Google’s AlphaGenome AI helps scientists study non-coding DNA, speeding research on genetic variants linked to diseases
PARIS: Google unveils AlphaGenome, an AI designed to decode the hidden 98% of human DNA, offering new insights into disease-linked genetic variants.
The deep learning model, called AlphaGenome, was presented as a breakthrough for studying the roots of difficult-to-treat genetic diseases.
Pushmeet Kohli, vice president of research at Google DeepMind, said the first complete human genome map in 2003 “gave us the book of life, but reading it remained a challenge.”
He explained that understanding the grammar of our DNA is the next critical frontier for research.
Only around 2% of our DNA contains instructions for making proteins, which build and run the body.
The other 98% was long dismissed as “junk DNA,” but is now believed to act like a conductor directing genetic information in our cells.
AlphaGenome aims to understand these non-coding sequences, which contain many variants associated with diseases.
The model was trained on public data measuring non-coding DNA across hundreds of human and mouse cell and tissue types.
It can analyse long DNA sequences and predict how each nucleotide pair influences different biological processes within the cell.
Lead study author Ziga Avsec said long sequences are required to understand the full regulatory environment of a single gene.
The model’s high resolution allows scientists to study the impact of genetic variants by comparing mutated and non-mutated sequences.
Co-author Natasha Latysheva said AlphaGenome can accelerate understanding by helping to map functional elements and their molecular roles.
The tool has already been tested by 3,000 scientists across 160 countries and is open for non-commercial use.
Ben Lehner, a researcher at Cambridge University who tested the model, said it “does indeed perform very well.”
He explained that identifying genomic differences linked to diseases is a key step towards developing better therapeutics.
However, Lehner cautioned that AlphaGenome is far from perfect because AI models are only as good as their training data.
Robert Goldstone of the UK’s Francis Crick Institute said the tool is not a magic bullet for all biological questions.
Gene expression is influenced by complex environmental factors that the model cannot see, he noted.
Despite this, Goldstone hailed AlphaGenome a
