Google DeepMind Launches AlphaGenome 2 to Decode Cellular Impacts of Genetic Variants

Image: Google’s AlphaGenome 2 zooms in on the cellular roots of disease · Source

Google DeepMind's new AI model, AlphaGenome 2, predicts how DNA changes affect specific cell types and states, advancing understanding of disease origins and aiding drug development. It improves on its predecessor by using over 100 million cell profiles for detailed genetic insights.

Google DeepMind has unveiled AlphaGenome 2, a sophisticated artificial intelligence model designed to interpret how genetic variations influence different human cell types and states. This breakthrough offers researchers unprecedented fine-grained insights into the cellular mechanisms underlying diseases, potentially guiding new therapeutic developments.

While all human cells share the same DNA, distinct cell types—including heart muscle, skin, or immune cells—activate different genes based on their function and condition. Consequently, a single genetic variant can exert diverse effects depending on the specific cell context, complicating efforts to connect genetics to disease.

AlphaGenome 2 addresses this challenge by predicting variant effects at the level of individual cell types and their dynamic states. Building on the original AlphaGenome model released last year, it was trained with data from more than 100 million single-cell profiles drawn from 68 tissues. This extensive dataset enables the AI to serve as a predictive tool, transforming static cell atlases—comprehensive maps cataloging individual cells—into actionable genetic insights.

DeepMind reports that AlphaGenome 2 identifies nearly twice as many gene expression-impacting variants compared to previous models. It can also generalize predictions to novel cell types or states without retraining and highlight gaps in existing experimental data, guiding future laboratory investigations.

According to DeepMind’s genetics lead Žiga Avsec, this level of specificity is crucial for drug discovery since treatments often target particular proteins within certain cell types. Research scientist Adam Gayoso emphasized that AlphaGenome 2 "ties human genetics to cell biology," enabling scientists to trace the impact of genetic variants down to cellular activity and disease relevance.

The development of AlphaGenome 2 has been made possible by recent advances in single-cell technologies that capture individual cellular states with high resolution, creating the rich datasets necessary for training. DeepMind is working to expand these datasets, including underrepresented cell states, and envisions future models trained on multi-species data to broaden biological understanding.

While AlphaGenome 2 offers greater detail, it requires higher computational resources and is not intended to immediately replace the original AlphaGenome, which still serves large-scale genetic scanning tasks.

Starting today, Google is providing access to AlphaGenome 2 for non-commercial research through the AlphaGenome API, with a limited private preview available for commercial use. This release marks a significant step in leveraging AI to unravel the complex genetics behind human diseases, complementing ongoing efforts in AI-driven science and medicine by DeepMind and other organizations.