Automated biological sequence description by genetic multiobjective generalized clustering

Recent advances in the accessibility of databases containing representations of complex objects - exemplified by repositories of time-series data, information about biological macromolecules, or knowledge about metabolic pathways - have not been matched by availability of tools that facilitate the retrieval of objects of particular interest and aid understanding their structure and relations. In applications, such as the analysis of DNA sequences, on the other hand, requirements to retrieve objects on the basis of qualitative characteristics are poorly met by descriptions that emphasize precision and detail rather than structural features. This paper presents a method for identification of interesting qualitative features in biological sequences. Our approach relies on a generalized clustering methodology in which the features being sought correspond to the solutions of a multivariable, multiobjective optimization problem with features generally corresponding to fuzzy subsets of the object being represented. Foremost among the optimization objectives being considered are measures of the degree by which features resemble prototypical structures deemed to be interesting by database users. Other objectives include feature size and, in some cases, performance criteria related to domain-specific constraints. Genetic-algorithm methods are employed to solve the multiobjective optimization problem. These optimization algorithms discover candidate features as subsets of the object being described and that lie in the set of all Pareto-optimal solutions - of that problem. These candidate features are then summarized, employing again evolutionary-computation methods, and interrelated by employing domain-specific relations of interest to the end users. We present results of the application of this two-step method to the recognition and summarization of interesting features in DNA sequences of Tripanosoma cruzi.