By Vivien Marx
CAMBRIDGE, Mass. — A Biogen Idec official said this week that the company is seeing encouraging results in the use of network modeling to gain insights into therapies for autoimmune diseases such as rheumatoid arthritis.
Speaking at the Network Biology 2.0 conference here this week, John Carulli, director of genetics and genomics at Biogen Idec, said that the company has partnered with Gene Network Sciences to look for new drug targets for rheumatoid arthritis patients who do not respond to tumor necrosis factor antagonists.
Anti-TNFs are only effective for about 50 percent of the population and the concept is to put networks behind the diagnostic algorithms physicians use, Carulli said.
The company worked with GNS to integrate and analyze genome-wide SNP data, expression data, and clinical data from three time points in more than 70 patients — a process that ultimately identified several targets, including one "that had not been on the radar screen," he said.
The modeling highlighted "known pathways and also some unexpected ones," and the genes that it identified "make sense," he said. "Building a model and perturbing the expression of every gene has led to the identification of pathways relevant in this disease," as well as new drug targets and mechanisms.
For example, the network identified a pathway that is a target for the arthritis drug Orencia (abatasept), a selective T-cell co-stimulation modulator that is already marketed as an alternative for patients for whom anti-TNF therapy has failed.
The modeling process was a "quick and promising way to identify targets" from the clinical data, he said.
Carulli added that he would "continue to try" the network modeling approach, but noted that the in silico method won't entirely replace empirical biology.
Carulli told BioInform that the company views the modeling project as a pilot study.
Building on GWAS
The partnership with GNS built upon a genome-wide association study involving 89 rheumatoid arthritis patients that the company published in Molecular Medicine in 2008 that identified several candidate SNPs associated with response to anti-TNF treatment.
"We had really been through the data a lot," Carulli said, but there were "unanswered questions." For example, he and his colleagues "found ourselves staring at gene lists from an expression study for the umpteenth time in the past 20 years and that wasn't satisfying."
The Biogen Idec researchers felt that if they could determine whether a gene was up- or down-regulated, "maybe we'd be staring at a different kind of gene list, one that showed impact," Carulli said.
First they needed to reduce the data size by choosing the most informative data. They selected the "best 6,000 hits" genes from GWAS data, as well as 5,000 "informative" transcripts, and broke down the clinical data into components for the modeling process.
The network modeling led to "one trillion building blocks" of information and the ability to look at clinical outcomes — for example, the degree of joint swelling in arthritis patients. In order to see "what happens to clinical data" if gene expression changes, they built "consensus models" with more than 1,000 networks in each model, pertaining to untreated and treated patients at different time points.
The models were "pressure-tested" to ensure that "they were sufficient to show us the models had predictive value," Carulli said
Once the models were completed, the researchers were able to use them to perturb every gene in silico and see whether it changed clinical outcome.
In response to a question from the audience, Carulli said that his team has worked with simpler modeling approaches such as expression quantitative trait locus mapping, but found that it delivers associations but not a model that can be queried and predict phenotypes.
GNS announced that it was working with Biogen Idec to apply its Reverse Engineering and Forward Simulation, or REFS, platform to drug discovery in 2008. Iya Khalil, executive vice president and co-founder of GNS, said that the rheumatoid arthritis project was the first in which the company's scientists wanted to "pull mechanisms from the data," and GNS helped to ask "what-if questions" of the data.
"They got to see that we were able to deliver and get them new hypotheses in a data-driven way," she said. "Without having put in any prior knowledge into the machine learning," the team was able to pull out a clinically validated mechanism.
Khalil noted that modeling projects need to make sure that "the experimental design matches the question," and then design a high-throughput assay to measure as much as possible. Nevertheless, the process is "still a fishing expedition — a data-driven, hypothesis-free method," she said.
A number of pharma researchers at the conference said the results from the Biogen Idec modeling project are encouraging. Some questioned, however, whether elaborate modeling was really necessary to achieve the results.
Lihue Yu from AstraZeneca's Cancer Discovery Group told BioInform that she views network modeling as a promising approach for inferring the interplay of pathways and disease. She noted, however, that such projects need to be well-defined in order to get the desired results.