AlphaGenome 2 Predicts How DNA Variants Act in Specific Cells

Introduction

Every cell carries essentially the same DNA, yet a cardiac muscle cell, an immune cell, and a developing blood cell read different parts of that blueprint. A disease-associated variant may therefore have little visible effect in a tissue-wide average while strongly changing regulation in one cell type or one biological state. Google DeepMind’s AlphaGenome 2 is designed to model that missing layer of specificity, linking DNA sequence variation to the cells in which its effects actually appear.

Key points

• A conditional architecture. The original AlphaGenome predicted many tissue and cell-line outputs in parallel from a DNA sequence of up to one million letters. AlphaGenome 2 instead receives a target tissue, cell type, or state alongside the sequence and uses that condition throughout the model. This lets it learn that the same regulatory region can be read differently by different cells.
• Single-cell training at scale. The model was trained on bulk and single-cell data from human and mouse samples, covering more than 107 million individual cell profiles across 68 tissues. Its outputs include gene expression, histone modifications, chromatin accessibility, and DNA methylation.
• Higher variant recall. On the TenK10K dataset, which maps 28 primary immune cell types, DeepMind reports that AlphaGenome 2 identified 30% of causal variants, compared with 16% for a single-cell-fine-tuned version of the earlier model. It captured the direction of gene effects at rates of 43% and 23%, respectively. The company also says the new system finds nearly twice as many variants that alter cellular gene expression.
• Generalization to unseen states. Researchers can provide a single-cell RNA-sequencing snapshot from only a few dozen cells. AlphaGenome 2 compresses that information into a cell-state representation and uses it to condition predictions without retraining. In an example based on an mRNA COVID-19 vaccine study, it distinguished activated from resting T cells and predicted a stronger effect for a severe-COVID-associated variant in the activated state.
• Cross-modality imputation. If an experiment measures only one property, such as single-cell gene activity, the model can estimate other properties, including chromatin accessibility. That capability may help researchers decide which experiments to run next.

Why it matters

The central shift is from asking whether a variant affects a broad tissue to asking which cell reads it, and under what circumstances. That distinction is important for immune and inflammatory diseases, where a regulatory change may be confined to a specialized cell population. DeepMind says AlphaGenome 2 computationally reproduced a reported inflammatory bowel disease mechanism in which a distant regulatory variant selectively increases ETS2 activity in gut macrophages while leaving other cell types and genes largely unaffected.

The model should still be viewed as a prioritization and hypothesis-generation system, not as a replacement for functional experiments. Predictions depend on the quality and definition of the single-cell state, and computational agreement does not by itself establish causality in patients. Even so, a model that can connect newly collected cell profiles with genome-wide variant effects could make the growing volume of single-cell atlas data more useful for human genetics. AlphaGenome 2 is available through an API for non-commercial research, while commercial research access is in private preview.

Source: Google DeepMind