Gene Network Sciences is partnering with Novartis on a project that extends the capabilities of its inference-modeling technology into several new areas, GNS said last week.
In one first for the company, CEO Colin Hill said the project is "outside of oncology" — where GNS has focused its efforts since it was founded in 2000. Hill didn't disclose the disease area that the companies are studying, but said that working in an area where "we didn't really have a starting foundation" has been "a good test" for the predictive ability of the GNS platform. "This was really about utilizing our methods for discovering new information," he said.
The project, which aims to understand the molecular mechanisms underlying toxicity of compounds for clinical trials, is also driving the company's modeling method downstream, well beyond its historical application in preclinical research. Hill said that this downstream progression will be a crucial requirement for the entire field of computational systems biology as it tries to gain a foothold in big pharma.
"At the end of the day, it's about what's being predicted and understood in humans — in clinical trials — that matters," Hill said. "I think a lot of these technologies have been applied and validated on preclinical samples and studies, but I think … as these things really start to make their way into clinical trials, that's when we're going to start to see some big home runs with this kind of approach."
GNS hasn't formally announced the collaboration with Novartis, but the companies recently co-authored a paper in the March 21 issue of FEBS Letters, and last week presented a talk together at Cambridge Healthtech Institute's Beyond Genome conference in San Francisco. The collaboration began around a year ago, Hill said.
Novartis researchers involved in the collaboration were not available for comment before press time, but the company has proven in the past to be open to computational systems biology. In April, Compugen announced that it was working with Novartis on a systems biology project under which Compugen would model signal transduction pathways using transcription factor binding information, expression profiles, and other data. [BioInform 04-25-05].
The Novartis project is the second collaboration with a major pharmaceutical company that GNS has disclosed. In March, the company announced an agreement with Johnson & Johnson to use its modeling technology in a preclinical oncology research project [BioInform 03-14-05].
Hill said that although the Novartis and the J&J projects differ in scope, they have enough in common to give inference modeling a try. "Companies need to have a better understanding of what their compounds are doing, what is inside this black box that drives whether a particular compound works really well or it fails — from either an efficacy standpoint or a safety standpoint," he said. "No matter who you are or what you're doing, that is the goal — discovering new biology around compound performance, and not throwing away previous knowledge."
One way that GNS and Novartis addressed this challenge in their collaboration was to combine inference modeling and mechanistic modeling in a two-pronged approach. As described in the FEBS Letters paper, the companies first used inference modeling to "reverse-engineer" regulatory models using high-throughput data from multiple platforms.
"… as these things really start to make their way into clinical trials, that's when we're going to start to see some big home runs with this kind of approach."
The researchers used two different animal models — monkey and rat — to study the toxicity of a "compound of current clinical interest." Tissue microarray experiments were conducted for a high-compound dose, a low-compound dose, a high dose of the compound's irreversibly active analogue, a high dose of the compound's irreversibly inactive analogue, and a control. Time series gene expression data from a separate animal study was also used to build the model, along with additional information, such as immunohistochemistry staining and RT-PCR data, tissue pathology scores, and animal physiology and blood biochemistry readouts.
"The modeling engine inferred, solely based on the various datasets, connections of expressed genes to additional molecular and physiological measurements," the authors wrote. The model "allowed scientists to examine classes of biological pathways that were affected by the drug in specific tissues, as well as to discern how such pathway-based effects were influenced by low vs. high dose, or by the compound vs. fully active or inactive analogues."
"We demonstrated that the method was able to differentiate between drug conditions that drive toxicity," Hill said. "This is critical in lead compound selection."
The modeling approach was "generally consistent" with results using "standard statistical methods for filtering and clustering," the authors wrote, and offered the further advantage of enabling "additional hypothesis testing via the study of specific recovered biological relationships, whereby expressed genes of unknown functions were linked to entities of known functions or known biological pathways."
One drawback of the approach, the authors noted, was its "significant number of false positive (also false negative) connections."
Hill said that this problem is "an aspect of high-throughput biology in general, and certainly in computation applied to high-throughput biology" — especially when using sparse data sets. "One is exploring such an enormous space of possible network topologies, so when we're actually trying to discover pathways and regulatory networks that drive compound performance from these high-throughput data sets, the number of possible network topologies grows astronomically large with the number of genes or proteins that you measure," he said.
One way the company is addressing this problem is by improving its computational methods to increase coverage of the network topology space — an approach that requires ever-more computational power. To meet this need, Hill said that GNS recently purchased a new Apple cluster.
In addition, Hill said, the company has recently "balanced" its approach "to use a combination of inference modeling to discover new biology, as well as using models and simulations of known pathways to provide kind of a starting point, a foundation, for the search through network space."
The FEBS Letters paper notes that this combined approach of inference modeling and mechanistic simulations of known pathways offers a number of advantages. Inference models, the authors wrote, "generally give a global view of biological drug effects, but do not reveal detailed biological mechanisms," while mechanistic simulations are more limited in scope, but "provide insights into the dynamical response of the system."
In the two-pronged approach that the authors recommend, key interactions predicted by the inference framework "can be used to inform the mechanistic simulations of the pathways underlying the disease state and drug interaction."
However, the authors note, linking these two approaches into an automated framework "will be challenging," especially for large-scale biological models that will require "both algorithmic advances (e.g., in global optimization) and computational power." Another limitation, they note, "will be acquiring the large-scale quantitative proteomic measurements required to constrain mechanistic models that integrate protein signal transduction with gene expression and metabolic networks."
Hill said that finding the right mix between the two approaches is a balancing act that differs from project to project. The goal, he said, is a platform "that can discover new pathway relationships from the high-throughput data, and which essentially connects the compound with changes in disease phenotype, and to do that while incorporating knowledge of existing pathways.
"Even before talking about what the platform is, or what the computational approach is, that is what [pharmaceutical companies] need to accomplish," he said.
— Bernadette Toner ([email protected])