Supervised Learning-Aided Optimization of Expert-Driven Functional Protein Sequence Annotation

The aim of this work is to use a supervised learning approach to identify sets of motif-based sequence characteristics, combinations of which can give the most accurate annotation of new proteins. We assess several of InterPro Consortium member databases for their informativeness for the annotation of full-length protein sequences. Thus, our study addresses the problem of integrating biological information from various resources. Decision-rule algorithms are used to cross-map different biological classification systems in order to optimise the process of functional annotation of protein sequences. Various features (e.g., keywords, GO terms, structural complex names) may be assigned to a sequence via its characteristics (e.g., motifs built by various protein sequence analysis methods) with the developed approach. We chose SwissProt keywords as the set of features on which to perform our analysis. From the presented results one can quickly obtain the best combinations of methods appropriate for the description of a given class of proteins. Introduction Availability of a wide variety of effective protein sequence analysis methods calls for an evaluation of their comparative performance and for development of approaches to integrated cross-method consistent annotation. A natural resolution of this problem came with the creation of InterPro [1]. The InterPro database is a single resource collecting sequence pattern data from PROSITE ([2], a repository of regular expressions and profiles), Pfam ([3], based on hidden Markov models), PRINTS ([4], provider of fingerprints) and from several other databases-participants. InterPro is a manually curated database, in which the curation process is supported by various automated procedures. One of the most straightforward approaches to characterizing a novel sequence is to compare it to the already annotated in InterPro proteins. While this potentially can produce high-quality functional predictions, the motif-focused nature of InterPro complicates the interpretation of such analyses, because in most cases it is impossible to find a single InterPro entry corresponding to the combination of motifs found in a given sequence. On the other hand, there are various systems for direct functional annotation of full-length protein sequences, such as GeneOntology (GO) [5] or SwissProt [6]. The quality of annotation that one can get based on similarity to InterPro entries/motifs can thus be improved by combining these two annotation paradigms. A correspondence between the annotation specific for a group of previously characterised full-length protein sequences and the domain/repeat architecture of a given sequence could help to achieve a more complete functional description of a protein related …