内容来源：https://news.microsoft.com/signal/articles/can-ai-learn-the-language-of-biology-to-reimagine-medicine/

内容总结：

人工智能能否“读懂”生命语言，重塑未来医疗？

人类基因组中约有两万个基因，这种多样性既造就了丰富的人类体验，也为疾病治疗带来挑战。目前，大多数医疗方案仍采用“一刀切”模式，仅少数癌症患者能获得靶向治疗。微软研究院首席研究员艾娃·阿米尼认为，若人工智能能学会“阅读与书写”生物学语言，将有望为每位患者量身定制治疗方案。

在近日于美国马萨诸塞州剑桥市举办的“酒吧讲座”活动中，阿米尼阐述了人工智能在生物学领域的五大应用前景：

一、解码生物复杂性
生物学系统极为复杂，而当前医疗往往基于群体平均值而非个体差异。阿米尼指出，人工智能能处理人类难以驾驭的海量生物数据，例如单次癌症活检即可产生近五千万个数据点。通过计算工具解读生物语言，AI可助力新药研发，推动个性化精准医疗。

二、精准医疗的困境与愿景
尽管精准医疗旨在根据患者独特的基因与细胞特征定制疗法，但现实中仅不足5%的癌症患者对现有靶向治疗产生有效反应，且耐药性与癌细胞演化常导致疗效难以持久。阿米尼强调，需超越群体共性，利用疾病异质性实现真正个体化治疗。

三、用“生物字母表”设计蛋白质
自1965年生物物理学家玛格丽特·戴霍夫用单字母代码表示氨基酸以来，蛋白质已被视作一种语言。微软基于此开发了EvoDiff与戴霍夫图谱等生成式AI模型，能够根据指令设计具有特定功能的新型蛋白质，堪称“生物学界的Copilot”。

四、AI设计蛋白质已见成效
实验证明，AI生成的蛋白质可精准靶向癌细胞或结合药物递送受体。微软模型将新蛋白质的成功生成率从传统方法的16%提升至50%，标志着该技术已从理论走向实践。

五、细胞建模的挑战与探索
构建能模拟人类细胞、预测药物反应的AI模型被视为科学界的“圣杯”。但阿米尼团队发现，现有细胞模型多仅能预测平均值，且数据量的增加并未带来性能提升。为此，微软与布罗德研究所等机构合作推进“离体项目”，通过实验与计算的深度融合，致力于构建全新的精准肿瘤学框架。

阿米尼坦言，数据质量与多样性仍是关键瓶颈，实现精准医疗仍需跨学科持续协作。但她坚信，这些挑战正推动技术向前发展：“所有评估与发现都将助我们更接近那个愿景——赋能人类迈向更健康的未来。”

中文翻译：

预计阅读时间：5分钟
人工智能能否学会生物学语言，重塑医学未来？
作者

我们每个人的基因组中都有大约两万个基因。这种多样性虽然让人类经验如此丰富，但在医学和疾病治疗领域，基因差异却可能带来更多挑战。

如今，绝大多数治疗方案仍采用“一刀切”的模式。例如，只有一小部分癌症患者能获得靶向治疗。但如果人工智能能学会阅读和书写生物学语言，它就能根据每位患者独特的生理构造，帮助定制个性化治疗方案。

微软研究院首席研究员艾娃·阿米尼正致力于实现这一愿景。近日，在美国马萨诸塞州剑桥市一家座无虚席的啤酒厂里，她探讨了人工智能在生物学领域的潜力。这场活动属于“酒吧讲座”系列——该系列活动将专家讲座与互动趣味相结合，在全美各地的休闲酒吧中举行。

从精准医疗的运作原理，到开发能预测细胞行为的人工智能宏伟蓝图，以下是阿米尼阐述的五个核心概念。

人工智能如何解读生物学
生物学极其复杂——每个人的基因构成和细胞行为都是独一无二的。当今医学往往基于群体平均值而非个体差异来治疗患者。阿米尼指出，人工智能能够解码海量生物数据集中人类无法独立处理的复杂模式。

“计算科学为我们提供了强大的工具包，用以理解我认为最复杂精妙的系统——生物系统及其语言。”她说道，“我们有机会构建计算系统和人工智能模型，利用正在生成的海量数据来学习这种生物语言，最终借此实现新发现、设计新药物，并更接近赋能人类健康未来的愿景。”

阿米尼举例说，单次癌症活检就能产生近五千万个独立数据点。人工智能可以帮助筛选这些海量数据，发现规律，从而实现个性化精准治疗，而非泛化医疗。

精准医疗如何造福人类
精准医疗旨在根据患者独特的基因、分子和细胞特征定制治疗方案。但阿米尼指出，目前大多数治疗仍属通用型，仅有少数癌症患者能获得靶向治疗，而获得持久疗效的患者更是凤毛麟角。

“现实是，基于当前的靶向疗法，只有不到5%的癌症患者能产生有效反应。”阿米尼坦言，“这是因为存在耐药性、癌细胞进化扩散等问题，这些患者实际上无法获得持久治愈的效果。”

精准医疗试图通过利用癌症等疾病的多样性和异质性来突破这些限制，超越群体平均水平，实现个体化治疗。

用生物学语言设计新蛋白质
早在1965年，美国生物物理学家玛格丽特·戴霍夫为生物学创造了一套字母表——用单字母代码表示20种天然氨基酸（蛋白质的基本构成单元）。她创立的这套氨基酸编码使蛋白质能够以语言形式呈现。

微软正基于此基础开发EvoDiff和戴霍夫图谱等生成式人工智能模型来设计新蛋白质。阿米尼将这一概念比作“生物学领域的Copilot”：输入指令，即可输出由该指令引导的新型蛋白质。

这些模型可通过生物学语言指令，设计具有特定功能的蛋白质。

人工智能设计蛋白质的进展与前景
阿米尼表示，人工智能设计的蛋白质可帮助靶向癌细胞，或与药物递送受体结合。

她透露，微软的EvoDiff和戴霍夫模型生成的蛋白质已在实验室测试中展现出成功功能。通过从更广泛多样的数据中学习，戴霍夫模型将新蛋白质的成功生成率从先前方法的16%提升至50%。这些进展表明，用于生物学的生成式人工智能已不仅是理论，而是正在成为现实。

“我们已在真实世界的实验室中进行测量测试，证明这些蛋白质具备我们预期设计的功能。”阿米尼强调。

然而，数据的质量与多样性仍是模型性能的关键，当前仍存在显著局限——尤其在完整细胞建模方面。

迈向人类细胞建模之路
通过学习生物数据中的模式来模拟人类细胞复杂性的AI模型，或能预测细胞对药物的反应，从而开启精准医疗的大门。阿米尼指出，许多人将其视为科学“圣杯”，并致力于构建能预测细胞行为的AI模型。她在微软的实验表明，现有细胞AI模型往往只能预测平均值，而非真实的生物差异。增加数据量并不能提升性能：模型会快速饱和，且无法按预期扩展。阿米尼团队等近期的重要研究揭示了这些局限性。

阿米尼仍满怀希望。她表示，虽然人工智能在生物学领域潜力巨大，但要实现个性化精准医疗，仍需持续的跨学科整合与协作。她共同领导了“离体项目”——这是微软与布罗德研究所在丹娜—法伯癌症研究所支持下开展的研究合作，致力于构建精准肿瘤学新框架，从零开始整合实验与计算，最终目标是改善患者疗效。

“作为技术研究者，我们将这些发现视为燃料，竭尽所能推动研究向前发展。”她说，“所有这些信息与评估，都帮助我们不断进步，更接近那个承诺。”

题图来自安德烈·奥努夫里延科/Moment/Getty Images。

英文来源：

– The estimated reading time is 5 min.
Can AI learn the language of biology to reimagine medicine?
Author
We all have about 20,000 genes in our genomes. While this diversity is what makes the human experience so rich, our genetic differences can make things more difficult when it comes to medicine and the treatment of diseases.
Today, most treatments are a one-size-fits-all approach. Only a small fraction of cancer patients, for example, receive targeted therapies. But if AI could learn to read and write the language of biology, it could help customize treatments for the unique makeup of each patient.
Ava Amini, a principal researcher at Microsoft Research, is working to make that happen. She recently spoke about the potential of AI for biology at a crowded brewery in Cambridge, Massachusetts, as part of “Lectures on Tap,” an event series that combines expert lectures with interactive fun in casual pub settings around the U.S.
Here are five of the concepts she covered, from how precision medicine works to the grand vision of developing AI that can predict how cells behave.
How AI can help make sense of biology
Biology is incredibly complex — each person’s genetic makeup and cellular behavior is unique. Today, medicine often treats patients based on averages, not individual differences. Amini says AI offers a way to decode patterns in massive biological datasets that humans can’t process alone.
“Computation gives us this incredibly powerful toolkit to understand what I think is the most complex and intricate system that we have, which is the system and the language of biology,” she says. “We have this opportunity to build computational systems, AI models, that can harness the scale of data that we’re generating, to learn this biological language and ultimately be able to use that to make new discoveries, design new drugs and hopefully get closer to that vision of empowering people to live a healthier future.”
Amini says a single cancer biopsy, for example, can generate nearly 50 million individual data points. AI could help sift through this massive data, find patterns and enable personalized, precise treatment rather than generalized care.
How precision medicine can help people
Precision medicine aims to tailor treatments to the unique genetic, molecular and cellular makeup of each patient. But most treatments are generic, and only a small fraction of cancer patients receive targeted therapies. Even fewer experience lasting success, Amini says.
“The truth is that based on today’s targeted therapies, less than 5% of this population is even going to respond effectively,” Amini says of cancer treatment. “That’s because there are things like resistance or the cancer evolves, it spreads and grows, and these patients will not actually see durable, lasting, curative outcomes.”
Precision medicine seeks to overcome these limitations by leveraging the diversity and heterogeneity of diseases like cancer, moving beyond population averages to individualized care.
Using the language of biology to design new proteins
Back in 1965, American biophysicist Margaret Dayhoff gave biology an alphabet — a one-letter code for the 20 natural amino acids, the building blocks of proteins. Her creation of this code for amino acids enabled the representation of proteins as a language.
Microsoft is building on this foundation with EvoDiff and The Dayhoff Atlas, generative AI models to design new proteins. Amini says the concept is like Copilot for biology: Input a prompt and output a novel protein guided by that prompt.
These models can be prompted in the biological language to design proteins with specific functions.
AI-designed proteins show progress and promise
AI-designed proteins could help target cancer cells or bind to receptors for drug delivery, according to Amini.
She says Microsoft’s EvoDiff and Dayhoff models have generated proteins tested in the lab with successful functional outcomes. By learning from a greater scale and diversity of data, the Dayhoff models improved the success rate of producing new proteins from 16% with previous methods to 50%. These advances show that generative AI for biology isn’t just theory; it’s happening now.
“We’ve actually gone and measured and tested in the lab in the real world to show that these proteins have the functions that we meant and sought to have,” Amini says.
However, the quality and diversity of data remain critical for model performance, and there are still significant limitations — especially in modeling entire cells.
Working toward modeling human cells
An AI model designed to simulate the complexity of a human cell by learning patterns in biological data could predict how cells respond to drugs, unlocking precision medicine. Many consider it to be a “holy grail” in science, Amini says, and have pursued the idea of building AI models to predict how cells behave. Amini says their experiments at Microsoft have shown that existing AI models of cells often predict only average values, rather than real biological differences. Increasing data volume does not improve performance: Models saturate quickly and do not scale as expected. Recent critical studies, including those by Amini and team, have exposed these limitations.
Amini still has hope. While the promise of AI in biology is immense, she says, realizing personalized, precise medicine will require continued integration and collaboration across disciplines. She co-leads Project Ex Vivo, a research partnership between Microsoft and the Broad Institute with support from the Dana-Farber Cancer Institute, which is building a new framework for precision oncology, integrating experimentation and computation from the ground up toward the ultimate goal of improving patient outcomes.
“As a technologist, we use these findings as fuel, and we want to take as much as we can to actually go further,” she says. “And all of this information, all of these evaluations, help us do better and get closer to that promise.”
Lead image by Andriy Onufriyenko / Moment / Getty Images.