AI may be able to predict which viruses will spread from animals to humans
According to the Centers for Disease Control and Prevention, three out of four emerging or new infectious diseases can be traced back in humans. The Centers for Disease Control and Prevention (CDC) defines zoonoses as diseases that are caused by harmful germs like viruses, bacteria, fungi and parasites being transmitted from animal to human. A new study in Cell Reports shows that scientists used artificial intelligence (AI), machine learning, to predict the spread of viruses from animals to people. This is a breakthrough that could one day prevent pandemics and viral infections.
“Our model recapitulates the known binding specificities of influenza virus and we show that these predictions are extensible to other viruses such as coronaviruses and rotaviruses,” said Daniel Bojar, PhD., an assistant professor at University of Gothenburg, Sweden. He was co-authored with John Quackenbush, PhD at Harvard School of Public Health, and Rebekka Burkholz (postdoctoral researcher at Harvard School of Public Health).
Researchers used glycans (also known as polysaccharides), which are chains-link structures made up of single sugar molecules (monosaccharides) to create a new approach. Glycans, which are carbohydrate-based polymers, are made by all living organisms.
Researchers wrote that glycocans are a special class of biological molecules because they have nonlinear, branching structure that allows them to perform a wide variety of functions. These include protein folding and destruction; stress response; cell interactions; cell migration patterns; self/non-self discrimination; and microbiome composition, composition and health.
They used AI deep learning to create a graph convolutional neural net (GCNN), which was able to learn a representation of glycans. The model was named SweetNet. Graph convolutional neural network are used for many purposes, including computer vision, bioinformatics and traffic detection.
The scientists wrote that SweetNet is a graph convolutional neural system that uses graph representation to facilitate a computational understanding glycobiology. “SweetNet explicitly reflects the nonlinearity of glycans, and provides a framework for mapping any glycan sequence into a representation.
SweetNet’s glycan representations outperformed other models, according to the researchers. The SweetNet-based models for glycans are more efficient and require less training.
According to the researchers, SweetNet was able to train models 30% faster than their equivalent SweetTalk models. “SweetNet also performed better than SweetTalk in terms of data efficiency, even though it used only a third the dataset for training.”
The researchers reported that SweetNet is superior to other computational methods when it comes to predicting glycan characteristics on all tasks. SweetNet’s glycan representations are predictive of both environmental and organismal properties. We use glycan-focused, machine learning to predict viral binding. This can be used to identify viral receptors.
Scientists have made a significant step in using AI deep learning to glycobiology to identify glycans that could be used as antivirals to help prevent pandemics and viral infections.