Last year, MIT researchers announced that they were turning the biochemical properties of proteins into music. Now, they’ve used those musical compositions to create entirely new proteins.
Proteins have repeated patterns, such as the alpha helices seen here in a keratin moelcule. … [+]
Turning music into proteins sounds more like a party trick than serious science, but it’s allowing researchers to create entirely new proteins, and that’s not a simple task.
Each naturally occurring protein is made up of a sequence of amino acids. There are 20 amino acids to choose from for each position in the protein (which can be hundreds or even thousands of amino acids long), but not all combinations of amino acids make functional proteins. The final structure of the protein is not simply a long chain of amino acids, but it’s folded and coiled up into a 3-dimensional structure that determines how the protein works. Without the right shape, a protein loses its function.
To be able to form these shapes, the amino acids need to follow certain patterns, but with 20 different ones to choose from for every location, there are a lot of different combinations, and it’s not easy to figure out which patterns are allowed and which are not.
By converting protein structures into music, Chi-Hua Yu and Markus Buehler of the Laboratory for Atomistic and Molecular Mechanics at MIT have created a library of music fragments that directly correspond to the kind of protein structures that you would find in real, existing proteins.
They then taught a neural network to create similar music patterns based on this music library. Without knowing anything about proteins and biochemistry, the neural network analysed the music patterns and made new music that resembled it.
Yu and Buehler then turned these new music patterns back into proteins. These were proteins that didn’t exist before, but they appeared to be stable, and folded in similar ways as naturally existing proteins.
Certain types of folds and turns occur regularly in proteins, and these repeating shapes were encoded in the music. One of these shapes is an “alpha helix” — a coiled section of a protein. The music segments that were used to train the neural network all came from proteins with many alpha helices, so the AI-generated music also corresponded to proteins with several alpha helices.
Why would you want to make entirely new proteins? In a press release accompanying the publication of this work in APL Bioengineering, Buehler suggested that this method could be used to create entirely new biomaterials, or to edit existing enzymes, for example.
If you’re curious what protein music sounds like, Buehler has listed several proteins’ music on Soundcloud, including a musical representation of the protein structure of the spike protein of the coronavirus responsible for COVID-19.