Oct. 16, 2006 | In 2000, a cadre of young physicists from Cornell, led by CEO Colin Hill and Iya Khalil, funded by angel investors and government grants, picked up the mechanistic modeling mantra and produced a remarkable computational model of the colon cancer cell. This feat quickly established start-up Gene Network Sciences' (GNS) ambitious science chops in the frothy young world of systems biology.
The idea was to combine physics and computational expertise to mine the relevant life sciences literature, do some experiments, and capture meaningful chunks of biology in silico. The resulting models would then be used to drive drug discovery and development. Profits would surely follow.
That was then...
Today, GNS is transforming itself, having learned the lessons all start-ups must to survive. It has evolved its technology platform and changed perspective. The big opportunity, GNS has decided, isn't turning known biology into algorithms and models. Instead, it is tackling the massive amount of unknown biology, much of it hidden in the wealth of accumulating omic data.
Colin Hill and crew now think they have the right tool to mine the treasure trove - inference technology that's not biased by past knowledge, that's powered by supercomputing capacity such as IBM's Blue Gene, and can chew through huge data sets.
"It's not that we don't like mechanistic models and that we're not big believers in them," says Hill. "We spent millions of dollars developing the mechanistic model for cancer and had our own wet lab generating data for it. We still do heavy duty mechanistic modeling."
But two issues dogged that approach. First, promising results trickled rather than gushed from mechanistic models, and they weren't always that novel. Second, the business model was a bear to scale - it takes a lot of time and effort (cash burn) to build decent models that often have narrow applicability.
During the past couple of years, GNS retooled its platform. It now has a front-end inference engine able to handle very large data sets and a back-end modeling and simulation engine that converts the inferences into mechanistic models. Hill says GNS now has a chance to start charting the 90-95 percent of biology he believes remains unknown and to reveal novel mechanisms of action and biomarkers in the process. Just don't try doing it on a PC.
"You have this astronomically large network topology space, especially when you're talking about 30,000 genes on an Affymetrix chip," says Hill. "It's not that we have [all] the computing power in the world to exhaustively go through that [exercise]. There are well-developed techniques from theoretical and statistical physics for sampling in a disciplined way."
"It's heavy, heavy-duty mathematics and computational physics," says Hill. "That's who we are. We're not hiring dozens of Ph.D. biologists to mine the literature. We did that. We know what we're good at and it's about fitting in this sort of large amount of data into mathematics, not just printing it in terms of overlaying it on top [of known pathways] and creating nice visuals. We're talking about something fundamentally different."
Recent work with Johnson & Johnson (J&J) and Janssen Pharmaceuticals illustrates the GNS approach. J&J/Jansen supplied omic data from treating cancer cell lines with a J&J compound. GNS "reverse engineered" models directly from the data (gene expression, proteomic, metabolomic) and conducted computational simulation to refine the models, identify modes of action, identify a variety of biomarkers (efficacy, prognostic, etc), and produce experimentally testable hypotheses.
Like other systems biology companies, GNS has worked with several pharmas, including Merck and Novartis. Likewise, its projects have tended to be modest proof-of-concept efforts. GNS also has an ongoing effort with the Mary Crowley Cancer Center (see "Gene Network Sciences Breaks New Ground," Bio-IT World.com) and an SBIR grant from NIH to further the company's cardiac modeling efforts.
Time for Change
But the proof of the new GNS concept will be in producing a profit. Founded just as the Internet boom was cratering, GNS "had to be creative" in attracting funds, raising some $12 million in grants and angel investments. Indeed, GNS has been criticized as a grant shop without a clear commercial vision.
"That has something to do with our academic heritage," acknowledges Hill, "and I have to take, at the end of the day, some of the responsibility for that. The grants served a very important function for GNS and allowed us to go after extremely ambitious science - science that other groups that had VCs banging on them every day weren't going to go after."
Now it's time for change. "We don't want to just be a grant shop. So we've been turning the corner over the last couple of years. We're now focusing this capability much more squarely on the marketplace."
To prepare for the commercial push, this year GNS hired a VP of business development, Zach Pitluk, and a seasoned CFO, Peter Rock. Both work out of the company's new Cambridge, MA, digs, bringing GNS much closer to potential customers. Hill says Cambridge is becoming the company's de facto headquarters, though GNS maintains its original Ithaca, NY, site. Staff size has grown to roughly 20.
"What I found was a company that had developed a basic tool set and approach and philosophy to allow it to navigate over the new wave technologies, whatever the data sets thrown at it," says Pitluk. "It was addressing the fundamental business problem which is this crisis in hypothesis testing. So you have a mismatch in scale between the data the customers are generating and their ability to, in a realistic way, take the data and test it in order to drive good discovery programs."
Pitluk's job is to sell pharma on GNS in a bigger way than earlier collaborations. With a Ph.D. in biophysics from Yale, research experience at Yale and Cellomics, and sales positions at Bristol-Myers Squibb and ISCO, Pitluk is already helping GNS become more customer-centric and escape preoccupation with its technology.
CFO Rock came from Dyax, where he spent five years and helped take the company public. "It doesn't mean we're doing an IPO or anything like that," says Hill quickly. "[We] want to have the operational and financial side of GNS running extremely smoothly. That part of operating a company, that's been done before, and should just not be an issue. You want to nail the easy things, 10 out of 10, because then you want to give your guys the best chance to hit the home run."
So far, no one has hit a home run in systems biology. Pathway software tool providers such as Ingenuity Systems, GeneGo, and Ariadne have all gained traction inside pharma, but they struggle with modest price points and a relatively small market size in terms of the number of available seats. The model makers and biosimulation technology providers such as Entelos, Genstruct, Optimata, Genomatica, and GNS have also had modest successes, but scaling those businesses has proven difficult.
Much as been written about these companies' desire to break out of the software license and fee-for-service business paradigms, but Hill insists the collaboration, fee-for-service model is a better route than trying to become a drug company.
Asked about the competitive landscape, Hill says, "I'm sure I'd give a different answer now than a year ago, or four or five years ago. Back when we started, Physiome was a lot closer in philosophy and approach to us than any of the current competitors. There's another company called Molecular Mining that in some ways also has some overlap."
At least the front end sounds a bit like Genstruct's approach (see "Predicting the Future of Systems Biology," Bio¥IT World September 2006) but Hill insists, "From what I can tell, they are not exploring any of the unknown biology. What they are trying to do is choose the most likely sets of interaction among the limited sphere of known biology."
And although GNS uses pathway tools sometimes, Hill says the problem with just overlaying microarray data on pathways is it often doesn't explain what the compound is doing. "I don't know what the exact percentage of known biology is, but it's small," says Hill. "The point is that probably a lot of the relevant biology for drug efficacy or drug toxicity has not been discovered. We've had companies come to us and say we plugged our stuff into one of these pathway databases and nothing showed up."
One of Hill's biggest concerns is demonstrating to customers sufficient value in the way GNS discovers and prioritizes drug mechanisms and biomarkers compared to current methods. But he says, "That doesn't scare me because clearly the current methods are inadequate."
Hill thinks both Entelos and Genstruct have "carved out a distinct enough space from us" so as not to be competitive. Nor does he run across Optimata very often. A bigger threat, says Pitluk, could come from the National Libraries Initiative and Molecular Libraries initiative, "where you have the opportunity to have academics in proximity with compounds and systems modeling and so forth. It will be interesting to see what comes out of that fermenter."
Pitluk also worries about pharma's impatience with new technologies. "They're still not succeeding, and so they continue to downsize silos. You're seeing a shift in resources kind of like the endgame is here. Lilly realized that when it shut down Sphinx (Pharmaceuticals). You'll see more of that kind of abandoning the high-throughput screening paradigm. Some companies will never do that, like Merck with a giant compound library and total automation, but many other companies are putting the high-throughput parts of their paradigm under the gun."
Hill won't divulge the number of GNS customers, but hopes news of the J&J collaboration will spread the word. A deal with IBM for on-demand access to its Blue Gene supercomputer should easily accommodate fluctuating high-capacity computing requirements. And to give potential clients a feel for the technology, GNS conducts "cooking shows" in its Cambridge office.
No doubt it would be interesting to sit in and see what's cooking at GNS. "It's really the best way to understand what we do, to actually see it," says Hill.
Hill, C. & Khalil, I. "Systems biology for cancer" Current Opinion in Oncology, 44-48 (2005).
Hill, C. et al. "Individualised cancer therapeutics: dream or reality?" Expert Opinion on Therapeutic Targets 9, 1189-1201 (2005).
Sidebar: The New Biz Dev Guy's Pitch
Zach Pitluk's job is to help turn GNS from a grant shop into a viable commercial enterprise.
Where's the Money?
"I came up with 17 different nodes in preclinical where with different omic data or combinations of functional and omic data, you could create a model that would help refine the discovery process. You're beginning to see significant investments at Novartis and Merck, etc., in terms of their systems biology capability. They appreciate that it's no longer going to be acceptable to take a principal component analysis approach to these data sets."
Safety: Why Compound-centricity Matters
"Our approach will be extremely helpful in safety in particular. Whereas the efficacy guys are kind of on the railroad track, and anything that contributes to them getting the train to the depot is cool. The safety guys basically want to know what else is happening. I threw out of this compound-centric view. You're really looking from a compound's perspective. Everyone else is starting from this biological knowledge that we believe we have. We're saying, "Well, no. Here's the compound and the data that support this association, and you end up with a network. It's like you're inside the compound driving through the cell."
Losing a Champion is Part of the Game
"We also suffered a champion loss [key advocate leaves a client company]. So at any technology company, and I've worked in enough of them, when your champion leaves, that really cuts a whole in you."
Is GNS a Virtual Screening Shop?
"In a sense, that's what we're going to do. We really offer the potential to unlock this treasure chest that every company with all these gene chips in there has. And they continue to use them. I mean Affymetrix is not going out of business. At the same time they're getting one percent of the value of data that they could get."
On the GNS Horizon
"We're not limited. The platform is domain-independent. I think that you will see our business model change over the next year, I'm not going to elaborate further on it."
Sidebar: GNS in Brief
CEO Colin Hill says tremendous impact need not require huge size.
How Big Can GNS Be?
"We're never going to be a 200-person entity. I actually think we'll double our size in a bit, you know 40-to-50 people. I see us having tremendous impact, partly because we leverage supercomputers. If you do things the way we used to do them a lot, just modeling pathways the Entelos way, you never have enough people because there's too much to cover."
Managing the Data Mountain
"Integrating data at this level means it's about supercomputing. So we have a relationship with IBM. We have a number of processors in-house, but we know we're never going to set up the infrastructure or ever want to bring evolving supercomputer capacity completely in-house. It doesn't make sense. And the computing world has been moving toward treating computing as a utility - something you buy like electricity. That is a key part of scale. If you're going to be serious about integrating data into models, I don't mean just overlaying data on models or simulating a single network picture of something which you can do on a laptop easily, then you need serious computing power and serious quantities of data, that need to be quantitative, and need to have error bars."
What Systems Biology Customers Want to Learn
"People really want from us a somewhat unbiased approach, not what did one particular modeler or biologist think was important to their drug. An unbiased approach, based on their data, this compound-centric view of the world that gets at the end of the day as Zach said [to answers] that have different levels of confidence. Sometimes we can't answer the question and we need to be able to say, "No, there was not enough data to really distinguish between these two different hypotheses."
What's the Barrier to Entry?
"We're not trying to have the lock on Bayesian inference algorithms. What we build is a specialized industrial scale platform for doing this inference to get at those kinds of answers - not just general models, but those kinds of answers that end up requiring a lot more work. But doing that in a very high-throughput fashion ends up being a lot of effort and energy that go into all the different parts. Being able to have your software code parallelized to work on the scale of thousands of processors is another barrier to entry. And we're not standing still. We know that this approach has major value that people are starting to really grab onto, and so there is going to be competition - and that's the way it should be." -- J.R.
Email John Russell.
Subscribe to Bio-IT World magazine.