DeepMind Unveils AlphaGenome 2 to Enhance Understanding of DNA Variants in Human Cells

DeepMind has announced AlphaGenome 2, an advanced model designed to predict how DNA sequences and their variations impact specific human cell types. This tool aims to assist scientists in uncovering the cellular drivers of diseases.

Cells serve as the fundamental building blocks of life, each carrying the same DNA blueprint. However, different cell types, such as heart muscle cells or immune cells, activate various genes to perform their specific functions. While initiatives like the Human Cell Atlas are mapping gene expression across millions of cells, predicting how subtle DNA variations influence this expression remains a significant challenge, crucial for understanding human traits and diseases.

Last year, DeepMind launched AlphaGenome, which predicted how DNA variants affect molecular activity across broad tissues. The subsequent AlphaGenome Atlas expanded this reach by providing predictions for all 9 billion possible single-letter changes in the human genome. These tools have enabled researchers to prioritize and interpret variant effects, facilitating discoveries in rare diseases and complex traits.

AlphaGenome 2 marks a significant improvement in predictive performance by focusing on individual cell types and states. This model can identify nearly twice as many variants that alter gene expression in cells compared to its predecessor. It has been trained on over 100 million individual cell profiles from 68 tissues, transforming static cell atlases into a predictive tool for human genetics.

Redesigning AlphaGenome for Specific Cell Types

The original AlphaGenome model processed a million-letter DNA sequence to predict properties across thousands of experiments simultaneously. While efficient, this approach diluted the model’s capacity to capture subtle differences between closely related cell types. In contrast, AlphaGenome 2 employs a conditional model, predicting DNA effects for one specific biological condition—defined as a tissue, cell type, or state—at a time.

This redesign allows the model to learn how the same DNA segment is interpreted differently across various cell types. Additionally, it enables cross-modality imputation, allowing AlphaGenome 2 to predict missing experimental data. For instance, if a scientist inputs a single measurement, such as gene activity, the model can predict unmeasured properties like chromatin accessibility.

Enhanced Predictions and Variant Recall

AlphaGenome 2 excels in predicting gene regulation within specific cells, aiding researchers in identifying disease-relevant variants. It demonstrates improved specificity in explaining gene expression differences across cell types within a tissue, including tracking changes as human blood stem cells mature into red blood cells.

On the TenK10K dataset, which maps 28 primary immune cell types, AlphaGenome 2 identified nearly twice as many causal variants and directional gene effects compared to the original model fine-tuned on single-cell data. This enhanced precision allows researchers to discover overlooked candidate variants.

In a recent study, AlphaGenome 2 computationally reproduced findings that a variant in a distant regulatory switch drives inflammatory bowel disease (IBD) by overactivating the ETS2 gene in gut immune cells. The model pinpointed the causal variant and demonstrated its selective effect on the ETS2 gene in gut macrophages.

Generalizing to New Cell Types and States

AlphaGenome 2 can generalize to new single-cell experiments without requiring retraining. Scientists can upload a snapshot of gene activity from a few dozen cells, which the model can then use to predict variant effects and gene regulation for previously unseen cell states. For example, it distinguished between vaccine-activated immune T cells and resting ones in a COVID-19 mRNA vaccine study.

Connecting Human Genetics and Cell Biology

Linking human genetics to cell biology is essential for understanding disease. AlphaGenome 2 aids researchers by precomputing predictions for trait-associated variants across various human cell and tissue conditions. This includes mapping enhancer-gene links in real human cell types and pinpointing causal variants from genome-wide association studies (GWAS).

As scientists continue to map the human body at the single-cell level, AI models like AlphaGenome 2 will evolve. DeepMind is collaborating with researchers at the Wellcome Sanger Institute to profile a broader range of human cell types and states. AlphaGenome 2 is currently available for non-commercial research through the AlphaGenome API and in private preview for commercial research.

AlphaGenome 2 and interactive tutorials for expression conditioning are accessible for research purposes, with plans to release AlphaGenome 2 Skills in Google Antigravity, DeepMind’s agentic development platform. This integration aims to accelerate research and facilitate new discoveries.

Source: Google DeepMind