Thread by @AdamRutherford

The @DeepMind protein folding result is really incredible, and incredibly important. But I know it’s pretty tricky to understand, so here’s a megathread, with a bit of Biology 101, for @holland_tom @thehistoryguy
et al.

bit.ly/2JnKQoF

@DeepMind @holland_tom @thehistoryguy Here we go: Genes are long strings of molecules made up of an alphabet of four ‘letters’

e.g.

gacgaagagcccatgatcaacgac

@DeepMind @holland_tom @thehistoryguy Each triplet of letters encodes an amino acid (which we code as capital letters)

gac gaa gag ccc atg atc aac gac

translates to

D. E. E. P. M. I. N. D.

@DeepMind @holland_tom @thehistoryguy But in the body, proteins only work as three-dimensional clumps, precisely folded. The way the protein folds up is determines by its amino acid sequence. And this is the bit we struggle with.
(that is a wretched viral protein btw)

Until now, it has been nigh on impossible to work out the 3D structure of a protein based on its amino acid sequence. It’s like knowing all the words, but them only being useful in a sentence.

So when we look at some DNA, we can translate it into an amino acid string, but not into a 3D – that is, functional - protein.

Bear in mind that *ALL* life is either made of, or by proteins. So understanding the process from the gene sequence to the working protein is fundamental to biology.

But it’s also very important for studying diseases. A genetic disease which is caused by a mutation in a DNA sequence can alter the amino acid sequence, and therefore the 3D structure.

A single letter change in the DNA of the globin gene will result in a single amino acid change, and a mis-shaped globin protein, and the result is Sickle Cell anaemia.

We have been working out how proteins are shaped in 3 dimensions using techniques such as X-Ray crystallography. But we can't predict its shape from the original DNA or amino acid sequence.

What DeepMind has done is use AI to make that prediction - an extraordinarily complex process of calculation, trial and error and guess work under very specific biological rules - and go from DNA to 3D protein.

It's radical cos it means we can predict now how mutations in DNA will change the function of a protein.

That will be incredibly important for fundamental biology, for studying diseases, and for designing drugs that treat diseases that are fundamentally born of misfolded proteins.

As @matthewcobb mentions, we still don't know how the process works, cos <black box AI mystery theatre 3000> but the prediction machine works.

Earlier attempts included gamifying the process, which let's face it, is really cool. FOLDIT was a way to make 3D structure competitive, and worked much better than any other technique then available.
I made a short film about this a few years ago bit.ly/2Jun4HL

This is all going to be on #BBCInsideScience on Thursday btw, presented by the awesome @thermoflynamics, who knows A LOT about proteins.