Uses computer reinforcement learning
April 28, 2022
When the original wave of COVID-19 hit Shanghai, China, it impacted Jason (Jiachen) Lu’s life.
“It was pretty severe and lots of people, including many people around me, were influenced by that,” the third-classman said. “They were having a pretty long period when there were no vaccinations for that.”
“And so I did some research on the vaccination-making process, and I found that the protein structure is a critical part of it. We need to decide the protein structure in order to make a vaccine,” he explained. “But we currently don’t have a really efficient method to solve this problem.”
“Since I’m pretty interested in computer science and AI (artificial intelligence), I wanted to see if this could work for solving this problem.”
Using Python, Lu developed a code that explores the protein folding problem with vaccines through reinforcement learning. Protein folding is how the proteins – amino acids – flow throughout the body, he said.
“It’s a very important question right now but the method currently used is not so efficient. I want to improve its efficiency by applying machine learning into it. And I chose reinforcement learning, which is one type of machine learning.”
He started his work while at home last summer and continued it after he came back to Culver Academies this fall. He submitted it as his Honors in Science project, “Explore the Protein Folding Problem using Deep Reinforcement learning,” last semester. He then submitted it to the Hoosier Science and Engineering Fair on April 2. His work was recognized with the Indiana University Purdue University Indianapolis Department of Biology Award for Excellence and Lu was selected to attend the International Science and Engineering Fair in Atlanta, starting on May 3.
The Atlanta event brings together over 1,800 students from more than 75 nations to compete for scholarships, tuition grants, internships, scientific field trips, and the grand prize, a $75,000 college scholarship.
Lu formulates a protein folding problem using a model called hydrophobic polarity. The goal is to get as many hydrophobic amino acids to be adjacent to each other as possible. “I find an organization of the amino acids so that the hydrophobic amino acids will locate each other and that’s an expected condition in real life. The purpose of this is to predict the protein structure from its primary structure to a structure in real life – that’s more three dimensional than two dimensional.”
The protein’s three-dimensional structures have several applications in many important processes, such as creating new drugs and vaccines, he said. “It is very important because every specific structure of a protein will define its specific function, so it is important for us to export this quickly.”
He is using the Markov Decision Process (MDP), which is a foundational element of reinforcement learning. Using this method, Lu can let “the agent learn by itself by giving it a reward and telling it what actions are more valuable so that, finally, it will find an optimal policy that formulates.”
He is using his personal MacBook to run the program, which can take up to 20 hours to train a model, he said. As his program becomes more complicated and goes from two dimensional to three dimensional, he’ll need more computing power, he said.
Along with Lu’s success at the Hoosier Science and Engineering Fair, Kevin (Minzhe) Yang ’24 (Shenzhen, China) received several awards for his project: “the study of the impact of the form of the transformable trimaran on the vessel’s speed and stability in water areas with waves.”
Yang was awarded the Office of Naval Research-Naval Science Award, the Yale Science and Engineering Association Most Outstanding Exhibit in STEM, and Society of Tribologosits and Lubrication Engineers Excellence in Science or Engineering Related Research.