January 20, 2010 | While working at Pfizer in the U.K. some 9 years ago, Andrew Hopkins and his colleague Colin Groom coined the phrase “the druggable genome”—a systematic analysis of the number of druggable targets in the human genome (see, “The Druggable Genome,” Bio•IT World, Oct 2002). In 2007, Hopkins made the bold decision to quit industry for academia. Weighing various offers, he chose to move north of the border, to the ‘City of Discovery’—Dundee, Scotland —where the ship in which Scott sailed to the Antarctic was built. Hopkins says it’s taken a good two years to adapt from industry, basically starting from scratch, but things are finally starting to come together. With patents filed, compounds being synthesized, biophysics and informatics operations up and running, Hopkins granted Bio•IT World this exclusive interview.
Bio-IT World: Andrew, Back in 2000-01, while you were working at Pfizer, you helped come up with the concept of the “druggable genome.” Where does that stand now?
Hopkins: Back in 2000 during the genome hysteria, we were thinking about the number of drug targets. How big is the genome going to be? It had huge commercial implications, because companies like Celera and Incyte were touting huge deals with pharma. We wanted to challenge this, because we’d started work with Colin Groom (as part of an internal consultancy at Pfizer called Target Analysis) and our job was to look across the entire [Pfizer] portfolio—about 50-60 projects a year—to give advice about the best lead-finding strategies: crystal structure, HTS assay design, etc. In doing that, we started to see similarities in targets. Why are some targets so difficult to get hits against and others not? It led to the concept of druggability.
Since we came from a crystallography background, we turned to ideas in molecular recognition, which could quite simply explain the whole concept of druggability… We thought about it in terms of binding sites, topology, surface accessibility, shape etc. It made perfect sense then that the lock-and-key idea could all be explained by molecular recognition theory. This really squared the circle of how you can think about the druggability of targets. A lot of work has gone on since then, people developing more sophisticated mathematical procedures to identify this. Schrodinger has new software called SiteMap—it’s good to see this type of analysis becoming established.
That was particularly inspired by Chris Lipinski’s work on Rule of Five, because he started to say drugs have a defined set of properties, then conversely the binding sites they bind to must as well. When we started to put together data about what sort of targets are likely to be druggable, we were shocked by the lack of data. We didn’t have systematic databases of what drug targets the industry has worked on. It was quite shocking considering the amount of money the industry is spending, that we didn’t know our ‘history’ in any formal way we could learn from… If we spent all this money sequencing the genome, wouldn’t it be good to know what targets we’ve already got?! That led to work with John Overington, published in 2006 (see, “Pfizer’s Global Survey of Pharmacological Space,” Bio•IT World, Sept 2006) on the number of drug targets.
We started thinking about comparative genomics. Once we looked across druggable targets in a comparative analysis of yeast and C. elegans say, we could predict the size of a genome back in around 2000. We came up with predictions of 23,000-27,000 genes in the human genome. I only wish then we’d put money down! It wasn’t such good news to those predicting 100,000. OK, so the genome’s going to be much smaller than expected, then we thought about predicting the druggability by homology, and came up with the figure of 3,000 targets. The latest figure is 3,600 or so. Even as we discovered new families and new targets, the power law structure of gene family populations suggests it was a pretty robust number…
How does chemogenomics contribute to rapidly identifying good drug targets?
It’s a common procedure in the drug industry now to consider the druggability of a target. We’ve become interested in rapidly identifying drug targets from pathogen genomes. Next generation sequencing means we have the capability to rapidly sequence pathogen genomes, yet it then takes us years to discovery new pharmacologies against that pathogen. Now we’re looking at not just the druggable genome but how to prioritize targets in a pathogen genome: for example what should be the top ten targets drug discoverers should go after, with the highest chance of success? When you look at a neglected disease pathogen, there may only be a few drug discovery groups in the world working on it that don’t have the luxury of doing hundreds of projects. So we are looking at network theory to infer which particular chemical attributes a protein might have and multi-parameter optimization to prioritize a druggable genome. Now we’re not just interested in the druggable genome in a pathogen but to really prioritize the genome to find what’s the best sort of target for small-molecule chemistry.
The other development in the past year is the establishment of the ChEMBL group at the European Bioinformatics Institute under John Overington. One of the key things missing in academia, when I moved, was the pharmacological/chemical SAR data, which is essential to link chemistry to biology. Traditionally, the pharma industry and American Chemical Society controlled this information. In my first week here, Mike Ferguson and I started lobbying the Wellcome Trust that they should think about investing in one of these large databases, such as Overington’s STARlight database at BioFocus, and make it an open access resource. Many academics including Janet Thornton supported the idea, and BioFocus was open minded. I have to commend Alan Schafer at the Wellcome Trust for having the vision and energy to create the Biofocus/EBI deal that lead to the creation of ChEMBL. Now it’s established as an excellent resource and it’s going to improve the prospect of chemoinformatics and chemogenomics and drug discovery methods development generally.
Once we can have high-quality data in chemoinformatics, it opens up all sorts of abilities in machine learning. If we are trying to increase efficiency in modern drug discovery, we need to think about how we learn from all the data that’s available. For example Brian Shoichet and Bryan Roth’s recent paper [Keiser M.J. et al. Nature Nov 12;462:175-81 (2009)], on the validation of a predictive polypharmacology method from large datasets is a great example of machine learning that can be developed on top of large databases. We ourselves are excited about this area of applying machine learning to chemical design. The challenge is how do you reduce costs of drug discovery and expand the opportunity space? We see polypharmacology or network pharmacology as an exciting new area for drug design. The druggable genome may be finite yet the combination of targets drastically expands the opportunity space. If we’re designing drugs against multiple drug targets, an idea inspired by systems biology and functional genomics is that we want to hit multiple nodes in the target. We need to think about new drug design tools to aid us. Hence in tackling the challenge of design polypharmacology we are faced with the problem of automating drug design per se due to the complexity of the problem.
We just filed a patent [last November] on a new automated drug design method that we believe shows a lot of promise. We’ve been inspired by Ross King’s work on the automation of science [King R.D. et al. Science 324, 85-89 (2009)] to think whether drug design can be automated in such a manner. Using this large public SAR data, we’re trying to automate the drug design cycle. We’ve got some exciting first results. We’re designing novel compounds, getting them synthesized and tested—and hopefully published in the near future. With a talented Ph.D. student, Jeremy Besnard, we’re writing new algorithms. The key is designing and modeling both the creative processes of drug design as well as incorporating predictive tools into the close loop cycle. If we think of medicinal chemistry as a creative process, how do we try to model that? Using a combination of techniques, including evolutionary multi-objective optimization methods, we’ve got some very promising results, going from a lead compound and optimizing it to desired profiles, or going from a drug that may have weak activity against a target and optimizing it. I’m convinced that for certain drug discovery problems, automating drug design is conceptually possible.
You seemed to have a lot of freedom at Pfizer, so why did you decide to leave?
A very good question! It was partly a strategic move on my part, considering the grander scale trends or structural changes taking place in pharma. There’s no doubt the patent cliff is causing major disruptions in employment and creating a lot of uncertainty. Academia gives you a much longer tenure and allows you to think about the deep problems rather than worry about whether your company will survive.
What we are witnessing now are not just M&As and company re-organizations but structural changes taking place in the industry. Structural change means there will be more opportunities for new ideas from outside the industry, whether from academia or small biotechs, to feed into the pharma. It will become more a customer and purchaser of research and new ideas and its own contribution of research share of the pie will potentially decline. The structural change happening now in pharma has an analogy with another industry that is dependent on the blockbuster model: a continuous pipeline of projects costing hundreds of millions of dollars, depending on a complex number of different skills—the movie business. With the breakup of studio conglomerates—by law—in 1948, a vertically integrated industry fragmented into a rich ecology of players. The major studios still exist today but their role more akin to VCs. The fundamental change that happen in the movie business was that the studios couldn’t command the same economic rents as they previously could. Those rents are increasingly shared with the creators of the product. That creates a great opportunity for people on the outside to contribute to the industry. So ironically, despite the turbulent times, it’s a great time to be outside the pharma industry and feeding ideas into it.
What do you think about the recent turmoil at Pfizer?
Ah, where to start?! It’s a real shame to see a great company like Pfizer go through these upheavals.. It was a great culture and a fantastic place to learn the craft, where I was lucky to have a lot of freedom to explore basic ideas about the science of drug discovery. The challenge comes from the impending loss of Lipitor and a huge seismic change in its revenue base. But you can’t overcome the fact that, ultimately, it’s a problem of innovation and productivity that is at the heart of the revenue gap, not just for Pfizer but for the industry. Ironically, as we respond to that, it creates a downward spiral where one potentially creates bad morale, then reduced capacity to innovate in the future as people are worried about maintaining their paycheck. It becomes a basic survival mode. That said, they have some exciting compounds—a JAK3 inhibitor for arthritis is a very exciting compound. But for the sheer amount of money Pfizer’s been spending on research ($7 billion/year), you should be getting something! When we are looking the cost of bringing a drug to market being greater than $6 billion, which it arguably is at Pfizer at present, one has got to challenge some fundamental assumptions of how innovation is managed. Tudor Oprea and I hope to publish some of our thinking on this topic shortly.
The changes in strategy with the BBC and the Wyeth merger were driven from a strong strategic perspective. From my understanding, Pfizer wanted to be a strong player in biologicals. The earlier ideas about establishing biotherapeutics, the BBC under Corey Goodman, were driven by the same strategic thinking, as was the purchase of Wyeth, a pharma with a strong biological presence. Unfortunately, the fallout in acquiring the new capability is that the nascent internal capability that was being grown organically becomes almost surplus to requirements, which is a real shame for the individuals involved. In the longer term, as many pharma companies want to move biological. There is still the question: will that produce the revenues required? The FDA approval of the biologicals isn’t growing exponentially, it’s pretty steady and still at a low level. That creates its own challenges. That’s why I believe we will see a renaissance in small molecule chemistry drug discovery.
Several biotechs are taking novel approaches to attacking chemical space, such as Ensemble Discovery (see, “Breaking the Rule (of Five)," Bio•IT World, April 2008) or binding models.
Yes, there are some very interesting developments taking place. There are new chemical and screening technologies coming online. We’re thinking about this in terms of automating design. It opens up medicinal chemistry, which for a long time has been seen as an art form. The other exciting area is in allosteric modulators—developing informatics and screening technologies to discover them.
This fits into another concept—the rediscovery of pharmacology and the importance of pharmacokinetics (PK) and pharmacodynamics (PD) as well as just the target. For a long time, we thought about chemicals and targets—as a 2-D space with binary read out on function such as gene knockouts to study target pharmacology. What we’re appreciating now is the concept of system dynamics: how one perturbs the target gives very different effects. For example, antipsychotic drugs often work when they have weak affinity. If you have too strong activity, you get side effects, but if you have too little affinity, you don’t get the beneficial effects. In some circumstances, it makes sense to have very slow offset to really affectively perturb the pharmacology of the target. The idea of insurmountable antagonism, which one can get from allosteric inhibitors. There are different ways to modulate the pharmacology and the perturbation of the system. Polypharmacology is another idea where small molecule chemistry is the toolbox for developing system biology based drugs. Interestingly, even though there’s a big head rush into biologics right now, this is where we can do things with small molecules. It’s not just about screening to look for molecules that bind to a target, it’s how will those compounds modulate the target.
Why did you choose to move to Dundee?
Dundee is a very good place to get started in translational research. Though it’s a small university, it’s very highly ranked in biology in terms of citations in Europe. Dundee has also been very serious in moving into drug discovery. It got a major grant from Wellcome Trust to set up a drug discovery unit—it’s like a small pharma company. It’s got HTS, industrial-class medicinal chemistry, DMPK (drug metabolism PK), dedicated to neglected diseases. It’s under the guidance of the head of research, Mike Ferguson, who is a parasitologist and Paul Wyatt (ex GSK and Astex). As well as neglected disease, it’s one of the few places in academia where you have a culture of wanting to do drug discovery. My interest is developing new methodologies, new screening methods and new informatics methods, so there are plenty of live projects we can work on.
Tell us more about what your group is working on, particularly in informatics?
I’m the chair of medicinal informatics. We wanted to combine cheminformatics, bioinformatics structural bioinformatics, text mining and clinical informatics later. We want to apply a plethora of informatics methods to the concept of information-driven drug discovery. The core problem we’re thinking about is the chemogenomics—how we can apply that to parasite and pathogen diseases, and quickly identify targets, particularly given that genomes can be sequenced so quickly. We’re also doing work on developing an ontology to describe drug targets and molecular interactions—collaborating with Ross King and Larisa Soldatova at Aberystwyth, It’s great to get into developing OWL and RDF technology. We want to apply that, the druggable characterization of drug targets, particularly to identify and ask questions, such as finding allosteric modulators.
The other side is the chemoinformatics, where we’re interested in applying AI methods to try to automate the drug design problem. I believe that drug discovery is an automatable and solvable problem. This could be the next revolutionary advance in drug discovery, with potentially tremendous implications for design work and cost reduction.
You’ve started a spin-off company?
We started a little company called Kinetic Discovery, based on fragment screening and GPCR biosensor analysis, and we’ve got some plans to form 1 or 2 other companies in the next year or two. I’d like to see if we can become a serial entrepreneur—that’s the dream anyway! We’ve got some exciting technology, developing biosensor technology for G-protein coupled receptors (GPCRs). There was great excitement with the advent of GPCR crystallography, with companies like Heptares being founded, but one of the limitations is the proteins are difficult to crystallize, and it doesn’t tell you anything about the actual kinetics or affinity of the compounds. We’ve been developing new biophysical techniques—my colleague Iva Navratilova developed this area when she was at Utah—and we’ve extended this to a new biophysical screening technology for GPCR allosterics, and recently filed a patent. This could open up this important target class. We’ve also been exploring biosensor approaches to fragment screening, which I’m a great fan of. The next areas for commercialization from our lab I see are allosteric screening and polypharmacology design.
This article also appeared in the January-February 2010 issue of Bio-IT World Magazine.
Subscriptions are free for qualifying individuals. Apply today.