Man vs. Machine: AI narrowly beats a human scientist in a test of scientific skill
PISCATAWAY, NJ — No invention signifies the ingenuity and intelligence of mankind like the computer. A wonder of the modern age, countless works of science fiction predict an inevitable conflict in the not-too-distant future: man versus machine. Now, according to researchers at Rutgers University, it appears that machines have already surpassed humanity in at least one scientific subject.
Professor Vikas Nanda of Rutgers University has spent more than two decades carefully studying the intricate nature of proteins, the highly complex substances found in all living organisms. He has devoted his professional life to considering and understanding the unique patterns of amino acids that make up proteins and determine whether they become hemoglobin, collagen, etc. Additionally, Prof. Nanda is an expert in the mysterious step of self-assembly, in which certain proteins come together to form even more complex substances.
So when the study’s authors set out to conduct an experiment pitting a human—someone with a deep, intuitive understanding of protein design and self-assembly—with predictive abilities AI computer programprof. Nanda made the perfect participant.
The authors of the study wanted to see who or what could do a better job in predicting which protein sequences would be most successfully combined – Prof. Nanda and a few other people or a computer. The published results show that the intellectual battle is close, but the AI program outperformed the humans by a small margin.
What can scientists use protein self-assembly for?
Modern medicine is investing heavily in protein self-assembly because many scientists believe that a complete understanding of the process can lead to a number of revolutionary products for medical and industrial use, such as artificial human tissue for wounds or catalysts for new chemical products.
“Despite our extensive expertise, the AI performed as well or better on several datasets, demonstrating the enormous potential of machine learning to overcome human bias,” says Nanda, a professor in the Department of Biochemistry and Molecular Biology at Rutgers Robert Wood Johnson Medical School, ua university edition.
Proteins consist of large amounts amino acids, connected end to end. These amino acid chains fold into three-dimensional molecules of complex shapes. The exact form is important; the exact shape of each protein, as well as the specific amino acids it contains, determine what it does. Some scientists, including prof. Nandu, regularly engage in an activity called “protein design,” which involves creating sequences that produce new proteins.
Recently, prof. Nanda and a team of researchers designed a synthetic protein that can quickly detect the dangerous nerve agent known as VX. This protein may lead to the development of new biosensors and treatments.
For reasons still unknown to modern science, proteins self-assemble with other proteins to form superstructures important to biology. Sometimes proteins seem to follow a design, such as when they self-assemble into the virus’s protective outer shell (capsid). In other cases, however, the proteins will seemingly self-assemble in response to something going wrong, ultimately creating deadly biological structures associated with diseases ranging from Alzheimer’s disease to sickle cells.
“Understanding protein self-assembly is fundamental to progress in many fields, including medicine and industry,” adds Prof. Nanda.
How did the AI program work?
During the test, prof. Nanda and five other colleagues were given a list of proteins and had to predict which ones were likely to self-assemble. A computer program made the same predictions, and then the researchers compared the answers man and machine.
Human participants made their predictions based on their previous experimental observations of the protein, such as patterns of electrical charges and degree of aversion to water. People eventually predicted that 11 proteins would self-assemble. The computer program, meanwhile, selected nine proteins through an advanced machine learning system.
The human experts were right about six of the 11 proteins they picked. Computer program earned a higher percentage of accuracywith six of the nine proteins he selected actually capable of self-assembly.
The study authors explain that human participants tended to “favor” certain amino acids over others, leading to incorrect predictions. The AI program also correctly identified some proteins that were not “obvious choices” for self-assembly, opening the door for additional research. Professor Nanda admits he was once skeptical of machine learning for protein assembly research, but is now much more open to the technique.
“We are working on a fundamental understanding of the chemical nature of the interactions that lead to self-assembly, so I was concerned that using these programs would prevent important insights,” he concludes. “But what I’m really starting to realize is that machine learning is just another tool, like any other.”
The study is published in the magazine Chemistry of nature.