Automatic discovery of cross-family sequence features associated with protein function.

backgroundMethods for predicting protein function directly from amino acid sequences are useful tools in the study of uncharacterized protein families and in comparative genomics. Until now, this problem has been approached using machine learning techniques that attempt to predict membership, or otherwise, to predefined functional categories or subcellular locations. A potential drawback of this approach is that the human-designated functional classes may not accurately reflect the underlying biology, and consequently important sequence-to-function relationships may be missed.

resultsWe show that a self-supervised data mining approach is able to find relationships between sequence features and functional annotations. No preconceived ideas about functional categories are required, and the training data is simply a set of protein sequences and their UniProt/Swiss-Prot annotations. The main technical aspect of the approach is the co-evolution of amino acid-based regular expressions and keyword-based logical expressions with genetic programming. Our experiments on a strictly non-redundant set of eukaryotic proteins reveal that the strongest and most easily detected sequence-to-function relationships are concerned with targeting to various cellular compartments, which is an area already well studied both experimentally and computationally. Of more interest are a number of broad functional roles which can also be correlated with sequence features. These include inhibition, biosynthesis, transcription and defence against bacteria. Despite substantial overlaps between these functions and their corresponding cellular compartments, we find clear differences in the sequence motifs used to predict some of these functions. For example, the presence of polyglutamine repeats appears to be linked more strongly to the "transcription" function than to the general "nuclear" function/location.

conclusionWe have developed a novel and useful approach for knowledge discovery in annotated sequence data. The technique is able to identify functionally important sequence features and does not require expert knowledge. By viewing protein function from a sequence perspective, the approach is also suitable for discovering unexpected links between biological processes, such as the recently discovered role of ubiquitination in transcription.