By Kevin Davies
March 1, 2008 | SHARON SHACHAM has a Ph.D. in molecular biophysics and is a brilliant computational biologist, but still she can’t seem to get the computer monitor to turn on. “I don’t have my glasses,” she says with a hint of embarrassment. Finally, the screen bursts into life, revealing a gorgeous 3-D model of a promising drug target, a G-protein-coupled receptor (GPCR), which Shacham twists and rotates with the mouse with dexterous familiarity. Software, not hardware, is more to her liking.
Michael Kauffman, CEO of Lexington, Mass.-based EPIX Pharmaceuticals, leaves no doubt that his chief assets are Shacham (senior VP product development) and the algorithm she wrote for her Ph.D. at Tel Aviv University. In 2000, Shacham and her thesis advisor, Oren Becker, founded a small biotech company with entrepreneur Silvia Noiman in Israel called Bio IT, later changing its name to Predix Pharmaceuticals. Kauffman left his position as head of Velcade clinical development at Millennium Pharmaceuticals to join Predix in 2002. “I thought this was going to be a game-changing technology,” he says simply.
Back then, Predix had just ten staff, including a solitary chemist in its U.S. office. The company has since grown to about 100 employees, including 25 computational scientists in Israel. Along the way, it acquired Physiome in 2003 although none of Physiome’s technology is currently being used (See A Virtual Pharmacopeia, Bio•IT World, November 2002). Remarkably, by 2006, Predix had propelled four new orally available drugs — all targeting members of the GPCR family — into the clinic. That year, Predix conducted a reverse merger with EPIX, a producer of MRI contrast agents with cash but no pipeline, thus securing a listing on NASDAQ.
“Every drug has made it into Phase II,” Kauffman. “They’ve all appeared to be safe and well tolerated. Every one has shown proof of biological activity in humans, and also clinical activity.” In 2006, he signed major deals with Amgen and GlaxoSmithKline (GSK), the latter around four compounds, including a clinical candidate for Alzheimer’s disease and three discovery programs. “We’re not perfect, but about 75% time we do a program, it makes it to animal studies,” he says.
EPIX’ early success relies on a suite of software and in silico programs, starting with PREDICT, the algorithm Shacham wrote for her Ph.D. to model the core structures of GPCRs. This is a huge gene family coding for up to 1,000 proteins in humans. Many currently approved drugs — beta blockers, anti histamines, beta agonists — target one GPCR or another to treat depression, asthma, allergies, hypertension, fertility — almost everything except cancer. Once EPIX scientists have the 3-D in silico model of a specific GPCR target, or lock, they can search for the right key, relying extensively on computational methods.
Protein folding prediction is notoriously difficult (See Improving Structure Prediction, Bio•IT World, November 2007), but for membrane-bound proteins such as GPCRs, there is little alternative. Only two GPCR proteins have been crystallized to date: rhodopsin in 2000 and the beta-adrenergic receptor last year. Because of these notorious difficulties, Shacham devised a de novo modeling approach that could be applied to the full GPCR family without relying on any homology data.
PREDICT focuses on the GPCR core — a trademark stretch of seven transmembrane domains that snake across the membrane. It models various theoretical conformations of the core purely from the receptor’s amino acid sequence. This suffices because most drug compounds would bind to a pocket in the GPCR core, exerting either an agonist or antagonist (inhibitory) action on the receptor.
Shacham explains: “We create many of the possible conformations in 3-D, and most of them are false. They are not stable energetically, so we call them ‘decoys.’ We’re trying to identify the most stable structure of the many options from those that are not energetically favorable. We’re trying to find out the energy minima.” These conformations are considered the most likely to exist naturally.
Shacham’s program first examines the entire conformational space of the GPCR target. Once she has identified a local energy minima area, the process repeats, zooming into this area to generate more structures to identify the global minima. PREDICT also allows Shacham’s team to build models of related receptors that might be the source of undesirable drug binding and thus side effects. Shacham says they’ve modeled some 25 protein structures on the computer for selectivity screening. (See sidebar below: PREDICTing Success.) The recently published structure of the beta-adrenergic structure provides a great test for PREDICT. “We are modeling this structure right now,” says Shacham. “We did rhodopsin; we had a very nice result.”
The 3-D model is typically produced with a drug candidate or ligand bound to the protein target. “Most of the commercial chemistry companies give you catalogs on CDs of their compounds,” says Shacham. “We don’t have to synthesize them.” From a virtual library of say three million drug-like compounds, Shacham’s group can rapidly dock them sequentially in the target’s binding pocket in the computer. They predict the best drug binders, select the best 100-200 compounds, then conduct a binding assay. “So far, the work is remarkable, almost 75% success,” she says.
Once they have a hit, they select the best compound from different chemical families. “Then we optimize it using our in silico models together with real wet chemistry and the biology team to come up with the drug candidate.”
The lead optimization process involves a suite of algorithms, including RISS for screening, SiteWise for scoring virtual hits, and ICELR-3D for 3-D optimization, to create a “virtual pharmacological profile.” These programs dock and design a relatively small number of drug candidates. Most pharmas take two million compounds and test one at a time. “It would be like going to the hardware store and buying two million keys!” says Kauffman. EPIX’ structure-based predictions to target and drug design reduces both lab work and time.
Says Kauffman: “Instead of having to make thousands of molecules in the laboratory and test them all, we make thousands of molecules on the computer, and use our structures to give us a real competitive advantage. We built a lot of our own software and we also use anything off the shelf. If it’s ‘not invented here,’ that’s better. It means we didn’t have to invent it.” Nor is he trying to discover any new disease biology. “Our targets tend to be clinically or biologically validated,” he says. “There has to be some data in humans that says this is a good target. Our job is to make better drug compounds, not to sort out human biology.”
EPIX relies on video systems to facilitate a regular transatlantic dialogue between Shacham’s team in the United States and the computational team in Israel. Hundreds are evaluated in this way — Shacham calls it a “chemistry gestalt.”
Together, they design and refine drugs that have a desirable pharmacological profile — about ten properties every drug needs to have. “There’s no assay for safety, but you need to hit the target receptor, and you’ve got to avoid these other targets,” says Kauffman. “The drug has to be taken by mouth, and stay in the blood a certain amount of time. The idea was to make this as rational as possible, from every aspect.”
The EPIX computational team is much smaller than a typical large pharma, and much less concerned by the sheer volume of compounds produced. Says Kauffman: “Nearly all companies build models, but they don’t usually use them to predict things. They use them to explain things. It’s like explaining the weather after it’s happened — it’s good, but no one’s really that interested.” At EPIX, “people aren’t allowed to make a compound … until it’s been through this iterative ying-yang matrix.”
Whereas big pharmas might have 10-20 medicinal chemists on a project, EPIX has four. And most medicinal chemistry groups in big pharma get paid by the number of compounds synthesized, Kauffman is more interested in quality than quantity. “If you make me one compound and that’s the drug, that’s enough!” he says. Besides, “The more compounds you make, the more you have to test, the more money you spend. Maybe they’ll find better compounds, but it hasn’t happened yet. We don’t find all the good compounds, but we only need to find a compound that’s good enough to be a drug.”
EPIX’ predictive prowess reveals answers much faster than the industry average. Kauffman says that most pharmas typically evaluate 1,000 compounds per new drug program but, “our average is less than 100.” EPIX’ internal numbering system tells the story: the depression drug candidate, PRX-00023, was the 23rd compound synthesized. Similarly, the Alzheimer’s candidate (PRX-03140) was number 40 in that series.
The most advanced of EPIX’ four drugs in the clinic is PRX-00023 for depression and anxiety (See, Molecular Evolution: PRX-00023). The average time from initial hit to the clinic is a mere 24 months instead of 3-5 years. That breaks down as follows: about two months to model the target; 4-6 months for screening; and 9 to 12 months for lead optimization and approximately 9 months for manufacturing and preclinical tests. “In most cases, we can have the [provisional] patent converted just before the drug candidate goes into the clinic. Literally the week before this drug hit the first human [in February 2004], we submitted the full patent [which now has been granted until 2024]. So we save about three to five years on average in patent life,” says Kauffman
PRX-00023 targets a serotonin receptor called 5-HT1A. The most prescribed drug against this receptor is Buspar, but Kauffman says PRX-00023 is more specific, “much better tolerated, and doesn’t have a lot of the commercial marketing problems” such as sexual dysfunction that most antidepressants have. He claims that PRX-00023 has many properties that are best in class for 5HT1A selective drugs.
EPIX’ Alzheimer’s drug, PRX-03140, is a 5-HT4 agonist that specifically enters the brain. Pfizer’s Aricept ($2 billion sales/year) increases acetylcholine by inhibiting cholinesterase, whereas PRX-03140 stimulates the release or production of the neurotransmitter implicated in memory. But acetylcholine also affects the stomach, resulting in side effects including nausea and diarrhea. The brain specificity of PRX-03140 comes down to “a bit of luck and very good planning,” says Kauffman. This improved selectivity allows EPIX to “hit the brain much harder, without affecting the gastrointestinal tract” raising the dose without causing side effects. The clinical data so far — despite a January revision blamed on the CRO running the trial — suggest a significant improvement in memory function in a period of weeks rather than months. “As a physician, doing something in my lifetime for Alzheimer’s would be phenomenal,” says Kauffman.
Ironically, Kauffman says EPIX has been a victim of its own early success. “A big bottleneck is moving drugs into Phase I, because we’re productive. We haven’t had the statistically expected attrition. It’s like having four kids who have to go to an Ivy League school, and you’ve no idea how you’re going to fund it!” He anticipates pushing another drug into the clinic this year.
Among ongoing projects are models of the cystic fibrosis transmembrane regulator (CFTR). This required Shacham to adapt PREDICT to model the CFTR, which is an ion channel (not a GPCR). With a model of the CFTR 3D-structure now in hand, she hopes to identify “correctors” — compounds that increase the levels of the protein expressed and potentially will also increase the activity from the reduced CFTR made in patients’ cells.
Shacham is also modeling the CCR2 chemokine receptor for inflammation. After the target is modeled in Israel, Shacham uses Benchware software from Tripos (“a very good program”) to view and manipulate the 3-D structure on her Dell flat panel (when it’s working). Nestled inside the CCR2 core is a ghostly image of the binding cavity. Shacham maneuvers the mouse for a closer look, showing just the target residues that directly interact with the drug candidate.
When comparing different drug candidates, Shacham superimposes the color-coded compounds, focusing on a specific residue or atom critical for activity. The drug compounds can be modified to strengthen the desirable interaction and improve potency. A list of compounds on screen represents the in silico modifications of EPIX’ compounds. This project is still “a work in progress,” she says.
There’s been considerable interest from big pharma in EPIX’s technology, but Kauffman says, “there’s also a lot of not-invented here.” Other than Arena Pharmaceuticals with a few compounds in the clinic, GPCRs are typically the purview of big pharma. Despite numerous conversations, “we refuse to do any type of service deal,” says Kauffman. “The entire GSK deal is success based. There’s some cost sharing, but it’s almost all milestone based.”
Kauffman says he would have to be offered “a heck of a lot” of money, for the algorithm itself because it would allow competitors to do more than just model GPCRs. “We’ve been approached about that in a couple of cases, but we see Millennium-deal sizes when we [consider] that.”
I ask Shacham if she ever thought that her Ph.D. project back in Israel would eventually become the prize ingredient of a dynamic drug development program. “Oren [Becker] always had confidence,” she answers. “I’m the most skeptical person in the company, which is good, because they have to show proof that it is working.”
Sidebar: PREDICTing Success
One of the keys to PREDICT’s success is the energy score. “We spend a lot of time trying to get the score that will work fast enough and enable us to differentiate between structures that are decoys and structures close to the energy minima,” says Shacham. The program starts simple and only “adds the complexities gradually through the algorithm. We start with the sequence, a set of amino acids, then we look at only two dimensions of the core structures in the membrane. Each structure is then expanded to a 3-D structure, however a reduced representation of the amino acids is used to facilitate the optimization of each structure. Once the most stable structure is recognized out of the thousands of decoy structures, it is finally expanded to a full atom 3-D structure. The approach enables us to mimic the folding process in a timeline that is reasonable to come up with a final model.”
The GPCR modeling is performed with a known molecule bound to the receptor. “We create a virtual complex, the receptor with a known ligand, and perform simulations with this. Then we remove the small molecule, and now we have a binding pocket that is ready for the next step.” As the GPCR models are not fully 3-D, Shacham uses software like CHARMm, from Accelrys, to “make it feel more comfortable.”
The entire modeling process for a specific protein takes about one month. The crunch processing period requires about 50-100 CPU for three days in a parallel processing manner. Shacham says that scientists are heavily involved: “It’s not a black box — we try to understand the sequence and the biology.” — K.D.
Sidebar: Molecular Evolution: PRX-00023
The optimization of EPIX’s first hit for anxiety and depression, resulting in PRX-00023, now in Phase III clinical trials, could be a case study for rational drug design. The company’s initial lead candidate (PRX-93009) was the result of a six-month screening process. But although it met potency criteria against 5-HT1A, it also tightly bound the alpha adrenergic receptor. (The same problem affects drugs such as the anti-anxiety drug Busperon, causing dizziness.)
Shacham describes the procedure EPIX used to modify 93009 to minimize the alpha-receptor binding, making it sound almost routine. She points to both sides of the molecule, as well as the linker region in the middle, as regions that can be sculpted and modified to optimize volume, conformation, and novelty. “The medicinal chemists say, ‘OK, we need to change these things. We need the computational guys to tell us what to synthesize.’”
Another major issue is selectivity. After docking the modified compounds to the target, they first check which molecules retain potency, which rules out certain scaffolds. She gestures to highlighted residues of the receptor that signal key interactions. Computational and medicinal chemists compare models of the target and secondary GPCRs, identifying regions or “volumes” in the former that will maximize selectivity. Although the binding pockets of the two receptors are very similar, Shacham points out, “here there is a volume [in the target] that doesn’t exist in the alpha adrenergic receptor, so we added something to the candidate molecule. It will still bind to H1A, but clashes with the alpha adrenergic receptor.”
After docking the new compounds using a platform called ICELR, they synthesized a couple of the most promising options. “Sometimes, the predictions are good, sometimes they’re wrong. It’s an iterative process,” says Shacham. The candidates are evaluated for a long list of properties including absorption, lipophilicity (Comp Log P), polar surface area, and more. They run additional models for blood brain barrier penetration and metabolism assessment.
After synthesizing 15 compounds, a new candidate showed minimal adrenergic activity, but had gained binding to HERG — a common cause of cardiovascular side effects and a “no go” compounds. So Shacham’s group had to generate another model, this time for the HERG channel. By docking freshly revised candidates to all three proteins, seeking to maintain potency while minimizing unwanted activity, the group identified PRX-00023 as a clean compound with no alpha activity or QT prolongation.
Because of the improved specificity, EPIX has not reached the maximum tolerated dose with this compound in the clinic. As an added bonus, “we also came up with another compound that still has a very good profile, but better pharmacokinetics, so we have a very nice backup,” says Shacham.
In all, the evaluation of 300 computationally profiled compounds, and 31 synthesized compounds, took less than six months. — K.D.
Becker, O.M. et al. 2004. G protein-coupled receptors: In silico drug discovery in 3D. PNAS 101, 11304-11309.
Becker, O.M. et al. 2006. An integrated in silico 3D model-driven discovery of a novel, potent, and selective amidosulfonamide 5-HT1A agonist (PRX-00023) for the treatment of anxiety and depression. J. Med. Chem. 49, 3116-3135.
Rosenbaum, D.M. et al. 2007. GPCR engineering yields high-resolution structural insights into b2-adrenergic receptor function. Science 318, 1266-1273.
This article appeared in Bio-IT World Magazine.
Subscriptions are free for qualifying individuals. Apply Today.