Predicting Protein Functions of Bacteria Genomes via Multi-instance Multi-Label Active Learning

Bacteria are workhorses in the fields of molecular biology, genetics, and biochemistry. Understanding the biological functions of proteins is important for bacteria studies in the post-genomic era. There are a large number of proteins with unknown biological functions in newly discovered bacteria species. Active learning can assist biologists in selecting the most valuable proteins as candidates for assays. Previously, we formulated the protein function prediction problem as a multi-instance multi-label (MIML) learning task. Here, we propose a MIML active learning algorithm MIMLAL-R for protein function annotation of bacteria genomes. MIMLAL-R minimizes an approximated surrogate loss by stochastic gradient descent with the most valuable bag-label pairs, which are chosen by a selection criterion that combines label cardinality inconsistency and diversity. We tested MIMLAL-R on four real bacteria, namely Geobacter sulfurreducens, Azotobacter vinelandii, Shewanella loihica PV-4, and Escherichia coli strain K12. The MIMLAL-R approach achieved excellent performances in all cases. The codes and data can be freely downloaded from.