By Arielle EmmettDec. 10, 2002 | WITH THE FASTEST known Beowulf supercomputing cluster in the civilian world, a venture capital war chest of $88 million, and a drug discovery algorithm company officials say will accurately forecast the structure and binding affinities of novel drugs in silico, Locus Discovery Inc. has the makings of a breakaway IT front-runner.
Founded three years ago as a spinoff of Sarnoff Corp., the Blue Bell, Pa.-based company boasts a drug discovery
|Software developed by Sarnoff Corp.'s Frank Guarnieri excels at identifying potential drug candidates
algorithm implemented in a Linux super- cluster that even competitors say could raise the bar in computational drug design, transforming it into predictive science. Locus officials believe the algorithm — implemented as a C++ program performing up to 2 trillion calculations per second (see "The Algorithm," right) — could shave off years from conventional drug discovery cycles, replacing structural screening methods with a computationally intensive approach.
In the past 18 months, Locus says it has produced several small- molecule drug leads, one reportedly blocking HIV entry. If the drugs prove clinically effective, the algorithm could change scientists' appreciation of how IT can be harnessed to predict the activity and binding of small-molecule ligands to their protein targets — without prior empirical knowledge of their location.
To date, few companies — if any — have claimed the ability to predict the character and location of novel binding sites based on bioinformatics alone. But that's just what Locus says it's doing. "For our therapeutic programs, the number of [virtual] molecules we're generating now from the algorithm is colossally large — up to 1020 for a given protein drug target," says Chief Scientific Officer William Moore Jr. "The challenge now is to sift through the data to decide which compounds to make."
Moore explains that the algorithm's uniqueness and power stem from the computation of not only all possible drug compounds that can bind to an active site but also the location of those sites based on quantum and statistical mechanical analyses of small-molecule (chemical fragment) affinities. "The number of chemical fragments the supercomputing cluster can utilize is extremely large. We presently start a protein project with more than 150 fragments and have a proprietary database of more than 10,000 fragments," Moore says. "The algorithm can also predict binding affinities of the novel small compounds it generates in rank order."
Since its launch, Locus has mounted an aggressive campaign of drug discovery rhetoric, claiming that its algorithm can compress years of structural screening and serendipitous lab adventures into just weeks of computational blitzkrieg.
The company's first president and CEO, Nicholas Landekic, a former business development vice president at Guilford Pharmaceuticals Inc., raised $88 million in several rounds of venture financing. "We never had to persuade any investors to come with us; we'd done our due diligence," says John H. Park, portfolio manager at Liberty Wanger Asset Management LP, who led Locus' third round of financing.
Park says Locus back-tested known protein receptors with its algorithm to predict novel compounds, identifying structures that were actually on the market. "We knew the compounds the computer is capable of generating are realistic," Park says. "But we weren't seduced by the cache of the supercomputer — that was more of a means to an end. What was enticing was the technology and computer algorithms that would generate thousands and thousands of drug compounds within weeks rather than years."
In March 2002, however, Landekic left the company for undisclosed reasons. Locus Chief Financial Officer Nancy Jean Barnabei, who worked with Landekic raising seed money, says that Landekic's departure was simply "a mutual decision between the board and Nick ... I'd call it the CEO shuffle." According to a Locus scientific advisor, who asked to remain anonymous, Landekic's exit likely arose because of a difference of opinion regarding the speed with which the company should push for a commercial product. Currently involved in another startup called PolyMedix Inc., Landekic says he is bound by a nondisclosure agreement, adding diplomatically, "I hope Locus has the most phenomenal success."
"Locus is an amazing business story," observes Edward Abrahams, former director of the Pennsylvania Biotechnology Association. "But it is also the biggest business mystery of this sector in 2002. We just do not know why the CEO left after raising nearly $100 million in a difficult market, but before proof of principle had been independently verified or an exit strategy clearly developed."
Publish or Perish?
Indeed, Locus has played many of its scientific and IT cards close to its chest. The company has not disclosed its potential clients or pharmaceutical partners. Moreover, it has yet to publish in a peer-review journal, although Moore says some articles are in the works. Even Landekic says that's normal. "Many companies in their early stages do not publish their results in peer-review journals," he says, arguing that informatics companies in particular must protect their algorithms and intellectual property.
Like academic institutions or tiny startup ventures, cheminformatics companies that publish their results too soon risk competitive infringements or anonymity. "If [our] algorithm was published instead of commercialized, you would never be interviewing me because it would be buried in some scientific journal," asserts Frank Guarnieri, the Sarnoff Corp. scientist who founded Locus Discovery in the late 1990s, and who now is the chief scientific advisor to the company.
"Creating a company and convincing [venture capitalists] to invest approximately $100 million requires infinitely more than the data needed for a publication. By setting the milestones to actually synthesize real compounds and showing biological activity on three different proteins, Locus can attract first-rate pharma pros and top-tier VCs to the endeavor," Guarnieri says.
Locus' position typifies a dilemma many bioinformatics companies now face: Either protect critical IT and drug discovery resources or risk credibility problems in the scientific community. "Locus is trying to keep the algorithm secret at this point because patents on algorithms are difficult to enforce," says Sandor Vajda, a professor in the department of biomedical engineering at Boston University. Vajda's laboratory filed a preliminary patent on a similar program but, he says, "we decided not to follow it up because of the difficulty of enforcing the patent. Someone can slightly modify it and go around it."
Vajda is a supporter of the Locus results. "We do similar calculations, and their binding free- energy evaluations seem to be more sophisticated than ours," he says. His own studies on the computational identification of protein binding sites, published in the Proceedings of the National Academy of Sciences in April 2002, parallel Locus' algorithmic approach.
Vajda also cites the solvent mapping work of Dagmar Ringe, a renowned biochemist at Brandeis University and a member of the Locus scientific advisory board. Ringe has developed methods to soak protein crystals in different solvents, which bind primarily at the active sites, and then determine the structure of the protein. Ringe's experimental results on porcine pancreatic elastase, for example, agree with data on a large number of drug targets presented by Matthew Clark, Locus' director of scientific computation, at the American Chemical Society meeting in August.
Ringe says the objective of this work is to find methods for determining how adding "synthetic bits and pieces" can improve the ligand-binding specificity at a particular site. Whether the ligand is a drug agonist (potentiator) or antagonist, the ability to predict ligand-binding behavior and to find active sites is invaluable, she says. "We'd like to find ligands that don't look like substrates. Now, with gene products being identified as targets for intervention, in many cases we don't even know where the active site is — we don't even know where to start!"
While intensive laboratory work can identify 3-D crystalline protein structures and the sites where solvents "cluster" or "stick" (indicating active sites), Ringe says a streamlined computational method is highly desirable. "To do everything experimentally is a difficult proposition; it's easier to do it with computers," she says.
"What Locus has done is to develop a computational method that determines substructures of small organic molecules bound to a protein surface," Ringe explains. The attraction of "small" means the possibility of developing oral rather than injectable medications. "If the ligands [acting as solvents] cluster, as they do in their computational method, as they do in our experimental one, then we get compounds which will have a particular specificity for the active site."
Computational synthesis can go one step further, she argues: Finding separate active sites where ligands cluster can result in new and more potent molecular configurations. This is the algorithmic way to "chemically connect the dots" between high-affinity ligands, she says. "This makes a compound that is specific and tightly bound."
Algorithmic attempts at "computing" small-molecule solutions are hardly new: Harvard chemist Martin Karplus' MCSS (multiple copy simultaneous search) program in 1991, and Peter Goodford's 1985 GRID program were pioneering achievements in the area. Used with some success to identify new chemical groups that improve ligand- binding specificity or activity, these programs required the location of the active site. By contrast, Ringe says Locus "expands on experimental verifications by starting with a naked protein, a naked active site, and calculating the pieces that fit into that active site, and then putting those together to form a large [and, in principal,] more effective compound."
David U'Prichard, CEO of 3-Dimensional Pharmaceuticals Inc., a cheminformatics company in Exton, Pa., that also develops libraries of virtual compounds, says the Locus algorithmic approach has merit. "It's very interesting and looks like an interesting piece of IT," he says. U'Prichard believes the preliminary claims are enticing. "Locus says it can take a 3-D structure of a protein and identify the biologically relevant binding sites for small molecules on that protein. That is quite a feat. [But] a lot of proteins have more than one active binding site, and I'm not clear how they will handle that in the software."
Vajda offers a final caveat: "I'm not sure that these high-affinity ligands will be meaningful as drugs. A high-affinity ligand is very far from a drug, and I'm not 100 percent sure that concatenating these fragments means they will be able to find molecules that are sufficiently druglike."
Ultimately, the merit of the Locus technology will be determined in the marketplace. In the meantime, Ringe has some sympathy for the company's publication predicament. Disclosure of the computer program code could allow someone to write the same program, tantamount to stealing a trade secret. But she says the results will ultimately be validated (or disproved) in a traditional, peer-review setting, because eventually, Locus will need to disclose enough drug discovery data and computational details to substantiate its claims.
"It's called proof of principle," she says. "This includes publishing an actual study in which a compound is identified by the computational method, then synthesizing it and showing its efficacy."
Arielle Emmett is a medical and IT writer based in Wallingford, Pa., and the editor of Wireless Data for the Enterprise: Making Sense of Wireless Business (McGraw-Hill, November 2001). She can be reached at firstname.lastname@example.org.