AlphaFold, DeepMind鈥檚 artificial intelligence that predicts the structure of proteins, is a gift to biologists.
To understand life, we need to understand proteins. Living things are molecular machines, and most of the key components are made of proteins. You are reading this article, for instance, with proteins in your retina that can detect light, various proteins that make your muscles move and so on.
In one way, proteins are simple. They are made of chains that can contain up to 20 different amino acids. The sequence of amino acids in a chain is determined by the gene coding for a protein. Thanks to the cheap genome sequencing ushered in by the Human Genome Project, we now know the amino acid sequences of hundreds of millions of proteins found in viruses, bacteria, fungi, plants and animals.
Advertisement
In most cases, however, we still don鈥檛 have a clue what most of these proteins do or how they do it. Techniques such as deleting a gene using CRISPR and seeing its impact on an organism can give us a good idea of what a protein does, but for the how we need to work out a protein鈥檚 shape, or structure.
Take kinesins, the cell鈥檚 beasts of burden. with two stubby 鈥渓egs鈥�, pulling along cargoes attached to the other end of the protein. You just can鈥檛 understand them without knowing their structures.
The problem is that a simple chain can fold up into a vast number of different structures. In the past, determining the 3D structure of a single protein experimentally could take many years. Now, biologists can look at the predicted structure of almost any known protein in minutes, thanks to the database of nearly 200 million such structures just released by DeepMind.
The protein structures and open-source tools previously released by DeepMind have already accelerated many areas of research. Already, more than 100 studies citing AlphaFold .
For instance, nuclear pores control the transport of molecules into and out of the DNA-containing nucleus of a cell. These pores are made of more than 1000 proteins and were poorly understood. But within months of the initial release of AlphaFold, researchers published of the human nuclear pore.
Knowing protein structure is especially valuable for developing drugs. Most drugs work by binding to specific parts of proteins. This binding happens where part of the drug鈥檚 structure fits into part of the protein鈥檚 structure, like a 3D jigsaw puzzle. So if you know the protein structure, you can design drugs to fit rationally, rather than relying on trial and error.
It is worth noting that DeepMind has made its database of protein structures freely available. It could have chosen to follow a similar route to Celera, the company that tried to sequence the human genome before the publicly funded project so it could commercialise the data. It isn鈥檛 clear if or how DeepMind intends to make money from AlphaFold, but what is clear is that the treasure trove of structures it has released will greatly accelerate many areas of research.
That said, it would be wrong to suggest the problem of determining protein structure is completely solved. The accuracy of AlphaFold鈥檚 predictions vary, and for drug development in particular, accuracy really matters.
And as with the Human Genome Project, it is likely to take many years before the advances made with the help of AlphaFold result in treatments and products that benefit people directly.
But we have made far more progress in solving one of the toughest problems in science than most researchers imagined was possible two years ago. Biology will never be the same again.
