DeepMind Catalogs 90 Quadrillion Mutations: AlphaGenome Atlas Goes Live

Google DeepMind today (September 8) released AlphaGenome Atlas, a functional prediction resource covering approximately 9 billion single-nucleotide variants (SNVs) in the human genome. Researchers can use the free portal to query the potential molecular biological effects of each single-letter DNA change.

This is not another model interface that merely “reads a stretch of DNA and produces a score.” Instead, DeepMind has precomputed the results of running AlphaGenome at scale and organized them into a searchable “genome-wide map of variant impacts.” The accompanying AlphaGenome Variant Impact (AVI) score attempts to compress complex gene-regulatory and protein effects into a metric more suitable for ranking.

For genomics researchers, the real value is not simply the addition of another model name. It is that variant screening is becoming more like “querying a database” for the first time, rather than repeatedly running inference from scratch and manually combining the outputs of multiple tools.

Experimentally Validating 9 Billion Variants One by One Is Not Realistic

The human genome contains approximately 3 billion base pairs. At any given position, replacing the current base with one of the other three bases produces an enormous number of potential single-nucleotide variants. DeepMind has summarized this scale as approximately 9 billion possible single-letter DNA changes.

These variants are not all equally important. Some are located in protein-coding regions and may directly alter amino acid sequences. More variants are located in noncoding regions. Although these regions do not directly “write” proteins, they may affect enhancers, promoters, transcription-factor binding sites, RNA splicing, and chromatin state.

The problem is that noncoding regions account for approximately 98% of the human genome, yet they have long been the most difficult regions to interpret. A variant may be thousands or even hundreds of thousands of bases away from its target gene, but still affect its expression through gene-regulatory networks. Relying solely on genomic position often makes it difficult to determine whether a variant is an inconsequential “spelling change” or a critical switch capable of altering cellular behavior.

Traditional research typically starts with candidate variants and screens them step by step using population genetics, disease cohorts, cell experiments, and functional validation. In rare-disease, cancer, and complex-disease research, the candidate set can easily grow to tens of thousands or even millions of variants. Experimentally validating them one by one is clearly impractical, so researchers must first prioritize them.

AlphaGenome Atlas is designed to address precisely this question of “which variants should be screened first.”

AlphaGenome: From Individual Predictions to a Genome-Wide Catalog

DeepMind previously developed AlphaGenome to predict how DNA sequences affect gene regulation. The model can process contexts of up to approximately 1 million base pairs and simultaneously predict multiple molecular features at single-base resolution, including:

• Gene expression levels;
• RNA transcription and CAGE signals;
• RNA splicing sites and their strengths;
• Chromatin accessibility;
• Histone modifications;
• Transcription-factor binding;
• Three-dimensional chromatin contacts.

Its basic use case is to compare the predicted differences for the same DNA segment in its “wild-type” and “mutant” forms. For example, a researcher can change an A at a particular position to G and then observe whether the expression, splicing, or regulatory signals associated with relevant genes change substantially.

A single inference run may not seem particularly complicated. But when the research target expands from dozens of candidate variants to approximately 9 billion SNVs across the genome, the computational workload quickly becomes a different kind of problem. AlphaGenome Atlas takes the approach of completing a large volume of predictions in advance and providing them to researchers as a dataset.

In other words, AlphaGenome is a “calculator,” while AlphaGenome Atlas is more like an “atlas that has already been calculated.” Researchers do not have to enter a sequence and wait for the model to perform inference each time. They can first search, filter, and compare variants in the catalog, then apply more detailed experiments or model analyses to the most valuable candidates.

A 1 PB Dataset: Size Is Not the Only Point

DeepMind says that AlphaGenome Atlas contains approximately 1 PB of data, more than 30 times the size of the AlphaFold Database. This figure first indicates that the bottleneck for genomic AI is shifting from “whether a model can make predictions” to “how to store, index, and use vast quantities of prediction results.”

Atlas does not provide only a single overall score. It consists of several interconnected resources:

1. Molecular Effect Predictions

Each variant is associated with thousands of molecular-effect predictions covering multiple dimensions of gene regulation and involving hundreds of human and mouse cell types and tissues.

This means researchers do not see a simple “harmful/benign” label. Instead, they receive an impact vector that more closely reflects biological processes: whether a variant might alter gene expression in a particular tissue, affect RNA splicing, change chromatin accessibility, or disrupt the binding of a particular class of transcription factors.

2. AVI Score

The AlphaGenome Variant Impact score combines predictions from AlphaGenome and AlphaMissense into a single value for rapidly ranking variants.

AlphaGenome is better suited to analyzing gene regulation, particularly potential effects in noncoding regions. AlphaMissense, by contrast, is primarily used to assess DNA variants that alter protein sequences. By combining the two, AVI attempts to cover different types of variants spanning coding and noncoding regions.

This step is important because research workflows typically do not require an “absolutely correct” score. Instead, they need to identify the few dozen candidates most worth experimentally validating from a large pool. A unified ranking metric can reduce the cost of switching back and forth between multiple model outputs.

3. AVI Feature Attribution

A score alone is not enough. Researchers also need to know why a particular variant was classified as having a high impact.

For this reason, Atlas associates AVI scores with corresponding feature-attribution information, indicating which biological predictions primarily drove the final result. For example, a variant may receive a high score because AlphaGenome predicts a strong change in gene regulation, or the score may be driven mainly by AlphaMissense’s assessment of its effect on a protein.

This makes AVI more than a black-box ranking system by retaining some interpretable clues. For designing follow-up experiments, knowing “what it affects” is often more useful than knowing “what its score is.”

4. DNA Sequence Motif Catalog

Atlas also includes more than 2,500 recurring DNA sequence motifs and their positions throughout the genome.

Motifs can be understood as the “vocabulary” of DNA. Some motifs may be recognized by transcription factors and thereby participate in regulating gene switches. If a single-base change disrupts or creates a new motif, it may alter how regulatory proteins bind. Placing motif locations and variant effects within the same resource system can help researchers further trace the regulatory logic behind a variant.

The Value of AVI: Turning “Finding the Cause of Disease” into a Ranking Problem

In rare-disease research, genome sequencing often identifies many variants of uncertain significance. Coding-region variants are relatively easier to handle because they may directly alter proteins. Noncoding variants are more challenging: they do not produce an obvious change in protein sequence, yet may cause disease by regulating gene expression.

AVI covers both coding and noncoding regions and attempts to place these two types of variants within a single prioritization framework.

A practical research workflow might look like this:

1. Obtain candidate SNVs from patient sequencing data;
2. Perform initial filtering using information such as population frequencies and family inheritance patterns;
3. Query the molecular effects and AVI scores for the corresponding variants in AlphaGenome Atlas;
4. Examine which cell types, tissues, or regulatory features drove the scores;
5. Send high-priority variants for validation through cell-based assays, reporter-gene assays, or gene-editing experiments.

This workflow cannot directly classify a variant as pathogenic, but it can substantially narrow the experimental scope. In cancer research, researchers can use a similar approach to search for noncoding mutations that may affect oncogenes, tumor-suppressor genes, or enhancer activity.

DeepMind says that AVI has shown leading performance on multiple benchmarks for variant pathogenicity and rare diseases. However, “leading” in this context should be understood as a relative advantage in computational evaluations, not as a clinical diagnostic conclusion. Benchmark data typically comes from existing annotations, while real-world variants may occur in cell types, tissues, or disease contexts that are insufficiently represented in the training data.

How It Differs from AlphaFold: Predicting Regulation, Not Structure

AlphaFold addresses the problem of protein structure: given an amino acid sequence, it predicts how the protein may fold.

AlphaGenome Atlas addresses a different layer of the problem: given a change in a DNA sequence, it predicts how that change may affect gene expression and other molecular regulatory processes.

If a living system is compared to a software system, AlphaFold is more like predicting the three-dimensional object produced when a particular function is compiled, while AlphaGenome analyzes whether a one-character change in the source code might alter a variable, trigger condition, or interaction between modules. The former focuses on “what the structure looks like,” while the latter focuses on “when and in which cells a gene is switched on, and to what extent.”

This is also one reason why the AlphaGenome Atlas contains far more data than the AlphaFold Database. Protein structures generally center on individual proteins, whereas gene regulation requires simultaneous consideration of sequence context, cell type, chromatin state, transcription factors, and long-range regulatory relationships.

The Biggest Change: From Calling Models to Querying a Resource

In the past, using genomic AI often required researchers to prepare reference sequences, construct mutant sequences, run models, and then standardize the outputs themselves. This process is familiar to teams experienced with Python, GPU environments, and bioinformatics toolchains, but the computational and engineering costs are substantial for large-scale screening.

Atlas moves part of this work forward and centralizes it. Researchers can first treat it as a high-dimensional variant index, then direct expensive experimental resources toward the candidates most worth validating.

This is similar to the difference between a search engine and a real-time calculator: a search engine crawls and indexes content in advance, allowing users to quickly find relevant results, but search results are not the final answer. Their sources and context still need to be evaluated. AlphaGenome Atlas works the same way. It improves retrieval and ranking efficiency, but it does not replace causal validation.

For developers and computational biology teams, the aspects most worth watching may include:

• Data access methods: Large-scale precomputed results require efficient indexing, not simply downloading and opening them locally;
• Result versioning: Changes to models, reference genomes, cell types, and prediction tracks can all affect result comparability;
• Multimodel integration: AVI suggests that future genomic analysis may increasingly depend on collaboration among multiple specialized models rather than a single model handling everything;
• Interpretability interfaces: Providing the features driving an impact score alongside the score itself makes the results more suitable for research workflows;
• Experimental feedback loops: Model outputs still ultimately need to be validated through experiments, clinical cohorts, or stronger causal evidence.

But It Cannot Replace Wet-Lab Experiments or Clinical Judgment

The most easily misunderstood aspect of AlphaGenome Atlas is that “covering 9 billion variants” can easily be interpreted as “explaining 9 billion variants.” In reality, coverage means that the system provides predictions for these variants. It does not mean that every result has undergone extensive experimental validation, nor does it mean that the model knows exactly what consequences each variant will produce in human patients.

Gene regulation is highly context-dependent. The same DNA variant may have different effects in different tissues, developmental stages, or disease states. Even if a model can predict a large number of molecular trajectories, it is difficult for it to fully reproduce the complex cellular environment inside the human body.

Furthermore, diseases are often not determined by a single variant. Interactions between multiple loci, environmental factors, epigenetic states, and lifestyle can all alter the final phenotype. AVI is better suited as a candidate-ranking tool than as a way to compress complex diseases into a single “risk switch.”

Accordingly, the most realistic role for AlphaGenome Atlas is not that of an “automated diagnostic system,” but of an infrastructure layer for genomic research: helping scientists identify which parts of an enormous possibility space are most worth investing resources in.

DeepMind Is Turning Genomic AI into Infrastructure

From AlphaFold to AlphaMissense and now AlphaGenome, DeepMind’s direction is becoming increasingly clear: first use models to solve high-value prediction problems, then turn their capabilities into infrastructure that the research community can repeatedly use through databases, scoring systems, and open portals.

The release of AlphaGenome Atlas is especially significant. The model itself answers, “What effects might this variant produce?” Atlas goes further by answering, “Which variants across the entire genome deserve priority attention?” The difference is like that between a high-performance compiler and an entire searchable software-package ecosystem: the former provides capability, while the latter determines whether that capability can enter everyday workflows.

Of course, a scale of 1 PB also creates new barriers. How research institutions can access the data efficiently, how results can be reproduced across different prediction versions, and how Atlas results can be integrated into existing VCF, annotation, and variant-analysis pipelines will all become practical challenges during deployment. For bioinformatics developers, the real value is not displaying a score on a webpage, but reliably integrating it into the complete chain from sequencing and annotation to experimental validation.

As of today, AlphaGenome Atlas provides free academic access. It is unlikely to immediately change clinical diagnostic workflows, but it will allow researchers to process noncoding variants that were previously among the most difficult to investigate more quickly. The next competition in genomics may no longer be limited to which model wins a few more categories on a benchmark. It may instead be about who can turn prediction results into data infrastructure that is searchable, interpretable, reproducible, and capable of entering an experimental feedback loop.

This time, DeepMind has not released only a new model. It has created an unprecedented map of the potential impacts of changes across the human genome. A map will not automatically tell you the answer, but it at least means researchers no longer have to feel their way through every possibility one by one.