Sept 15, 2005 | Researchers at the BioSciences Group of Fujitsu Computer Systems Corporation, working with a commercial partner, have developed a new technique for generating enhanced in silico ADME/Tox predictions.
The work started out as part pure research, part a project to demonstrate the application of high-performance computing to a life sciences problem, and part a collaboration to help a customer solve a real-world problem.
"The goal was to conduct a study and prove the capabilities of Fujitsu in this space and show we could produce a commercial offering in the ADME/Tox [arena]," says Ian Welsford, manager of application science at Fujitsu.
The effort took a look at in silico ADME/Tox by exploring the use of some computational techniques that previously might not have been considered because of their huge computing requirements.
The idea was to use the new techniques to get around some of the shortcomings of the commonly used approaches. For example, many approaches use database or quantitative structure activity relationship (QSAR) modeling techniques. The database approaches often rely heavily on text-mining searches of literature, looking for entities with similar pathways. The limitation to that approach is the in silico predictions are only as good as the curation process.
Welsford notes that the shortcomings of QSAR approaches are that they are molecule-centric and often poorly represent the underlying biology. Additionally, the QSAR mathematical techniques and algorithms frequently are proprietary and are hard to integrate into an organization's computational workflow.
The Fujitsu project used a docking-based approach that combined off-the-shelf and purpose-built software components and required very large amounts of computer processing power. "We needed massive processing and an intelligent pipeline," says Welsford.
To that end, the BioSciences Group developed a computational workflow that builds metabolite models, identifies and models active sites, and then runs the results through a high-throughput docking algorithm that docks the database of predicted metabolites against all Cytochrome P450 (CYP) models. This produces a metric for relative distances of the docked configurations to determine how much a compound and its derivative classes interact with differing CYP models.
To accomplish these steps and perform the various tasks, the group used a handful of existing Fujitsu technologies including life sciences and mathematical modeling software such as ADMEWorks, BioMedCAChe, a neural network program called Neurosim L, and a curated database of CYP interactions called BioFrontier P450. Additionally, some software was developed specifically for this project.
The researchers used Fujitsu's BioServer, the company's massively parallel grid computing device. The BioServer is used to perform the numerous, simultaneous high-throughput docking simulations.
When evaluating a drug's ADME/Tox properties, the process starts with an initial assessment of a drug candidate's properties using ADMEWorks. Results are assessed and promising variations are rerun with CYP interaction information.
Next, a knowledge base of the metabolites is developed using neural net techniques and custom pattern recognition technology based on a class of algorithms called support vector machines§.
Next CYP data is used to build homology models. This step uses crystal structure data from the Protein Data Bank to identify and model possible sites for dockings.
Then the database of predicted metabolites is docked against all of the CYP models. The computational burden to perform these numerous dockings is huge, according to Welsford. "Essentially, you are running all-against-all," he says. That is where the processing power of the BioServer comes in, running the jobs in parallel.
A point of differentiation with this method is it combines two docking approaches. One, which is commonly used, uses a rigid target and allows flexibility in the ligand being tested. The BioSciences group also uses so-called flex-flex docking where both the target and ligand have flexibility. Often, this yields more realistic results.
At this stage, an in-house-developed tool called the Ligand Interaction Distance (LID) Score is applied. This application quickly derives a composite score of the relative atom distances for a compound across active sites.
Initial work for this project was done with a client-partner investigating cancer and drug-resistant infectious diseases. To validate the new predictive ADME/Tox modeling work, Fujitsu used several datasets including the National Cancer Institute diversity data sets for initial calibrations and error estimates and a proprietary database that included loading dose and efficacy data from animal models (from the partner).
The results are promising. The key, says Welsford, is the combination of the varied in silico techniques, the high-throughput docking approach, and wet lab validation.
At the time the new technique was announced during the late summer, Michael McManus, vice president of business development for the BioSciences Group, expanded on this point: "Conventional ADME/Tox 'black box' approaches alone are not capable of telling researchers everything they need to know about the multifactorial processing involved in drug reaction in a living system. The Fujitsu view is that different problems are going to involve different levels of in silico modeling support and different amounts of wet lab validation."
The BioSciences Group is now offering its expertise in this area as a form of consulting service to help companies develop in silico ADME/Tox predictive results.