I had already done it with Alpha Fold, an artificial intelligence that predict how a structure will be folded (a characteristic that determines its abilities) and that was done with the Nobel Prize in Chemistry in 2024. And now Google looks for something similar with Alphagenome.
It is, according to Google’s own statement, “a new artificial intelligence tool (AI) that predicts more completely and precisely how individual variants or mutations in human DNA sequences impact a wide range of biological processes that regulate genes ”.
Alphagenome, covers both the coding regions and the non -coding of the genome and offers a unified vision of the effects of the variants as never before. Provides information resolution to the long -range genomic analysis, decoding the impact of mutations with unprecedented speed, scale and depth.
The model processes up to one million base pairs in a single pass and predict thousands of molecular propertiesincluding gene expression, union patterns, protein binding sites and chromatin accessibility (the substance that composes chromosomes) in various types of cells. It is the first time that such a wide range of regulatory characteristics can be jointly modeled using a single AI system.
Alphagenome architecture first uses convolutionary layers (“digital neural networks” that detect patterns) to detect similarities in the DNA sequence and then apply transformers to Share information throughout the genetic code. A final set of layers converts these patterns learned into predictions for various genomic characteristics.
During training, all single -sequence calculations are distributed among multiple interconnected tensioning processing units (TPU), allowing large -scale efficient processing. The surprising thing is that The entire model trained in just four hours (using public data) and with half of the computational budget required by its predecessor, enforce.
Currently, Alphagenome is The only model capable of jointly predicting all the molecular modalities evaluatedexceeding or matching specialized models in 24 of the 26 reference tests.
Unlike the previous models that prioritized the length of the sequence in favor of the resolution, Alphagenome manages both precisely. Captures the long -term genomic context and It offers base predictions, which reveals information on disease biology, research of rare variants, synthetic DNA design And more.
An outstanding feature of the new model is its variant score system, which effectively compares mutated and not mutated DNA to evaluate the impact on different modalities. But its impact can also be seen in other sectors: in synthetic biology, for example, Alphagenome could help design regulatory sequences that activate genes selectively, for example, in nerve cells, but not in muscle cells.
In a trial case, Alphagenome predicted precisely how a mutation linked to leukemia introduces a reason for union to myb DNA that activates the Tal1 gene, imitating the mechanisms known in acute lymphoblastic leukemia of T and T lymphoblastic demonstrating its ability to connect non -coding variants with disease genes.
Although Alphagenome represents an important advance, it is not designed or validated for personal genomic interpretation or for clinical use. It also faces challenges when modeling very distant regulatory interactions, especially those to more than 100,000 DNA lettersand the time to fully capture the specific patterns of cells and tissues.
Even so, those responsible claim that it feels a solid base for future expansion, Con potential to adapt to other speciesModalities and Specific Laboratory data sets.
Alphagenome is now available in preliminary version for non -commercial use through the Alphagenome API. Google invites researchers from all over the world to explore cases of use, ask questions and share comments. IA -based tool predictions are destined exclusively for research purposes.
“We hope Alphagenome helps us better understand the complex cellular processes encoded in DNA and drive new discoveries in genomics and medical care”, Concludes the statement.