医生兼免疫学家德里亚·乌努特马兹多年来一直对人工智能感兴趣。但他的“顿悟时刻”出现在2025年底，当时GPT-5 Pro帮助他和他的实验室重新审视了一个持续三年的谜题，该谜题围绕一种帮助人体对抗癌症和其他疾病的特殊免疫细胞展开。
这个谜题的核心是免疫学中一个基础但重要的问题：葡萄糖如何影响T细胞的发育和特化？T细胞是帮助身体对抗病毒、杀死癌细胞、应对某些细菌和寄生虫，并区分健康细胞与威胁的免疫细胞。在发育过程中，它们承担不同的任务，包括影响癌症、自身免疫性疾病和感染的角色。了解是什么促使T细胞向某一特化方向分化，可以帮助研究人员更好地理解，并最终更好地治疗这些疾病。
如今，杰克逊实验室和康涅狄格大学的教授乌努特马兹表示，人工智能已成为他工作的核心，以至于他无法想象没有它如何从事科学研究。“那就像失去了双手，或失去了半个大脑，”乌努特马兹说。
这个谜题始于2022年，当时乌努特马兹进行了一项实验，试图了解一种名为葡萄糖的糖类如何影响T细胞的发育。细胞将葡萄糖作为燃料来源，同时也用于构建蛋白质和执行其他功能。
乌努特马兹的实验结果可能对癌症、自身免疫性疾病和感染等疾病产生影响。但当时，乌努特马兹和他的实验室无法理解他们所观察到的现象。
用GPT-5 Pro解决问题
先前的研究提供了强有力的证据，表明葡萄糖代谢影响T细胞的特化方式。为了更好地理解这种关系，乌努特马兹和他的团队将早期发育中的T细胞暴露于低葡萄糖环境，或含有一种名为脱氧葡萄糖的类葡萄糖分子的环境中。脱氧葡萄糖会干扰细胞利用葡萄糖的能力，破坏能量产生和蛋白质构建。蛋白质之所以重要，是因为它们协调细胞内的活动，并充当在细胞外发送和接收信息的信使。
团队预期这两种条件会产生相似的结果。在这两种情况下，葡萄糖以及T细胞功能所需的能量都会受到限制。但事实并非如此。
暴露于脱氧葡萄糖的T细胞大量产生了参与身体炎症反应的细胞。一些暴露于低浓度葡萄糖的T细胞特化为炎症反应细胞，但数量远不及脱氧葡萄糖组。即使研究人员移除了类葡萄糖分子，早期暴露于脱氧葡萄糖的影响仍然持续。
这种差异不能仅归因于能量缺乏。还有其他因素在起作用。但乌努特马兹和他的实验室无法弄清楚发生了什么，于是他们搁置了这项实验，转而处理其他需要关注的紧急任务。
随后，GPT-5 Pro于2025年底发布，乌努特马兹决定重新审视这项实验。他将结果上传到模型中，并要求其分析数据。
GPT-5 Pro提出，脱氧葡萄糖干扰了一种名为IL-2的蛋白质的构建。这种蛋白质可以阻止T细胞成为被称为Th17的炎症反应细胞。脱氧葡萄糖实质上移除了T细胞成为Th17细胞的一个障碍。这可能是为什么低葡萄糖环境中的T细胞成为Th17细胞的数量远不及脱氧葡萄糖环境中的原因。
“GPT-5提出了一个非常了不起的见解，事后看来，这完全合理，”乌努特马兹说。这个见解恰好超出了他自己的专业领域，以至于他自己没有看到这种联系，他实验室里的其他人也没有。
随后，乌努特马兹决定测试GPT-5是否能预测实验的结果。这位免疫学家从一个他已经完成的实验开始，该实验针对一种靶向某种淋巴瘤的T细胞。他的实验表明，这些被称为CD8+的特定T细胞具有更强的杀死淋巴瘤细胞的能力。
当乌努特马兹要求GPT-5 Pro模拟相同的实验时，它正确预测了CD8+细胞杀死淋巴瘤细胞能力的提升。该模型不可能从互联网上获取这些结果，因为乌努特马兹尚未发表这些结果。
“那一刻我觉得，好吧，这些模型现在已经到了真正理解的地步，”他说。
这对科学研究意味着什么
乌努特马兹表示，像GPT-5 Pro这样的模型现在更像合作者。它们可以简化文献综述，处理每周发表的数百篇新学术论文，并帮助科学家识别尚未解答的问题。它们还可以帮助研究人员完善假设，减少确定最值得进行的实验所需的时间。
“你可以用来验证假设的方法数量是巨大的，”乌努特马兹说。“你有无数种方法，却不知道哪一种是最佳策略。”因此，他使用GPT-5 Pro模拟实验并预测结果，以帮助缩小哪些实验值得在实验室中重复进行。这可以为研究人员节省数周到数月甚至数年的时间，极大地加速生物学领域的发展。
尽管如此，专业知识仍然至关重要。人工智能可能产生见解，但人们仍然必须评估其重要性和合理性。例如，没有乌努特马兹专业知识的人无法判断GPT-5 Pro在其免疫细胞实验中指出的机制性见解是否重要。
产生见解和加速工作的能力意味着这些能力需要负责任地处理。人工智能可以帮助研究人员在生物学和医学领域更快地前进，但这些能力也可能降低滥用的门槛，包括那些试图设计或使用生物或化学武器的恶意行为者。OpenAI的准备框架概述了我们追踪这些风险并针对可能造成严重伤害的人工智能能力建立保障措施的方法。
乌努特马兹对人工智能的发展方向持乐观态度。他说，这与以往的任何事物都不同——无论是互联网还是工业革命。最近，乌努特马兹尝试了先进的人工智能工具，包括Codex和GPT-5.2 Deep Research，以帮助汇编大规模癌症突变数据集并生成研究材料——包括一本以T细胞为重点的广泛草稿教科书——旨在加速精准免疫疗法的工作。
乌努特马兹为自己能成为这个发现时代的一部分而感到幸运。“不仅能够历史性地见证它，还能参与其中，我深感幸运和荣幸。”
Doctor and immunologist Derya Unutmaz has been interested in artificial intelligence for years. But his “aha” moment came in late 2025, when GPT‑5 Pro helped him and his lab revisit a three-year-old puzzle centered on a special type of immune cell that helps the human body fight cancer and other illnesses.
The mystery centered on a basic but consequential question in immunology: how does glucose affect the way T cells develop and specialize? T cells are immune cells that help the body fight viruses, kill cancerous cells, respond to some bacteria and parasites, and distinguish healthy cells from threats. As they develop, they take on different jobs, including roles that can shape cancer, autoimmune disease, and infection. Understanding what pushes T cells toward one specialization or another could help researchers better understand, and eventually, better treat those diseases.
Today, Unutmaz—a professor at The Jackson Laboratory and the University of Connecticut—says AI has become so central to his work that he can’t imagine doing science without it. “That would be like taking both of your hands away, or half of your brain away,” Unutmaz said.
The puzzle began in 2022, when Unutmaz performed an experiment trying to understand how a type of sugar called glucose affected the development of T cells. The cells use glucose as a fuel source, but also to build proteins and carry out other functions.
The results of Unutmaz’s experiment could have implications for ailments like cancer, autoimmune disease, and infections. But at the time, Unutmaz and his lab couldn’t make sense of what they were seeing.
Solving a problem with GPT‑5 Pro
Previous studies provided strong evidence that glucose metabolism influenced how T cells specialize. To better understand this relationship, Unutmaz and his team exposed T cells early in their development to either a low-glucose environment or to one containing a glucose-like molecule called deoxyglucose. Deoxyglucose interferes with a cell’s ability to use glucose, disrupting energy production and protein construction. Proteins matter because they coordinate activity within a cell and act as messengers that send and receive information outside the cell.
The team expected the two conditions to produce similar results. In both cases, glucose, and therefore the energy the T cells needed to function, would be limited. But that’s not what happened.
The T cells exposed to deoxyglucose overwhelmingly produced cells involved in the body’s inflammatory response. Some of the T cells exposed to low concentrations of glucose specialized as inflammatory-response cells, but not at the numbers seen for deoxyglucose. The effects of early exposure to deoxyglucose persisted even when researchers removed the glucose-like molecule.
This difference couldn’t be attributed to a lack of energy alone. Something else was going on. But Unutmaz and his lab were unable to figure out what was happening, so they shelved the experiment and moved on to other urgent tasks that needed their attention.
Then GPT‑5 Pro came out in late 2025 and Unutmaz decided to resurface the experiment. He uploaded the results into the model and asked it to analyze the data.
GPT‑5 Pro suggested that deoxyglucose interfered with the construction of a protein called IL-2. This protein can prevent T cells from becoming an inflammatory-response cell known as Th17. Deoxyglucose essentially removed a barrier to a T cell’s ability to become a Th17 cell. That’s potentially why T cells in the low-glucose environment didn’t become Th17 cells at nearly the numbers seen in the deoxyglucose environment.
“GPT‑5 came up with this really remarkable insight that retrospectively, makes perfect sense,” Unutmaz said. It was just enough outside of his own area of expertise that he didn’t see the connection himself, and neither did anyone in his lab.
Unutmaz then decided to see if GPT‑5 could predict the outcome of an experiment. The immunologist started with one he had already conducted on a T cell that targets a type of lymphoma. His experiment showed that these particular T cells, called CD8+, had an enhanced ability to kill the lymphoma cells.
When Unutmaz asked GPT‑5 Pro to simulate the same experiment, it correctly predicted the boost in the CD8+ cells’ ability to kill lymphoma cells. The model couldn’t have gleaned the results from the internet because Unutmaz hadn’t yet published the results.
“That was the moment that I felt like, okay, these models have now come to a point where they really, truly understand,” he said.
What this means for scientific research
Unutmaz said that models like GPT‑5 Pro function more like collaborators now. They can streamline literature reviews, processing hundreds of new academic papers published every week and helping scientists identify questions that remain unanswered. They can also help researchers hone their hypotheses, reducing the amount of time it takes to identify the most worthwhile experiments to conduct.
“The number of things you can do to address your hypothesis is vast,” Unutmaz said. “You have countless approaches, and you don’t know which one will be the best strategy.” So he uses GPT‑5 Pro to simulate experiments and predict outcomes to help narrow down which experiments are worth repeating in the lab. This can cut out weeks to months, even years, of work for researchers, drastically accelerating the field of biology.
Despite this, subject matter expertise is still key. AI may generate an insight, but people must still evaluate its significance and plausibility. For instance, someone without Unutmaz’s expertise wouldn’t have been able to tell if the mechanistic insight GPT‑5 Pro flagged in his immune cell experiments was important or not.
The ability to generate insights and accelerate work is why these capabilities need to be handled responsibly. AI could help researchers move faster in biology and medicine, but those capabilities could also lower barriers for misuse, including by bad actors seeking to design or use biological or chemical weapons. OpenAI’s Preparedness Framework outlines our approach to tracking these risks and building safeguards against AI capabilities that could create severe harm.
Unutmaz is optimistic about where AI is headed. It’s unlike anything that has come before, he says—not the internet or the industrial revolution. Most recently, Unutmaz has experimented with advanced AI tools, including Codex and GPT‑5.2 Deep Research, to help compile large-scale cancer mutation datasets and generate research materials—including an extensive T-cell-focused draft textbook—aimed at accelerating efforts in precision immunotherapy.
Unutmaz feels fortunate to be part of this time of discovery. “To not only be able to witness it historically but participate a little bit, I feel truly lucky and privileged to do that.”
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文