Cooperative Metaheuristics for Exploring Proteomic Data

Gras, Robin ; Hernandez, David ; Hernandez, Patricia ; Zangge, Nadine ; Mescam, Yoann ; Frey, Julien ; Martin, Olivier ; Nicolas, Jacques ; Appel, Ron

In: Artificial Intelligence Review, 2003, vol. 20, no. 1-2, p. 95-120

Add to personal list
    Most combinatorial optimization problems cannotbe solved exactly. A class of methods, calledmetaheuristics, has proved its efficiency togive good approximated solutions in areasonable time. Cooperative metaheuristics area sub-set of metaheuristics, which implies aparallel exploration of the search space byseveral entities with information exchangebetween them. The importance of informationexchange in the optimization process is relatedto the building block hypothesis ofevolutionary algorithms, which is based onthese two questions: what is the pertinentinformation of a given potential solution andhow this information can be shared? Aclassification of cooperative metaheuristicsmethods depending on the nature of cooperationinvolved is presented and the specificproperties of each class, as well as a way tocombine them, is discussed. Severalimprovements in the field of metaheuristics arealso given. In particular, a method to regulatethe use of classical genetic operators and todefine new more pertinent ones is proposed,taking advantage of a building block structuredrepresentation of the explored space. Ahierarchical approach resting on multiplelevels of cooperative metaheuristics is finallypresented, leading to the definition of acomplete concerted cooperation strategy. Someapplications of these concepts to difficultproteomics problems, including automaticprotein identification, biological motifinference and multiple sequence alignment arepresented. For each application, an innovativemethod based on the cooperation concept isgiven and compared with classical approaches.In the protein identification problem, a firstlevel of cooperation using swarm intelligenceis applied to the comparison of massspectrometric data with biological sequencedatabase, followed by a genetic programmingmethod to discover an optimal scoring function.The multiple sequence alignment problem isdecomposed in three steps involving severalevolutionary processes to infer different kindof biological motifs and a concertedcooperation strategy to build the sequencealignment according to their motif content