"I bought the book, but I didn't know how to read it."

On September 8th, Google DeepMind unveiled a new platform called "AlphaGenome Atlas." It's a database that provides pre-calculated predictions for all 9 billion possible single-nucleotide variants (Snucleotide variations) in the human genome—that is, all combinations of "what happens when you replace one letter of DNA with another." Pushmith Kohli, head of DeepMind's AI for Science division, reportedly said at a press briefing, referencing the "unfinished business" of the Human Genome Project, completed in 2003, that "we bought the book, but we didn't know how to read it." While the genome sequence itself was deciphered more than 20 years ago, understanding the "meaning" of that sequence remained an incredibly time-consuming task.

What is AlphaGenome Atlas?

The foundation is "AlphaGenome," an AI model released by DeepMind last year. AlphaGenome can predict how mutations in DNA sequences containing non-coding regions (regions that do not code for proteins but regulate the on/off state of genes) will affect gene expression, splicing (RNA editing), and chromatin structure. Previously, researchers had to either input each mutation they wanted to know about into the model and calculate the results, or painstakingly verify them in the laboratory. AlphaGenome Atlas pre-calculates these predictions for all approximately 9 billion possible single-nucleotide substitutions in the human genome and packages them in a form accessible via a web browser. According to a DeepMind paper, each mutation is associated with an average of approximately 27,000 individual predictions, calculated across hundreds of human and mouse cell and tissue types.

The Invention of the AVI Score

A new metric, the "AlphaGenome Variant Impact (AVI) score," is practically important for researchers. Simply looking at the vast amount of prediction data makes it difficult to determine which mutations are truly important. The AVI score consolidates these predictions into a single numerical value, allowing for a ranking of how strong the biological impact of each mutation is likely to be. For protein-coding regions, it combines insights from DeepMind's existing model, "AlphaMissense," while for non-coding regions, it relies on predictions from AlphaGenome. Furthermore, a "feature contribution" is calculated, indicating which molecular processes (such as RNA splicing or gene expression) are predicted to be most disrupted, ensuring researchers don't perceive the score as a mere black box.

How was it validated?

According to the validation results presented in the paper, the AVI score demonstrated top-tier performance across multiple benchmarks in distinguishing between pathogenic and harmless mutations in clinical genomics databases. A team at the University of Exeter in the UK validated it using whole-genome data from the UK Biobank, and a team at the Broad Institute in Massachusetts has also independently utilized the predictions from this atlas. The fact that multiple independent validations by non-DeepMind researchers are underway is crucial for evaluating the reliability of this type of large-scale prediction tool. On the other hand, Martin Kircher, a bioinformatician at the Max Delbrück Center for Molecular Medicine in Berlin, praised it as "a useful and generous way to greatly expand access to powerful models," but cautioned that "it will not replace experiments or the consideration of individual differences in disease diagnosis."

Researchers are already using it

Immediately after the paper's publication, there have been reports of it being used in actual analyses. One researcher identified 19 genomic regions that appear to be related to BMI (Body Mass Index) from non-coding variants narrowed down to the top 1% in terms of predictive impact. Multiple access routes are available, including a web portal for researchers without coding experience, a GitHub API that can be called from programs, and even skills to integrate it into DeepMind's agent-based development platform, Google Antigravity. Non-commercial academic use is free, while commercial use involves licensing a 1-petabyte dataset via Google Cloud.

Limitations Clearly Stated

An important note clearly stated by DeepMind itself is that AlphaGenome has not been validated or approved for clinical use. It is purely a research tool for hypothesis generation and prioritization, and is not intended for use in diagnosis itself. Jonathan Sebat of the University of California, San Diego, who specializes in psychiatric genetics, also supports a cautious approach, while acknowledging its potential to accelerate this type of research. While the impressive figure of "9 billion predictions" is eye-catching, it's clear that this means "the speed at which hypotheses are narrowed down has increased," not "answers to diagnosis or treatment."

What Researchers Should Consider

The essential value of this atlas lies not so much in the accuracy of individual predictions themselves, but in the fact that it eliminates the research bottleneck of "running the model from scratch for each mutation" through pre-calculation and indexing. The comparison between "experiments in a drying room" and "computer-based screening," often cited in the field of genomics, has become practical for the first time on this scale. What we should keep an eye on going forward is how independent groups other than the UK Biobank and the Broad Institute will validate data of this magnitude and what kinds of disease research they will apply it to. If the Human Genome Project was at the stage of "reading" the sequence, then the AlphaGenome Atlas is the first step in the attempt to "index its meaning," and the evaluation continues to be in its early stages.