Glycan Image Recognition Using Deep Learning
Open AccessGlycans refer to the carbohydrates (oligosaccharides and polysaccharides) attached to proteins, lipids, and also RNA. They are known not only as a source of energy but also play essential roles in metabolism, inflammation, cancer, pathogen infection, cellular communication and more. With the increasing attention to glycobiology in the research field, it is important to have an integrated glycoinformatics database for ease of use. GlyGen ( https://www.glygen.org/) is one of the resources that integrates and harmonizes information from publications and trusted databases such as UniProt, GlyTouCan, ChEBI and PubChem. However, the challenge remains between mapping glycan images from articles to the corresponding accessions from databases. To overcome this challenge, deep learning was utilized due to its success in image pattern recognition including medical image analysis, optical image studies, and more. We adapted this technique to map glycan images to GlyTouCan accessions on GlyGen. Glycan images in publications and databases are usually present in Symbol Nomenclature for Glycans (SNFG) format. A pipeline was developed using transfer learning of deep learning for glycan image recognition. In this study, transfer learning retrains a pre-trained network that originally recognizes different images to recognize various classes of glycans. We created 136 distorted images from each of the 33,081 glycan images from GlyGen, resulting in a total of 4,499,016 initial images for training. After testing and comparing 20 different pre-trained networks from Deep Network Designer Toolbox on MATLAB using 10 glycan classes (1370 images) on local machines, four of the networks obtained higher than 95% validation accuracy. Furthermore, we scaled up the testing to 100 glycan classes and DarkNet-19 outperformed all the other networks with the highest accuracy. Therefore, we trained the entire dataset with the DarkNet19 algorithm on Pegasus (High-Performance Computing) with four GPUs running in parallel. As a final result, we achieved over 80% accuracy on the top result and over 99% accuracy in the top 5 results matched to the GlyTouCan accessions. The data used for training can be accessed and further evaluated from the GlyGen website at ( https://data.glygen.org/GLY_000339).With the implementation of glycan image recognition, researchers can simply search for a glycan of interest by uploading an image from any resource. Currently, the algorithm only supports SNFG images. This application produces promising results and is expected to increase the searchability of GlyGen. A beta version of the software is available at https://data.glygen.org/upload.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.