Bioinformaticians from the Skolkovo Institute of Science and Technology posed a problem to the artificial intelligence AlphaFold, which they could not solve. The research was published in the journal PLOS ONE.
Structural bioinformatics is the science that studies the structures of proteins, RNA, DNA and their interactions with other molecules. The data obtained form the basis for the development of drugs and the creation of proteins that do not exist in nature. It was assumed that by knowing the sequence of amino acids in the composition of the protein, it is possible to calculate what three-dimensional form this protein will take in the body, and therefore how it will work.
50 years later, AlphaFold, the artificial intelligence created by Google DeepMind, has learned to solve this central problem of structural bioinformatics. This success led to the suggestion that the neural network somehow learned the physics of proteins and should not just work for the task for which it was designed. Some hoped that AI would soon answer the remaining questions of structural bioinformatics.
At the School of Bioinformatics for High School Students at Skoltech, scientists and schoolchildren decided to put this question to the test. He assigned AlphaFold, installed on the Zhores supercomputer, another structural bioinformatics task: predicting the effect of single mutations on protein stability. The task involved replacing one amino acid in a protein with another. The AI had to predict whether and to what extent the resulting mutant was more or less stable.
AlphaFold failed to do so, as evidenced by its inconsistent estimates with experimental data.
The authors of the study stressed that the creators of AlphaFold never claimed the applicability of AI to other tasks, other than predicting the structure of proteins based on amino acid sequences.
“But some machine learning enthusiasts have been quick to predict the end of structural bioinformatics. So we thought it would be a good idea to go ahead and test it, and we now know that it can’t predict the effect of single mutations,” the scientists said.