Prediction of Protein Subcellular Localization from Microscopic Images via Few-Shot Learning

Abstract

Benefitting from the breakthrough development of microscopy imaging techniques, various bio-microscopic images have accumulated rapidly for the past decade. Using computer vision and machine learning methods, biological activities and molecular functions can be interpreted from these images, thus image analysis has become more and more important in current life science research. A prominent difficulty in biological image analysis is the lack of annotation, and the test set even contains data from unseen classes, i.e. the open-set issue. The image-based protein subcellular localization is a typical open-set problem. There are tens of subcellular compartments in cells, while the labeled data may only consist of proteins from several major organelles. Till now, the open-set problem has been rarely studied for biomedical image data. The main goal of this study is to train a few-shot learning model for the recognition of protein subcellular localization from immunofluorescence images. We conduct experiments on a data set collected from Human Protein Atlas (HPA) and the results show that the introduced system can provide accurate results even with a small handful of images for an unknown class in a multi-instance learning scenario.