Subcellular Location Prediction Research Of Human Protein Images Based On Multi-scale Monogenic Signal Analysis And Deep Learning

Proteins need to be in the right place at the right time and bind to the appropriate molecules to perform their functions.Therefore,accurately predict subcellular localization of proteins plays an irreplaceable role in understanding protein function,the development of cancer targeted drugs and the screening of cancer markers.In the early stage,the protein subcellular location was obtained by employing traditional biomolecular experiments,and then researchers annotated the protein subcellular location according to the experimental data.However,the method results in a significant loss of manpower,funds and time.Therefore,researchers devoted to developing automated subcellular location prediction models based on existing massive biological data,pattern recognition and machine learning theory.Although the existing prediction models for predicting humans,animals and plants protein subcellular location achieve an encouraging result,these prediction models are almost developed based on protein sequence data,and then these models are unable to accurately capture the protein expression situation when biochemical environments occur change,namely the tissue cancer.To contrary,the protein image can accurately reflect the protein expression situation in multiple human tissues,such as protein texture distribution,contour,edge and color information,and encouraging image-based prediction model can accurately capture the distribution pattern of protein in normal and cancer tissues so that it can effectively find out the cancer biomarker and complete clinical medical experiments.Therefore,this paper adopts the protein image data as the research object,and focusing on mining the frequency domain information of protein image and utilizing the deep learning improve the prediction accuracy of protein subcellular location.The main content of paper is summarized in two aspects as follows:(1)A novel multi-label subcellular location prediction model MIC?Locator is proposed based on multi-scale monogenic signal analysis and image intensity coding strategy.Firstly,the benchmark dataset is collected and collated,which conforms to the newly published guidelines for subcellular location labeling on the HPA database.Secondly,Fast Fourier transform(FFT),Riesz transform,Log-Gabor filter and image intensity coding strategy are used to obtain the frequency feature of monogenic signal based on various frequency scales.Finally,Classifier Chains(CC)is proposed to handle the distribution of multi-label protein image.The experimental results show that MIC?Locator can achieve 60.56% subset accuracy,and it can achieve higher performance than state-of-the-art prediction models.(2)Using the deep learning model to extract the deep feature of IHC image and combining shallow features construct the prediction model AR?Locator,and its construction processes can be divided into three stages.Firstly,in the image preprocessing step,the canny edge detection algorithm with multi-scale is utlized to cut the protein cluster region of IHC images,and the protein and DNA channel of IHC image is obtained by linearly separation method.Secondly,in the extraction of deep and shallow feature step,the Resnet50 is utilized as the backbone network,and multiple attention mechanism modules are embedded into it to build the AR?Network;the cut IHC is fed into AR?Network as input image,and AR?Network is trained by an end-to-end fashion.The deep feature is extracted from the attention mechanism module 2 and the global average pooling layer(GAP)of AR?Network respectively.Furthermore,SLFs feature and local binary descriptors(LBP)are extracted as shallow features.Thirdly,in the decision-level fusion step,because deep and shallow features reflect the abstract and texture information of IHC image respectively,fusing these single prediction model trained by the two features and BR classifier constructs AR?Locator prediction model.The experimental results demonstrate that AR?Network can achieve 72.09% prediction accuracy,which is superior to state-of-the-art prediction models.