The technology that powers Siri, Amazon Echo, facial recognition software and more is coming soon to the realm of protein structure prediction.
MU Electrical Engineering and Computer Science Professor Jianlin Cheng and Professor Amarda Shehu of George Mason University recently landed a three-year, $845,283 grant from the National Science Foundation to support their project, “Guiding Exploration of Protein Structure Spaces with Deep Learning.”
Deep learning is a form of artificial intelligence that uses nonlinear processing to pull out important features and information and processes it much the same way our brain does. It might be most easily recognized in software that uses voice recognition such as the personal assistant in most smartphones and smart speakers. It’s also a frequently used tool in facial and photo recognition software, among a multitude of uses.
Put simply, it’s a powerful tool utilized by a computer to allow it to think and recognize patterns similar to how human brains work. And this new project will utilize it to help in the vital biomedical realm of protein structures.
Knowing the structure of a given protein is key because the way a protein folds into its structure determines its function, including in proteins that protect our immune systems, allow us to move our bodies and much, much more. More accurate knowledge of these structures could help with breakthroughs in several areas, including drug discovery, protein engineering, new protein design, disease diagnosis, precision medicine and more.
“You can make many different predictions. What is the quality of this one, the quality of that one? We’re trying to assess the quality even though we don’t know the true structure,” Cheng explained.
Less than 1 percent of protein structures are known. Predicting how proteins will fold and therefore their function can be time consuming and expensive, which is why computer programs have been created to predict these shapes based off previous data.
That’s where this project comes in. Cheng has worked for years in this field and hopes to use deep learning to even further automate the process and improve the quality of protein structure prediction. Shehu’s team will create the predictions, and Cheng’s team will utilize deep learning to assess the viability and quality of those simulations.
How can deep learning help? Think of the system as akin to how human beings recognize faces. Past a certain point, the human brain can recognize a face in a split second. It can do this because it’s been trained to recognize the features — nose, eyebrows, lips, etc. — and can recognize those shapes and features rapidly.
Cheng’s deep learning method will train the computers by feeding them data on known protein structures. Eventually, the program will learn to recognize the features of good, viable protein structures and use them to quickly recognize both protein structures as a collective and determine viable structures from their nonviable counterparts.
“This pattern is of good quality, this pattern is of higher quality,” Cheng said. “Deep learning can learn these patterns from known structures to assess other structures. Once it is trained, you can apply this model to new structures.”