基因组图谱如何帮助研究人员更好地理解疾病

内容总结:
谷歌DeepMind正利用人工智能模型AlphaGenome加速人类对基因组的解读,帮助科研人员更快锁定致病基因变异,推动罕见病诊断、药物研发等领域的研究进程。
人类基因组由约30亿个碱基对组成,堪称人体的“指令手册”。其中任何一个字母的改变,都可能导致疾病。然而,由于可能的单字母变异多达90亿种,科学家难以逐一通过实验验证。为此,谷歌DeepMind开发了AlphaGenome模型,可对长达100万个碱基对的DNA序列进行数千项预测,并通过比对这些预测结果,帮助研究人员更快找到关键变异。
该团队负责人、研究科学家日加·阿夫塞克表示,许多遗传疾病仅由单个字母的改变引起,但科学家没有时间开展数十亿项研究来确定哪些变化真正产生影响。为此,团队利用AlphaGenome预测了全部90亿种单字母变异的后果,并构建了大规模数据集“AlphaGenome Atlas”,供科学家优先选择研究方向。
研究人员指出,尽管科学界已能读取全部30亿个碱基对,但仍未完全破译基因组的语义,尤其是调控基因表达和疾病风险的非编码“暗物质”区域。传统湿实验仍是取得突破的最佳途径,但成本高、耗时长,一项突破可能需要数年。以罕见病患者为例,其DNA中可能存在约400万个变异,即便缩小到罕见变异范围,仍可能剩下成百上千个候选目标。借助AlphaGenome Atlas,科学家可获得一个易于解读的AVI评分,快速判断某个变异对整个基因组的潜在影响,从而优先跟进最具破坏性的变化。
在实际应用中,埃克塞特大学研究员加雷思·霍克斯利用该工具从10亿个遗传变异中筛选出影响最大的1%,最终确定了19个与体重指数相关的变异组用于后续研究。布罗德研究所研究员劳拉·科维尔则利用AVI评分优先研究未解决的罕见病,AlphaGenome Atlas在其中识别出一个与神经发育障碍相关基因的关键变异,为后续研究提供了重要证据,并最终帮助破解了病例。
AlphaGenome于2025年6月发布,AlphaGenome Atlas于同年9月推出。谷歌DeepMind近期又发布了AlphaGenome 2,针对初代模型在细胞类型特异性方面的局限进行改进,从组织层面的平均值进一步细化到预测遗传变异如何影响特定细胞类型,帮助研究人员锁定疾病的细胞驱动因素。团队还在研究更优的方法来衡量模型预测的置信度,以便研究人员判断每项结果的可信程度。
阿夫塞克表示,短期目标是利用人工智能更好地理解基因组的“语法”,准确预测突变影响;长期目标是在此基础上构建完整的人类细胞数字模拟,使研究人员能够通过计算建模研究疾病机制、理解罕见遗传病并测试候选疗法。
中文翻译:
基因组图谱如何帮助研究人员更好地理解疾病
你的基因组是你身体的说明书:是指导身体运作的蓝图。它包含30亿个DNA碱基对。由于这些DNA部分负责开启和关闭基因,它可能决定了一个人是患病还是免于患病。确定哪些DNA变化会导致疾病(以及其他结果),就像大海捞针。
示意图。你的基因组,一点一滴地解读。
解码DNA序列正是Google DeepMind模型AlphaGenome的目标。AlphaGenome可以接收长度为100万个碱基对的DNA片段序列,并做出数千项预测。通过比较这些针对DNA变化的预测结果,科学家可以更快地找到那根“针”。
我们采访了Google DeepMind研究科学家Žiga Avsec,以了解更多信息。
你在Google做什么工作?
ŽA:我是一名研究科学家,负责监督我们的基因组学计划。基本上,我带领一个团队构建模型,帮助我们更好地理解基因组。
你的团队如何让科学家更容易预测和理解疾病?
归根结底,一切都取决于基因组如何运作。它包含30亿个DNA碱基对。每个碱基对包含两个碱基,我们称之为“字母”。这些字母——A(腺嘌呤)、T(胸腺嘧啶)、C(胞嘧啶)、G(鸟嘌呤)——就像一套生物字母表,拼写出造就每个人之所以为人的基因说明书。
由于你可以将任何一个字母替换为另外三种字母,因此存在90亿种可能的单字母变化。许多遗传疾病仅由其中一个字母的改变引起。但科学家没有时间开展数十亿项研究来确定哪些变化会产生影响。
因此,我们用AlphaGenome预测了全部90亿种单字母变化的结果,并创建了一个庞大的数据集——AlphaGenome Atlas——以便科学家更快地决定研究什么。
示意图。示例中的字母为虚构。
是什么阻碍了科学家更好地理解基因组?
虽然科学可以读取人类DNA全部30亿个碱基对,但我们仍未完全破译遗传密码的语义——尤其是调控基因表达和疾病风险的非编码“暗物质”。实验研究(也称为湿实验室研究)是帮助我们在理解上取得突破的最佳方式,但它既昂贵又耗时。一项突破可能需要数年时间。
假设你对一名罕见病患者的DNA进行了测序。你最终可能会得到400万个变异——即他们的DNA与标准参考人类基因组之间存在400万种差异。但问题是:哪个变异导致了他们的病症?也许你可以将列表缩小到仅包含罕见变异。但即便如此,你可能仍然面临数百甚至数千个选项。
当科学家使用AlphaGenome Atlas时,他们会得到一个单一、易于读取的评分——AVI(AlphaGenome变异影响)评分——它能立即显示某个变异在整个基因组中的潜在影响。这有助于他们确定优先级,更轻松地找到最具破坏性的变化以进行后续研究。
示意图。每个点代表多个变异。评分为虚构。
你能举一些AlphaGenome如何影响普通人的例子吗?
像AlphaGenome这样的工具让我们从仅仅读取DNA,迈向帮助我们更好地理解疾病背后的分子机制,为研究人员提供新的线索供其验证,并朝着有朝一日对完整人体细胞进行数字化模拟迈出重要一步。
这将通过多种方式产生影响。它可以加速研究,帮助社会提高对患有罕见或此前未被表征的遗传疾病患者的诊断和个性化治疗能力。它还可以帮助制药研究人员更快地识别和验证高概率治疗靶点,从而加速药物发现流程。在医疗健康之外,AlphaGenome的模型架构还可以帮助研究人员应对更广泛的社会挑战,包括培育气候适应性作物。
示意图。AlphaGenome可能发挥作用的途径。
AlphaGenome促成了哪些类型的突破?
AlphaGenome还相当新。我们于2025年6月发布,今年9月刚刚推出AlphaGenome Atlas。但研究人员已经在使用它来推动科学进步。
埃克塞特大学医学研究委员会研究员Gareth Hawkes使用AlphaGenome Atlas将10亿个遗传变异的列表缩小。通过聚焦于影响最大的1%,他识别出19个与体重指数相关的变异组,并在研究中优先关注。
布罗德研究所研究员Laura Covill使用AVI评分在其未解罕见病研究中确定遗传变异的优先级。AlphaGenome Atlas识别出一个与神经发育障碍相关的基因中的关键变异,提供了关键的支持性证据,推动了他们的后续研究,包括破解了该病例。
你接下来希望做什么?
我们刚刚发布了最新模型AlphaGenome 2,它解决了AlphaGenome的一个关键局限:细胞类型特异性。虽然AlphaGenome处于最先进水平,但它主要是在整块组织或细胞系的数据上训练的。我们知道这些组织中的单个细胞类型具有非常特定的作用,并且在疾病中可能发生功能障碍。AlphaGenome 2超越了组织水平的平均值,能够预测遗传变异如何破坏这些单个细胞类型,帮助研究人员精确定位疾病的细胞驱动因素。
我们还在积极研究更好的方法来衡量模型对其预测的置信度,以便研究人员知道每个结果的可信程度。从早期分析来看,我们发现预测效应强烈的变异表现相当不错,达到了较高的准确度。与往常一样,任何使用AlphaGenome的人都应结合其他信息来源来解读其预测结果。
你在生物学领域工作的最终目标是什么?
短期内,我们希望利用AI更好地理解基因组的语法,并准确预测突变的影响。长期来看,我们希望在此基础上构建一个功能完整的人体细胞数字模拟。
这将使研究人员能够通过计算模拟疾病机制、理解罕见遗传病,并测试候选疗法。
英文来源:
How genome mapping can help researchers better understand disease
Your genome is your body’s instruction manual: the blueprint that informs how your body functions. It contains 3 billion base pairs of DNA. Because this DNA is responsible in part for switching genes on and off, it could be the difference between developing a disease and avoiding one. Determining which DNA changes cause diseases (and other outcomes) is like finding a needle in a haystack.
Illustrative. Your genome, one point at a time.
Decoding the DNA sequence is the goal behind AlphaGenome, a model from Google DeepMind. AlphaGenome can take a partial DNA sequence of 1 million base pairs in length, and make thousands of predictions. By comparing these predictions for DNA changes, scientists can find that needle faster.
We talked to Google DeepMind research scientist Žiga Avsec to learn more.
What do you do at Google?
ŽA: I’m a research scientist overseeing our genomics initiative. Basically, I lead a team that builds models to help us better understand the genome.
How is your team making it easier for scientists to predict and understand diseases?
It really all comes down to how the genome works. It contains 3 billion DNA base pairs. And each pair contains two bases, which we call “letters.” These letters — A (adenine), T (thymine), C (cytosine), G (guanine) — act like a biological alphabet, spelling out the genetic instruction manual that makes someone who they are.
Since you can replace any of those letters with three other letters, there are 9 billion possible single-letter changes. Many genetic diseases are caused by just one of these letters changing. But scientists don’t have time to run billions of studies to figure out which changes make a difference.
So we used AlphaGenome to predict the outcomes of all 9 billion single-letter changes, and we created a massive dataset — AlphaGenome Atlas — so scientists can decide what to research faster.
Illustrative. The letters are invented for this example.
What’s holding scientists back from better understanding the genome?
While science can read all 3 billion base pairs of human DNA, we still haven’t fully deciphered the semantics of the genetic code — particularly the non-coding "dark matter" that regulates gene expression and disease risk. Experimental research, or as it’s also called, wet-lab research is the best way to help us reach breakthroughs in our understanding, but it’s expensive and time-consuming. A single breakthrough could take years.
Let’s say you sequence DNA from a patient with a rare disease. You might end up with 4 million variants — 4 million ways their DNA differs from the standard reference human genome. But the question is: Which variant is causing their condition? Maybe you narrow the list to only include rare variants. But even then, you could be left with hundreds or thousands of options.
When scientists use AlphaGenome Atlas, they get a single, easy-to-read score — the AVI (AlphaGenome Variant Impact) score — which instantly indicates the potential impact of a variant across the full genome. This helps them prioritize and more easily find the most disruptive changes to follow up on.
Illustrative. Each dot stands for many variants. Scores are invented.
Can you give me some examples of how AlphaGenome impacts everyday people?
Tools like AlphaGenome move us from merely reading DNA to helping us better understand the molecular mechanisms behind a disease, giving researchers new leads to test and taking an important step towards someday digitally simulating a complete human cell.
There are several ways that will have an impact. It can accelerate research that could help society improve our ability to diagnose and personalize treatment for patients suffering from rare or previously uncharacterized genetic disorders. It could also speed up drug discovery pipelines by helping pharmaceutical researchers identify and validate high-probability therapeutic targets faster. Beyond healthcare, AlphaGenome’s model architecture could help researchers tackle broader societal challenges, including breeding climate-resilient crops.
Illustrative. Possible ways AlphaGenome can help.
What kinds of breakthroughs has AlphaGenome enabled?
AlphaGenome is still pretty new. We launched it in June 2025, and we just introduced AlphaGenome Atlas in September this year. But researchers are already using it to advance science.
Gareth Hawkes, a Medical Research Council fellow at the University of Exeter, used AlphaGenome Atlas to narrow down a list of 1 billion genetic variants. By focusing on the most impactful 1%, he identified 19 variant groups linked to body mass index to prioritize in his research.
And Laura Covill, a Broad Institute research fellow, used the AVI score to prioritize genetic variants in their research of unsolved rare diseases. AlphaGenome Atlas identified a critical variant in a gene linked to neurodevelopmental disorders, providing crucial supporting evidence that drove their subsequent research, including solving the case.
What are you hoping to do next?
We just launched our latest model, AlphaGenome 2, which tackles one of the key limitations of AlphaGenome: cell type specificity. While AlphaGenome is state-of-the-art, it was trained mostly on data from whole tissues or cell lines. We know that individual cell types in those tissues have a very specific role and can malfunction in disease. AlphaGenome 2 moves beyond tissue-level averages to predict how genetic variants disrupt these individual cell types, helping researchers pinpoint the cellular drivers of disease.
We’re also actively researching better ways to measure how confident our model is in its predictions so researchers know how much to trust each result. From early analyses, we see that variants with strongly predicted effects are quite good and achieve high levels of accuracy. As always, anyone using AlphaGenome should interpret its predictions with other sources of information.
What's the ultimate goal of your work in biology?
In the short term, we’d like to use AI to better understand the grammar of the genome and accurately predict the effects of mutations. In the long term, we’d like to build on this foundation towards a fully functioning digital simulation of a human cell.
This will allow researchers to computationally model disease mechanisms, understand rare genetic conditions, and test candidate therapies.