MULTI-LABEL CLASSIFICATION BASED ON SUBCELLULAR REGION-GUIDED FEATURE DESCRIPTION FOR PROTEIN LOCALISATION

In this paper, we present a multi-label classification pipeline and a novel feature descriptor for the protein subcellular localisation. The challenge here is the development of a computational model that can classify multi-site proteins on a highly unbalanced dataset with a long-tail distribution and multi-label images. To address this challenge, we design a Location-Sorted Random Projections feature descriptor to represent image intensity and gradient of the protein of interest in reference to the correlated cellular region. Multilabel Synthetic Minority Over-sampling Technique is optimised to generate synthetic features with labels to handle class imbalance. Our method achieves the state-of-the-art performance on a large-scale public dataset and demonstrates excellent performance for the minority classes.

Automated summary

This paper presents a multi-label classification pipeline and a novel feature descriptor for protein subcellular localization. By utilizing a Location-Sorted Random Projections feature descriptor and Multilabel Synthetic Minority Over-sampling Technique, the computational model achieves state-of-the-art performance on a highly unbalanced dataset with long-tail distribution and multi-label images. Additionally, the method shows excellent performance for minority classes.