Human Protein Subcellular Localization using Convolutional Neural Network as Feature Extractor
Protein subcellular localization of the human cell is an important step to analyze the distribution and functioning of proteins. Over the decades different sources have been utilized to gain insights into protein function. One of them is HTI (High Throughput Imaging) microscopy used by Human Protein Atlas(HPA). Anti-body staining techniques using fluorescent markers provide a high-resolution spatial visualization of proteins in human cells. HPA provides 4 channels for every image sample out of which green depicts protein distribution. This work is focused on localizing proteins into 14 distinct organelles in human cells. The task is to identify in which cell organelle are the protein present. Deep learning methods are validated for pattern recognition tasks in microscopy images. The architecture of Mobilenet is light-weight and results in reduced number of computation. The convolutional part of the network was used to extract image features while the top or fully connected layers of the model were replaced with a single softmax layer. Validation accuracy of 96% while a test accuracy of 91% was achieved.
Paper
Full text
Human Protein Subcellular Localization using Convolutional Neural Network as Feature Extractor
Semantic Scholar · Computer Science · 2019
Abstract
Protein subcellular localization of the human cell is an important step to analyze the distribution and functioning of proteins. Over the decades different sources have been utilized to gain insights into protein function. One of them is HTI (High Throughput Imaging) microscopy used by Human Protein Atlas(HPA). Anti-body staining techniques using fluorescent markers provide a high-resolution spatial visualization of proteins in human cells. HPA provides 4 channels for every image sample out of which green depicts protein distribution. This work is focused on localizing proteins into 14 distinct organelles in human cells. The task is to identify in which cell organelle are the protein present. Deep learning methods are validated for pattern recognition tasks in microscopy images. The architecture of Mobilenet is light-weight and results in reduced number of computation. The convolutional part of the network was used to extract image features while the top or fully connected layers of the model were replaced with a single softmax layer. Validation accuracy of 96% while a test accuracy of 91% was achieved.