By Kevin Davies
June 14, 2005 | A quartet of speakers provided a vivid snapshot of the promise of in silico biology during the IT Solutions conference — although not without a few words of caution.
Jeremy Gunawardena, a member of the faculty at Harvard Medical School’s Department of Systems Biology, said that the challenge of building models of “virtual cells is to describe their behavior, not merely their structure.” Such models are dynamical systems, and cell behavior may be very different in populations. “I’m somewhat allergic to the term ‘predictive biology,’” said the eloquent Gunawardena. “Is this what we’re about? We’re telling better stories. We need to think in a more subtle way... Models must become modular and sharable to avoid becoming a cottage industry.”
To that end, Gunawardena cited systems biology markup language (SBML), a format that is being accepted by leading science journals. His colleague, former Millennium Pharmaceuticals programmer Aneil Mallawarpu, is developing b (pronounced “little bee”), which could be the equivalent of HTML for virtual cell modeling.
Entelos co-founder and chief technology officer Alex Bangs leads a group of programmers, engineers, and life scientists in creating “virtual patients” for a number of customers, including Pfizer, Johnson & Johnson, and Organon. In one case,
Entelos’ models of “theoretical people” predicted that even in the best-case scenario, a promising drug based on animal model data would barely cross the efficacy threshold in humans. The biopharma customer elected not to pursue the drug — results that were subsequently validated by clinical trial data. In another study, Entelos’ dosing data allowed Johnson & Johnson to reduce the number of patients by 60 percent.
Iya Khalil, co-founder of Gene Network Sciences (GNS), echoed the notion that computational biology strategies offer a promising route to systematically capturing the effect of a given drug on complex molecular networks. She discussed recent work with Novartis, in which GNS modeled data on hundreds of gene-gene, gene-drug, gene-pathology, and gene-expression data points to reveal the likely mechanism of toxicity of a promising drug candidate at high doses.
Keith Elliston, co-founder of Genstruct, argued: “Drugs act at the molecular level, whereas drug development is assessed at the animal level.” “Complexity is the cognitive barrier,” he said. “Parts must be described. Circuit diagrams won’t work because of crosstalk.” Elliston said that models must first describe systems, make predictions, describe perturbed biology, develop hypotheses — and do it all in real time. Five programs have been carried out with pharma partners to date in areas such as cancer and safety-toxicology. Genstruct’s Jack Pollard presented a human knowledge assembly model for diabetes that consists of 136,000 “causal assertions,” more than 30,000 genes and proteins, and 5,800 small molecules.