Researchers at Google DeepMind have introduced AlphaGenome 2, an artificial intelligence model designed to understand how DNA variations alter the functioning of different cell types. This notoriously difficult problem could help identify the causes of certain diseases and guide new treatments.
- AlphaGenome 2 prédit comment les variants génétiques modifient l’activité des gènes selon le type cellulaire.
- Entraîné sur plus de 100 millions de profils cellulaires, il identifie près de deux fois plus de variants à effet que son prédécesseur.
- Son coût de calcul élevé l’empêche de remplacer AlphaGenome à court terme.
- Il est accessible pour la recherche non commerciale via une API, et une version commerciale est prévue en accès limité.
AlphaGenome 2 succeeds AlphaGenome, a genomic prediction model previously published by DeepMind. Almost all cells in the human body contain the same DNA, but their functions differ: a heart muscle cell, a skin cell, or an immune cell fighting an infection do not activate the same genes. These activation patterns also vary depending on what the cell is doing at a given time. A single genetic variant, meaning a change in a single letter of DNA, can therefore have very different effects depending on the cell’s type and state.
A model trained on millions of cellular profiles
The new model is designed to predict the effects of these variants on specific cell types and states. According to Google, it offers a much finer-grained view than more general models like AlphaGenome, and can highlight disease-related effects in specific cells that might have gone unnoticed.
AlphaGenome 2 was trained on over 100 million individual cellular profiles, covering 68 tissues. DeepMind claims the model transforms static cellular atlases—vast databases cataloging cells in detail—”into a predictive engine for human genetics.” According to the company, AlphaGenome 2 identifies nearly twice as many variants capable of altering gene expression in cells as its predecessor, which was presented as the state-of-the-art benchmark model.
The researchers also found that the model could make predictions about new cell types and new biological states without needing to be retrained. It can also identify gaps in experimental data, which could help scientists decide what to study next in the lab.
Adam Gayoso, a research scientist at Google DeepMind involved in the project, believes the major breakthrough lies in “connecting human genetics to cellular biology.” He told The Verge that the model allows for examining variants identified in large populations and seeing their effects at the cellular level, before potentially linking them to a disease.
Žiga Avsec, genomics lead at Google DeepMind, highlights the importance of this precision for drug discovery. “When developing drugs, you target specific proteins and specific cell types,” he explains. He noted that models of this type have only recently become feasible, thanks to technologies capable of capturing individual cells and their states at high resolution. DeepMind is now working to generate more data, including for cell states that are still poorly represented. Future versions could also be trained on data “across multiple species.” In the longer term, these models could fuel efforts to create virtual cells, digital simulations intended to replicate biological systems.
According to Avsec, AlphaGenome 2 is not expected to quickly replace its predecessor. Its more detailed predictions come with a higher computational cost and are “more expensive” to produce. This constraint would make it impractical for generating the genome-wide atlas published by DeepMind in September, which covers “every possible single-letter change in human DNA.”
Google states that AlphaGenome 2 is being made available for non-commercial research via the AlphaGenome API, an interface allowing developers to access the model. Limited access, in a “private preview,” is planned for commercial research.
This model is part of a series of efforts by Google and other AI labs to tackle major problems in science and medicine. Pushmeet Kohli, head of AI for Science research at DeepMind, hinted that other scientific results would follow. Anthropic is experimenting with drug design and has given Claude a wet lab, while OpenAI has solved a significant number of long-standing problems in mathematics and related fields.
