In February 2017, we released a special perspective on the history of computational drug discovery, recent trends in the space, and insight into Cyclica’s approach. That publication has proven timely, in light of the recent article for the FDA Blog “FDA Voice” on the impact of the 21st Century Cures Act, posted by Scott Gottlieb, commissioner of the USA Food & Drug Administration. The act, signed into law on December 13, 2016, authorised $6.3 billion in additional funding, part of which is dedicated to research and drug development. Gottlieb remarks: “Modeling and simulation play a critical role in organizing diverse data sets and exploring alternate study designs. This enables safe and effective new therapeutics to advance more efficiently through the different stages of clinical trials.” The purpose of this special perspective is to shed light on a newly practical type of modeling and simulation that we believe can play a large role in this effort and may well be the next big thing: in silico polypharmacology. - Naheed and Andreas
In the early days of drug discovery, new drug molecules were usually identified through their physiological effects, observed either by serendipity (like in the case of penicillin), or through ancient tradition (like in the case of Aspirin). There was really no other way to find drugs, because very little was known about the molecular underpinnings of biology. After the advent of molecular biology, it became possible to examine the machinery of life in detail, and to identify parts (usually proteins) that, when broken in some way, caused disease. These proteins became targets for intervention, and the hunt for drugs focused on finding small molecules that would specifically bind to, and interfere with them. This approach is known as target-based drug discovery.
Target-based drug discovery rests on the “lock and key” model of drug binding, which goes back to the work of Emil Fischer in 1894 (Fischer, E. (1894) Einfluss der Configuration auf die Wirkung der Enzyme. Ber. Dtsch. Chem. Ges. 27, 2985–2993). The lock and key model posits that a drug molecule will bind to a protein in a very specific fashion. Like a key will only fit one lock, a drug molecule will only bind to one particular protein. This model holds up quite well, and is the foundation of modern pharmacology, which stands alongside hygiene and vaccination as one of the fundamental achievements of modern medicine. It is hard to overemphasize the vast improvements in life expectancy and public health that have been built on this principle. Just think of antibiotics, pain relievers, anticancer drugs, as well as the many widely prescribed preventive medications for metabolic, cardiovascular and psychiatric disease. Many of these life-changing drugs were discovered in a target-based fashion, most often systematically using high throughput screening (HTS), i.e. the brute force testing of large “libraries” of thousands to millions of chemical compounds for activity against one selected target protein.
As with most models, though, the lock and key model is not a perfect description of reality. There are more than 20,000 distinct proteins in the human body, and drug molecules that were designed to bind one of them (i.e. “on-target”) commonly also affect multiple others (“off-target”). Often, these “off-target” interactions cause unintended and unanticipated side effects, necessitating a thorough, lengthy and expensive process to ensure safety and efficacy of newly discovered drugs. This phenomenon of one drug acting on multiple receptors is called polypharmacology (Figure 1).
A toxic side effect found late in the process, after tens or hundreds of millions of dollars have been spent, can cause the derailment of the entire program. For example, Pfizer’s torcetrapib had $800 million invested in it when it failed spectacularly in 2006 for increased mortality in patients taking it. The market value of the company dropped by $21 billion when the negative results were announced. A post mortem review of the data suggests that polypharmacology played a large role in torcetrapib’s adverse effects. Numerous similar stories can be told, and some recounted in the publication linked above. If there was a way to quickly and efficiently find all the proteins in the human body that interact with a molecule, potential adverse effects could be flagged early in the process, when there is still flexibility about the choice of the molecule, allowing the selection of a safer molecule and reducing the chance of a costly failure.
Target-based drugs are only as good as the relevance of the selected target to the desired effect, which has led to some disappointment with the results of target-based drug discovery. Together with a number of scientific and technological advances, this disappointment has led to a recent renaissance of phenotypic screening methods. Phenotypic screening is, in some ways, a return to ancient history, when drugs were identified by their physiological action rather than through molecular understanding of biology. In other ways, though, phenotypic screening is new and much different today, in that the discovery of a suitable molecule is much less serendipitous, through the use of HTS with an assay that measures a physiological effect rather than target binding.
Phenotypic screening circumvents the requirement of target validation, because no target needs to be identified to perform it. It has the great advantage of allowing for the discovery of drugs with unknown mechanism of action. Because we do not understand everything about biology, this greatly increases the chances of finding an effective molecule. However, in order to chemically optimize discovered molecules, and to advance drugs through the development process, it is still necessary to identify the relevant target(s) of a candidate after screening. This is done through the laborious and expensive process of target deconvolution, which is somewhat like the proverbial search for a needle (the target) in a haystack (all proteins). The difficulty of target deconvolution has stopped many phenotypic screening efforts dead in their tracks. If there was a way to quickly and efficiently find all the proteins in the human body that interact with a molecule, an unknown target would be flagged and could readily be validated in vitro, circumventing the risky and expensive in vitro deconvolution process.
As discussed in our perspective on computational drug design, much work has been done to use computational modeling for drug discovery. Computational chemistry is most often ligand-based, i.e. the properties of a drug molecule are predicted from its structure in conjunction with known properties of other similar drug molecules. A drug’s target is not explicitly taken into account in this approach, which is exemplified by Quantitative Structure Activity Relationship (QSAR) modeling. Since the most important property of a drug molecule, its biological effect, is almost always mediated through a protein, QSAR is limited in predictive power. Structure Based Drug Design (SBDD) goes one step further by modeling the physical interactions between a molecule and its intended target protein, but it still fails to account for off-target interactions, or polypharmacology. To create a full polypharmacological profile of a given drug molecule, a proteome-wide screening approach is required.
There have been several efforts at in silico methods for proteome-wide screening, with very limited success. One reason for this is the computational power required for proteome-wide screening. There are around 20,000 human proteins, with many potential binding sites on each. Simply extending the same approaches used for single target SBDD would require million-fold larger computers for a proteome-wide screen. Another, more pertinent reason for the lack of in silico proteome-screening in the past is that, to be useful, it requires the molecular structure of a substantial fraction of all proteins to be known. For a long time, the number of protein-structures available for screening was limited to only a small percentage of the proteome. However, with advancement in x-ray crystallography and cryo-electron microscopy (cryo-EM), the structural representation of protein structures available in the protein databank (PDB) is exponentially increasing, and has reached the point where most structures are available with only some gaps left to be filled in. Figure 2 illustrates the confluence of these two factors in making in silico proteome-wide screening possible.
Besides increased computing power and sufficient coverage, in silico proteome screening is sufficiently different from other, target-based, SBDD methods that a novel algorithm is required. Cyclica, in the 4 years since its founding, has developed and perfected its PROBEx surface matching technology, which compiles the publicly available protein structures deposited in the PDB into a proprietary index database and allows all the surfaces of all the proteins to be queried with a drug molecule for any suitable binding sites. PROBEx is now integrated in the Ligand Express platform, which also provides the SWITCHx and DIVEx components to connect proteins with outcomes using rich systems biology data sets.
With Ligand Express, we can uniquely address several key pain points that we have heard of again and again from partners in the pharmaceutical industry. Foremost amongst them is the deconvolution of targets from phenotypical screens. As described above, phenotypical screens result in hits, i.e. specific molecules with a desired biological or physiological effect, but without any knowledge about the mechanism of that effect. PROBEx can provide a short list of candidates for the protein target (or targets) that the molecules bind to, thereby providing the elusive first step of the mechanism of action that makes deconvolution so difficult and expensive.
Another obvious pain point PROBEx can address is the prioritization of lead candidates, i.e. drug molecules under consideration for further development. Ligand Express can produce and help analyze polypharmacological profiles for each of several lead candidates, which leads to insights into their potential adverse effects and allows scientists to make smarter decisions about which ones to advance through the risky, billion dollar development process.
In Ligand Express, the confluence of modern computing power, sufficient coverage of protein structure data, and the PROBEx algorithm has finally led to an efficient computational approach to target deconvolution and polypharmacology prediction to help make drug discovery faster, cheaper, and safer.
Naheed Kurji, President and CEO
Andreas Windemuth, Chief Science Officer
With thanks to our awesome team!