塞萨尔·德·拉富恩特（César de la Fuente）实验室利用 Codex 和 ChatGPT，在现存与已灭绝生物的基因组中搜索抗菌候选分子，以应对耐药性感染。

César de la Fuente’s lab uses Codex and ChatGPT to search living and extinct genomes for antimicrobial candidates to fight drug-resistant infections.

塞萨尔·德拉富恩特（César de la Fuente）及其实验室致力于从现存与已灭绝生物的基因组中挖掘可用于对抗耐药性感染的分子。包括细菌、真菌、寄生虫和病毒在内的耐药性微生物正构成日益严峻的全球威胁。2021年，约五百万人的死亡与细菌抗菌素耐药性相关（在新窗口中打开），而这一数字预计将在2050年翻倍。发现具备成为抗菌剂潜力的分子往往需耗时数年；研究人员正借助人工智能加速这一早期发现阶段。 “在我看来，抗菌素耐药性是人类面临的最重大的生存威胁之一，”生物工程师塞萨尔·德拉富恩特（在新窗口中打开）表示。他所领导的跨学科实验室专注于寻找抗菌候选分子。“然而，我们已有五十年未出现全新类别的抗生素。”当前多数抗菌药物研发集中于改造现有药物或搜寻已知化学类别，但该路径正面临收益递减的困境。 该实验室开发的深度学习模型经训练可识别生物序列中的模式，从而在海量基因组与蛋白质数据集中高效筛选潜在抗菌分子。该方法可将初始候选分子筛选周期从数年缩短至数小时。然而，仅基因组中极小一部分具有明确功能注释，其中更仅有极少部分编码可对抗病原微生物的分子。挑战在于：识别赋予分子功能活性（即生物学活性）的序列模式，并进一步判断哪些分子具备抗感染潜力。 如今，数字化基因组与蛋白质数据库使科学家得以横跨生命之树系统性地搜寻新型化合物，极大拓展了可供探索的数据库广度。但信息爆炸亦带来新挑战：如何在海量可能性中甄别出真正有前景的信号。人工智能尤其适用于此类“大海捞针”式任务——它能快速扫描庞大数据集，识别研究人员难以察觉的隐含模式，并为实验验证优先筛选出数量可控的候选分子集。 然而，识别出有前景的候选分子，并不意味着其必然转化为有效药物。科学家须首先确认该分子能否杀灭靶标微生物，测定其有效剂量，并评估其对人类细胞的影响。化学家随后可能对其结构进行优化，以提升疗效、安全性或稳定性。后续研究还需测定其毒性剂量阈值、微生物对该分子产生耐药性的难易程度，以及其在体内的药代动力学行为。研发团队还需确立该分子的可靠规模化制备工艺。即便通过上述所有环节，候选分子仍须经历监管审批与临床试验，方能作为获批抗菌药物应用于患者。 对德拉富恩特而言，这正是人工智能与实验生物学必须协同并进的原因。“真实世界的实验验证对于确认AI预测至关重要，”他表示，“若要持续深化我们对生物学——这一世界上最复杂系统的理解，这种‘实证校准’将在未来生命科学领域发挥关键作用。” 其实验室系统性地解析现存与已灭绝生物的基因组，以发掘候选分子。解读这些基因组、阐明其所编码蛋白质的结构与功能、并确定相应分子的生物学效应，需要横跨多个学科的专业知识。“我们的ChatGPT工作空间正接收来自不同背景成员的输入——他们以迥异视角思考我们正致力攻克的难题，”德拉富恩特指出。实验室成员将各自的好想法与坏点子均输入其中，使其成为一种协作式思维共振板。但他同时警示不可过度依赖AI：“显然，你始终需反复核查其输出的准确性。” 尽管如此，他高度认可AI助力科研人员跨越学科边界的潜力：“突破性发现正等待在那里——本质上，它们就潜藏在各学科交界处，而那里恰恰是鲜有人涉足的前沿地带。”

César de la Fuente and his lab probe the genomes of living and extinct organisms for molecules that could help fight drug-resistant infections. Drug-resistant microbes including bacteria, fungi, parasites, and viruses are a growing global threat. About five million deaths in 2021 were associated⁠(opens in a new window) with bacterial antimicrobial resistance—an annual toll projected to roughly double by 2050. It can take years to find molecules with the potential to become antimicrobials. Researchers are using AI to accelerate this early stage of discovery. “Antimicrobial resistance is one of the greatest existential threats to humanity in my opinion,” said César de la Fuente⁠(opens in a new window), a bioengineer whose cross-disciplinary lab searches for antimicrobial candidates. “And yet, we haven’t had a new class of antibiotics for 50 years.” Much of modern antimicrobial development focuses on modifying existing medicines or searching familiar classes of chemicals. But that approach offers diminishing returns. The lab’s deep-learning models are trained to recognize patterns in biological sequences, allowing them to search vast genome and protein datasets for potential antimicrobials. The approach can reduce the initial search for candidate molecules from years to hours. Only a fraction of a genome has a clearly understood function, and fewer still encode molecules that can fight infectious microbes. The challenge is to identify patterns that make a molecule functional, or biologically active, then determine which have the potential to combat infectious microbes. Digital genome and protein databases now let scientists search across the tree of life for new compounds, dramatically expanding the breadth of databases available for exploration. But that abundance of information comes with its own challenges: finding promising signals among an enormous number of possibilities. AI is particularly well suited to this needle-in-a-haystack task. It can scan huge datasets, identify patterns that might be difficult for researchers to spot, and prioritize a manageable set of candidates for experimental testing. But identifying a promising candidate doesn’t necessarily mean that it will become an effective medicine. Scientists must first confirm that a candidate molecule kills the target microbe, determine the amount needed in order to be effective, and test how it affects human cells. Chemists may then optimize it to improve its effectiveness, safety, or stability. Further tests assess the dose at which the candidate becomes toxic, how readily microbes develop resistance to the molecule, and how the candidate moves through the body. Teams also determine a reliable way to manufacture the molecule. Candidates that clear these hurdles still face regulatory review and clinical trials before they can reach patients as approved antimicrobial drugs. For de la Fuente, this is why AI and laboratory biology must advance together. “Ground-truth experiments are essential to validate AI predictions,” he said. “This will be critical in the life sciences in the years to come if we are to continue scratching the surface of our understanding of biology, which is the most complex thing out there.” His lab explores the genomes of living and extinct organisms for candidate molecules. Searching those genomes, understanding how their encoded proteins form and function, and determining what those molecules do requires expertise spanning several fields. “Our ChatGPT workspace is receiving input from all these different people that think differently about the problems that we’re trying to tackle,” said de la Fuente. Lab members feed it their good and bad ideas, making it a kind of collaborative sounding board. But he cautions against relying on AI alone. “Obviously you have to always double-check for accuracy,” de la Fuente said. Still, he appreciates AI’s ability to help researchers explore the boundaries between scientific disciplines. “That’s where the breakthroughs are waiting to be discovered. They’re essentially at the edges between fields where very few people go.”