A team of US researchers used a bioinformatics platform to screen roughly 6,000 human proteins to identify a number of off-target interactions that could explain why an HIV protease inhibitor has been found to have efficacy as a cancer therapy.
In a paper published this week in the online edition of PLoS Computational Biology, the scientists, from the University of California, San Diego, and Hunter College of the City University of New York, proposed that the anti-cancer effects of the drug, Pfizer's Viracept, stem from weak binding to multiple protein kinases upstream of the PI3K/AKT signaling pathway.
The drug, known generically as nelfinavir, is currently in a number of cancer clinical trials sponsored by the National Cancer Institute. The researchers are not association with the studies at this point, but PLoS CB lead author Philip Bourne said he plans to wait to see what kind of reaction the paper generates before possibly contacting NCI about further investigating the results.
The results of the PLoS CB study also suggest that multiple weak drug-target interactions could have a larger therapeutic effect than generally thought, particularly in the case of complex diseases like cancer, said Bourne, who is professor at UCSD's Skaggs School of Pharmacy and Pharmaceutical Sciences.
"The common wisdom has been to make something against one target – one drug, one target, one disease," said Bourne, who is also the associate director of the Protein Data Bank and the editor-in-chief of PLoS CB. "But clearly with something like cancer, given how much energy has been put into [studying it], and what we know about the disease, the idea that you have to have a sort of broad-scale approach to dealing with it just seems intuitively quite attractive."
To investigate the effects of drugs on such a broad scale, Bourne developed a so-called "structural proteome-wide off-target pipeline" — a computational technique for screening compounds against large numbers of proteins to determine their likely binding affinities.
The process is "like reverse-engineering drug discovery," he told ProteoMonitor. "Instead of taking one receptor and a large library of small molecules and doing high-throughput screening to find the right small molecule, we're taking a known small molecule and a known protein and asking what other proteins it binds to."
Starting with an initial drug-protein pair, the researchers screen a structural proteomic database for proteins with binding sites similar to the initial target's. Upon finding potential matches, they identify the intrinsic ligands associated with these similar proteins and perform "cross-docking studies" — statistical analyses that define the likelihood that the ligand of one protein will fit in the binding site of the other protein. If the match looks promising in silico, they then investigate the compound-protein binding in vitro.
In the case of nelfinavir, the researchers identified off-target binding to a number of receptor tyrosine kinases involved in upstream regulation of the PI3K/AKT pathway. Their modeling also predicted that the drug would inhibit insulin-like growth factor 1 receptor, which could explain the insulin resistance that has been observed as a side effect of nelfinavir, the authors noted.
According to Bourne, his team "regard[s] everything we do in the computer as a hypothesis," noting the difficulties inherent in translating results of an in silico model to an actual biological system.
"The attrition rate is huge," he said. "That's one of the main reasons why [pharmaceutical companies] don't have more drugs on the market."
He gave the example of work his team did looking at repositioning Valeant Pharmaceuticals' Parkinson's disease drug Tasmar, known generically as tolcapone, and Novartis AG's tuberculosis therapy Comtan, known generically as entacapone.
Based on his lab's in silico modeling, the drugs looked like good agents for targeting the TB protein inhA. In vivo, however, the level of the drugs needed to effectively inhibit TB turned out to be toxic.
"So you might have something that looks good, but when you're actually dealing with absorption and toxicity and everything else, it's just highly problematic," Bourne said.
The method is also limited by its dependence on protein structural data, which Bourne's lab draws from the Protein Data Bank and the ModBase catalogue of comparative protein structures, which is maintained by Andrej Sali at UC San Francisco.
"The number of known three-dimensional [protein] structures in humans is only around 1,200 out of about 20,000 to 25,000 [proteins]," he said. "With homology modeling we can get that up to around 6,000 to 7,000, so we can end up having around 30 percent of the human proteome."
Another limitation, Bourne noted, is the fact that relatively few of the proteins with well-characterized structures are membrane proteins, which comprise the majority of drug targets.
"I've talked about this [platform] quite a bit, and [Nobel Laureate] Roger Tsien, who is in my department, has basically said, 'This is all great, Phil, but until you show me a soup-to-nuts example, I'm not really sold,'" Bourne said. "What he's saying is [we need a case] where something goes into trials as a direct result of [the bioinformatic] work we've done.
"I can't claim that yet, but that's what we're doing all this work to try to do, and that would be a big success from my point of view," he added.
Bourne said he has taken his lab's work to several large pharmaceutical companies, but "basically we haven't gotten very far," he said. Most of his collaborations with industry have been with smaller biotech firms, and he is currently performing pre-clinical trials with several of them.
A small biotech company "might be based around one lead molecule, and so they're very interested in knowing in the early stages before they spend a lot of money and time on clinical trials whether in fact there are significant off-targets," he said. "If we can predict [off-targets], then they can test for them. We do a very broad-based targeting. Much more broad-based than you would do experimentally."
In addition, these collaborations provide Bourne's team with feedback on the platform, which they can use to enhance its performance, he said. For instance, the team is currently applying machine-learning techniques to improve the system's docking scores.
The researchers are also trying to develop better methods to model membrane proteins so they can include more of them in their screens.
"There's definitely a feedback loop [from the collaborations] in terms of computational development," he said.
Currently his lab runs screens "basically gratis," typically charging collaborators just to cover the cost of the work, Bourne said. He said that he has considered commercializing the service, but at the moment he is focused on an online scientific media firm he's launched called SciVee.
He's co-founded three other companies in the past, including the bioinformatics firm Protein Vision, which was acquired in 2002 by Quorex Pharmaceuticals.
"I can't do two companies at once and be a university professor," he said. "It just doesn't scale."
Have topics you'd like to see covered in ProteoMonitor? Contact the editor at abonislawski [at] genomeweb [.] com.