Just over a year ago, Fujitsu launched its BioSciences group in the Boston area to sell its hardware and software to the life science market. Now, the company is adding a services component to the mix with the rollout of a multi-step computational method for ADMET prediction.
The launch of the services offering sets the firm apart from other IT firms in the life science market, who tend to build their marketing strategies around hardware sales. Those IT firms that do provide life science services, such as IBM and HP, have so far set their sights on early discovery or the clinical arena. None of Fujitsu's competitors in the IT market provides a service focused specifically on computational predictive ADMET (absorption, distribution, metabolism, excretion, toxicity).
Fujitsu believes that the offering will also distinguish it from software and database firms who are targeting the ADMET market. According to Ian Welsford, manager of application science in Fujitsu Computer Systems' BioSciences Group, the company's technique incorporates a number of methods that are similar to those available from other vendors, but combines them in a workflow that is not currently available from any single source.
"There's a lot of activity in the ADME/tox space right now, but in our analysis, we didn't see any that completely solved all the problems for the researchers."
The company is entering an increasingly crowded market, however. A number of informatics vendors have launched new products targeting the predictive ADMET market this year. Most recently, Bio-Rad released eight new databases containing ADMET properties as part of its KnowItAll informatics system in March.
In February, Inpharmatica launched its Admensa Interactive platform, which includes a set of ADME models and compound-prioritization tools. The company has so far announced licensing agreements for the system with Serono and Daiichi Pharmaceutical.
GeneGo, meanwhile, launched a pathways-based ADMET informatics system called MetaDrug in January.
"There's a lot of activity in the ADME/tox space right now," Welsford told BioInform, "but in our analysis, we didn't see any that completely solved all the problems for the researchers." The company's multi-step services approach is consistent with the "ethos for Fujitsu's global strategy," he said. "We don't just provide one-off things. We're looking to build relationships and solve problems for our customers."
The offering is based on a workflow that combines a number of Fujitsu's predictive ADMET tools, including structure-activity-relationship modeling, database mining, 3D homology modeling, virtual screening, and statistical analysis. Welsford said that some of these steps are already available as standalone tools from other vendors, but that Fujitsu's combined approach offers a predictive capability that surpasses that of single-point solutions.
Fujitsu also developed several new informatics tools, including a support-vector-machine-based database-classification tool and a ligand-interaction scoring method, to tie the workflow together, Welsford said.
The workflow covers eight separate steps, from building the initial metabolic models, to identifying and modeling active sites, and then docking the set of predicted metabolites against 3D models of the cytochrome p450 metabolism enzyme. It begins with two methods developed by Fujitsu subsidiary FQS Poland: ADMEWorks, a SAR modeling program; and BioFrontier P450, a curated database of CYP 450 interactions. "Each of these products had a piece of the puzzle," Welsford said, "but in and of themselves, they didn't complete the entire process."
The Fujitsu researchers use these two tools to derive a "large surrogate database" of predicted interactive metabolites that are then docked against 3D CYP structures on the company's BioServer grid computing platform. Very few 3D CYP structures are publicly available in the Protein Data Bank, Welsford said, so the company had to use homology modeling methods to generate the 3D structures for those molecules before the virtual screening step.
Virtual screening "adds a strong and useful reality check" to the process, Welsford said, because it indicates the degree to which a specific compound interacts with particular binding sites for different CYP models. Because of this, he said, Fujitsu found that it had to account for "flex/flex" docking, in which both the protein and the ligand are considered to be flexible — an approach that requires a great deal more computational power than virtual screening with rigid molecules.
In an initial proof-of-concept collaboration with an undisclosed drug-discovery partner, Fujitsu found that its predicted rankings offered "very good correlation" with the company's proprietary toxicity data, Welsford said.
Standard linear modeling approaches for predicting LD50 scores, for example, tend to have an average R2 value, which measures the fit of the model, of around 0.64, Welsford explained via e-mail. "In contrast, when one incorporates a knowledgebase coupled with derivative docking into CYP models, the fit goes up significantly, above 90 percent for an R2 (0.946 R2)," he said.
Fujitsu plans to publish a paper describing the method, but Welsford did not disclose further details on the company's publishing plans.
Hot Market, Complicated Problem
Recent toxicity concerns spurred by high-profile drug recalls in the past year have driven the pharmaceutical industry to seek new ways to assess the efficacy and toxicity of compounds earlier in the drug-discovery process. The "Critical Path" whitepaper that the US Food and Drug Administration released last March estimated that a 10-percent improvement in the prediction of drug failures could lead to a $100-million reduction in development costs per drug.
But methods for improving ADMET prediction have yet to prove their worth, according to Richard Okita, program director in the Division of Pharmacology, Physiology, and Biological Chemistry at the National Institute of General Medical Sciences. Predictive ADMET is a "hot area right now," Okita said, but the problem is a "very complicated one to solve."
In addition, he said, many firms developing predictive methods are finding it difficult to validate their technology because "it's hard to get [toxicity] data to test against." Since most toxicity information is generated by pharmaceutical firms, who are historically stingy with their data, "there hasn't been a lot of sharing" in the field, he said.
In an effort to advance the state of the art a bit, the NIH has launched a Roadmap initiative called "Novel Preclinical Tools for Predictive ADME-Toxicology" under which it plans to fund four to seven projects this year with a total of $2 million. Okita, the program director for the initiative, said that response to the RFA, issued in late October [BioInform 11-08-04], was "far more than we expected." NIGMS has received about 50 applications, "but we were expecting about 30," he said. Around 20 percent of the applications were for in silico methods, he estimated.
Although the pharmaceutical industry has been struggling with better ways to assess efficacy and toxicity for decades, Okita said that the recent influx of new data from genomics, proteomics, and gene expression analysis is likely to revitalize the field of predictive ADMET. "Some people might disagree, but I still look at this as a very young field," he said.
Welsford said that Fujitsu is still "in the process of completing market analysis" for its predictive ADMET service, but the company believes that it's "quite sizable in terms of its ability to streamline the [drug development] process." Some informatics vendors have already "captured market share in particular points in the process," he admitted, but Fujitsu believes that "there is still an opportunity for a more comprehensive solution."
As for market demand, there is still some resistance to computational approaches from wet lab researchers who demand proof that in silico ADMET works, he said. "We need to validate these types of approaches with real-world chemicals, and without that validation, it's hard to get traction," he said.
Nevertheless, the company is optimistic. Welsford said that it is currently expanding the docking step of the process to include 3D models of other proteins that play a key role in metabolism in order to better predict off-target effects.
— Bernadette Toner ([email protected])