Structured Output Prediction of Novel Enzyme Function with Reaction Kernels

Abstract

Enzyme function prediction is an important problem in post-genomic bioinformatics, needed for reconstruction of metabolic networks of organisms. Currently there are two general methods for solving the problem: annotation transfer from a similar annotated protein, and machine learning approaches that treat the problem as classification against a fixed taxonomy, such as Gene Ontology or the EC hierarchy. These methods are suitable in cases where the function of the new protein is indeed previously characterized and included in the taxonomy. However, given a new function that is not previously described, these approaches are not of significant assistance to the human expert. The goal of this paper is to bring forward structured output learning approaches for the case where the exactly correct function of the enzyme to be annotated may not be contained in the training set. Our approach hinges on fine-grained representation of the enzyme function via the so called reaction kernels that allow interpolation and extrapolation in the output (reaction) space. A kernel-based structured output prediction model is used to predict enzymatic reactions from sequence motifs. We bring forward several choices for constructing reaction kernels and experiment with them in the remote homology case where the functions in the test set have not been seen in the training phase.