As the number of sequenced genomes rapidly grows,there are a large number of proteins with unknown function in new genomes.Active learning methods can assist biologists for selecting the most valuable ones as candidates for biological experiments.Previously,it is proved that the protein function prediction task is naturally Multi-Instance Multi-Label (MIML) learning problem.In this paper,we formulate the problem of selecting the most valuable proteins in annotating genome-wide protein functions as a MIML active learning task,Then,we propose a MIML active learning framework named MIMLAL and design two algorithms MIMLAL-A and MIMLAL-R for genome-wide protein function prediction.