The relationship among these targets, which arguably for drug action is often irrelevant and even confusing, is meaningful in other contexts and reflects shared evolution and signaling. Correspondingly, the NMDA ion channel is related to the κ-opioid receptor only through the similarity of the drugs that modulate it; for many other biological questions, this similarity is as meaningless as the lack of sequence and structural similarity between them suggests. When the carrier of information is itself a small molecule, then that molecule may illuminate the bases of diseases in which the target is involved and sometimes also treat them. The chemical view of biology had a feeling-around-in-the-dark aspect to it and was often deeply frustrating to its practitioners, but it was necessarily focused on reagents that might themselves become drugs.
We do not pretend that drug discovery should return to this chemical approach to pharmacology or even that the chemical approach, despite its age, is mature. We do not understand the physical basis for the binding of related ligands to unrelated targets; we have merely exploited that observation. We lack a theoretical basis for information flow in the chemical view, and often the most pragmatic representations of chemical information, such as topological fingerprints, are deeply unsatisfying. Developments in these areas will move the chemical view from pattern recognition to a theoretically grounded science, helping to reintegrate the bicameral house of pharmacology. This will be crucial to meet the promise of both chambers. The chemical approach will benefit from adopting the substantial statistical, physical, and evolutionary theory that has gone into molecular biology and protein structure. Correspondingly, the sharp focus on small molecules that the chemical view brings will give the molecular biology view ready access to reagents that can modulate the receptors that it has done so much to illuminate.
In 1941, the first edition of Goodman and Gilman’s The Pharmacological Basis of Therapeutics appeared. Its title revealed a program of research based on the specific modulation of receptors by organic molecules, the behavior of which in the body could be monitored, understood, and exploited. What was then provocative is today so well accepted that the book’s title seems antique. At the time, the chemical basis of pharmacology needed no emphasis, so focused was the field on the actions of molecules that were not only its goals but also its primary classifiers and biological informants. Whereas Goodman and Gilman continues to be a central text, the classical pharmacology that it once represented has largely disappeared, as has the chemical view of biology. As primitive as that view remains, the opportunities for its exploitation are clear. The new methods sketched here systematically and comprehensively compare those targets for which ligand information is available. This remains a blinkered perspective, requiring preexisting ligand information, and so it retains some of the frustrations of classical pharmacology that made that field so receptive to the molecular biology wave that broke upon it in the mid-1980s. In its favor, the chemical view retains the integrationist program implicit in classical pharmacology. Classical pharmacologists often worked on whole organs and were systems biologists avant la lettre. As cipherlike as organic molecules can seem as information carriers, they can report on the similarities of pathways and systems as easily as individual receptors. What is new in the past few years is the quantitative restatement of classical ideas, allowing formal comparisons among targets and ligands at a scale not previously attempted. This has suggested unexpected relationships among receptors, identified targets active in phenotypic screens (58), and predicted off-targets and new disease indications for drugs (40). The new techniques remain largely the domain of specialists, but at least some of them are accessible to general investigators interested in bringing new chemical tools to targets and systems of vital interest to themselves (see, for instance, http://sea.bkslab.org).
We thank Bryan Roth and Randy Peterson for insightful conversations and collaboration in the experimental part of our research program. We thank R. Shoichet for reading this manuscript and A. Rochester and D. Barker for the table of contents graphic. The authors declare a competing financial interest.
National Institutes of Health, United States
†Research supported by National Institutes of Health Grants GM71896 (to J.J.I. and B.K.S.) and AGG002132 (to S. Prusiner and B.K.S.) and by the Rogers Family Foundation (to M.J.K., J.J.I., and B.K.S.).