In 2024, two scientists, Demis Hassabis and John Jumper from Google DeepMind, won the 2024 Nobel Prize in Chemistry for their work in creating AlphaFold 2. An AI system that significantly enhanced scientists’ understanding of how proteins work and fail, AlphaFold 2 has become a tool used worldwide to facilitate scientific research. In fact, it’s led to the creation of new, disease-fighting proteins, some of which are currently undergoing clinical trials. Initially, AlphaFold 2 was trained on data from a 2019 GoogleDeepmind project which involved the collection of three billion base pairs, building blocks which make up genes that are used to make proteins, and thousands of experiments involving genes. AlphaFold 2 has significantly enhanced the pace of research, expediting new understanding that would have previously taken scientists numerous, painstaking experiments to determine.
More recently, researchers have unveiled AlphaGenome, a similar AI model that instead focuses on DNA, specifically the human genome. AlphaGenome, originally the Enforcer AI model introduced in 2021, has specific applications for understanding how cells splice genes and how they choose which genes to make proteins from. Splicing is when cells skip over certain parts of genes while creating proteins, allowing for the creation of hundreds of different proteins from the same gene. Incorrect splicing can lead to diseases, highlighting the importance of understanding splicing. Since there aren't clear indicators of where splicing will occur, scientists have painstakingly created databases of areas where splicing will occur over years of work. By training on pre-existing results, AlphaGenome can somewhat accurately make predictions that significantly enhance progress and our understanding of splicing.
AlphaGenome has also been able to help scientists understand how cells choose genes to make proteins. Within human cells, special molecules stretch DNA into intricate loops, sometimes exposing DNA to protein-making machinery and sometimes choosing to instead coil it aside. AlphaGenome has been able to help untangle this extremely complicated and intricate process for scientists.
So far, AlphaGenome has outperformed other related AI models in terms of accuracy. For instance, the AI model was able to successfully identify the effect of a mutation that was previously known to cause leukemia. This may be due to the fact that AlphaGenome is trained to handle 11 different processes, whereas most AI models only focus on one, an exhaustive understanding from different perspectives potentially enhancing accuracy. Some scientists however, criticize the technology, citing its unreliability. In addition, they point to base differences in scientists' data based on interpretation, potentially leading to inherent inaccuracies with AlphaGenome.
In reality, just between two people, there can be millions of genetic differences, severely complicating AI’s ability to make predictions. These issues, along with others, will mean it will likely be years before technologies such as AlphaGenome can actually be used in the healthcare industry. This AI model does represent a major achievement in science though, a reminder of the potential of AI.