DeepMind Precomputed 9 Billion DNA Variants. The Lookup Layer Is the Breakthrough

AlphaGenome Atlas looks like another spectacular DeepMind biology release: 9 billion possible single-letter changes scored across the human genome, a dataset DeepMind says is roughly 1 petabyte, and a headline that invites comparison with AlphaFold.

That obvious reading is incomplete. The important change is not that DeepMind trained a smarter genomics model. AlphaGenome already existed. The change is that Google has converted expensive, repeated model inference into a queryable prediction layer. Researchers and software teams can now look up precomputed molecular effects instead of running a one-million-base-pair neural model for every candidate variant.

That is an infrastructure move. It changes the feasible product surface for rare-disease triage, cohort studies, variant browsers, target discovery, and biotech analysis pipelines. The scarce resource shifts from GPU time toward good filtering, biological context, version control, and experimental validation.

It also creates a dangerous temptation. AlphaGenome Variant Impact, or AVI, is a rank-based compression of many model outputs plus other genomic features. It is not a calibrated probability that a variant causes disease. Atlas has not been validated or approved for clinical use. The right mental model is a versioned hypothesis-ranking and mechanism-exploration resource—not a clinical verdict and not a complete account of personal-genome effects.

The model breakthrough happened earlier. Atlas changes who can afford to ask genome-scale questions, how quickly they can ask them, and where the new errors enter the workflow.

What DeepMind actually shipped

DeepMind's launch bundles several layers that are easy to blur together. The Atlas is the precomputed resource. The base AlphaGenome model remains the system that predicts molecular tracks from DNA sequence. AVI is the compact prioritization score built on top. The motif maps are an interpretation resource. The portal, Python API, and an Antigravity skill are access surfaces.

The underlying peer-reviewed AlphaGenome model accepts sequences up to 1 million base pairs and predicts thousands of tracks spanning gene expression, splicing, chromatin accessibility, protein binding, and 3D genome contacts. Atlas precomputes variant-effect scores derived from that model over the GRCh38 reference genome.

At launch, the no-code portal and Atlas API are available for non-commercial research. The open-source client exposes single-variant, multi-variant, interval, and scorer-metadata queries. The launch-day PyPI package is now version 0.9.0, which adds the Atlas API surface.

The model stopped being the bottleneck

The ordinary AlphaGenome API is useful for thousands of predictions, but its own documentation warns that the live model is not suited to analyses requiring more than one million. Genome-wide research lives far beyond that boundary. Precomputation changes the unit economics.

old workflow: candidate variant → model inference → scores
new workflow: candidate variant → Atlas lookup → scores

cohort workflow: millions of variants → local/remote ranking → targeted model calls → experiments

According to the Atlas manuscript, each variant has about 27,000 experiment-specific scalar predictions on average when active-allele outputs are included. Multiplying that average by 9 billion variants suggests roughly 243 trillion scalar values before compression. That figure is an arithmetic inference, not a number published by DeepMind, but it explains why a one-petabyte indexed resource is a different product from an inference endpoint.

This is the first non-obvious consequence: Atlas moves differentiation up the stack. When raw prediction becomes a lookup, the valuable work is deciding which signals matter for a tissue, disease, cohort, or experiment; combining them with population and clinical evidence; preserving provenance; and routing a small candidate set into expensive follow-up.

A variant browser can return richer answers. A rare-disease pipeline can rank a whole genome without scheduling a GPU farm. A cohort study can form groups based on predicted mechanisms rather than only genomic distance. But none of those products becomes trustworthy merely because the lookup is fast.

AVI is a compression codec, not an oracle

AVI is what makes Atlas operationally approachable. It reduces the high-dimensional prediction space to 10 aggregated AlphaGenome features, combines those with AlphaMissense, conservation, protein loss-of-function, and indel indicators, then produces one rank score and 18 additive feature attributions. In other words, AVI is not AlphaGenome's pure opinion. It is an ensemble-like compression layer.

The distinction matters because AVI is trained with a proxy target: it learns to separate very rare gnomAD variants from common ones, using 0.1% filtering allele frequency as the dividing line. Rarity is a useful signal of selection, but it is not the same as pathogenicity. A rare variant can be benign; a common variant can still affect a trait; and underrepresented ancestry can distort what counts as rare.

This creates a subtle circularity risk. AVI contains a population-selection prior, then some downstream tasks reward it for prioritizing variants expected to be rare or deleterious. The authors balance sequence context and remove coordinate overlaps between training and held-out benchmarks, which helps. Builders should still keep the molecular evidence separate from the prior. If a hit is strong, reviewers need to know whether the score is driven by splicing, expression, conservation, protein change, or another feature—not just that the percentile is high.

A single score is useful because it hides complexity. It is risky for exactly the same reason.

The evidence is promising—and narrower than the launch headline

DeepMind reports strong results across several evaluations. The useful way to read them is not as one victory lap, but as evidence for a specific claim: Atlas can improve candidate prioritization and mechanism-aware grouping. It does not yet establish prospective diagnostic yield or universal biological accuracy.

The rare-disease case study shows the workflow at its best. In an unsolved epileptic encephalopathy case, AVI ranked a brain-relevant intronic DNM1 variant first. AlphaGenome predicted that it created a cryptic splice acceptor, and a minigene experiment supported an abnormal 13-amino-acid extension. The evidence contributed to a Likely Pathogenic recommendation.

That is impressive, but the causal chain matters. The result came from ranking, tissue-specific interpretation, prior gene knowledge, and a dedicated wet-lab assay. Atlas generated a sharp hypothesis. It did not independently diagnose the patient.

The 22% result is about search-space compression

In 54,189 UK Biobank participants, Atlas-filtered rare non-coding aggregate tests across 2,028 circulating proteins produced a 22% increase in conditionally independent discoveries. At the PLA2G7 locus, a top-1% AVI filter shrank 526 upstream variants to four variants in a 15-base-pair window. That is a striking example of making an intractable search space experimentally legible.

But it is not the same as proving every added association. The population analysis used an older internal Atlas snapshot because of platform downtime. Release-version association results correlated strongly, but imperfectly, with the reported build. Of 31 highlighted complex-trait associations, 25 could be tested in All of Us; four replicated at nominal P below 0.05, and none survived Bonferroni correction.

My read: the 22% figure is best understood as evidence that model-guided, mechanistically homogeneous grouping can improve statistical power. It is not yet evidence that Atlas has delivered 22% more generalizable biological truths. The durable near-term value is candidate compression—turning hundreds or millions of possibilities into a short list worth testing.

“Every possible DNA letter change” has a coordinate contract

The headline is accurate within a precise definition: Atlas enumerates the three alternate A, C, G, or T alleles at each non-N base in GRCh38. That is enormously useful. It is not every possible genetic state a person can carry.

• It does not cover arbitrary non-reference-to-non-reference substitutions.

• It does not model linked variants or a person's full haplotype as one genome-wide state.

• Most structural variation sits outside the 9-billion-SNV headline.

• Effects that depend on distant trans-acting changes, environment, development, or missing cell states are not directly represented.

• The base model's one-megabase cis horizon cannot capture every long-range interaction in one prediction.

This is why reference-build provenance is product correctness, not bioinformatics housekeeping. A service must normalize inputs to GRCh38, verify the reference allele, left-normalize indels, and reject or explicitly transform incompatible builds. A swapped allele or stale coordinate can quietly return no result—or worse, a plausible result for the wrong variant.

The same caution applies to biological coverage. The manuscript flags uneven cell-type and assay data, missing non-polyadenylated RNAs from standard poly(A)-selected RNA-seq, no direct modeling of trans effects, and known weakness around some enhancer effects. A low score can mean low predicted impact. It can also mean the relevant context is missing.

The licensing split reveals Google's product boundary

DeepMind is making the Atlas portal and API free for non-commercial research and says commercial Atlas access is coming to Google Cloud. No Atlas price, launch date, quota, or service-level commitment has been published. The commercial base AlphaGenome model is already generally available with allowlisting as a paid, self-deployed Cloud model. It requires an 80 GB A100- or H100-class GPU; the 40 GB A100 series is unsupported.

The interesting exception is the static AVI SNV download. DeepMind labels it a permissive-use artifact for commercial and non-commercial use, while splicing scores, feature-importance downloads, and richer API-delivered Atlas data remain non-commercial unless separately licensed. That is not a footnote. It sketches the likely ecosystem.

My prediction is a two-tier product pattern: run permissive AVI locally across the broad candidate set, then pay for rich mechanism-level analysis on a much smaller subset. Google can let AVI become a common ecosystem primitive while retaining differentiation in the one-petabyte data layer, hosted inference, and future versions. It resembles a broader infrastructure strategy I discussed in Mistral's move from open weights toward controlled infrastructure, though the genomics licensing and validation stakes are very different.

What builders should test before they ship

1. Treat AVI as one ranking feature. Combine it with allele frequency, phenotype fit, inheritance, segregation, established clinical evidence, and domain-specific predictors. Never display AVI PHRED as a probability of disease.

2. Preserve the explanation layer. Store attributions and selected raw modality scores so a reviewer can distinguish a splicing signal from conservation, expression, chromatin, or protein effects.

3. Validate on your actual population and task. A retrospective ClinVar benchmark does not answer how the score behaves in a prospective rare-disease cohort, a specific ancestry, an enhancer screen, or a particular tissue.

4. Pin the full coordinate and version contract. Record the reference build, Atlas release, API/client version, annotation release, score definition, normalization method, and license tier for every result.

5. Separate lookup from live inference. Use Atlas for cohort-scale filtering. Use the live model for selected variants, custom sequences, alternative genetic backgrounds, haplotype-oriented experiments, and visual track inspection.

6. Make missing context visible. Flag enhancer-heavy, trans-effect, absent-cell-type, structural-variant, and multi-variant cases. Silence or a low score should not be rendered as proof of no effect.

7. Keep humans and experiments in the loop. For rare disease, drug targets, or other high-stakes claims, route candidates through qualified review and orthogonal validation. Accessibility is not auditability, a distinction that also matters for deployable frontier models.

reference_build: GRCh38
atlas_release: <pinned release ID>
alphagenome_client: 0.9.0
annotation_release: <pinned GENCODE version>
variant_normalization: <documented method>
score: AVI_PHRED
license_tier: <permissive-static | research-api | commercial-cloud>
review_status: <machine-ranked | expert-reviewed | experimentally-tested>

That manifest may look bureaucratic. It is the minimum needed for reproducibility. The launch paper itself found strong but imperfect agreement between an older internal Atlas build and the released data in its population analysis. Every future model improvement or annotation update can reorder candidates. In a lookup-driven workflow, dataset version is part of the model.

What happens next

AVI spreads faster than the full Atlas

The compact score is downloadable, rank-based, easy to add to existing variant tables, and more permissively licensed than the rich Atlas layer. Those properties matter more for adoption than scientific elegance. Expect AVI to reach annotation and triage pipelines first, with raw Atlas features reserved for investigations that survive the initial filter.

Reproducibility becomes a release-management problem

DeepMind calls the initial Atlas a starting point. Better models, more assays, more cell states, new reference resources, and improved annotations should make future versions more useful. They will also make old analyses harder to reproduce. Teams will need migration reports that show which variants changed rank, which mechanism features moved, and whether downstream cohort associations survive the re-index.

The first durable wins will be upstream of diagnosis

Atlas is most convincing when it narrows a search, groups variants by a plausible mechanism, or suggests an experiment. It is least convincing when a percentile is asked to carry a clinical conclusion. The immediate winners are likely to be candidate prioritization, cohort aggregation, assay design, and mechanism exploration—not autonomous interpretation of a patient's genome.

Frequently asked questions

Is AlphaGenome Atlas a new AI model?

Not primarily. AlphaGenome is the underlying model, first introduced in 2025 and published in Nature in January 2026. Atlas is the newly released, precomputed genome-scale prediction resource built from AlphaGenome outputs, with AVI and motif interpretation layered on top.

Does Atlas really cover every possible DNA mutation?

It covers the three possible single-nucleotide substitutions from each non-N position in the GRCh38 reference, producing roughly 9 billion reference-relative SNVs. That does not include every haplotype, structural variant, multi-variant interaction, or non-reference-to-non-reference change. The team also scored more than 100 million observed indels, but launch-day availability differs by access surface and should be verified before product commitments.

Is AVI a pathogenicity probability?

No. AVI PHRED is a genome-wide rank transform. A score of 20 means a variant is in the top 1% of AVI predictions, not that it has a 99% chance of causing disease. Thresholds should be specific to the region, application, and validation data.

Can companies use AlphaGenome Atlas today?

The portal and Atlas API are offered for non-commercial research. DeepMind says commercial Atlas access is coming through Google Cloud, but it has not published timing or pricing. The base AlphaGenome model already has a paid, allowlisted Cloud deployment. The manuscript identifies the static AVI download as a permissive-use artifact; teams should review the exact current terms for every dataset and API field before commercial use.

Can Atlas results be used clinically?

No. DeepMind explicitly says AlphaGenome has not been validated or approved for clinical use and that predictions must not be used on their own for diagnosis or medical decision-making. Research workflows still need qualified interpretation and orthogonal evidence.

What should an engineering team try first?

Start with a retrospective, well-labeled cohort in the exact disease area or assay you care about. Compare AVI alone, your current pipeline, and a combined pipeline. Measure recall at a fixed review budget, calibration within relevant variant classes, ancestry-specific behavior, explanation usefulness, and rank stability across Atlas versions. Then decide whether the lookup layer saves enough analyst and compute time to justify integration.

The useful way to think about AlphaGenome Atlas

AlphaFold made a huge prediction problem available as a database. AlphaGenome Atlas applies a related distribution idea to regulatory genomics, but the object is messier. A protein structure is not the same kind of claim as a context-dependent molecular effect, and neither is automatically a clinical conclusion.

The release matters because it changes the default workflow from compute first to query first. That will make ambitious genomic analysis cheaper to prototype and easier to embed in software. It may also standardize AVI as a common first-pass signal across the ecosystem.

The teams that benefit most will resist the seductive shortcut. They will use Atlas to shrink candidate sets, expose mechanisms, and design better experiments. They will pin versions, preserve feature-level evidence, validate on their own populations, and keep human judgement at the decision boundary.

DeepMind did not release a genomic oracle. It released something more practical: a massive, queryable prior over which single-letter changes may matter and why. That is enough to reshape the tooling layer around genomics—provided builders remember that a faster hypothesis is still a hypothesis.