Multi-kernel transfer learning based on Chou's PseAAC formulation for protein submitochondria localization

► We propose multi-kernel transfer learning model for protein submitochondria localization. ► We conduct a comprehensive survey of GO-based model performance. ► We quantitatively survey how much GO knowledge should be transferred. ► The noise from homologous GO knowledge transfer is properly depressed. ► The risk of performance overestimation for novel proteins is greatly reduced.