Deep Learning-based Research And Application Of Protein Subcellular Localization Prediction From Immunohistochemistry Images

Subcellular structures refer to the fine cellular compartments separated by membranes within a cell,such as nucleus,mitochondria,etc.In general,proteins perform function in specific structures,and subcellular location of a protein is closely related to its function.Abnormal localization may lead to cell dysfunction and metabolic disorder.For example,some proteins appear to translocate or mislocate in cancer tissues,which serve as the basis for using those proteins as biomarkers for tumor diagnosis and treatment monitor in clinical practice,and can help to estimate cancerous process in clinical practice.Biological experimental methods are highly professional and costly for obtaining subcellular location of proteins,so developing automated systems to extract information from protein data for location identification is crucial.Protein amino acid sequences are commonly used for this purpose,but they are not sensitive to detect protein distribution changes.In recent years,thanks to the rapid growth of imaging technology,protein microscopic images which can directly reflect the distribution and translocation of proteins in subcellular level,are becoming increasingly popular in subcellular localization research.Among them,immunohistochemistry(IHC)images can show the distribution of proteins in normal and cancer tissue cells,and play an important role in cancer detection and tissue property evaluation.IHC-based protein subcellular localization is an important direction for studying protein functional properties and related diseases.In this work,considering the low accuracy and poor availability of current automated prediction models constructed by traditional features methods,we proposed an automated classifier for predicting protein subcellular locations in IHC images by combining traditional feature engineering and deep learning methods,and validated its ability in detecting location changes of biomarker proteins for colon cancer.Our experimental results demonstrated that the proposed classifier can effectively detect cancer biomarkers.In addition,to adaptively adjust the number of small patches extracted from each IHC image,we constructed a GFNet model that can determine patch numbers according to protein expression level in IHC images.Finally,since the tumor biomarkers can be identified not only according to the changes of subcellular location,but also the changes of protein expression level,we extracted color features from the IHC images and constructed classification models for protein expression.Our results showed that the use of both subcellular location and expression methods can improve the performance of detecting colorectal cancer biomarkers.