AI creates protein symphony

Researchers have found that proteins – building blocks for the human body found in everything from hair to muscle and tendons – make their very own music.

The unusual breakthrough, published in the journal APL Bioengineering, was prompted by something of a thorn in the side of the research team, engineers Markus Buehler from MIT in the US and Chi Hua Yu from the National Cheng Kung University in Taiwan.

Nature makes miraculous proteins in abundance, from silk to human cells, but scientists have no way to automatically decipher their design and use the information to make new proteins.

Intent on plugging the gap, Buehler and Yu came up with a workaround that started with a single, very left field observation: each of the twenty amino acids that make a protein has its own vibration frequency. Which is a necessity if you want to make musical notes of different pitches.

Then the pair came up with another insight – proteins and music share a hierarchical structure.

The basic structure for proteins is the varied ordering of the amino acids, such as leucine, alanine and cysteine, in a chain. As you climb the ladder of complexity there is a panoply of twists and folds, including helical arrangements and pleated structures called beta sheets, that are integral to the protein’s function, whether that be the strength of a tendon or the catalytic properties of an enzyme.

The hierarchy of music is surprisingly similar.

At base there is pitch, the notes of the C major scale for example. But for composers to work their magic the length of notes must be altered, multiple pitches are played simultaneously to form chords, and then hundreds of other variations are added until you have an Eroica or Bittersweet symphony.

Buehler and Yu leveraged that likeness by assigning each protein trait a musical analogue. For example, a certain amino acid order might be given the note “C”, or a particular way the chain is folded could indicate the duration of a crotchet.

The result was a musical translation for each protein, recorded on actual staves as a real music score. That could be played.

Click this SoundCloud link, for example, and you can listen to the protein 1akg, sourced from the venom of the predatory sea snail Conus pennaceus, scored for the koto (a Japanese stringed instrument), cello, double bass, viola, flute, and flugelhorn.

So far, so out there. There is, however, more.

Buehler and Yu set about training a deep learning algorithm on data sets that included a host of proteins and their linked musical scores. Training complete, they prompted the AI with snippets of music to “seed” production of brand new musical compositions. 

The result was never before heard music. Which translated back to never before invented proteins that could, say Buehler and Yu, form part of future targeted protein design, to make new enzymes for example.

The researchers even built in a knob to dial up or down the creativity of their AI.

Increasing the temperature during protein design ramped up the randomness of the musical creations and their partner proteins. Reducing the temperature pulled back the manufacturing process in a more orthodox direction.

It is not clear if or when this cross-discipline harmony will bring about real word benefit, but Buehler is up-tempo on the prospect.

“This paves the way for making entirely new biomaterials,” says Buehler. “Or perhaps you find an enzyme in nature and want to improve how it catalyses or come up with new variations of proteins altogether.”

And even if the research does remain “blue sky” it has, at the very least, created some novel additions to our Spotify playlists.