June 26 News.Google DeepMind The official blog published a blog post yesterday (June 25) announcing the launch of the AlphaGenome models that are capable of multiple gene regulatory processes during theMore comprehensive and accurate predictions of human DNA single variant or mutation effects in the sequence and is planned to be made available to the scientific community through an API preview.
Note: AlphaGenome is a new artificial intelligence tool that can process DNA sequences up to one million letters long and predict thousands of molecular properties to characterize their regulatory activities. The model scores the effect of a genetic variation or mutation by comparing the predictions of mutated sequences with those of unmutated sequences.
AlphaGenome's training data comes from large public consortia such as ENCODE, GTEx, 4D Nucleome and FANTOM5, which cover important patterns of gene regulation in hundreds of human and mouse cell types and tissues.
The model uses a convolutional layer to detect short patterns in the genome sequence, a transformer to convey information about all positions in the sequence, and a series of final layers to transform the detected patterns into predictions of different patterns.
AlphaGenome has several distinguishing features: the ability to process long sequences of up to 1 million letters and make predictions at single letter resolution; the ability to make comprehensive multi-mode predictions; the ability to efficiently score variants; and, for the first time, the modeling of splice junctions.
The predictive power of the AlphaGenome is expected to help in a variety of fields such as disease understanding, synthetic biology, and basic research. Despite the model's remarkable progress, some challenges remain, such as accurately capturing the effects of distant regulatory elements.