OHMEGA/MONSTER

Article

Google DeepMind releases searchable map of every possible DNA letter change

A 1-petabyte precomputed Atlas puts billions of regulatory predictions behind a free browser to filter the 98% non-coding genome, but its ranking score is uncalibrated for clinical use and every hit still needs lab proof.

United KingdomUnited States

Pictures in this frame are AI-generated, not photographs.

What happenedOn 8 September 2026 DeepMind published a 1-petabyte precomputed Atlas covering about 9 billion possible single-letter DNA changes and more than 100 million short insertions or deletions, searchable in a browser without code.

Why it mattersIt lets any lab filter the vast 98% regulatory genome for likely functional variants in seconds, accelerating rare-disease hypothesis generation and drug-target triage.

Still openWhether top-ranked hits become real diagnoses, with no prospective trials, no formal clinical calibration of the ranking score, and only retrospective pilots so far.

a modern genetics laboratory bench with pipettes and sample tubes
AI-generated illustration — not a photograph of this story

On 8 September 2026 Google DeepMind, the London-based AI lab, released a free, browsable Atlas that predicts what every possible single-letter change in human DNA might do to gene regulation. Any researcher can now look up a variant in a browser and get a ranked hypothesis in seconds — a real shift for rare-disease and population genetics — but a high score is not a diagnosis and still needs laboratory proof.

Human DNA is about 3.2 billion letters long, written in A, C, G and T. At each position there are three possible swaps, so about 9 billion possible single-letter changes, plus short insertions or deletions. Only about 2% codes directly for proteins. The other 98% is regulatory: it controls when, where and how strongly genes are turned on. Most variants linked to common traits and diseases in large genome studies lie there, and it has been the hardest part to interpret.

The Atlas is an attempt to make that 98% searchable. DeepMind ran its sequence-to-function model across hg38, the current standard human reference genome, comparing each reference base to its three alternatives — exhaustive computer simulation — and stored the results. It covers all ~9 billion possible single-letter changes plus more than 100 million short insertions or deletions seen in large datasets such as gnomAD, UK Biobank and All of Us in the United States. Each variant carries thousands of predictions across hundreds of cell types and tissues, totalling more than 240 trillion values. At 1 petabyte it is more than 30 times larger than the AlphaFold Database as expanded in 2022 from about 190,000 experimental protein structures to more than 200 million predicted ones. The 9 billion are all possible changes, not variants already seen in patients.

The Atlas is not a new model. It is a precomputed version of AlphaGenome, DeepMind's generalist model released in 2025 and described in Nature in early 2026.

AlphaGenome takes up to 1 million letters of DNA context at once and predicts effects at single-base resolution across 11 molecular readouts — including gene expression, splicing, chromatin accessibility and 3D genome contacts — across hundreds of human and mouse cell types. It was trained on data from four public consortia that catalogue gene regulation. In that Nature evaluation it was reported as state-of-the-art on 22 of 24 track-prediction tasks and 25 of 26 variant-effect tasks, with a preprint version reporting 24 of 26. Those numbers depend on task and comparator choice and do not amount to universal superiority. DeepMind and independent researchers note overall accuracy remains below AlphaFold's protein-structure accuracy, and performance is stronger for promoters and splicing than for distant enhancers.

Making exhaustive precomputation feasible took engineering. DeepMind distilled a smaller, faster model from a slower ensemble and optimised it to run in less than a second per variant on a leading AI chip, where earlier estimates had deemed a full-genome run impossible.

The Atlas condenses those thousands of predictions into layers a biologist can scan. First, raw molecular predictions per variant. Second, a single-number rank called the AlphaGenome Variant Impact score, or AVI. AVI merges 10 aggregated regulatory signals with protein-impact scores, conservation and other annotations — 18 features in total — trained to distinguish rare from common variants as a proxy for impact. It is scored like the PHRED scale used in genetics: 10 means top 10% most impactful genome-wide, 20 top 1%, 30 top 0.1% — one in a thousand. Each AVI comes with attributions showing which process drives it, such as splicing or accessibility. Third, those attributions make the score inspectable. Fourth, a compendium of 2,601 recurrent short DNA words — including sites where regulatory proteins bind — mapped across the genome, some matching known biology, others labelled broad, unresolved or unknown and not all validated. The Atlas is locked to hg38 and does not update in real time for new cell types or a future reference without recomputation.

Before Atlas, using AlphaGenome meant calling an interface and running graphics-processor compute, which capped users at about 9,000 globally. Atlas removes that friction with a free web portal requiring no coding, plus interface access and an agent-assisted workflow. EMBL-EBI, the European bioinformatics institute that co-runs Ensembl, is embedding AVI scores into Ensembl's Variant Effect Predictor, the widely used annotation tool many clinical pipelines already run.

Access is free for non-commercial academic research. Commercial access via Google Cloud is described only as coming soon, with no disclosed launch date, price, licensing terms, rate limits or bulk-download costs for the 1-petabyte dataset. DeepMind's sister drug-discovery company Isomorphic Labs will also require a commercial licence. That creates a two-tier structure unlike AlphaFold DB's fully open release, with risks of reference-genome bias, staleness and asymmetry between well-funded companies and academic labs. No ancestry breakdown or false-positive rate for AVI was disclosed in the launch materials.

DeepMind highlighted two retrospective pilots selected by its collaborators, not prospective trials. They illustrate hypothesis generation, not proven diagnostic yield.

What the Atlas changes for finding causes in the 98% non-coding genome — and what it still does not do. AlphaGenome Atlas improved two incumbent filters, and still left most rare-disease causes unfound. In retrospective tests on solved GREGoR cases, AVI placed the known causal variant in the top 50 candidates 29.5% of the time versus 12.5% for CADD (about 2.4×), which means 70.5% of those cases still sat outside the top 50. Grouping rare variants by Atlas-predicted effect in 54,189 UK Biobank genomes across 2,028 proteins found 728 significant non-coding associations versus 595 by the standard method (+22%, +133), including a PLA2G7 locus narrowed from 526 candidates to 4 — both tests are reported, retrospective, and not a prospective diagnostic yield; Atlas is not approved or validated for clinical use. — AI-assisted analytic, built only from real cited or sourced data. Source: Fortune, TechTimes, WinBuzzer. As of 2026-09-09.

At the Broad Institute of MIT and Harvard in the United States, researchers Laura Covill and Anne O'Donnell-Luria working with the NIH-funded GREGoR Consortium for unsolved rare diseases used AVI to reprioritize variants that conventional filters had overlooked. The top hit was a deep intronic variant in DNM1, a gene strongly linked to epileptic encephalopathy. AlphaGenome predicted it creates a brain-specific cryptic splice site that extends the protein by 13 amino acids. Because that brain form is barely expressed in blood, earlier RNA sequencing from blood had been inconclusive. A laboratory minigene assay — a cell-based test of splicing — across five cell lines reproduced the predicted effect, and the variant was reclassified as likely pathogenic.

In retrospective testing on solved GREGoR cases, AVI placed the known causal variant in the top 50 candidates 29.5% of the time versus 12.5% for CADD, the widely used genome-wide ranker — about a 2.4-fold improvement. That figure is retrospective on solved cases selected by DeepMind collaborators, with no independent replication outside those cases, no prospective diagnostic yield, and no CADD version specified. It should be read as illustrative prioritization recall, not the rate at which unsolved patients would now be solved.

At the University of Exeter in the United Kingdom, Medical Research Council fellow Gareth Hawkes applied Atlas to whole-genome data from more than 54,000 UK Biobank participants — 54,189 in the preprint analysis across 2,028 circulating proteins. By grouping rare non-coding variants by their predicted molecular effect, the analysis found 728 significant non-coding rare-variant associations with protein levels that remained distinct after overlapping signals were separated, up from 595 without those aggregated features — 22% more signals under that specific grouping method. Examples included PLA2G7, linked to aging, and EGLN1, a cellular oxygen sensor, and in one locus the approach narrowed 526 candidates to four. A separate body-mass-index analysis identified 19 associated regions after focusing on the top 1% most impactful non-coding variants. Those are additional association signals in that cohort and method; they measure neither predictive accuracy nor causation, and the supporting research remains a preprint without peer review.

DeepMind states explicitly that Atlas is not approved or validated for clinical use and is not a substitute for professional medical advice; predictions are hypotheses requiring independent laboratory validation.

That caution is built into clinical genetics rules. Under the 2015 joint guidelines from the American College of Medical Genetics and Genomics and the Association for Molecular Pathology in the US, computational predictor evidence is capped at supporting-level only, with an explicit warning not to overestimate it because different algorithms may rely on similar data and most have not been validated against well-established pathogenic variants. If predictions disagree, the evidence cannot be counted as supporting.

AVI has no published calibration to exceed that ceiling. The ClinGen Sequence Variant Interpretation Working Group, which calibrates how strongly a tool can be counted, found in 2022 and expanded in 2025 that only 4 of 13 evaluated tools could reach strong support, nine topped out at moderate or supporting, and one failed to meet even supporting. CADD's developer-recommended threshold of 20 falls within moderate for benign evidence, meaning it has commonly been misapplied as pathogenic evidence. REVEL, another widely used predictor, is the only tool covering strong to very strong at specific thresholds. AVI, trained on an allele-frequency threshold as a rarity proxy, is not among ClinGen-calibrated tools and has no published thresholds; clinical labs therefore cannot apply it above supporting without formal calibration. A low AVI score likewise must not falsely reassure.

Other limits are structural. The 1-million-letter window misses enhancers acting over longer distances, single-variant scores ignore interactions between multiple variants and environmental or developmental context, and variant-effect benchmarks can be inflated when the same variant or same gene appears in both training and test data. AVI's hybrid design inherits both power and that leakage risk, and the 22/24 versus 25/26 versus 24/26 discrepancies reflect task and comparator choice.

The practical shift is a bottleneck move. Before Atlas the limiting factor was compute — which variants to test and how to run the model. After Atlas the limiting factor is experimental validation capacity: labs can only follow up a small number of candidates in cells or animals.

That makes Atlas infrastructure, not an answer. It collapses the cost of filtering the non-coding genome the way AlphaFold in 2020 — recognised with a share of the 2024 Nobel Prize in Chemistry — collapsed the cost of getting a protein shape and AlphaMissense in 2023 collapsed coding missense scoring for 71 million variants. The Human Genome Project in 2003, in this telling, bought the book without knowing how to read it.

What to watch is whether that filter translates into real prospective yield: Ensembl integration going live, Google Cloud commercial terms and bulk-access pricing, independent replication beyond the two DeepMind-selected vignettes, formal ClinGen calibration of AVI thresholds, and a future model that sees beyond 1 million letters to capture distal enhancers and multi-variant effects. Until then, the Atlas is best understood as a searchable dictionary for non-coding DNA — powerful for generating the next experiment, not for settling the diagnosis on its own.

Source recordSources / claims / limits

How this piece is framed: A searchable map of the dark genome — infrastructure that makes the 98% non-coding genome filterable for anyone, powerful as a hypothesis engine but not a diagnosis

Charts & tables — AI-assisted; provenance on each line

  • Better at surfacing the needle, still far from a diagnosis — from claims clm_8067ce8025, clm_915bbf80a3, clm_798c5c176d, clm_3e029cf387, clm_36dcd9b32b · as of 2026-09-09

Visuals not shipped — planned but not fulfilled

  • A petabyte to map every DNA typo — 30 times larger than AlphaFold (req_alphagenome_scale): integrity_check_failed — figure check: numbers not found in cited evidence — 33000.0, 1000000.0

Sources

Claims, and how far we tracked each down

  • [confirmed] DeepMind states Atlas is not approved or validated for clinical use and is not a substitute for professional medical advice; predictions are hypotheses requiring independent laboratory validation before informing diagnosis. · read in full (as of 2026-09-17)
  • [confirmed] GREGoR/Broad 29.5% vs 12.5% CADD top-50 recall figure is retrospective testing on solved GREGoR cases selected by DeepMind collaborators, with DNM1 cryptic splice validated by minigene assay across 5 cell lines; no independent replication outside DeepMind-selected cases, no prospective diagnostic yield, and CADD version not specified. · read in full (as of 2026-09-18)
  • [confirmed] ACMG/AMP 2015 guidelines (Richards et al., Genet Med 17:405-423) limit computational (in silico) predictor evidence to supporting-level only (PP3 for pathogenic, BP4 for benign), explicitly warning 'Not overestimating computational evidence is important, particularly given that different algorithms may rely on the same (or similar) data to support predictions and most algorithms have not been validated against well-established pathogenic variants' and that 'algorithms can have vastly different predictive capabilities for different genes.' If predictions disagree, evidence cannot be counted as supporting. This constrains AVI to supporting at most without independent ClinGen calibration. · read in full (as of 2026-09-18)
  • [confirmed] Google DeepMind released AlphaGenome Atlas on 8 September 2026 containing precomputed predictions for all ~9 billion possible single-nucleotide variants in the human genome (every position × 3 alternative bases). · read in full (as of 2026-09-17)
  • [confirmed] Atlas dataset is ~1 petabyte, more than 30 times larger than the AlphaFold Database as expanded in 2022 (200M+ structures). · read in full (as of 2026-09-17)
  • [confirmed] Atlas was built by in silico saturation mutagenesis: running AlphaGenome across every position in the hg38 human reference genome and comparing each reference base to its three alternatives, plus scoring >100 million short insertions/deletions observed in gnomAD, UK Biobank and All of Us. · read in full (as of 2026-09-17)
  • [confirmed] AlphaGenome model takes up to 1 million base pairs (1 Mb) of DNA as input and predicts thousands of functional genomic tracks at single-base-pair resolution. · read in full (as of 2026-09-17)
  • [confirmed] Predicted modalities include gene expression (RNA-seq, CAGE, PRO-cap), transcription initiation, chromatin accessibility (DNase-seq, ATAC-seq), histone modifications, transcription factor binding, chromatin contact maps (Hi-C/Micro-C), and splice site usage plus splice junction coordinates and strength, across hundreds of human and mouse cell types/tissues. · read in full (as of 2026-09-17)
  • [confirmed] AlphaGenome achieved state-of-the-art on 22 of 24 genome-track prediction tasks and 25 of 26 variant-effect prediction tasks in the Nature peer-reviewed evaluation (blog preprint version reported 24 of 26). · read in full (as of 2026-09-17)
  • [confirmed] Each variant in Atlas carries on average ~27,000 scalar predictions spanning hundreds of cell types/tissues, totaling >240 trillion predictive values. · read in full (as of 2026-09-17)
  • [confirmed] Atlas releases the AlphaGenome Variant Impact (AVI) score, a single number per variant that integrates AlphaGenome regulatory predictions with AlphaMissense protein-impact predictions plus conservation and loss-of-function features, with per-variant feature attributions (splicing, expression, accessibility, etc.). · read in full (as of 2026-09-17)
  • [confirmed] AVI score works across both the ~2% protein-coding genome and the ~98% non-coding regulatory genome, unlike AlphaMissense which only scores coding missense variants. · read in full (as of 2026-09-17)
  • [confirmed] Before Atlas, accessing AlphaGenome required writing code against an API and GPU compute, limiting use to ~9,000 researchers globally. · read in full (as of 2026-09-17)
  • [confirmed] Atlas is available from launch via an intuitive free web portal requiring zero coding skills, plus API and as a skill in Google Antigravity, free for non-commercial academic research. · read in full (as of 2026-09-17)
  • [confirmed] Commercial access to Atlas will be via licensing through Google Cloud (timing 'soon'); DeepMind's sister company Isomorphic Labs will also require a commercial license. · read in full (as of 2026-09-17)
  • [confirmed] Broad Institute researchers Laura Covill and Anne O'Donnell-Luria with the GREGoR Consortium used AVI to reprioritize unsolved rare-disease cases and surfaced a deep intronic DNM1 variant creating a brain-specific cryptic splice site that extends the protein by 13 amino acids, causing epileptic encephalopathy; the prediction was experimentally validated and the variant reclassified as likely pathogenic. · read in full (as of 2026-09-17)
  • [confirmed] In retrospective testing on solved GREGoR cases, AVI placed the known causal variant in the top 50 candidates 29.5% of the time versus 12.5% for CADD, a ~2.4× improvement. · read in full (as of 2026-09-17)
  • [confirmed] MRC fellow Gareth Hawkes at University of Exeter applied Atlas to whole-genome data from >54,000 UK Biobank participants; grouping rare variants by predicted molecular effect uncovered 22% more non-coding associations with circulating protein levels (e.g., PLA2G7 linked to aging, EGLN1 oxygen sensor), narrowing one locus from 526 candidates to 4. · read in full (as of 2026-09-17)
  • [confirmed] University of Exeter MRC fellow Gareth Hawkes applied AlphaGenome Atlas to whole-genome data from >54,000 UK Biobank participants (54,189 in preprint analysis across 2,028 circulating proteins), grouping rare variants by predicted molecular effect to uncover 22% more non-coding genetic associations (728 vs 595 significant associations) including PLA2G7 and EGLN1, and identified 19 BMI-associated regions by focusing on top 1% most impactful non-coding variants. · read in full (as of 2026-09-18)
  • [confirmed] AlphaGenome Variant Impact (AVI) score is an 18-feature model: 10 aggregated AlphaGenome predictions + AlphaMissense + 2 conservation features + 3 protein loss-of-function annotations + 2 indel indicators, trained to distinguish variants above vs below population allele-frequency threshold (rarity as proxy for impact), output as PHRED-like rank (10=top 10%, 20=top 1%, 30=top 0.1%) with per-variant feature attributions. · read in full (as of 2026-09-18)
  • [confirmed] Motif compendium comprises 2,601 recurrent DNA sequence motifs derived from patterns the model learned, mapped across genome and cell types with inferred roles (activating, repressing, altering accessibility); some match known biology with experimental support, others labeled broad/unresolved/unknown. · read in full (as of 2026-09-18)
  • [confirmed] AVI score calibration: 10=top 10% most impactful genome-wide, 20=top 1%, 30=top 0.1% (1 in 1,000), as PHRED-like rank. · read in full (as of 2026-09-18)
  • [confirmed] ClinGen Sequence Variant Interpretation Working Group calibration (Pejaver et al. 2022 Am J Hum Genet 109:2163-2177, expanded Bergquist et al. 2025 Genet Med 27:101402) found only 4 of 13 evaluated tools could reach PP3 strong, 9 topped out at moderate/supporting, and GERP++ failed to meet even PP3 supporting; CADD's developer-recommended threshold of 20.0 falls within BP4 moderate range, meaning it was being misapplied as pathogenic evidence. REVEL is the only tool covering PP3 strong (≥0.932) to BP4 very strong (≤0.003). AVI (18-feature model trained on allele-frequency threshold as rarity proxy) is not among ClinGen-calibrated tools and has no published PP3/BP4 thresholds; clinical labs therefore cannot apply it above supporting without formal calibration. · read in full (as of 2026-09-18)
  • [likely] Variant-effect predictor benchmarks are inflated by Type 1 circularity (same variant in train and test) and Type 2 circularity (same gene in train and test even with distinct variants), especially where genes have all-benign or all-pathogenic labels; ensemble methods (REVEL, MetaLR) and supervised models are prone to this bias. In a 46-method comparison, only 19 methods met criteria for unbiased clinical evaluation (not trained on ClinVar/HGMD, not using features from such methods, not using allele frequency), because 'the vast majority... have been trained on clinical databases like ClinVar, leading to circularity issues.' AVI's training on allele-frequency threshold and inclusion of conservation/LoF features risks similar allele-frequency leakage, and the 22/24 and 25/26 (vs preprint 24/26) SOTA claims depend on comparator and task selection. · read in full (as of 2026-09-18)
  • [confirmed] Independent genomic-model developer Carl de Boer (University of British Columbia, author of framework for comparing models like AlphaGenome) called AlphaGenome the 'field's leading model' but 'very slow and computationally intensive,' and warned the Atlas single-number AVI score 'has a clear use, but it also is probably going to be easily misinterpreted. We're talking about a very complex system, and there's a lot of moving parts.' He and IEEE Spectrum note the 1 Mb receptive field misses enhancers acting over longer distances, and many diseases involve multiple variants, limiting single-variant scores. · read in full (as of 2026-09-18)
  • [confirmed] Precomputed 1 PB Atlas trades on-demand freshness for convenience: locked to hg38 reference, no ancestry breakdown or false-positive rate for AVI disclosed in DeepMind post, no bulk-download/egress pricing, and commercial access requires separate Google Cloud license (non-commercial free) — a two-tier structure unlike AlphaFold DB's fully open release. This creates equity and staleness risks: reference-genome bias, inability to update for new cell types or hg38 successors without recomputation, and pharma/academic access asymmetry. · read in full (as of 2026-09-18)
  • [likely] Hybrid predictors that integrate protein structure, multiple sequence alignments, or population genetics data gain accuracy but add complexity, limited applicability where data are unavailable, and circularity risk; sequence-only PLM versions of TranceptEVE and SaProt 'decrease substantially' without MSA/3D structure, and 'the use of population genetics data can lead to data circularity concerns in clinical genetics applications.' This frames AVI's design choice (10 AlphaGenome + AlphaMissense + conservation + LoF + indel features trained on rarity) as inheriting both the power and the circularity/interpretability tradeoffs of hybrid models. · read in full (as of 2026-09-18)
  • [confirmed] Limitations include: 1 Mb input window misses distal enhancers acting over longer distances; performance is stronger for promoters/splicing than enhancers; overall accuracy is below AlphaFold's protein-structure accuracy; single-variant predictions do not capture multi-variant combinatorial effects or environmental/developmental context. · read in full (as of 2026-09-17)
  • [confirmed] Human genome comprises ~3-3.2 billion base pairs of four letters (A, C, G, T); 9 billion possible single-letter substitutions; ~2% codes for proteins, ~98% is non-coding regulatory DNA where most trait-associated variants lie. · read in full (as of 2026-09-17)
  • [confirmed] AlphaGenome was trained on ENCODE, GTEx, 4D Nucleome and FANTOM5 data, uses a U-Net-style encoder-transformer-decoder with sequence parallelism across TPUs (pre-training on 256 TPUv3), and was distilled to run all-modality predictions for a variant in <1 second on an NVIDIA H100, achieving ~80× speedup needed to make 9B-variant precomputation feasible. · read in full (as of 2026-09-17)
  • [confirmed] Atlas follows the AlphaFold Database precedent, which in 2022 expanded from ~190,000 experimental structures to >200 million predicted structures. · read in full (as of 2026-09-17)

Where we hit a limit / what to double-check