Scientists announce that a long-standing and too complex scientific issue concerning the structure and behaviour of proteins has been solved successfully by a new method of artificial intelligence (AI).
For many years the UK-based AI company DeepMind has been wowing us with its parade of ever more advanced neural networks that keep pushing people on complex games like chess and go. However, all these radical developments were much more than recreational diversions.
In the background, DeepMind researchers tried to coax their IAs to solve even more basic scientific puzzles – such as new approaches to tackle diseases through the prediction of the infinitesimal, but essential aspects of human biology. As we know in disease biology the structure and behaviour of proteins are very important. Now with the current update of their AlphaFold AI engine, they seem to have achieved this very ambitious goal – or at least got us closer than scientists have ever had before.
Structure and Behaviour of Proteins
For about five decades, researchers have been working to predict how proteins can achieve their three-dimensional structure, and this is not an easy problem to solve.
The astronomical number of potential layouts is so mind-bogglingly massive that researchers have argued that it would take longer than the age of the Universe to sample all possible molecular patterns.
However, if we can solve this puzzle, known as a protein folding problem, it would represent a significant scientific breakthrough and would accelerate research efforts in areas such as drug discovery and disease modeling, and also lead to new applications far beyond health.
For this purpose, researchers have been collaborating for years to make gains in providing a solution to the protein-folding issue despite the scale of the challenge.
In the 1990s, a rigorous experiment called CASP (Critical Assessment of Protein Structure Prediction) began, challenging scientists to devise systems capable of predicting protein folding’s esoteric enigmas.
Now in its third decade, the CASP study seems to have generated its most viable solution – with DeepMind’s AlphaFold providing 3D protein structure predictions with unprecedented accuracy.
We have been stuck with this one problem – how to structure and behaviour of proteins fold – for almost 50 years,” says John Moult, co-founder of CASP at the University of Maryland.
Throughout the experiment, DeepMind used a new deep learning structure for AlphaFold that was able to interpret and calculate the ‘spatial graph’ of 3D proteins, forecasting the molecular structure underpinning their folded configuration.
The system, which was trained by analyzing a database of about 170,000 protein structures, brought its unique ability to the CASP challenge of this year, called CASP14, achieving a median score of 92.4 GDTT in its forecasts (Global Distance Test).
This is above the ~90 GDT threshold, which is generally considered viable with the same results obtained through experimental methods, and DeepMind says its forecasts are only about 1.6 angstroms on average (about the width of an atom).
“I nearly fell off my chair when I saw these results,” says Ewan Birney from the European Molecular Biology Laboratory, a genomics researcher.
I comprehend how rigorous CASP is-it practically guarantees that the difficult task of ab initio protein folding must be performed by computational modelling. It was humbling to see that they could do that so accurately with these models.
Many aspects will have to be understood, but this is a huge advance for science.
It should be noted that the study has not yet been peer-reviewed, nor published in a scientific journal (although researchers from DeepMind say that is on the way).
However, experts who are familiar with the field already recognize and applaud the breakthrough. This supercomputing work represents a stunning advance on the problem of protein folding, a tremendous biological challenge of five decades,” says Venki Ramakrishnan, President of the Royal Society.