Google DeepMind 发布 AlphaGenome Atlas，为 90 亿个人类 DNA 变异提供预计算分子效应预测和 AVI 评分

谷歌 DeepMind 发布了 AlphaGenome Atlas：https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/，这是一个预计算预测人类基因组中每一个可能的单核苷酸变异分子效应的目录。大约有 90 亿个单字母变化。此次发布还引入了 AlphaGenome 变异影响（AVI）评分，这是一个按预测影响对变异进行排名的单一数字，并提供了每个变异的特征归因以及全基因组的基序集合。该资源作为免费的网页门户提供学术使用：https://alphagenome.google/atlas，通过 AlphaGenome API 提供使用：https://github.com/google-deepmind/alphagenome，并作为 Google Antigravity 的技能提供：https://antigravity.google/use-cases/science。

它可以部署吗？部分可以。Atlas 今天可通过门户和 API 查询用于非商业研究，商业接入在 Google Cloud 上显示为“即将推出”。基础的 AlphaGenome 模型已经可在 GitHub 上用于学术用途：https://github.com/google-deepmind/alphagenome_research，并可在 Google Cloud 的 Model Garden 上用于商业用途：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/alphagenome。

AlphaGenome：https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/，于 2025 年 6 月发布，预测 DNA 变异如何改变诸如基因表达和 RNA 剪接等分子过程。它已被广泛使用，但总是一次预测一个变异或一个区域。Atlas 改变了工作单位。DeepMind 团队对所有 90 亿个单核苷酸变异运行了 AlphaGenome 并存储结果，生成了 1 PB 的数据集。这比 AlphaFold 数据库：https://deepmind.google/blog/alphafold-reveals-the-structure-of-the-protein-universe/，其包含超过 2 亿条蛋白质结构预测，大约大 30 倍。

在实验室测试 90 亿个突变是不可能的，而且针对每个候选变异按需运行大型模型在基因组规模研究中也很慢。附带解释的查找表消除了这两个瓶颈。

该地图集展示了4个关联资源：

DeepMind团队报告称，AVI评分在许多变异致病性和罕见疾病基准测试中表现出最佳性能。技术报告：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 包含基准详情。

该图谱已经为人类基因组中每一个90亿个单字母变化预先计算了预测。本演示展示了单次查询返回的结果。

点击任意字母以进行替换。编码碱基会获得AlphaMissense蛋白术语；非编码碱基仅依赖AlphaGenome。

发布前，三组研究团队使用了该图谱，他们的结果支撑了此次公告：

需要与我们合作推广您的GitHub仓库、Hugging Face页面、产品发布或网络研讨会等吗？请联系我们：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq是Marktechpost Media Inc.的首席执行官。作为一名有远见的企业家和工程师，Asif致力于利用人工智能的潜力造福社会。他最近的项目是推出人工智能媒体平台Marktechpost，该平台因深入报道机器学习和深度学习新闻而脱颖而出，这些报道既技术过硬又易于广大观众理解。该平台每月浏览量超过200万，显示了其受欢迎程度。

以实践者为核心的人工智能/机器学习新闻和分析，每月有超过100万开发者和研究人员阅读。

Google DeepMind has released AlphaGenome Atlas：https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/, a catalogue of precomputed predictions for the molecular effects of every possible single-nucleotide variant in the human genome. That is roughly 9 billion single-letter changes. The release also introduces the AlphaGenome Variant Impact (AVI) score, a single number that ranks variants by predicted impact, plus per-variant feature attributions and a genome-wide motif collection. The resource ships as a free web portal：https://alphagenome.google/atlas for academic use, through the AlphaGenome API：https://github.com/google-deepmind/alphagenome, and as a skill in Google Antigravity：https://antigravity.google/use-cases/science.

Is it deployable? Partially. The Atlas is queryable today for non-commercial research via the portal and API, and commercial access on Google Cloud is listed as “coming soon”. The underlying AlphaGenome model is already available for academic use on GitHub：https://github.com/google-deepmind/alphagenome_research and for commercial use on Model Garden on Google Cloud：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/alphagenome.

AlphaGenome：https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/, released in June 2025, predicts how a DNA variant changes molecular processes such as gene expression and RNA splicing. It has been used widely, but always one variant or one region at a time. The Atlas changes the unit of work. DeepMind team ran AlphaGenome across all 9 billion single-nucleotide variants and stored the outputs, producing a 1-petabyte dataset. This is more than 30 times larger than the AlphaFold Database：https://deepmind.google/blog/alphafold-reveals-the-structure-of-the-protein-universe/, which holds over 200 million protein structure predictions.

Testing 9 billion mutations in a lab is not feasible, and running a large model on demand for each candidate variant is slow for genome-scale studies. A lookup table with attached interpretation removes both bottlenecks.

The Atlas exposes 4 linked resources:

DeepMind team reports that the AVI score delivers best-in-class performance across many variant pathogenicity and rare disease benchmarks. The technical report：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 carries the benchmark details.

The Atlas already holds a precomputed prediction for every one of the 9 billion single-letter changes in the human genome. This demo shows what a single lookup returns.

Click any letter to swap it. Coding bases get an AlphaMissense protein term; non-coding bases rely on AlphaGenome alone.

Three research groups used the Atlas before launch, and their results anchor the announcement:

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.

Practitioner-first AI/ML news and analysis, read by 1M+ developers and researchers every month.