Google DeepMind Unveils AlphaGenome 2: AI Predicts Genetic Variants Across 637 Cell Types

Google DeepMind has unveiled AlphaGenome 2, a breakthrough AI for genome mapping that transforms how scientists understand genetic variants and their role in disease1

1

2

2

. The DNA sequence model represents a fundamental leap in predicting genetic variants across hundreds of specific cell types, offering unprecedented precision in genetic research.

Replicating 18 Months of Lab Work in 30 Minutes

The power of AlphaGenome 2 became evident when it replicated a major inflammatory bowel disease discovery in just 30 minutes—work that originally took researchers 18 months of sophisticated molecular biology experiments1

1

. In 2024, scientists at the Francis Crick Institute identified ETS2 as a driver of inflammatory bowel disease, tracing it to a DNA variant in non-coding DNA sequences. "They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments," says James Lee, the immunologist who led the original study1

1

.

Training on 100 Million Individual Cells

Google DeepMind trained AlphaGenome 2 on more than 100 million individual cell profiles across 68 tissues, a massive expansion from its predecessor2

2

3

3

. This training enables the model to predict how genetic variants affect 637 human cell types and 311 mouse cell types. The model identifies nearly twice as many variants that change gene expression prediction in cells compared to previous state-of-the-art models3

3

.

Solving the Cell-Type Specificity Problem

The original AlphaGenome model had a critical limitation: it couldn't specify which type of cell a variant would affect1

1

. "Having something that cannot predict cell type specificity is a bit of a problem, which is why it hasn't been hugely useful for us," Lee explained1

1

. AlphaGenome 2 addresses this by fundamentally redesigning how the model treats cell types, using a conditional architecture that predicts DNA effects for one specific condition at a time3

3

.

Bridging Single-Cell Biology and Human Genetics

For Adam Gayoso, a Google DeepMind research scientist on the project, the breakthrough is "tying human genetics to cell biology"2

2

. The model gives researchers a way to drill down on genetic variants identified in large populations and see their effects at the cellular level, connecting those effects to the cellular roots of disease. Žiga Avsec, who leads the AlphaGenome team, emphasized this specificity matters for drug discovery: "When you develop drugs, you're targeting specific proteins and specific cell types"2

2

.

Predicting New Cell States Without Retraining

AlphaGenome 2 can make predictions about entirely new cell types and biological states without needing retraining2

2

3

3

. Scientists can provide single-cell gene activity data from disease states, and the model generates variant predictions. In testing with COVID-19 mRNA vaccine data, AlphaGenome 2 successfully distinguished vaccine-activated immune T cells from resting ones3

3

. This capability helps researchers identify variants underlying disease more quickly.

Decoding the 98% of DNA That Remains Mysterious

Non-coding DNA sequences comprise 98% of the human genome but remain among the hardest to decipher1

1

. These sequences don't code for proteins but regulate gene activity in specific cell types. The inflammatory bowel disease variant Lee's team studied sits in a "gene desert" far from protein-coding regions, making it particularly challenging to identify through traditional methods1

1

. AlphaGenome 2's ability to predict effects of these sequences represents a major advance in understanding genetic regulation.

Computational Costs and Practical Tradeoffs

The increased precision of AlphaGenome 2 comes with higher computational demands. Avsec noted the detailed predictions are "more expensive" to produce than its predecessor2

2

. This made it impractical to use AlphaGenome 2 for the genome-wide AlphaGenome Atlas that DeepMind released in September, which contains predictions for all 9 billion possible single-letter DNA changes in the human genome2

2

. Both models will continue serving different research needs.

Availability Through AlphaGenome API

Google DeepMind is making AlphaGenome 2 available for non-commercial research via the AlphaGenome API, with a limited private preview for commercial research2

2

3

3

. The model also performs cross-modality imputation, predicting unmeasured properties like chromatin accessibility when given just single-cell gene activity data, helping researchers prioritize which experiments to conduct next3

3

.

Future Directions in AI-Powered Biology

DeepMind plans to improve its genetics models by generating more data for underrepresented cell states2

2

. Avsec suggested future versions could be trained on data across multiple species, giving models a broader view of biology. Longer term, models like AlphaGenome 2 could feed into efforts to create "virtual cells"—digital simulations designed to mimic biological systems2

2

. Pushmeet Kohli, who heads DeepMind's AI for science research, has hinted that more scientific breakthroughs are coming2

2

.