Did AI find a new gene-editing tool? Why Anthropic’s ‘autonomous discovery’ is sparking controversy

AI executives are betting big on biology. They say their large language models that have already dramatically changed the fields of coding and software development and mathematics can upend biological research in the same way.

But one company’s attempt to do so has already set off a wave of skepticism and controversy.

Anthropic, the AI and research giant behind Claude, last month said its biology research lab had used AI agents to discover an unusual pattern of DNA in the genetic code of viruses, specifically a “ novel enzyme system. ” The signature is similar to the microbial system harnessed in gene-editing tools such as CRISPR Cas-9, which is used to modify the DNA of living organisms. The technology is commonplace in labs around the world and won the Nobel Prize in chemistry for its inventors in 2020.

Dario Amodei, the chief executive of Anthropic, said on social media the finding offered great scientific potential. “Today we announced the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism,” he wrote on X.

Anthropic said the discovery had been made “with only high-level direction from our scientists” and shares of companies developing gene-editing technology fell shortly after the announcement.

OpenAI says it has solved one of math’s ‘Millennium Problems’

However, some experts not involved in the work said the finding was early and incremental, and far from a genuine breakthrough.

“These are very interesting preliminary results, but at this point they don’t demonstrate gene editing or provide a mechanistic picture of exactly what’s happening,” said Aaron Engelhart, an associate professor at the University of Minnesota’s Department of Genetics, Cell Biology, and Development.

The Anthropic researchers have not determined what the newfound viral sequences actually do and thus whether they represent a mechanism that has practical applications similar to those used in CRISPR tools, according to a document detailing the research. The research also has not been published in a peer-reviewed scientific journal.

Kent Nishimura/AFP/Getty Images

Scientists, including Engelhart, said Anthropic’s finding definitely warrants further investigation, and a news release from the company quoted Feng Zhang — a professor at MIT and the Broad Institute in Cambridge, Massachusetts and a pioneer of CRISPR genome editing for sickle cell disease treatment among other advances — who called the work “genuinely intriguing.”

But just days after Anthropic’s announcement, Mario Rodríguez Mestre, who said he used Claude extensively for his recently completed doctoral research at the University of Copenhagen, claimed his unpublished research describes the same viral signatures.

“It is essentially the same finding. This is not simply a case of two groups studying related protein families or similar biological systems,” Mestre told CNN via email.

Mestre’s claims echo those made by mathematicians after OpenAI, which built ChatGPT, announced it had solved a longstanding and high-profile math problem . They also raise questions about how Claude’s find unfolded.

The biologists working at the Anthropic lab said they gave Claude a prompt to search a massive database of DNA sequences for “interesting new examples” of reverse transcriptase, or RT, an enzyme that copies RNA into DNA and thus plays a role in the spread of genetic information in an organism.

Samyukta Lakshmi/Bloomberg/Getty Images/File

AI creates 16 new viruses from scratch, showing promise for drug resistance and drawing warnings about potential for misuse

The company said 950 AI agents, which can independently plan and execute tasks, spent 21 hours searching the data, picking out 3,500 RTs. It narrowed the pool down to the 20 most compelling. For an experienced scientist, this type of analysis can take weeks to months, according to a statement from Anthropic.

Then, “one of the agents spotted something remarkable: a repeating pattern of DNA sequences that occurs next to the gene for an odd-looking RT,” the company noted in the statement. “After a thorough analysis it was convinced that it had found a new biological system, and filed a report for human review.”

Anthropic said scientists at its biology lab then analyzed and tested the finding, which they described as a “previously uncharacterized” biological system in bacteriophages, a type of virus that attacks bacteria, with a similar structure to CRISPR, which is naturally found in bacteria.

Mestre, however, said he had shared unpublished research on his private Claude account, including drafts of his dissertation, analyses and material describing these systems.

“I am not claiming that Anthropic deliberately took our work. I cannot demonstrate that,” he told CNN via email.

“Either information related to our work somehow reached the model and it was trained on it. The second possibility, which I currently think is more plausible, is that there was considerably more prior scientific knowledge and human direction behind the search than the phrase ‘autonomous discovery’ suggests,” Mestre added, referring to language used in the release.

Anthropic did not respond to an emailed request to comment.

It’s not unusual in the scientific process for different research groups to converge on the same result, but Mestre’s allegation is serious, and many scientists may think twice in the future about inputting their unpublished results into these tools, said Ilya Finkelstein, a professor in molecular biosciences at University of Texas at Austin.

Still, Anthropic’s announcement does suggest that AI is beginning to work with more autonomy in science, said Gustavo Sudre, a professor of genomic neuroimaging and artificial intelligence at King’s College London.

“It received a broad direction and then had to use its own judgement about what was interesting in the data. That is a real shift from the usual ‘analyze this dataset for me,’” he said in an email.

AI excels at the “needle in a haystack search,” rather than the “imaginative breakthrough,” Finkelstein added. AI tools would help explore domains where students would get too tired or bored with finding an interesting pattern worth further investigation, he noted.

“They can pursue a very large number of unproductive paths rapidly,” Finkelstein told CNN. “They are really good at grunt work when orchestrated by experts in the field.”

Finkelstein also noted that the Anthropic paper’s lead author is Dr. Peter Yoon , who used to work in the lab of Jennifer Doudna, who is one of CRISPR Cas-9’s architects and shares the 2020 Nobel Prize with Emmanuelle Charpentier. So, Anthropic’s AI agents were guided by people at the top of their fields, he said.

“Four of six authors are senior molecular biologists and domain experts. Give that group a heap of compute and frontier models (without safeguards, I’m guessing), and I’m sure we’ll be seeing more interesting bioinformatic discoveries,” Finkelstein wrote in a blog post on his lab’s website.

However, other researchers say talk of autonomous AI-driven discovery is hype that gets ahead of the actual science, and models aren’t yet able to generate meaningful results on their own — despite the claims to the contrary.

Letting a model comb through data at a scale no human can match does move a field forward, but it doesn’t necessarily herald a big advance, Sudre said. “From experience I can tell you that the gap between ‘interesting pattern’ and ‘useful knowledge’ is where most of the work lives,” he noted.

Anthropic launched Claude Science, its tool designed for research, earlier this year. Its rivals have all recently unveiled similar tools. OpenAI has GPT-Rosalind, Microsoft has Quine and Google DeepMind has Co-Scientist. They operate like large language models but also harness specific scientific datasets such as DNA sequences.

However, models are only as reliable as the data they are trained on. For example, the strength of Google DeepMind’s AlphaFold, which won the 2024 Nobel Prize in chemistry for decoding protein structures, comes from training the AI system on a narrow, specialized data source collected over decades .

AI developers are under intense public scrutiny amid growing concern over rogue AI agents and whether AI companies can adequately police them. Biological discoveries can offer company executives an opportunity to craft a more positive narrative that counterbalances some of the doomsaying surrounding their products.

But moving toward significant and tangible findings in the field likely requires integrating the AI models into hands-on lab work through robotics.

“Biology is not maths and at some point the intelligence has to be expressed in the physical world through a robot that can do the fiddly one-off experimental tasks that a good student would normally do almost instinctively,” said Yuval Elani, an associate professor in biochemical technologies at Imperial College London. “From what I have seen of robotics and automation we aren’t really close. So it will be a while until we can really justify the hype around AI in biology.”

Illustration by Sarah Hashemi/CNN

Parents confront a new fear: Could AI end humanity before their kids grow up?

Anthropic says it understands there is a limit to what artificial intelligence can do alone and is investing in “ model hardware ” that will allow AI agents to safely operate physical devices.

Sudre at King’s College, whose work integrates genomic, clinical, neural and cognitive data to predict how mental health symptoms will develop in children, is emphatic that the role of humans in science isn’t shrinking. The risk isn’t that scientists will get pushed out of discovery, but that they’ll get lazy and won’t use their judgment to understand whether a result is meaningful, he said.

The next step for Anthropic’s scientists is to do the slow work of physical lab experiments to understand the pattern they have identified, and whether the signature might have a practical use in biotechnology.

“Sure, in future some of that can go to robots, too,” Sudre said. “But here is the thing, the questions worth asking in science don’t stay hidden in databases waiting to be found. They come from us — from what we care about, what worries us.”

人工智能领域的高管们正大力投资生物学。他们表示，他们开发的大型语言模型已经极大地改变了编程、软件开发和数学领域，同样也能颠覆生物学研究。

但一家公司试图这样做已经引发了一阵质疑和争议。

上个月，开发出人工智能机器人Claude的科技巨头Anthropic公司宣布，其生物学研究实验室利用人工智能代理在病毒的遗传密码中发现了一种不寻常的DNA模式，具体来说是一种“新型酶系统”。这种特征与基因编辑工具（例如CRISPR-Cas-9）中使用的微生物系统类似，后者用于修改生物体的DNA。这项技术在世界各地的实验室中已得到广泛应用，其发明者也因此荣获2020年诺贝尔化学奖。

Anthropic公司首席执行官达里奥·阿莫迪在社交媒体上表示，这一发现具有巨大的科学潜力。他在X上写道：“今天，我们宣布了克劳德领导的一项分子机器的发现，我们怀疑它可能代表一种新的基因编辑机制。”

人猿科技表示，这一发现“仅在我们科学家的高层指导下”取得，消息公布后不久，开发基因编辑技术的公司的股价下跌。

OpenAI 声称已解决数学界的“千年难题”之一

然而，一些未参与这项研究的专家表示，这一发现尚处于早期阶段，属于渐进式发现，远非真正的突破。

“这些是非常有趣的初步结果，但目前它们并不能证明基因编辑的存在，也不能提供关于究竟发生了什么的机制描述，”明尼苏达大学遗传学、细胞生物学和发育系副教授亚伦·恩格尔哈特说。

根据一份详细介绍该研究的文件，Anthropic公司的研究人员尚未确定新发现的病毒序列的实际功能，因此也无法确定它们是否代表了一种类似于CRISPR工具的、具有实际应用价值的机制。此外，该研究也尚未在同行评审的科学期刊上发表。

Kent Nishimura/AFP/Getty Images

包括恩格尔哈特在内的科学家们表示，Anthropic 的发现绝对值得进一步调查。该公司发布的新闻稿援引了麻省理工学院和马萨诸塞州剑桥市博德研究所的教授、CRISPR 基因组编辑技术在治疗镰状细胞病等方面的先驱张锋的话，称这项工作“确实引人入胜”。

但就在 Anthropic 宣布这一消息几天后，马里奥·罗德里格斯·梅斯特雷 (Mario Rodríguez Mestre) 声称，他最近在哥本哈根大学完成的博士研究中大量使用了 Claude，他未发表的研究描述了相同的病毒特征。

“这本质上是相同的发现。这并非仅仅是两个研究小组在研究相关的蛋白质家族或类似的生物系统，”梅斯特通过电子邮件告诉CNN。

梅斯特的说法与OpenAI（ChatGPT的开发者）宣布解决一个长期存在的、备受瞩目的数学难题后，数学家们的说法不谋而合。这些说法也引发了人们对克劳德发现过程的疑问。

在 Anthropic 实验室工作的生物学家表示，他们给 Claude 布置了一个任务，让他搜索庞大的 DNA 序列数据库，寻找逆转录酶（RT）的“有趣的新例子”。逆转录酶是一种将 RNA 复制到 DNA 的酶，因此在生物体中遗传信息的传播中发挥作用。

Samyukta Lakshmi/Bloomberg/Getty Images/文件

人工智能从零开始创造出16种新病毒，这显示出其在耐药性方面的潜力，同时也引发了人们对潜在滥用风险的警告。

该公司表示，950个能够独立规划和执行任务的人工智能代理花费了21个小时搜索数据，筛选出3500条实时反馈（RT）。随后，他们从中筛选出20条最具说服力的反馈。Anthropic公司在一份声明中指出，对于经验丰富的科学家来说，这类分析可能需要数周甚至数月的时间。

随后，“其中一名特工发现了一些非同寻常的东西：在一种外形奇特的逆转录酶基因旁边，存在重复的DNA序列模式，”该公司在声明中指出。“经过彻底分析，他们确信发现了一个新的生物系统，并提交了报告以供人类审查。”

人类学研究所表示，其生物实验室的科学家随后分析并测试了这一发现，他们将其描述为噬菌体（一种攻击细菌的病毒）中“以前未被描述过的”生物系统，其结构与天然存在于细菌中的 CRISPR 类似。

然而，梅斯特表示，他曾在自己的私人 Claude 账户上分享过未发表的研究成果，包括他的论文草稿、分析以及描述这些系统的材料。

“我并没有声称Anthropic公司故意窃取了我们的成果。我无法证明这一点，”他通过电子邮件告诉CNN。

“要么是与我们工作相关的信息以某种方式进入了模型，并被用于训练模型。第二种可能性，也是我目前认为更合理的，是这项探索背后存在着比‘自主发现’一词所暗示的更多的先前科学知识和人为指导，”梅斯特补充道，他指的是新闻稿中使用的措辞。

Anthropic公司未回复记者的邮件置评请求。

德克萨斯大学奥斯汀分校分子生物科学教授伊利亚·芬克尔斯坦表示，在科学研究中，不同的研究小组得出相同的结果并不罕见，但梅斯特的指控很严重，许多科学家将来可能会三思而后行，不会再将他们未发表的结果输入到这些工具中。

不过，伦敦国王学院基因组神经影像学和人工智能教授古斯塔沃·苏德雷表示，Anthropic 的声明确实表明，人工智能正在科学领域开始拥有更大的自主权。

他在一封电子邮件中说：“它得到了一个大致的方向，然后必须运用自己的判断力来判断数据中哪些内容是有趣的。这与通常的‘帮我分析这个数据集’截然不同。”

芬克尔斯坦补充说，人工智能更擅长“大海捞针”，而不是“富有想象力的突破”。他指出，人工智能工具可以帮助探索那些学生会因为寻找值得进一步研究的有趣模式而感到疲惫或厌倦的领域。

芬克尔斯坦告诉CNN：“他们可以迅速地走上大量毫无成效的道路。如果由该领域的专家进行统筹安排，他们非常擅长从事繁琐的工作。”

芬克尔斯坦还指出，Anthropic公司论文的第一作者是彼得·尹博士，他曾在詹妮弗·杜德纳的实验室工作，杜德纳是CRISPR-Cas-9基因编辑技术的设计者之一，并与埃玛纽埃尔·沙尔庞捷共同获得了2020年诺贝尔奖。因此，他说，Anthropic公司的人工智能代理是由各自领域顶尖的专家指导的。

“六位作者中有四位是资深分子生物学家和领域专家。如果给他们提供大量的计算资源和前沿模型（我猜是没有安全措施的），我相信我们会看到更多有趣的生物信息学发现，”芬克尔斯坦在他的实验室网站的博客文章中写道。

然而，其他研究人员表示，关于人工智能驱动的自主发现的说法是一种炒作，它超越了实际的科学发展，而且尽管有人声称模型可以独立产生有意义的结果，但实际上模型还无法做到这一点。

苏德雷表示，让模型以人类无法企及的规模梳理数据确实能推动某个领域向前发展，但这并不一定预示着重大突破。“根据我的经验，我可以告诉你，大部分工作都存在于‘有趣的模式’和‘有用的知识’之间的差距之中，”他指出。

今年早些时候，Anthropic推出了专为科研而设计的工具Claude Science。其竞争对手近期也纷纷发布了类似的工具。OpenAI推出了GPT-Rosalind，微软推出了Quine，谷歌DeepMind推出了Co-Scientist。这些工具的运行方式类似于大型语言模型，但也能利用特定的科学数据集，例如DNA序列。

然而，模型的可靠性取决于训练数据的质量。例如，谷歌DeepMind的AlphaFold模型因其解码蛋白质结构而荣获2024年诺贝尔化学奖，其强大的性能源于使用数十年来收集的狭窄而专业的数据源对人工智能系统进行训练。

人工智能开发者正面临着公众的密切关注，人们越来越担心失控的人工智能代理以及人工智能公司能否对其进行有效监管。生物学发现或许能为公司高管提供一个契机，让他们构建一个更为积极的叙事，以抵消围​​绕其产品的种种负面舆论。

但要在该领域取得重大而切实的发现，可能需要通过机器人技术将人工智能模型整合到实际的实验室工作中。

“生物学不是数学，在某种程度上，智能必须通过机器人以物理方式在现实世界中体现出来，这些机器人能够完成那些繁琐的、一次性的实验任务，而优秀的学生通常几乎是凭直觉就能完成这些任务，”伦敦帝国理工学院生物化学技术副教授尤瓦尔·埃拉尼说道。“就我所见，机器人和自动化技术的发展还远未达到目标。因此，我们还需要一段时间才能真正证明人工智能在生物学领域的应用前景是合理的。”

插图：Sarah Hashemi/CNN

父母们面临着一种新的恐惧：人工智能会不会在孩子们长大成人之前终结人类？

Anthropic公司表示，他们明白人工智能单独能做的事情是有限的，因此正在投资“模型硬件”，使人工智能代理能够安全地操作物理设备。

伦敦国王学院的苏德雷（Sudre）的研究整合了基因组学、临床、神经学和认知数据，旨在预测儿童心理健康症状的发展趋势。他强调，人类在科学中的作用并没有缩小。他指出，风险不在于科学家会被排挤出探索领域，而在于他们会变得懒惰，不再运用判断力去理解研究结果的意义。

人类学研究所的科学家们下一步要做的，就是通过缓慢的物理实验室实验来了解他们所发现的模式，以及这种特征是否可能在生物技术领域有实际用途。

“当然，未来其中一些工作也可以交给机器人，”苏德雷说。“但关键在于，科学中真正值得提出的问题不会一直藏在数据库里等待被发现。它们源于我们自身——源于我们关心的事情，源于我们担忧的事情。”