Google DeepMind has unveiled Alphagenome 2, a genome artificial intelligence (AI) model that predicts how DNA variants affect cells depending on their type and state. The model is expected to help pinpoint the causes of disease and find new treatments by precisely analysing effects that can appear differently across cells even with the same genetic variant.

Google said on Nov. 11 local time that Alphagenome 2 is the follow-up to Alphagenome, a genome prediction AI model released last year. Most cells in the human body have the same DNA, but heart muscle cells, skin cells and immune cells activate different genes depending on their type and operating state. As a result, the impact of the same genetic variant can vary depending on which cell shows it and in what state.

Alphagenome 2 is designed to reflect those differences by predicting the effects of genetic variants by individual cell type and biological state. It is designed to capture disease-related effects at a finer level than models that analyse a broad range like the existing Alphagenome.

The model learned from more than 100 million single-cell profiles secured from 68 tissues. DeepMind described it as turning a static cell atlas into a predictive engine for human genetics. The company said Alphagenome 2 identified nearly twice as many genetic variants that could change gene expression within cells as the previous best-performing model.

Researchers said Alphagenome 2 was able to predict new cell types and biological states without additional retraining. It can also help researchers choose targets for follow-up experiments by finding gaps in experimental data.

Adam Gayoso (애덤 가요소), a Google DeepMind research scientist who participated in the project, described the key to the achievement as "connecting human genetics and cell biology". He said it allows researchers to analyse how genetic variants found in large population studies affect cells and link them to disease.

Žiga Avsec (지가 아브섹), head of genomics at Google DeepMind, also stressed that precision analysis at the cell level is important to develop treatments for specific diseases. "When developing drugs, you target specific proteins and specific cell types," he said, explaining the need for fine-grained predictions.

Alphagenome 2 does not immediately replace the existing model. Avsec said the computing costs needed for more precise predictions are higher than for the existing model. That makes it difficult to use right away for tasks that analyse vast ranges, such as the atlas of possible DNA base changes across the entire human genome that DeepMind released in September.

Google will provide the model for non-commercial research through the Alphagenome API starting on the day. For commercial research, it will be released in the form of a limited, private preview.

DeepMind is improving its genomics model by securing additional data by cell state, and it is also considering ways to train on multiple types of data together. Over the long term, models such as Alphagenome 2 could also be used to develop "virtual cells" that digitally simulate biological systems. As AI research companies, including Google, expand the use of AI to solve difficult problems in science and medicine, DeepMind's new model is also drawing attention as progress in related research.