Protein Scientists and AI Go Head-to-Head
A team of protein scientists at Rutgers University went head-to-head against a computer program.
Spoiler alert: the AI won. But only by a hair.
Matching Humans Against AI
Scientists decided they wanted to conduct an experiment matching a human with a deep understanding of protein design and self-assembly against an artificially intelligent computer program with predictive capabilities. Topping the list of potential scientists was Vikas Nanda, a researcher at the Center for Advanced Biotechnology and Medicine (CABM) at Rutgers.
The experiment set out to see whether the human or AI could do a better job at predicting which protein sequences would combine most successfully.
The results were published in Nature Chemistry.
Nanda, researchers at Argonne National Laboratory in Illinois, and various colleagues around the U.S. say that the battle was “close but decisive.” The competition put Nanda and several colleagues against the AI program, which one by a small margin.
Scientists are seeking more knowledge around protein self-assembly, believing that by understanding it better, they could design new and innovative products for medical and industrial uses. One of these products could be artificial human tissue for wounds while another could be catalysts for new chemical products.
Nanda is a professor in the Department of Biochemistry and Molecular Biology at Rutgers Robert Wood Johnson Medical School.
“Despite our extensive expertise, the AI did as good or better on several data sets, showing the tremendous potential of machine learning to overcome human bias,” Nanda said.
Protein Design and Self-Assembly
Proteins consist of large numbers of amino acids joined end to end, and the chains fold up to form three-dimensional molecules with complex shapes. The shape of each protein, and the amino acids contained in it, determine its behavior. Researchers such as Nanda are involved in “protein design,” meaning they create sequences that produce new proteins. The team has recently designed a synthetic protein that can quickly detect VX, a dangerous nerve agent. This new development could have big implications for new biosensors and treatments.
Proteins self-assemble with other proteins to form superstructures that are important in biology. In some cases, it appears that proteins are following a design, such as the case when they self-assemble into a protective outer shell of a virus. Other times, they self-assemble when forming biological structures associated with certain diseases.
“Understanding protein self-assembly is fundamental to making advances in many fields, including medicine and industry,” Nanda said.
Nanda and five other colleagues were provided a list of proteins and asked to predict which ones were likely to self-assemble. The predictions were then compared to those of the computer program.
The human experts used rules of thumb based on their observation of protein behavior in experiments, including patterns of electrical charges and degree of aversion to water. They selected 11 proteins they predicted would self-assemble while the AI chose nine proteins.
Their experiment showed that the humans made six correct predictions out of the 11 proteins while the computer program chose nine.
The experiment also demonstrated that the human experts “favored” certain amino acids over others, which led to incorrect choices. The AI correctly chose some proteins with qualities that didn’t make them obvious.
“We’re working to get a fundamental understanding of the chemical nature of interactions that lead to self-assembly, so I worried that using these programs would prevent important insights,” Nanda said. “But what I’m beginning to really understand is that machine learning is just another tool, like any other.”