Unravelling the Mysteries of Life Itself
DeepMind’s incredible achievement with Alpha Fold is a sign of good things to come.
Beneath almost every biological process, proteins carry out the essential work that gives rise to the function of cells. Their name quite literally means ‘primary’ or ‘holding the first place’ due to the fundamental nature of their role in living things. While DNA may be the code through which life is written, proteins are the product through which living beings gain action. Whether it’s the activity of a pathogen in a host, muscular movement or even the creation of a thought inside your head, it is through proteins that these goals are met.
Developing our knowledge of proteins is of long standing significance for both medicine and our understanding of life. But for a great number of years, scientists have struggled with a particular issue. Known as ‘the protein folding problem’, it concerns the 3-D structure of proteins and how this shape specifies their function. Every protein is made up of chains of simpler molecules known as amino acids and, using DNA sequencing, we can determine the particular order of these for a protein. However, simply understanding the order doesn’t tell us what shape the chain will fold into – these large, complex models seem to almost randomly pick and choose which rules of nature they want to follow. With current methods, predicting the overall shape from this order is a lengthy, expensive and often woefully inept process. In 2018, DeepMind first rocked the field with the success of the AlphaFold system – an AI model capable of predicting protein structures and far outperforming others in the field.
Building upon their past success, DeepMind have developed their AI further, such that it can now predict a structure with astonishing accuracy. “This is a big deal,” John Moult, who co-founded and oversees CASP (Critical Assessment of Protein Structure Prediction), told Nature – “In some sense the problem is solved.”
To get a sense of the progress, consider the performances that have been recorded at the annual CASP competition. In 2018, Alpha Fold won with an average score of about 68% accuracy, a figure that was a substantial improvement from recent years and greatly exceeded expectations. The problem is not considered solved until an accuracy of 90% has been met, which as of November 2020 at CASP14, the latest iteration of Alpha Fold achieved.
The AI works by feeding the neural network at the heart of Alpha Fold examples of known protein structures and their sequences. Using deep learning techniques, it then creates a model from which it can predict the overall shape using two key factors: the distances between pairs of amino acids and the angle of the bonds that connects them.
The potential for gaining a greater understanding of how many diseases occur is just one example of the benefits down the line that are likely to come from this work. For many, the results came somewhat as a shock, as the protein folding problem had seen slow progress and a number of scientists predicted it would not be solved for years. What this means, is that perhaps one of the most exciting points to take is the clear effect that AI and technology can have on the life sciences.
This breakthrough can be taken alongside a number of other exciting developments that spell an optimistic future. Earlier in the year, the first human trial of a drug discovered by AI was launched, thanks to the AI platform Centaur Chemist, led by Exscientia. The project went from the discovery stage to the end of preclinical testing in only 12 months, a rate that is significantly faster than average, suggesting that AI could be used to reduce the eye-wateringly high spending used in drug discovery.
Similarly, pharma and biotech companies are quickly realising the potential for cloud-technology in their research. Sharing clinical trial data through cloud computing is something a number of firms have engaged to improve research methods. Cloud-computing is also being harnessed by Folding@Home, who are simulating protein dynamics using the processing power from volunteers’ computers.
Here at Hadean, we’re committed to using our platform to contribute to this wave of innovation in the life sciences, and we have already begun. In partnership with the Francis Crick Institute, we scaled a protein-docking technique called Cross-Docking using our Aether Engine. We achieved a 10% uplift in docked quality structure and a reduction in computational time and effort, solving a difficult 2018 CAPRI target (the sister competition of CASP for docking). We look forward to seeing how Hadean can be used to combine the same massive scale with the power of tools like DeepMind AI so can we attack even more complex problems.
With a new year on the horizon, the stage has certainly been set for what is likely to be an incredibly exciting union between technology and the life sciences.