Google DeepMind has unveiled AlphaGenome 2 \u2014 a new artificial intelligence model that predicts how changes in DNA affect gene expression in specific cell types. The system was trained on over 107 million cell profiles obtained from 68 types of human and mouse tissue. During testing, it identified almost twice as many genetic variants that alter gene activity as the previous version. The developers hope that the technology will help to identify the molecular causes of diseases more quickly and find promising targets for new medicines.<\/p>\n
Most cells in the human body contain the same DNA, but use it in different ways. For example, a genetic mutation may affect the function of an immune cell, but have almost no effect on heart muscle cells. It is difficult to identify such differences based solely on DNA sequence.<\/p>\n
Previous AlphaGenome<\/strong> previously analysed genetic variants mainly at the tissue level. The new model is more accurate: it takes into account the specific cell type and its biological state, including changes that occur during disease or an immune response.<\/p>\n The system analyses DNA sequences up to one million nucleotides in length and predicts how a particular genetic change will affect gene activity. It can also predict cellular characteristics that researchers have not directly measured.<\/p>\n When tested on the TenK10K dataset, which covers 28 types of immune cells, AlphaGenome 2 identified 30 per cent of causative genetic variants affecting gene activity. The previous model, which had been further adapted for single-cell data, achieved a result of 16 per cent.<\/p>\n There has been an even more marked improvement in determining the direction of a mutation\u2019s effect \u2014 whether it enhances or suppresses gene activity. In this respect, the result has risen from 23 per cent to 43 per cent.<\/p>\n The researchers also tested the model against a previously known genetic mechanism underlying inflammatory bowel diseases. AlphaGenome 2 was able to reproduce the link between a specific DNA region and the over-activity of the ETS2 gene in the gut\u2019s immune cells.<\/p>\n This is an important finding, as genetic changes associated with diseases are often located far from the genes whose function they affect.<\/p>\n One of the key features of AlphaGenome 2 is its ability to handle previously unknown cellular states. Researchers can upload the results of studies involving even a few dozen cells, after which the system uses this data to generate new predictions without the need to retrain the model.<\/p>\nThe new model identifies almost twice as many significant genetic variants<\/h2>\n
AI can analyse new cells without the need for retraining<\/h2>\n