How AI Changed the World of Protein Science
20th December 2022 - Last modified 4th July 2024
20 years of Alto. 20 years of science. #19
By Kelly Buggy PhD, Senior Manager
As part of Alto Marketing’s 20 year celebrations, we’re looking back at some of the most important advances in science over this time in our blog series “20 years of Alto. 20 years of science.” In this blog, one of Alto’s Senior Managers, Kelly Buggy, looks at the story of DeepMind’s AlphaFold algorithm and how AI has changed the world of protein science.
AlphaFold reveals the structure of the protein universe, DeepMind [1]
A.I. Predicts the Shape of Nearly Every Protein Known to Science, New York Times [2]
DeepMind uncovers structure of 200m proteins in scientific leap forward, The Guardian [3]
On 28th July 2022, the scientific world awoke to one of the most exciting developments of our time. AlphaFold – an artificial intelligence (AI) system developed by DeepMind – had deciphered the structure of almost every protein known to science.
In partnership with EMBL’s European Bioinformatics Institute (EMBL-EBI), DeepMind announced the expansion of its AlphaFold database to 200+ million protein structures. This 200-fold increase (up from 1 million structures) covered structures across the whole protein universe, including predicted structures for plants, bacteria, animals, and other organisms [1] and was the latest development in an exciting project that began for DeepMind in 2016.
On the release date, these 200+ million predicted protein structures became available for download and were added to the main protein database, UniProt. This wealth of freely accessible information holds incredible promise and potential for our understanding of biology, with applications in pretty much every branch of science. From antibiotic resistance to climate change, scientists now have another great tool in their research arsenal.
“It will change everything”
As a structural biologist in a previous life, the emergence of a leading protein structure prediction program was pretty exciting stuff for me! And of course, it was a turning point for scientists around the world, who had predicted the impact that AlphaFold from DeepMind would have as soon as it properly entered the scene in November 2020.
Following the 14th biennial protein-structure prediction challenge, CASP14 (Critical Assessment of Structure Prediction) in 2020, it was announced that the DeepMind team using AlphaFold had outperformed over 100 other teams. AlphaFold even managed to solve the structure of membrane proteins, which are one of the most challenging protein structures to determine.
Although DeepMind had excelled in the competition before in 2018, the 2020 results showed it to be miles ahead of the rest of the entrants. Articles in Nature [4] and Science [5] included quotes from experts in the field, heralding AlphaFold and its potential impact as revolutionary.
“This is a big deal,” says John Moult, a computational biologist at the University of Maryland, who co-founded CASP in 1994 to improve computational methods to help solve the problem of accurately predicting protein structures. “In some sense the problem is solved.” [4] “This is a 50-year-old problem,” he further explained. “I never thought I’d see this in my lifetime.” [5]
“It’s a game changer” said Andrei Lupas, an evolutionary biologist at the Max Planck Institute for Developmental Biology in Tübingen, Germany. “This will change medicine. It will change research. It will change bioengineering. It will change everything.” [4]
“What the DeepMind team has managed to achieve is fantastic and will change the future of structural biology and protein research,” said Janet Thornton, director emeritus of the European Bioinformatics Institute. [5]
You don’t get much higher praise than that.
The next year, DeepMind had more to share. In July 2021, they released and open-sourced the AlphaFold code and launched the AlphaFold Protein Structure Database. At the same time, peer-review research demonstrated high-quality predictions for the structure of every protein in the human body – offering a clear picture of the human proteome. In addition, 3D protein structures from other key organisms critical for scientific research like yeast, the fruit fly, and the mouse were also predicted using AlphaFold.
What is AlphaFold?
AlphaFold is fast on its way to becoming one of the most integral tools of protein science – and biological research itself. Because a protein’s structure is so heavily linked to its function, gaining a deeper understanding of how a protein works and its role in health and disease (and other wide-ranging applications!) through structure prediction helps scientists immensely.
But what exactly is AlphaFold? And how does it achieve these results?
DeepMind is the creator of AlphaFold, and is a London-based division of Alphabet, Inc. which is a collection of companies that includes Google. The aim of AlphaFold is to develop general purpose AI technology through deep learning on a convolutional neural network. This means that it learns directly from data for high levels of accuracy.
We know that proteins are made up of amino acids that assemble in a particular sequence, according to their corresponding sequence of DNA. Amino acids interact with each other as a result of different repulsion and attractive forces, which then determines how the amino acid sequence folds to create the distinct three-dimensional structure of each individual protein.
The idea of computationally predicting a protein’s structure by determining all of the individual amino acid interactions within its sequence is not new and was first proposed in the 1960’s. But it’s probably fair to say that progress in this field has been slow. That is, until the AlphaFold algorithm, which was taught using the sequences and structures of around 100,000 known proteins. By following the rules learnt from these known structures, AlphaFold predicts the shape of a protein in minutes at atomic accuracy.
From defeating the hardest board game in the world, Go, and completing the whole library of Atari games to the development of the AlphaFold algorithm for protein structure prediction, revolutionary AI technology from DeepMind has been paramount to many major accomplishments.
Then and now?
Solving a protein structure experimentally is an expensive, laborious task that can often result in little success or reward – and I should know! I spent three years studying a particular protein interaction between calmodulin (a protein that mediates the calcium regulation of a wide range of physiological processes) and a protein kinase involved in protein translation. This was to try and better understand the role of this protein interaction and downstream signalling pathway in disease.
X-ray crystallography and cryo-electron microscopy (EM) are popular experimental methods to determine a three-dimensional protein structure, as well as nuclear magnetic resonance (NMR) to study proteins in solution. But it’s a tricky business – and according to one estimate, only 170,000 of the more than 200 million proteins known to exist have an experimentally determined structure [5].
The availability of accurately predicted protein structures is a significant breakthrough that demonstrates the impact that AI can have on science. The applications of this new knowledge and insight into individual protein structures (and therefore functions) is seemingly endless.
What now for protein science?
The 200+ million protein structures available thanks to AlphaFold are already contributing to our scientific knowledge, becoming an essential tool for researchers around the world. You can find out more about the different projects underway on the DeepMind website but some highlights include:
—A partnership with the Drugs for Neglected Diseases initiative (DNDi) is advancing drug discovery for neglected diseases like Chagas disease and leishmaniasis, which impact millions of people in often the poorest and most and vulnerable communities.
—Researchers at the Centre for Enzyme Innovation (CEI) are discovering and engineering enzymes for breaking down single-use plastics.
—At ETH Zurich, research is underway to study the evolution of proteins by looking at how changes in our DNA alter protein structures to result in changes to our traits.
These are just some specific examples of the exciting scientific research inspired and facilitated by AlphaFold structure predictions, so let’s finish with a quote that sums up how AI really has changed the world of protein science:
“This will be one of the most important datasets since the mapping of the Human Genome.” Professor Ewan Birney, EMBL Deputy Director General and EMBL-EBI Director.
References
(1) https://www.deepmind.com/blog/alphafold-reveals-the-structure-of-the-protein-universe
(2) https://www.nytimes.com/2022/07/28/science/ai-deepmind-proteins.html
(4) Callaway, E., ‘IT WILL CHANGE EVERYTHING’: AI MAKES GIGANTIC LEAP IN SOLVING PROTEIN STRUCTURES. Nature. 2020; 588 203-204
(5) Service, R.F., ‘The game has changed.’ AI triumphs at solving protein structures. Science. https://www.science.org/content/article/game-has-changed-ai-triumphs-solving-protein-structures
