As Nobel prize season approaches, there’s nothing scientists like more than to speculate about who is going to get a gong. One possibility mooted this year – as is so often the case these days, it could be for either the chemistry or the medicine/physiology prize – is a reward for the development of a computer algorithm called AlphaFold that uses artificial intelligence to predict the shape of protein molecules based on nothing more than a knowledge of the sequence of building blocks (amino acids) it is made from.
AlphaFold was devised by the AI company DeepMind, a subsidiary of Google led by computer scientist Demis Hassabis. First unveiled in 2018, the algorithm seemed to finally crack one of the most challenging and coveted goals in modern molecular biology: to figure out the connection between a protein’s sequence (which to a first approximation can be deduced from the corresponding gene sequence in DNA) and its three-dimensional structure.
That structure holds vital clues to the molecule’s biological function. Many proteins are enzymes that facilitate biochemical reactions in our cells and steer them towards the required outcome. If we know an enzyme’s structure, we might be able to devise drug molecules with the right shape to bind to the protein and intervene in its function, for example stopping it from carrying out some process injurious to health. Traditionally, working out a protein’s structure has depended on a technique called X-ray crystallography, which is laborious and only works if the protein can be crystallised (not all of them can). Other methods now also exist for protein structure determination, but still it’s painstaking work.
AlphaFold seemed to render all that effort obsolete. In 2022 researchers used the algorithm to predict the structures of more than 200m proteins – a thousand times more than had been found experimentally, and encompassing virtually every known protein. Better still, the DeepMind team has now tweaked AlphaFold so that it can predict how proteins interact with each other and with other biomolecules such as DNA and lipids.
So can all of molecular biology now be elucidated on the computer at the push of a button? Not exactly. For one thing, AlphaFold’s structural predictions vary in quality, and don’t always match the experimental results in fine enough detail to be reliable for drug discovery. What’s more, structure isn’t everything: the functions of some proteins depend crucially on their wobbles and floppiness. More importantly, finding a candidate drug molecule that binds to a protein is only the first step in a long and difficult path to a working drug; more often than not, such binding doesn’t turn out to make a difference at the physiological level. AlphaFold looks sure to be a useful tool in drug discovery, but far from a panacea.
It might turn out that its biggest value lies elsewhere. One of the most exciting prospects is to use the algorithm in reverse, as it were: to choose a particular protein shape and then figure out which sequence will produce it. In this way researchers can design protein molecules very different from those found in nature that might be able to conduct entirely new enzymatic tasks or to act as building blocks for new materials (silk, for example, is a natural protein material). Biochemist David Baker of the University of Washington is a leader in this field and has been tipped to join Hassabis on the Nobel list: both researchers have been awarded a Breakthrough Prize, the international awards worth $3m each and funded by billionaire entrepreneur Yuri Milner. Baker’s team have used AlphaFold in conjunction with their own software to design proteins with arbitrary, non-natural shapes such as rings – and then to show that real proteins with these tailored sequences do indeed fold up into the targeted shape.
AlphaFold might also help to answer fundamental questions in biology. A team at the University of Glasgow has used it to work out the structures of proteins used by flaviviruses – a family that includes those responsible for hepatitis C and dengue fever – and thereby to spot evolutionary links between different viral pathogens that aren’t obvious from looking at their genomes (very different genome sequences can turn out to encode proteins with similar functions). The researchers deduced that some of these viruses apparently captured a crucial infection-enabling enzyme from bacteria way back in their evolutionary history. This sort of information might help for developing vaccines against these potentially fatal conditions.
Personally I doubt that AlphaFold will be judged Nobel-ready just yet. But it’s likely that its time will come.
