Google DeepMind Unveils AlphaGenome 2 to Predict Genetic Variant Effects in Specific Cell Types

Google DeepMind has announced the launch of AlphaGenome 2, a next-generation DNA sequence model designed to predict how genetic variants and DNA sequences impact specific human cell types and disease states [1]. The system succeeds last year's original AlphaGenome model, aiming to provide researchers with a detailed tool to uncover cellular drivers of disease [1][2].

Bridging Single-Cell Biology and Human Genetics

Every cell in the human body contains the same DNA blueprint, but individual cell types and states—such as a heart muscle cell or an immune cell—activate different genes to perform their functions [1]. While global initiatives like the Human Cell Atlas map gene expression across millions of cells, predicting how subtle DNA variations alter this expression remains challenging [1]. AlphaGenome 2 addresses this by zooming in from broad tissue averages to individual cell types and states [1].

Trained on more than 107 million individual cell profiles across 68 tissues from human and mouse samples, AlphaGenome 2 uses a redesigned architecture [1]. Unlike its predecessor, which predicted properties across thousands of experiments simultaneously, AlphaGenome 2 functions as a conditional model [1]. It takes a million-letter DNA sequence alongside a target biological condition—such as a specific tissue, cell type, or state—as input at every stage [1]. According to Google DeepMind, this conditional approach enables the model to identify nearly twice as many variants that modify gene expression compared to the previous state-of-the-art model [1].

Applications in Disease Research and Novel Cell States

AlphaGenome 2 is capable of cross-modality imputation, meaning it can predict missing experimental properties—such as chromatin accessibility—when provided with a single measurement like single-cell gene activity [1]. Crucially, the model can generalize to entirely new single-cell experiments and states without requiring retraining [1]. For instance, when provided with unseen single-cell data from a COVID-19 mRNA vaccine study, the model distinguished resting immune T cells from vaccine-activated ones, indicating that a severe COVID-19-linked variant disrupts gene activity more strongly in activated cells [1].

Google DeepMind researchers have also precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions [1]. These precomputed scores assist researchers in mapping enhancer-gene links in primary human cell types and pinpointing causal genome-wide association studies (GWAS) variants for complex traits ranging from height to type 2 diabetes [1]. Adam Gayoso, a research scientist at the project, noted that the breakthrough lies in tying human genetics to cell biology [2]. Žiga Avsec, genomics lead at Google DeepMind, stated that such high specificity is vital when targeting specific proteins and cell types for drug development [2].

Availability and Computational Demands

While AlphaGenome 2 offers higher granularity than its predecessor, it comes with increased computational demands [2]. Avsec explained that the model is more expensive to produce than the original version, making it impractical for generating sweeping genome-wide atlases that cover every single-letter change [2]. Google has made AlphaGenome 2 available immediately for non-commercial research through the AlphaGenome API, alongside a private preview for commercial research [1][2].

Understanding how DNA variations influence cellular mechanisms is a foundational challenge in human genetics and drug discovery. Large-scale global initiatives like the Human Cell Atlas have worked to map gene expression across millions of cells, while genome-wide association studies (GWAS) routinely flag broad stretches of DNA linked to specific traits and diseases. However, connecting these association signals to precise cellular targets and disease mechanisms has historically required extensive experimental validation.

For researchers and the broader biotechnology sector in India and globally, genomic AI models that operate at single-cell resolution offer a powerful method to bridge human genetics and cellular biology. By accurately identifying causal genetic variants and the specific cell types where they operate, models like AlphaGenome 2 can assist scientists in prioritizing laboratory experiments and advancing targeted drug discovery programmes.

Predictive DNA models with single-cell resolution could significantly accelerate target identification and validation in pharmaceutical research, potentially reducing the time required to translate genetic insights into clinical treatments. Over the longer term, improvements in genomic modeling and cross-species data training could contribute to wider computational simulations of biological systems, such as virtual cells.

Because AlphaGenome 2 requires higher computational resources than its predecessor, broad-scale commercial adoption and extensive deployment may be constrained by computing costs. Furthermore, while the model generalizes to unseen cell states, its precomputed predictions and computational inferences still require empirical validation in physical laboratories.