本报讯 人类基因组极其复杂。在约30亿个碱基对中，仅有2%负责编码蛋白质，剩余部分极难破译。9月8日，美国谷歌旗下的“深层思维”公司公布了一份由人工智能（AI）生成的“人类基因组图谱”，意在帮助科研人员破解人类生物密码。
单个核苷酸或碱基对的改变是人类基因组中最普遍的变异形式之一。这类变化造就了人与人之间在疾病风险等诸多方面的差异，其中一些罕见的单个核苷酸变异甚至能直接引发疾病。
最新问世的“AlphaGenome图谱”基于DeepMind去年发布的AI模型AlphaGenome的预测数据，描绘了90亿个DNA单碱基变化对人类基因组产生的影响，也就是所有潜在的突变。科研人员指出，这份图谱能辅助诊断罕见及不明病因的疾病，揭示常见疾病与生物学特征背后的机制，甚至有望解开DNA序列调控基因活性的某些隐藏规律。该图谱支持非商业用途免费使用。
DeepMind的机型经理Dhavi Hariharan指出，自AlphaGenome推出以来，约有9000名研究人员通过自动化编程接口（API）提取了该模型的预测数据，不过这种方法要求编写软件代码，给部分生物学家带来了障碍。
为构建AlphaGenome图谱，DeepMind对人类基因组中每个DNA碱基可能出现的3种核苷酸变化进行了预测，生成了1拍字节的数据量。同时，该图谱还收录了在人类基因组中观测到的超1亿个短片段插入或缺失。据DeepMind透露，这项工作的灵感来源于其AlphaFold数据库，该数据库内含超2亿条蛋白质结构预测结果，已被数百万使用者访问。
AlphaGenome图谱针对变异的多种潜在影响给出了数千条预测结果。但Hariharan提到，API使用者不断提出的需求是操作简便——希望能有一个简单的评分，来辅助他们判断是否要关注某个变异、是否有必要深入探究。
为此，AlphaGenome团队负责人Ziga Avsec带领科研人员研发了一种评估变异对生物学影响的指标，名为AlphaGenome变异影响（AVI）评分。DeepMind与学术研究人员在一篇预印本论文中报告称，仅凭这一数值，就能在临床基因组数据库里精准区分致病突变与无害变异。AVI评分及图谱中的其他预测结果已协助美国布罗德研究所的一个团队，将某项非编码变异确认为引发严重癫痫病例的潜在病因。
西班牙基因组调控中心的计算生物学家Mafalda Dias和Jonathan Frazer指出，罕见病研究人员通常依靠计算需要较低的模型来解析此类变异，因为对多数人来说，将AlphaGenome等模型应用于整个人类基因组并不现实。“通过消除这一计算壁垒，该图谱理应成为一项极具价值的资源。”
Avsec对借助该图谱探索人类基因组中普遍存在的短DNA片段——“基序”感到格外兴奋。科学家已明确其中部分基序能协助调控信使RNA（mRNA）的合成，mRNA携带着特定蛋白质的指令；另外一些基序则能吸引转录因子蛋白，进而调控特定基因的表达。但目前学界对这一机制的了解仍不够全面。
科研人员以该图谱的预测数据为指引，在人类基因组中锁定了数千个DNA基序，进而推断出这些基序在不同细胞类型中可能具备的功能，为科学界提供了一部可检索的非编码DNA“词典”。
英国埃克塞特大学的统计遗传学家Gareth Hawkes表示，该图谱还能帮助在庞大的全基因组和健康数据库中，挖掘“非编码”DNA 变异与生物性状及疾病之间的联系。Hawkes利用预测数据，在5.4万名英国生物样本库参与者的全基因组数据中，识别并解析了与数千种蛋白质血液水平相关的罕见非编码变异。
Avsec强调，由于该图谱每次只描绘单个突变对“参考”基因组产生的影响，因此无法直接捕捉多个变异的叠加效应，但可以借助一些统计方法来尝试解释此类影响。
英国威康桑格研究所的分子生物学家Ben Lehner认为，该图谱会非常有用，但不会带来革命性的突破。“这并非‘AlphaFold时刻’。”他表示，“像AlphaGenome这样的基因组模型，其预测表现还称不上出色，不应仅凭这些模型来制定临床决策。未来的重心应放在生成实验数据上，这将有助于基因组模型达到与AlphaFold同样出色的水平。”
德国马克斯·德尔布吕克分子医学中心的生物信息学家Martin Kircher表示，该图谱既不能取代实验，也无法在疾病诊断维度替代对个案细节的考量。“这是扩大对强大模型使用的一种实用的方式。”（文乐乐）
本文来源：https://finance.sina.com.cn/tech/roll/2026-09-10/doc-inirhrrh3252870.shtml
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
This report states that the human genome is extremely complex. Of the approximately 3 billion base pairs, only 2% are responsible for encoding proteins, and the remainder is difficult to decipher. On September 8, Google's DeepMind announced the release of an "Artificial Intelligence (AI)-generated Human Genome Atlas" aimed at helping researchers crack the human biological code.
Changes in single nucleotides or base pairs are among the most common forms of genetic variation in the human genome. These changes create differences between people in terms of disease risk and other aspects, with some rare single-nucleotide variants capable of directly causing disease.
The newly released "AlphaGenome Atlas" is based on predictive data from the AI model AlphaGenome, which DeepMind published last year. It maps the effects of 9 billion DNA single-base changes on the human genome, covering all potential mutations. Researchers indicate that this atlas can assist in diagnosing rare and unexplained diseases, revealing the mechanisms behind common diseases and biological traits, and may even help uncover hidden rules for DNA sequence regulation of gene activity. The atlas is available for free use for non-commercial purposes.
Dhavi Hariharan, product manager at DeepMind, noted that since the launch of AlphaGenome, approximately 9,000 researchers have extracted predictive data via the automated programming interface (API). However, this method requires writing software code, posing a barrier for some biologists.
To build the AlphaGenome Atlas, DeepMind analyzed the potential effects of three different nucleotide changes for every DNA base in the human genome, generating a petabyte of data. The atlas also includes over 100 million short insertions or deletions observed in the human genome. According to DeepMind, the inspiration for this work comes from the AlphaFold database, which contains over 200 million protein structure predictions and has been accessed by millions of users.
The AlphaGenome Atlas provides thousands of predictions regarding the potential impact of variations. Hariharan mentioned that a recurring request from API users is for a simpler scoring system to help them determine whether to focus on a specific variation or investigate it further.
To address this, the AlphaGenome team, led by Ziga Avsec, developed a metric to assess the biological impact of variations, known as the AlphaGenome Variation Impact (AVI) score. In a preprint paper, DeepMind and academic researchers report that this score alone can accurately distinguish between pathogenic and harmless variants in clinical genomic databases. The AVI score and other predictions from the atlas have assisted a team at the Broad Institute in identifying a non-coding variation as a potential cause of a severe epilepsy case.
Mafalda Dias and Jonathan Frazer, computational biologists at the Center for Genomic Regulation in Spain, noted that researchers studying rare diseases often rely on less computationally intensive models to interpret these variations, as running models like AlphaGenome across the entire genome is not feasible for most. "By removing this computational barrier, the atlas should be a valuable resource."
Avsec is particularly excited about using the atlas to explore the "motifs" — short DNA fragments — that are common in the human genome. Scientists already understand that some motifs help regulate the synthesis of messenger RNA (mRNA), which carries instructions for specific proteins; others attract transcription factor proteins to regulate gene expression. However, the field's understanding of these mechanisms is still incomplete.
Researchers have used the atlas's predictive data to locate thousands of DNA motifs within the human genome, inferring their potential functions in different cell types and creating a searchable "dictionary" of non-coding DNA for the science community.
Gareth Hawkes, a statistical geneticist at the University of Exeter, stated that the atlas can also aid in discovering correlations between non-coding DNA variations and biological traits and diseases within large-scale genomic and health databases. Hawkes utilized the predictive data to identify and analyze rare non-coding variations associated with thousands of protein levels in the blood of 54,000 participants in the UK Biobank.
Avsec emphasized that because the atlas plots the impact of individual mutations relative to a "reference" genome, it cannot directly capture the additive effects of multiple variations, though this can be attempted using statistical methods.
Ben Lehner, a molecular biologist at the Wellcome Sanger Institute, believes the atlas will be useful but not revolutionary. "This is not an 'AlphaFold moment,'" he said. "Genomic models like AlphaGenome do not yet perform well enough for their predictions to be the sole basis for clinical decisions. The focus should remain on generating experimental data, which will help genomic models reach the level of accuracy seen with AlphaFold."
Martin Kircher, a bioinformatician at the Max Delbrück Center for Molecular Medicine in Germany, stated that the atlas cannot replace experiments or account for the individual details required in disease diagnosis. However, he praised the project as a "practical way to extend the use of a powerful model."
The original article was published on September 10, 2026, by the People's Daily, authored by Wen Lele.