Dynamically predicting protein functions from semantic associations of proteins

Abstract

Predicting functions of un-annotated proteins is a significant challenge in the post-genomics era. Among existing computational approaches, exploiting interactions between proteins to predict functions of un-annotated proteins is widely used. However, it remains difficult to extract semantic associations between proteins (i.e. protein associations in terms of protein functionality) from protein interactions and incorporate extracted semantic associations to more effectively predict protein functions. Furthermore, existing approaches and algorithms regard the function prediction as a one-off procedure, ignoring dynamic and mutual associations between proteins. Therefore, deriving and exploiting semantic associations between proteins to dynamically predict functions are a promising and challenging approach for achieving better prediction results. In this paper, we propose an innovative algorithm to incorporate semantic associations between proteins into a dynamic procedure of protein function prediction. The semantic association between two proteins is measured by the semantic similarity of two proteins which is defined by the similarities of functions two proteins possess. To achieve better prediction results, function similarities are also incorporated into the prediction procedure. The algorithm dynamically predicts functions by iteratively selecting functions for the un-annotated protein and updating the similarities between the un-annotated protein and its neighbour annotated proteins until such suitable functions are selected that the similarities no longer change. The experimental results on real protein interaction datasets demonstrated that our method outperformed the similar and non-dynamic function prediction methods. Incorporating semantic associations between proteins into a dynamic procedure of function prediction reflects intrinsic relationships among proteins as well as dynamic features of protein interactions, and therefore, can significantly improve prediction results.