Function Prediction Using Neighborhood Patterns

Traditionally, function prediction is performed using sequence/structure homology and then manual verification in the wet lab. The first step, also called computational function prediction, directs the experimental design to a narrow set of possible annotations for unstudied proteins. Recently, large interaction networks have been constructed using data from high-throughput techniques like Microarray co-expression analysis and Yeast2Hybrid experiments. These networks contain more information compared to sequence or structure alone. A molecular function is performed in the context of a biological process with multiple interacting agents. Hence, a natural next stage of computational function prediction includes the use of a protein’s interaction context within the network. A recent survey [1], classifies most existing function prediction methods in two groups: module assisted and direct methods. Methods from the first group detect network modules and then perform a module-wide annotation enrichment [2]. They all differ in the manner they identify modules— some use graph clustering [3] while others use hierarchical clustering based on network distance [2]. Direct methods assume that similar functional annotations are neighbors in the network. The Majority method [4] predicts prevailing functions among the direct interactors of a target protein. This idea has later been extended to higher levels in the network [5]. Indirect Neighbor [6], utilizes direct and indirect functional associations, considering level 1 and level 2 associations. The Functional Flow method [7] simulates an annotation network flow from known to target proteins. A common drawback of the methods in the two groups is their hypothesis that proteins with similar functions are always close in the network. Functional annotations in actual protein networks do not always corroborate this hypothesis. The direct methods are also limited to use information about neighbors up to a certain level. Thus, they cannot predict the functions of proteins surrounded by unannotated interaction partners.