Artificial intelligence solved a protein puzzle; why it matters

There’s long been speculation if artificial intelligence (AI) could perform scientific tasks beyond the understanding of human beings. A paper in Nature indicates this may now have happened. 

Since 1994, bio-scientists have held a biennial contest called CASP (Critical Assessment of Structure Prediction). Teams use computers to predict the ways in which given proteins will fold. The 100 or so problems presented at each CASP are those of proteins where the structure has already been determined, but not yet made public. The participating computer programs are judged on the basis of their accuracy in solving those structures where the answers are already known. 

Let’s try to understand this problem and why it’s important. Imagine giving 20 differently coloured pieces of cloth to a blind man and asking him to make a patchwork quilt. Now, imagine trying to guess the pattern in which he’s likely to have placed those patches. This is a very simplified analogy to the protein-folding problem. 

Proteins are large strings of amino acids, stitched together and folded over in peculiar ways in three-dimensions (not in two, like a quilt).  DNA contains the information needed to produce proteins. That information is transferred to RNA, which takes it to the cytoplasm of a cell and helps to build the protein.

While there are 20 amino acids, a protein molecule could consist of 100,000 odd strands of those amino acids put together in an insanely large number of possible combinations. The way in which a protein acts depends on the way in which those amino acids are stitched together and folded. Hence, even if we know the constituents of a given protein, it is hard to know how it will be folded and hence, what it will do. 

Proteins act in many different ways. They can be antibodies to a virus or bacteria. They can be enzymes performing a multitude of different biological functions. They can be messengers carrying signals between different cells. They provide the “muscle” in the “muscle” helping us move around (perhaps the most familiar function for laypersons). They store and transport different types of cells across various parts of the body. 

When a protein is folded “wrong”, it can misbehave causing diseases that can be life-threatening. 

Scientists use many experimental methods to figure out protein structures. This includes Nuclear Magnetic resonance imaging (NMRI), X-ray crystallography and cryo-electron microscopy (Cryo-EM), which can determine the individual atoms in a protein. These are all incredibly laborious and expensive. 

In CASP, teams use algorithms run on high-end computers to predict the ways in which a protein will fold. The latest CASP results indicate DeepMind’s AlphaFold program hit unprecedented levels of accuracy of close to 90 per cent in many predictions. An earlier AlphaFold version also topped the CASP scores in 2018 but it was much less impressive. 

Coronavirus and protein shape

Protein functions depend on protein shape. The interactions between a protein and another molecule is due to the proximity of certain amino acids in the protein to the molecule. If the shape changes, the function alters, or the protein becomes inactive. If researchers can predict the shape of a protein, they can predict how it will function. This is critical in drug development. For example, the spike in the coronavirus, SARS COV-2, is a protein shaped to adhere to and penetrate cells. Altering the shape could render the virus inactive.  

If you’re wondering why that name sounds familiar, DeepMind is the British company that cracked Go and Chess with its self-learning “Alpha” algorithms. It was subsequently bought by Google and is now a subsidiary of the IT giant. But it continues to work on this self-learning curve. 

The first 2018 iteration of AlphaFold applied the method of “deep learning” to structural and genetic data to predict the distance between pairs of amino acids in proteins. Then it tried to build a “consensus” model of what the protein should look like, says John Jumper who leads the project at DeepMind. 

The second iteration, AlphaFold2 developed an AI network using more information about physical and geometric constraints on how a protein folds. It was also more ambitious: Instead of predicting relationships between amino acids, the network directly predicts the final structure. 

Apart from the test problems, AlphaFold has also predicted structures of some of the SARS-CoV-2 proteins though these have not been confirmed experimentally yet. The program has also helped find the structure of a protein that stumped the Max Planck Institute for a decade. 

Assuming this CASP score wasn’t a fluke (which seems unlikely), scientists can now try to work out protein structures in less laborious ways. This should lead to far quicker drug development and an understanding of certain genetic diseases. Alphazero was hailed as a game changer by chess players. AlphaFold could change the real world.