Automated annotation in UniProt

UniProt is a high quality, comprehensive protein resource in which the core activity is the expert review and annotation of proteins where the function has been experimentally investigated. At the same time, the UniProt database contains large numbers of proteins which are predicted to exist from gene models, but which do not have associated experimental evidence indicating their function. UniProt commits significant resources to developing computational methods for functional annotation of these predicted proteins based on the data in entries that have gone through the expert review process.

We will describe the two main automated annotation systems currently in use. First, UniRule, which is an established UniProt system in which curators manually develop rules for annotation. Second, ARBA (Association-Rule-Based Annotator), which is a multi-class learning system which uses rule mining techniques to generate concise annotation models. ARBA employs a data exclusion algorithm that censors data not suitable for computational annotation, and generates human-readable rules for each UniProt release. As part of our interest in engaging with the machine learning community, we will also introduce the contribution of ProtNLM (Protein Natural Language Model), from Google Research, which annotates proteins which have "uncharacterised" names.