> ## Content Index > Fetch the complete content index at: https://www.implicator.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Google Launches AlphaGenome Atlas With Predictions for 9 Billion DNA Changes - URL: https://www.implicator.ai/google-alphagenome-atlas-9-billion-dna-variants/ - Published: 2026-09-09T06:32:00.000Z - Updated: 2026-09-09T06:32:00.000Z - Description: Google DeepMind has released AlphaGenome Atlas, a searchable database of predictions for 9 billion possible DNA changes. Its ranking score helps researchers choose variants for laboratory testing, with early work spanning a rare-disease case and UK Biobank data. - Author: Marcus Schuler - Tags: AI News Google DeepMind released [AlphaGenome Atlas](https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/?ref=implicator.ai) on Sept. 8, 2026, a searchable database containing predictions for 9 billion possible single-letter changes in human DNA. It moves calculations that once had to be run variant by variant into a precomputed resource. An accompanying score ranks which changes researchers should examine before committing time and laboratory work to follow-up experiments. What Changed - Google released AlphaGenome Atlas with precomputed predictions for 9 billion possible single-letter changes in human DNA. - The AVI score combines several signals to help researchers rank variants for closer study and laboratory testing. - Collaborators tested a predicted splicing effect in a rare-disease case involving the DNM1 gene. - Outside tests show limits in the underlying model. AlphaGenome has not been validated or approved for clinical use. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). ## A searchable genome A single-nucleotide variant is one DNA letter replaced with another. The [AlphaGenome model, published in January 2026](https://www.nature.com/articles/s41586-025-10014-0?ref=implicator.ai), can estimate how such changes affect molecular processes including gene expression and RNA splicing, a cellular process that cuts and joins sections of RNA before it is used to make a protein. The Atlas is a new layer: its September 2026 release stores those calculations across the genome in a dataset of about 1 petabyte, instead of making researchers request each prediction separately. The accompanying [AlphaGenome Variant Impact score](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/alphagenome-atlas.pdf?ref=implicator.ai), or AVI, reduces multiple signals to a ranking for each variant. It combines AlphaGenome predictions with AlphaMissense, evolutionary conservation and protein-coding features. Researchers can use the rank to decide which candidates deserve closer study, then inspect the contributing features to choose an experiment, such as testing whether a suspected change disrupts how RNA is assembled. Laura Covill, Anne O'Donnell-Luria and their Broad Institute colleagues used that process in research conducted with the Atlas team. FREE AI BRIEFING · WEEKDAYS Follow the evidence behind AI in biology. Get the AI stories shaping the day, with concise context from San Francisco. The briefing takes about five minutes and arrives at 4:45 a.m. Pacific, 7:45 a.m. Eastern. Email address Keep me informed Check your inbox for the confirmation link. Free. No hype. Unsubscribe anytime. ## A collaborator case In one unresolved case involving a patient with epileptic encephalopathy, AVI ranked a non-coding change in the DNM1 gene first. The prediction indicated that the change created a false splice site, causing cells to attach an abnormal extension to a brain-specific version of the protein. The collaborators tested that proposed mechanism in laboratory screens. Their experiments reproduced the abnormal splicing and found nearby variants with similar effects. Covill and O'Donnell-Luria are coauthors of the Atlas manuscript, so the findings are collaborator research rather than unaffiliated replication. The research team also applied Atlas features to protein measurements from 54,189 UK Biobank participants after quality control. In the Sept. 8 manuscript, that method produced a 22% increase in discoveries of statistically independent groups of rare non-coding variants associated with protein levels. It was not a gain in diagnostic accuracy. The analysis used slightly outdated Atlas releases because the UK Biobank computing platform was temporarily unavailable. Know someone who'd find this useful? [✉️ Email it to a friend in one click](mailto:?subject=A%20newsletter%20I%20think%20you%27d%20like&body=This%20is%20one%20of%20maybe%20three%20newsletters%20I%20actually%20read.%20The%20rest%20just%20pile%20up%2C%20unread%2C%20judging%20me.%0A%0AAnd%20yes%2C%20this%20email%20mostly%20wrote%20itself%2C%20which%20is%20a%20little%20on%20the%20nose%20for%20an%20AI%20newsletter.%20Doesn%27t%20make%20it%20wrong.%20implicator.ai%20is%20good.%0A%0ASubscribe%20free%3A%20https%3A%2F%2Fwww.implicator.ai%2Fsubscribe%2F%3Futm%5Fsource%3Dnewsletter%26utm%5Fmedium%3Dforward%26utm%5Fcampaign%3Demail%5Fforward), or they can [subscribe free here](https://www.implicator.ai/subscribe/?utm%5Fsource=newsletter&utm%5Fmedium=forward&utm%5Fcampaign=forward%5Fto%5Fcolleague). ## What outside tests show In an [independent CRISPRi benchmark](https://genomicsxai.github.io/blogs/2026-007/?ref=implicator.ai) published June 25 and updated Aug. 3, 2026, Alan Murphy and Peter K. Koo of Cold Spring Harbor Laboratory found that AlphaGenome recorded the highest correlations on both tests at predicting how disabling distant gene-control regions would alter gene activity, although it was in a near tie with Borzoi on the Fulco test. The predictions still understated the size of the measured effects, with the gap widening for more distant regions. The benchmark used K562 cells, which are well represented in model training, so it does not establish performance in other cell types. It tested the underlying AlphaGenome model, not the new Atlas or AVI. ## Research access and limits Atlas and AVI predict molecular effects. Those predictions can support a chain of evidence, but they do not establish a diagnosis on their own. AlphaGenome has not been validated or approved for clinical use, and the peer-review status of the Atlas manuscript released Sept. 8 is unconfirmed. The Atlas is available now for noncommercial academic research through a [web portal](https://alphagenome.google/atlas?ref=implicator.ai) and API. Commercial access to the Atlas is due through Google Cloud, although Google has not given a launch date; separate commercial access to the base AlphaGenome model already exists. Frequently Asked Questions What is AlphaGenome Atlas? It is a searchable database of precomputed predictions for 9 billion possible single-letter changes in human DNA. Its September 2026 release contains about 1 petabyte of data, letting researchers look up predictions instead of requesting each calculation separately. How does the AVI score help researchers? The AlphaGenome Variant Impact score combines AlphaGenome predictions with AlphaMissense, evolutionary conservation and protein-coding features. Researchers can rank variants, inspect the signals behind a score and choose which candidates to test. What did the DNM1 research show? Atlas collaborators used AVI to prioritize a non-coding variant in an unresolved rare-disease case. The model predicted abnormal splicing, and laboratory screens reproduced the effect and identified nearby variants with similar effects. The researchers were coauthors of the Atlas manuscript. Has AlphaGenome Atlas been independently validated? The article describes collaborator experiments and an independent benchmark of the underlying AlphaGenome model. That benchmark found useful performance alongside underestimated effects, but it tested one cell type and did not evaluate the new Atlas or AVI. Can AlphaGenome Atlas establish a diagnosis? Its molecular predictions can contribute evidence, but they do not establish a diagnosis on their own. AlphaGenome has not been validated or approved for clinical use. The Atlas is available for noncommercial academic research, with commercial Atlas access planned through Google Cloud. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). ## Related stories [Anthropic Routes Fable 5's Blocked Biology Requests to OpusAnthropic turned a safety result into an access decision. Biology requests that Fable 5 blocks now land on Opus 5, a model priced at half Fable 5's rates, on the strength of an alignment audit the company ran on itself and scored at 2.3.Implicator.ai![](https://www.implicator.ai/content/images/2026/07/2026-07-25-06.49.48-anthropic-opus-5-biology-routing-alignment-audit@2x.webp)](https://www.implicator.ai/anthropic-reroutes-biology-requests-blocked-on-fable-5-to-claude-opus-5/) [OpenAI's Miles Wang Holds $200M AI Drug Startup TalksOpenAI researcher Miles Wang is reportedly discussing $200 million for an unnamed AI drug startup at a $2 billion valuation. He disputes the figures and company description. With Chai and Isomorphic raising billions, investors are pricing a venture whose model and terms remain unsettled.Implicator.ai![](https://www.implicator.ai/content/images/2026/07/2026-07-15-10.01.06-wang_drug_ai@2x.webp)](https://www.implicator.ai/openais-miles-wang-holds-200m-ai-drug-startup-talks/) [Fudan's HealthClaw Posts 45.7% on Its Own Health BenchmarkFudan researchers released HealthClaw and reported 45.7% answer accuracy on a synthetic year-long benchmark they built and graded themselves. Plain full-history prompting scored higher. The paper's own tables trace the largest gain to a tool tested on the same public data it came from.Implicator.ai![](https://www.implicator.ai/content/images/2026/07/2026-07-16-03.18.15-fudan-healthclaw-open-source-health-agent-benchmark@2x.webp)](https://www.implicator.ai/fudans-open-source-health-agent-posts-45-7-accuracy-on-its-own-synthetic-benchmark/)