斯坦福与 Arc Institute 用 AI 设计全新病毒基因组，16 种在实验室成功杀死细菌

斯坦福大学和Arc研究所的一个团队让一个人工智能模型从零设计完整的病毒基因组，然后在实验室中构建了16种功能性病毒，这些病毒在自然界中并不存在。这项工作此前仅以预印本的形式发布。现在，它已通过同行评审并发表在《科学》杂志上：http://www.science.org/doi/10.1126/science.aec2657。

《纽约时报》报道了新的细节，特别是关于命中率：https://www.nytimes.com/2026/08/06/science/ai-viruses-bacteria-arc.html。该模型名为Evo，提出了700,000种可能的基因组。团队只追踪了最有前景的候选基因组，对285条序列进行了DNA化学合成，并将其插入细菌中。其中16条产生了能够复制的病毒。预印本中提到合成了302条基因组。

训练过程也变得更加清晰。Evo首先从数百万种动物、植物、微生物和病毒中约九万亿个核苷酸中学习，获取贯穿整个生命树的模式。只有在此之后，才进行第二轮、专门的训练，使用噬菌体Phi X-174的11个基因及约15,000个其最接近的亲属。对于博士生兼共同作者Samuel King来说，这是合乎逻辑的步骤。他说：“这感觉就是显而易见的下一步。”

最终得到的病毒也不仅仅是弱化的复制品。它们证明同自然病毒一样强健，有些甚至比Phi X-174复制得更快。“它们不仅仅是现有病毒的病弱版本，”牛津大学未参与该研究的蛋白质化学家Oliver Crook说。曼彻斯特大学的合成生物学家Patrick Cai称这项工作是“一个重要的里程碑”。

不过，Crook提醒不要过度期待。人工智能并没有发明任何基理上全新的东西。这些病毒与自然种类非常相似：https://www.biorxiv.org/content/10.64898/2026.06.12.731871v1，并依赖相同的生物学原理。Evo是否在其他病毒组中同样成功仍是一个悬而未决的问题。如果成功，结果可能为医学和生物技术提供有用工具。Crook说：“我们的很多科学研究都依赖病毒作为技术。”

生物安全监管的漏洞现在比以往任何时候都更明显。美国国立卫生研究院（NIH）在七月底发布了一项关于高风险生命科学研究的政策：https://www.nih.gov/about-nih/nih-director/statements/announcement-release-us-government-policy-stopping-high-risk-life-sciences-research。该政策禁止将病原体变得更具危险性的实验。但该机构表示，纯计算性工作，即在计算机上设计病毒 DNA 的工作，并不在其涵盖范围内，“除非它涉及关注的实体”。

问题显而易见。对于天花病毒，这种分类非常明确。而对于从 AI 模型生成的病毒，则不然。约翰霍普金斯大学健康安全中心的 Moritz Hanke 问：“我从未见过的东西，其风险是什么？”他认为研究进程与相关防护措施之间存在很大差距。“这之间存在巨大的脱节。”他设想的一种滥用情景是：“你可能会说，‘嘿，基因组语言模型，为我生成一个可以增加传播性或致死性的流感基因组。’”

该团队自行采取了预防措施。在训练过程中，Evo 没有使用感染人类的病毒的数据，也没有使用来自动物、植物或真菌的相关病原体的数据。这意味着该模型一开始就无法生成这些基因组。斯坦福大学的计算生物学家及研究共同作者 Brian Hie 表示：“我们只是想格外小心。”Hanke 称这是“相当值得称赞的”，尤其是因为没有任何官方规定要求这样做。他说：“因为他们没有从任何地方得到关于应该做什么的指导。”

2025年9月21日的原始文章：

加利福尼亚的一支研究团队使用人工智能设计了能够杀死细菌的活病毒，他们称之为“完整基因组的首次生成设计”。根据《麻省理工科技评论》的报道，该项目标志着迈向 AI 设计生命形式的早期一步。

这项工作由斯坦福大学和非营利组织Arc Institute的科学家完成。在一篇预印本论文中：https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1，他们描述了一个人工智能系统如何为病毒提出新的基因密码。团队随后将其中302个设计化学打印为DNA链，并将其暴露于大肠杆菌中。由AI生成的16种病毒成功复制并破坏了其细菌宿主。

“这非常惊人，实际上亲眼看到这个AI生成的球体，”负责Arc Institute实验室的Brian Hie说道，这间实验室正是病毒产生的地方。

该项目的核心是一种名为Evo的人工智能：https://the-decoder.com/evo-2-an-ai-model-that-understands-the-language-of-life/，它的功能类似大型语言模型，但训练对象是生物学而非文本。Evo没有从书籍和文章中学习，而是从大约两百万个噬菌体基因组中接受训练。在这项研究中，研究人员让它设计phiX174的变体，这是一种只包含11个基因、约5000个DNA字母的简单噬菌体。

纽约大学朗格健康中心的生物学家Jef Boeke将该项目描述为迈向AI设计生命的“令人印象深刻的第一步”，即使病毒本身在技术上并不算活体。他说AI的表现“出乎意料的好”，其设计“令人意外”，在基因顺序和排列上作出了人类科学家未曾考虑的改变。

并非所有人都信服。合成DNA的先驱J. Craig Venter称这种方法“只是加快了的试错实验”。他的实验室曾通过类似过程创建合成细胞，但过程更加缓慢，需要手工查阅科学文献。

该技术可能具有重大应用。医生长期以来尝试过噬菌体疗法，以治疗多药耐药的细菌感染。病毒也是基因治疗中的关键工具，用于将新基因传递到人体细胞中。AI设计的病毒可能使这两种方法更加有效。

但风险同样明显。团队刻意避免让Evo接受人类病原体的训练。即便如此，Venter对如果用同样的方法对危险病毒如天花或炭疽进行实验可能发生的情况表示“严重关切”。他表示：“我特别呼吁在任何病毒增强研究中保持极度谨慎，尤其是随机增强的时候，因为你根本不知道会得到什么。”

将该方法扩展到活细胞也要复杂得多。像大肠杆菌这样的细菌，其DNA数量大约是phiX174的1000倍。“复杂性将从令人震惊的程度跃升到远远超过宇宙中亚原子粒子数量的水平，”Boeke警告道。

尽管如此，Ginkgo Bioworks的首席执行官Jason Kelly认为，追求AI设计的细胞应该成为国家优先事项。他设想了自动化实验室，可以持续测试AI生成的基因组设计，并将结果反馈到模型中。“这将是国家级的科学里程碑，因为细胞是所有生命的组成基础，”他说。“美国应该确保我们率先实现这一目标。”

保持对AI的关注。清晰、有用、没有废话。

关注The Decoder获取AI新闻、背景故事和专家分析。

解码器：https://the-decoder.com/

A team from Stanford University and the Arc Institute had an AI model design complete viral genomes from scratch, then built 16 functional viruses in the lab that don't exist in nature. The work had previously only been available as a preprint. It has now been peer-reviewed and published in the journal Science：http://www.science.org/doi/10.1126/science.aec2657.

The New York Times reports：https://www.nytimes.com/2026/08/06/science/ai-viruses-bacteria-arc.html new details, particularly about the hit rate. The model, called Evo, proposed 700,000 possible genomes. The team pursued only the most promising candidates, had 285 sequences chemically synthesized as DNA, and inserted them into bacteria. Sixteen of those produced viruses capable of replicating. The preprint had mentioned 302 synthesized genomes.

The training process is also coming into sharper focus. Evo first learned from roughly nine trillion nucleotides drawn from millions of animals, plants, microbes, and viruses, picking up patterns that run through the entire tree of life. Only then did a second, specialized training round follow, using the 11 genes of the phage Phi X-174 and about 15,000 of its closest relatives. For doctoral student and co-author Samuel King, it was the logical move. "It just felt like the obvious next step," he said.

The resulting viruses weren't just weak copies, either. They proved as robust as natural ones, and some replicated even faster than Phi X-174. "They're not just sickly versions of stuff that already exists," says Oliver Crook, a protein chemist at the University of Oxford who wasn't involved in the study. Patrick Cai, a synthetic biologist at the University of Manchester, calls the work an "important milestone."

Crook tempers expectations, though. The AI didn't invent anything fundamentally new. The viruses are very similar to natural species：https://www.biorxiv.org/content/10.64898/2026.06.12.731871v1 and rely on the same biology. Whether Evo would be equally successful with other virus groups remains an open question. If it is, the results could yield useful tools for medicine and biotech. "A lot of our science rests on viruses as technology," Crook says.

A gap in biosafety regulation is now more visible than ever. The U.S. National Institutes of Health released a policy on high-risk life sciences research：https://www.nih.gov/about-nih/nih-director/statements/announcement-release-us-government-policy-stopping-high-risk-life-sciences-research in late July. It bans experiments that make pathogens more dangerous. But purely computational work, meaning designing viral DNA on a computer, isn't covered "unless it involves an entity of concern," the agency said.

The problem is obvious. With smallpox, that classification is clear-cut. With a virus that came out of an AI model, it isn't. "What is the risk of what I've never seen before?" asks Moritz Hanke of the Johns Hopkins Center for Health Security. He sees a wide gap between the pace of research and the guardrails around it. "There's just a huge disconnect." His misuse scenario: "You could say, 'Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal.'"

The team took precautions on its own. During training, Evo received no data on viruses that infect humans, nor on related pathogens from animals, plants, or fungi. That means the model can't generate those genomes in the first place. "We just wanted to be extra careful," says Brian Hie, a computational biologist at Stanford and co-author of the study. Hanke calls that "quite commendable," especially because no official rules required it. "Because they don't get any guidance from anywhere on what they should be doing," he says.

Original article from September 21, 2025:

A research team in California has used artificial intelligence to design working viruses that kill bacteria, in what they describe as the "first generative design of complete genomes." The project marks an early step toward AI-designed life forms, according to a report in MIT Technology Review.

The work was carried out by scientists at Stanford University and the nonprofit Arc Institute. In a preprint paper：https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1, they describe how an AI system proposed new genetic codes for viruses. The team then chemically printed 302 of these designs as DNA strands and exposed them to E. coli bacteria. Sixteen of the AI-generated viruses successfully replicated and destroyed their bacterial hosts.

"That was pretty striking, just actually seeing this AI-generated sphere," said Brian Hie, who runs the Arc Institute lab where the viruses were created.

At the center of the project is an AI called Evo：https://the-decoder.com/evo-2-an-ai-model-that-understands-the-language-of-life/, which functions like a large language model but is trained on biology instead of text. Instead of learning from books and articles, Evo was trained on about two million bacteriophage genomes. For this study, the researchers tasked it with proposing variants of phiX174, a simple bacteriophage containing only 11 genes and about 5,000 DNA letters.

Jef Boeke, a biologist at NYU Langone Health, described the project as an "impressive first step" toward AI-designed life, even though viruses themselves are not technically alive. He said the AI's performance was "surprisingly good" and its designs "unexpected," with changes to gene orders and arrangements that human scientists hadn't considered.

Not everyone is convinced. J. Craig Venter, who helped pioneer synthetic DNA, called the method "just a faster version of trial-and-error experiments." His lab once created synthetic cells through a similar process, but with much slower, manual searches through scientific literature.

The technology could have major applications. Doctors have long experimented with phage therapy as a treatment for multidrug-resistant bacterial infections. Viruses are also a key tool in gene therapy, where they deliver new genes into human cells. AI-designed viruses could make both approaches more effective.

But the risks are equally clear. The team deliberately avoided training Evo on human pathogens. Even so, Venter raised "grave concerns" about what could happen if the same approach were used on dangerous viruses like smallpox or anthrax. "One area where I urge extreme caution is any viral enhancement research, especially when it's random so you don't know what you are getting," he said.

Scaling the method to living cells is also far more complex. A bacterium like E. coli has about 1,000 times more DNA than phiX174. "The complexity would rocket from staggering to way, way more than the number of subatomic particles in the universe," Boeke warned.

Despite this, Jason Kelly, CEO of Ginkgo Bioworks, argues that pursuing AI-designed cells should be a national priority. He imagines automated labs that could continuously test AI-generated genome designs, feeding results back into the model. "This would be a nation-scale scientific milestone, as cells are the building blocks of all life," he said. "The US should make sure we get to it first."

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/