The field of biological sciences witnessed a major milestone this month, as Google DeepMind announced the creation of the AlphaGenome Atlas. By leveraging the DeepMind developed AlphaGenome , the research team was able to pre-compute predictions for all ~9,000,000,000 single-letter changes to the 3-billion-letter human genome.

The 1 petabyte genomic data set, made available to academic researchers via web portal and API, has the potential to fundamentally change how genome-scale analyses are performed by enabling rapid, data-driven insights rather than relying on traditional, often time-consuming methods.

1. Decoding the 98%: From Protein-Coding DNA to Regulatory “Switches”

While genome-wide association studies have traditionally focused on the ~2% of human DNA that encodes instructions for making proteins, most disease-associated variants lie outside of these protein-coding regions, in regulatory DNA elements that govern the process of gene regulation. The AlphaGenome Atlas predicts the impact of all possible single-letter substitutions for each DNA position in the genome, considering both protein-coding and regulatory DNA.

Long-range analysis: The underlying predictive model uses sequences of up to 1,000,000 letters to associate specific single-letter substitutions with their downstream effects on genome function across cells and tissues.

Context-dependent predictions: For each position in the genome, predictions were made for all three possible single-letter substitutions in the context of regulatory, structural, and coding DNA.

Multi-track analysis: Beyond single-letter substitutions, the AlphaGenome Atlas evaluates thousands of genome-related tracks, including transcription factor binding, RNA splicing, chromatin accessibility, and more, to fully characterize the impact of each substitution.

2. Variant Impact Score Unifies Protein + Regulatory Analysis

In order to aid both basic biological discoveries and therapeutic development, researchers at DeepMind have created a unified genome variant impact score, called AlphaGenome Variant Impact (AVI). The AVI score integrates the regulatory impact predictions described above with similar protein-centric predictions from AlphaMissense, another DeepMind model, to prioritize variants based on their potential impact on biology.

The AVI score enables researchers to quickly understand how a specific variant may disrupt normal genome function, and which specific mechanisms are involved.

For example, an AVI score could indicate that a particular variant creates a “splice site error,” which could affect the production of multiple proteins and lead to disease. By acting as a “triage,” the AVI score can save researchers valuable time by focusing their efforts on variants with the highest potential impact.

3. Discovery of Unsolved Rare Diseases + Trait-Associated Variants

At the announcement of the AlphaGenome Atlas, researchers highlighted some initial success cases in which variants previously linked to disease had been analyzed using the new tool.

For instance, when working with the GREGoR Consortium and Broad Institute researchers to analyze non-coding mutations associated with rare diseases, the AVI score was instrumental in prioritizing variants for further testing. By doing so, the research team was able to uncover a novel non-coding mutation in the DNM1 gene that caused severe epileptic encephalopathy by disrupting RNA splicing, after confirming the prediction experimentally in the lab. Additionally, researchers analyzing complex traits found that using AlphaGenome predictions helped identify 22% more trait-associated non-coding variants compared to standard approaches, leading to the discovery of 19 new regions associated with body mass index.

4. Prioritization, Not Replacement: The Importance of Triage in Genomics

While many researchers consider the release of the AlphaGenome Atlas to represent a major step towards making powerful genomic analyses more accessible, they also note that the tool should be regarded as a prioritization aid, and not a replacement for traditional laboratory testing.

Google DeepMind notes that the data set should be used to triage and prioritize variants for further laboratory investigation, and not as a diagnostic tool. As thousands of researchers begin exploring the genome variant data set, the AlphaGenome Atlas will serve as a valuable resource that will accelerate the process of decoding the genome, ultimately expanding humanity’s understanding of the biological mechanisms that underpin health and disease.

Sources

Reference links