Making virtual screening a reality

The discovery of new bioactive compounds for specific biomolecular targets represents a significant hurdle in the early stages of drug discovery. Advances in automation and bioanalytical methods have provided high-throughput screening (HTS) techniques that can perform individual biochemical assays on as many as a million compounds or more. Even with HTS, the discovery of new lead compounds largely remains a matter of trial and error. Although the number of compounds that can be evaluated by HTS methods is seemingly large, these numbers are small in comparison to the astronomical number of possible molecular structures that might represent potential drug-like molecules (1). Often, far more compounds exist or can be synthesized by combinatorial methods than can be reasonably and affordably evaluated by HTS. As the costs of computing decreases and as computational speeds increase, many researchers have directed efforts to develop computational methods to perform “virtual screens” of compounds (2–4). Because the cost of performing screens in silico can be faster and less expensive than HTS methods, virtual screening methods may provide the key to limit the number of compounds to be evaluated by HTS to a subset of molecules that are more likely to yield “hits” when screened. For the practical advantages of virtual screening to be realized, computational methods must excel in speed, economy, and accuracy. Striking the right balance of these criteria with existing tools presents a formidable challenge. In this issue of PNAS, Schapira et al. (5) present an inspiring example of structure-based virtual screening applied to a challenging problem of developing new thyroid hormone receptor antagonists when only a related receptor structure is available.

Receptor-based virtual screening uses knowledge of the target protein's structure to select candidate compounds with which it is likely to favorably interact. Even when the structure of the target molecule is known, the ability to design a molecule to bind, inhibit, or activate a biomolecular target remains a daunting challenge. Although the fundamental goals of screening methods are to identify those molecules with the proper complement of shape, hydrogen bonding, and electrostatic and hydrophobic interactions for the target receptor, the complexity of the problem is in reality far greater. For example, the ligand and the receptor may exist in a different set of conformations when in free solution than when bound. The entropy of the unassociated ligand and receptor is generally higher than that of the complexes, and favorable interactions with water are lost on binding. These energetic costs of association must be offset by the gain of favorable intermolecular protein–ligand interactions. The magnitude of the energetic costs and gains is typically much larger than their difference, and, therefore, potency is extremely difficult to predict even when relative errors are small. Although several methods have been developed to more accurately predict the strength of molecular association events by accounting for entropic and solvation effects (6, 7), these methods are costly in terms of computational time and are inappropriate for the virtual screening of large compound databases. The challenge in developing practical virtual screening methods is to develop an algorithm that is fast enough to rapidly evaluate potentially millions of compounds while maintaining sufficient accuracy to successfully identify a subset of compounds that is significantly enriched in hits. Accordingly, structure-based screening methods typically use a minimalist “grid” representation of the receptor properties and an empirical or semiempirically derived scoring function to estimate the potency of the bound complex (8). Several programs now employ a range of scoring functions, but it is often difficult to assess their effectiveness on difficult “real-world” problems. In this study, computational chemists Matthieu Schapira and Ruben Abagyan team up with molecular endocrinologists and organic chemists from Herbert Samuels' and Stephen Wilson's groups to evaluate their methods in identifying new thyroid hormone receptor (TR) antagonists.

The ability to design a molecule to bind, inhibit, or activate a biomolecular target remains challenging.

Although the structure of the TR bound to its natural agonist triiodothyronine (T3) is known (9), antagonists of TR that may be useful for the treatment of hyperthyroidism are believed to bind TR in a different but related conformation. By structural analogy to known antagonist-bound structures of related hormone receptors, a hypothetical model of the antagonist-bound form of the receptor was created through molecular modeling (10). This model was used to virtually screen drug-like molecules from the available chemicals directory (ACD) as potential TR antagonists. The selected candidate compounds could be further narrowed to just a few hundred compounds by using virtual screening to eliminate the compounds that bind the agonist-bound form of the receptor. Of 75 analogs that were purchased and evaluated in cell-based assays, 14 (19%) showed low micromolar antagonist activity, far more than would typically be obtained from random-based screening or screening methods that only consider the structure of known ligands. Future studies will ultimately determine whether such high hit rates are general for different receptor types. However, even hit rates below 1% can represent a substantial savings of time and costs for lead identification compared with blind HTS.

Several groups have used structure-based design to convert known hormone receptor agonists to antagonists. Because the difference between the agonist-bound and unliganded or antagonist-bound receptor structures primarily involves the displacement of helix-12 of nuclear receptors, analogs of ligands that bear molecular extensions that prevent helix-12 from adopting an agonist-like conformation often behave as antagonists (11). Although this approach has been successfully applied to the discovery of other hormone receptor antagonists, including TR, it limits the discovery of new antagonists to direct analogs of known agonists (12). Virtual screening based on receptor structure therefore has the distinct advantage of aiding the discovery of new antagonist structural classes or pharmacophores that may ultimately yield compounds with improved pharmacokinetic properties or help in the discovery of leads that are not restricted by existing patents. The 14 compounds identified by virtual screening importantly represent new TR antagonist pharmacophores; however, these potential lead compounds have notably lower potency than antagonists previously identified by “extension” modification of known TR agonists (12). To help improve the potency of their best hit, Schapira et al. (5) synthesized a series of analogs to optimize their lead compound by using virtual screening to select from among the possible compounds that could be synthesized directly from commercially available reagents. The eight new analogs were all active antagonists, some of which showed modest improvement over the initial lead.

Virtual screening methods continue to develop in sophistication and accuracy. These results serve to illustrate that structure-based virtual screening provides a rapid and powerful tool for the discovery of new bioactive pharmacophores and may additionally benefit the process of lead optimization, even when the available structure represents an alternative conformational state of the biomolecular target.