DeepMind unveils AlphaGenome to predict effects across thousands of genes

Google DeepMind researchers on Wednesday, Jan. 28, 2026, published AlphaGenome in the journal Nature, saying the A.I. was trained on a vast set of molecular data to make predictions about thousands of genes, including whether mutations can shut genes off or turn them on at the wrong time.

The paper follows DeepMind’s earlier success with AlphaFold2, for which two DeepMind scientists shared the 2024 Nobel Prize in Chemistry after the program transformed how researchers predict protein structures and spurred broad adoption in biology. AlphaGenome evolved from a 2021 A.I.

called Enformer and was trained at larger scale to predict 11 different genomic processes. The authors and outside researchers described it as state of the art and said it performed as well or better than other programs across those tasks. In reported tests the team said AlphaGenome accurately predicted the effect of mutations near the TAL1 gene that are known to activate the gene abnormally and can contribute to leukemia; researchers who uncovered those mutations called the result striking.

The AlphaGenome team acknowledged, and outside scientists stressed, that the model’s predictions still require laboratory validation and can weaken the farther they are from a focal gene.

alphagenome, google deepmind, alphafold2, enformer, tal1 gene, mutations near tal1, predict gene effects, splicing prediction, 11 genomic processes, trained on molecular data, laboratory validation required, leukemia-associated mutations, gene regulation loops, three billion base pairs, protein structures prediction, training at larger scale, ziga avsec, peter koo, mark gerstein, steven salzberg, katherine pollard, cold spring harbor laboratory, gladstone institutes