AI narrowly wins in a test of knowledge about self-assembly

A researcher was defeated by an artificial intelligence computer program, when both were asked to predict protein sequences that would combine most successfully.

The research and the AI program

Vikas Nanda, a researcher at the Center for Advanced Biotechnology and Medicine (CABM) at Rutgers in the U.S. has been studying the format of proteins for over 20 years and is knowledgeable on the matter. He often examined how distinctive patterns of amino acids that create proteins determine whether they become anything, such as hemoglobin or collagen. He often thought about the reasoning behind why certain proteins gather to form more distinctive compounds. His expertise is in self-assembly of proteins that clump together and self-assemble into different structures.

With Nanda’s expertise on the subject matter of proteins spanning over two decades, he was the perfect candidate to test the artificial intelligence of a machine’s knowledge on protein sequences. Researchers decided that he was an excellent choice for the study. The research team wanted to conduct an experiment that would have a human compete against a machine, specifically an artificially intelligent computer program. The person would have instinctive knowledge and comprehensive information about protein design and sequences, while the AI would have predictive capabilities based on datasets and programing.

The research team wanted to see if the human or the machine would do a better job at predicting protein sequences. Also, the team wanted to see if either the man or machine could combine the sequence of proteins most successfully. It was a close call, but the artificial intelligence program beat Nanda by a small margin.

The study was published in the journal Nature Chemistry.

The importance of self-assembly

The researchers are interested in self-assembly because they believe that fully understanding the concept could help them create numerous products for medical and industrial uses, such as artificial human tissue for wounds.

“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,” said Nanda, a professor in the Department of Biochemistry and Molecular Biology at Rutgers Robert Wood Johnson Medical School.

Background on proteins and Nanda’s knowledge on protein design

Proteins consist of large numbers of amino acids joined together. The chains fold to form three-dimensional molecules with various shapes. The unique shape of each protein, along with its amino acid, determines what it does. Nanda studies “protein design”, which involves sequences that make new proteins. Recently, he and his team designed a synthetic protein that can identify a dangerous nerve agent called VX. This discovery could help create new biosensors — devices that detect chemicals — and treatments.

Scientists are unable to explain the reason why proteins self-assemble with other proteins to form superstructures. Sometimes, proteins self-assemble because they seem to be following a specific layout, such as creating a protective outer shell of a virus called a capsid. Other times, it’s done when something goes wrong, such as forming structures associated with diseases like Alzheimer’s disease or sickle cell. “Understanding protein self-assembly is fundamental to making advances in many fields, including medicine and industry,” Nanda said.

The study

For the research, Nanda and five other collaborators were given a list of proteins and asked to predict which ones would self-assemble. The AI computer program was given the same list to make its guesses as well. The predictions were then compared with each other, human vs machine.

The human participants made their predictions based on observation of protein behavior, such as patterns of electrical charges. The machine based its knowledge on an advance machine-learning system. The human participants predicted that 11 proteins would self-assemble, while the AI computer program predicted that nine proteins would self-assemble.

The results

Overall, the human experts were correct for six out of the 11 proteins they selected, and the AI computer was correct with six out of the nine proteins it selected, therefore, beating the researchers.

The study showed that the experts favored and selected some amino acids over other proteins, which would sometimes lead to erroneous results. The computer, on the other hand, would select proteins with qualities that weren’t considered obvious choices for self-assembly, leaving the researchers with additional questions on how the selections were made by AI.

Nanda came to the realization that artificial intelligence is another useful tool like many others but needs to be understood and studied more. “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.”

AI could potentially help researchers in the future with identifying proteins that self-assemble, leading to new discoveries and cures for certain diseases.