| | | | The Discovery Bus is a new workflow automation technology, the “Discovery Bus” has been developed to automate workflows in molecular design, and much work has been done on the QSAR modelling workflow in particular. It has been shown to automatically generate multiple QSAR models for multiple data types and has numerous other advantages including automated updating when new data or QSAR modeling methods are added to the system. A recent paper (Cartmell, J., Leahy, D.E. and Krstajic, D., J Comput Aided Mol Des. 2005 Nov;19(11):821-33) explains the technology in more detail and gives some validation results. A summary article (Cartmell, J., Leahy, D.E. and Krstajic, D., Curr Opin Drug Discov. Devel. 2007 10(3):347-52) also gives more background.
The Discovery Bus facilitates structural evolution in specialist capabilities, such as the introduction of new components and strategies as well as a continual reappraisal and adjustment to newly acquired knowledge such as the arrival of new screening data, new chemical source databases, synthetic methods or competitor patents. The architecture positively supports competition arising from the availability of multiple algorithms, strategies and data sources.
The most obvious metaphor for automation of problem solving strategies, is that of workflow. Applications of workflow have been available for a long time and there are many commercial toolkits available, with at least one available for cheminformatics (www.scitegic.com). There are also open source alternatives under development. In addition, the workflow concept is central to many developments around the use of Web and Grid Services.
The use of autonomous collaborating software agents to handle the complexity and dynamics required by some workflow management applications is an area of active research and autonomous software agents have been applied to distributed learning, tool-integration, bioinformatics workflows and systems biology modelling.
Studies aimed at the automation of the QSAR modelling process to compare the performance of descriptors and both descriptors and correlation methods have also been carried out previously. Chemical descriptors are features of a chemical structure or molecular properties that usually can be calculated rapidly by computational methods. There are many thousands of these available from computational chemistry software covering features related to shape, size, molecular complexity, charge distribution and solvation. The challenge in modern QSAR is to identify which of these are significant inany model and to eliminate the “noise” as well asguard against spurious correlation.
The Discovery Bus described here takes advantage of concepts from both workflow and multi-agent systems and has significant additional advantages. In particular, all possible combinations of components are explored leading to exhaustive evaluation of potential solutions. Since requests are always regarded as questions to which there are multiple answers and no request is ever regarded as complete, improved results emerge naturally as new strategies, components or additional data are made available.
The Discovery Bus therefore has the following distinguishing features: different technologies naturally compete with each other (and with any human agent), all combinations of components available are explored, and, results are automatically updated as new components become available. In conditions where there are insufficient resources to explore all possible combinations of strategies, which we expect to be the norm as more agents are added, then the most promising pathways are explored first by a reinforcement mechanism. Pathways leading to successful results are reinforced through increasing priority for CPU resource, creating a mechanism by which the community of agents can learn from experience.
As a consequence of these features, the Discovery Bus provides a mechanism for implementing knowledge acquisition and use in a highly efficient, comprehensive and continuously improving manner. This project runs parallel to other research programs exploiting the discovery bus in multi-objective optimization of chemical structures where nature-inspired algorithms are used to evolve novel chemical structures that meet multiple activity and property criteria.
|
| |
|
