Towards Protein Function Annotations for Matching Remote Homologs

Motivation. Identifying functional homologs from their protein structures is an important problem in biology. One way to approach this is to discover their common local structures (i.e. motif) among protein families. Unfortunately, most of the motif models are inadequate to characterize the structural diversities, especially when the proteins are distantly related. In this study, we first introduce a statistical model, together with a semi-supervised refinement method, to perform post-processing on the motif obtained from a motif discovery algorithm or from motif database. Our model makes use of Markov Random Fields (MRF), which provide a large search space to optimize the motif while preserving the dependencies among neighbor elements. The resulting model not only can better represent the underlying patterns for different protein families, but also can enable the power of functional annotation from a set of unknown proteins using probability approximations. In addition, we develop two filter approaches (three methods) to further eliminate the false positives introduced by any motif models. By considering the local environment around the active sites of each family, the filters reject proteins to match with the model without similar environment profiles. Results. Our experimental results, as evaluated in five sets of enzyme families with less than 40% sequence identity, demonstrated that our methods can obtain more remote homologs that could not be detected by traditional sequence-based methods. At the same time, our method could reduce large amount of random matches which were originally introduced by the motif representations. On average, our methods could improve about 13% of the functional annotation ability (measured by their AUCs). In certain experiments, our improvement went even up to 70%.