新智元报道
Claude真下场造药了！
前一天，Anthropic刚交完IPO招股书，今天扭头就杀进药物研发圈，而且这波操作极其硬核——
直接把「一人药企」Mithril Medicine收入囊中。
创始人Cody Tranbarger单枪匹马，唯一的工具就是Claude。
就靠这个组合，他一人调度了11家外部实验室，不到18个月，把一个罕见病疗法从想法直接干到小鼠实验，结果还真成了。
如今，这一实验pipeline已正式并入Anthropic，预计最快2028年底进入人类临床试验！
不仅如此，Anthropic还联手6大顶尖生物医药及风投，让Claude直接杀入抗癌，食物过敏治疗及临床试验的核心流水线。
结果非常令人震撼——
Claude一个月狂搓约3.5万种抗体设计，把CAR-T一个关键组件缩短约70%，还让药品生产文档的审核时间减少89%。
如今的Claude已穿上白大褂，向癌症和罕见病发起总攻！
在这份长长的官方报告中，最让人掉下巴的操作，莫过于Anthropic对Mithril Medicine的直接收购。
说出来你可能不信，从成立到被巨头收购，全公司上下只有创始人Cody Tranbarger一个人。
他要死磕的疾病，是极其棘手「LAMA2型先天性肌营养不良症」（LAMA2-CMD）。
这是一种极其残酷的罕见儿科绝症：患儿从出生起就肌肉无力，42名重症患儿的追踪数据显示，仅1人能独立行走。
六分之一活不到成年，且目前全球没有任何获批的根本性疗法。
这一问题的根源，出在肌肉纤维与外部支撑结构的连接上。
健康肌肉的每根纤维外面，都包着一层叫「基底膜」的结构。其中一个关键部件，叫层粘连蛋白α2。它参与组成的蛋白复合物，既帮助织起这张网，也负责把网锚定在肌肉细胞表面。
α2缺失后，肌肉纤维失去重要支撑，逐渐受损。
Cody把除了核心决策之外的几乎所有科学和运营工作，全权委托给了Claude。
在不到18个月的时间里，Cody与Claude组成了一支极致的「碳基+硅基」二人特种部队。
论文地址：https://www-cdn.anthropic.com/52e1f5d50c47ca3818f56704c27cf19e59b62e77.pdf
他们设计出了一种「双特异性桥接抗体」，思路就是去造一枚「分子订书钉」。
既然替补蛋白就在那儿，肌膜上的「铆钉底座」β-dystroglycan也还在，缺的只是中间那根连接——
那就造一个双特异性抗体：一只手抓住层粘连蛋白-α4的LG4-5结构域，另一只手抓住β-dystroglycan，硬把基底膜重新「订」回肌膜上。
最终胜出的两个分子，在细胞实验里能把细胞牢牢粘在目标蛋白的涂层上，扛住两轮离心冲洗。
随后，Tranbarger在患病小鼠（dy^W小鼠）体内实验，彻底验证了这个神仙设计。
从出生起连续给药六周后：治疗组的三头肌诱发力量，是对照组的1.95倍；握力是对照组的1.94倍；作为阳性对照的现有基因疗法，只做到1.31倍。
抗体也确实到了该去的地方：末次给药几天后，仍能从肌肉里回收到同时抓着两个靶点的抗体，浓度恰好落在最容易「搭桥」的区间。
那么，在这场硬核研发里，Claude到底是怎么帮一个人管住11家外包实验室（CRO）的？
技术报告专门用了一个章节，复盘了进入大规模筛选阶段后「极客造药」的真实一周：
在团队交接细胞实验的4个工作日里，面对与CRO之间悬而未决的11个实验细节问题（孔板排布、包被蛋白浓度、运输条件等），人类只需敲下两句话并甩过去一个文件夹链接。
Claude在读完所有文件后，仅用96秒给出了核心结论，18分钟内掏出了完整的推进计划和5封可直接发送的英文邮件草稿。
在那周的第4天，负责生产的CRO发来了一份厚达341页的113个分子的生产与质检数据包。
2分40秒后，Claude给出核心结论；7分53秒后，Claude把341页文档一字不落地读完，直接画成了一张14×8的可视化决策矩阵图。
在这场实验中，从底层AlphaFold与晶体结构的叠加对比、Python数据处理脚本编写，到整篇38页学术级论文图表的绘制与正文起草，Claude几乎扛了所有的重体力活。
如今，一个人就能联手AI造出新药，确实令所有人大为震撼。
除了收购Mithril亲自下场造药，Anthropic在过去几个月里，还派出了自家的生命科学团队，与6大机构展开了深度合作。
其中包括，Generate Biomedicines、Divergence Labs、BigHat Biosciences、Enveda、The Column Group以及百时美施贵宝（BMS）。
过去几个月里，他们用Claude干出了超硬核的战绩。
CAR-T疗法在血液癌症上屡建奇功，但面对实体瘤一直力不从心。
Generate Biomedicines想改造它，思路是把疗法遗传指令中的某一段「压缩」，腾出空间装载对付其他癌症的新指令。
6位科学家用Claude Code提出新设计。
其中一个比标准版短了约70%，实验室里表现反而更好：免疫细胞表面用来识别癌细胞的受体数量提升约10倍，杀伤肿瘤细胞更高效，活跃时间也更久。
又短又强，这在生物工程里可不常见。
食物过敏的「罪魁」是一种叫IgE的分子，一部分飘在血液里，一部分牢牢扒在免疫细胞上。现有药物主要中和血液里游离的那部分，扒在细胞上的，基本拿它没办法。
Divergence Labs想走一条更狠的路：直接把IgE从免疫细胞上「撬」下来。这需要一种带两条「手臂」的精密分子：一条死死抓住IgE，另一条负责把它撬离。
Claude先把目标拆成一份详细的设计方案，然后在很大程度上自主运行，一个月左右生成了约3.5万种抗体设计，供团队筛选测试。
结果两条「手臂」都找到了苗子。3个设计能结合IgE，20个能撬下IgE，其中3个的撬离效果比标准药物还强。下一步，就是把两条手臂合体。
一个好药，不光要和靶点结合得紧，还得耐得住生产储存、别结块。
BigHat Biosciences把Claude接入自家实验数据系统，针对两个抗体片段提出改造方案，实验室10周内造出并测试了约1010个变体。
此外，Enveda的生产化学家，审核时间砍掉89%，把人类从文书解放出来；The Column Group筹备了一家新公司Pleiades，用Claude让一个小团队跑更多、更复杂的药物项目；百时美施贵宝正用Claude评估一款早期药物的临床试验计划，并预估试验成功的概率。
过去的AI制药，大多停留在“单点突破”：预测蛋白结构、筛选化合物、生成分子。
AI是一把好用的刀，但握刀的、决定往哪儿切的，始终是人。这一次不一样。Claude出现在了药物研发链条的几乎每个环节：
- 上游：评估靶点值不值得做；
- 设计：生成抗体、重构CAR-T组件；
- 优化：迭代候选药，甚至纠正实验指标；
- 生产与申报：审核质量文档、起草监管文件；
- 临床：规划试验、预估成功率；
- 运营：签合同、管物流、协调11家实验室。
它不只回答问题，还会把目标拆解成方案、自主执行、发现方法论上的漏洞。
这已经很接近一位初级科研人员加项目经理的综合体了。
而Anthropic也不再满足于「卖铲子」给淘金者，而是亲自收购药企、亲自下场做药。
「双十定律」统治制药业几十年，它背后是无数患者等不到的新药，是无数因为「不划算」而被搁置的罕见病。
现在，一个人带着一个AI，18个月走完了过去要好几年的路。
当然，真正的考卷，在人体临床试验里。接下来，还需拭目以待。
参考资料：
https://x.com/AnthropicAI/status/2115178675050307942
https://x.com/i/trending/2115195261421011324
编辑：桃子
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Claude's making drugs!
The day before, Anthropic had just turned his head into a drug research and development ring, and the wave was extremely hard-core.
One-man pharmaceutical company, Mistril Medicine, is in the bag.
Cody Tranbarger, the founder, is alone. The only tool is Claude.
With this combination, he moved 11 outside laboratories alone, less than 18 months ago, to dry a rare treatment directly from thought to a mouse experiment, and the result was a real success.
Today, this experiment, Pipeline, has been officially integrated into Anthropic and is expected to enter human clinical trials by the end of 2028!
Moreover, Anthropic joined forces with six top biopharmaceuticals and wind surges to get Claude directly into the core stream of cancer resistance, food allergies and clinical trials.
The result was very shocking.
Claude razed some 35,000 antibody designs a month, reducing a key CAR-T component by about 70 per cent and reducing the processing time for pharmaceutical production files by 89 per cent.
Now Claude has put on a white coat and launched a general attack on cancer and rare diseases!
In this long official report, there is no better operation than the direct acquisition of Mithril Medicine by Anthropic.
You may not believe it. From being founded to being bought by giants, there's only Cody Tranbarger in the company.
He is suffering from the extremely difficult problem of "LAMA2 congenital myocardial malnutrition".
This is an extremely cruel and rare paediatric terminal illness: children have been muscular from birth and tracking data for 42 seriously ill children show that only one person can walk independently.
One in six will not live as an adult, and there is currently no approved underlying therapy globally.
The root causes of the problem are the connection of muscle fibres to external support structures.
Each fibre of healthy muscles is covered by a structure called "the base film". One of the key components, called lacquered protein alpha2. It was involved in the formation of protein complexes, which helped to woven the net and was responsible for anchoring it on the surface of muscle cells.
When alpha2 is missing, muscle fibres lose significant support and are gradually damaged.
Cody delegated full authority to Claude for almost all scientific and operational work, with the exception of core decision-making.
In less than 18 months, Cody and Claude formed an extraordinary "carbon plus silicon" special unit.
Address: https://www-cdn.antropic.com/52e1f5d50c47ca3818f56704c27cf19e59b62e77.pdf
They've designed a "two-specific bridge resistance" the idea of building a "molecule stapler."
Now that the replacement protein is there, the beta-dystroglycan on the membrane is still there, and all that's missing is the middle link.
Then create a double antibodies: one hand grabs the adhesive protein-alpha4 LG4-5 structural domain, the other hand grabs the beta-dystroglycan, and hard "set" the substrate back to the membrane.
The two molecules that eventually won were able to bind the cells to the target protein coating in cell experiments, carrying two rounds of centrifugal wash.
Tranbarger then thoroughly validated this fairy design in an experiment in sick mice (dy^W mice).
Six weeks after the delivery of the drug from birth: the three-headed muscle force of the treatment group is 1.95 times greater than that of the control group; the grip is 1.94 times greater than that of the control group; the current genetic therapy, which is positive, is only 1.31 times greater.
The antibodies did reach the place to go: a few days after the last delivery of the drug, they could still be recovered from the muscles to the antibodies that caught two targets at the same time, at concentrations that fell in the areas that were the easiest to “bridge”.
So, how exactly did Claude manage 11 contracted laboratories for one person in this hard nuclear development?
The technical report was devoted to a chapter replaying the true week of "pods" after the mass screening phase:
During the four working days of the team-to-cell experiment, in the face of 11 outstanding test details with the CRO (pore sheeting, encapsulation, transport conditions, etc.), humans need only knock down two sentences and throw a folder link.
After reading all the documents, Claude gave the core conclusions in only 96 seconds, with a complete advance plan and five drafts of direct English mail available within 18 minutes.
On the fourth day of that week, the CRO, which was responsible for production, sent a production and quality control package of 113 molecules of 341 pages thick.
Two minutes and 40 seconds later, Claude gave the core conclusion; after seven minutes and 53 seconds, Claude finished reading 341 pages of documents without a word, and drew a visualized decision matrix of 14 x 8.
In this experiment, Claude carried almost all of the heavy work from the supercomparison of the AlphaFold at the bottom with the crystal structure, from the Python data processing script to the drawing and drafting of the entire 38 pages of academic paper charts.
Now that one person can work with AI to make new drugs, it's really a shock to all.
In addition to the purchase of Mistril's own drugs, Anthropic has sent his own life science team over the past few months, working in depth with six major institutions.
These include Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group, and the Universal Treasure (BMS).
Over the past few months, they have used Claude to make super-hard nuclear scores.
CAR-T therapy has repeatedly built wonders on blood cancer, but has not been able to cope with real cancer.
Generate Biomedicines wants to change it by "compressing" one of the therapeutic genetic instructions and freeing space for new instructions against other cancers.
Six scientists proposed new designs with Claude Code.
One of them is about 70 per cent shorter than the standard version, but it is better performed in the laboratory: the number of receptors whose immunocellular surfaces are used to identify cancer cells has increased by about 10 times, anti-tumour cells have become more efficient and active.
It's short and strong, and it's not very common in bioengineering.
The food-insensitive "guilty" is a molecule called Ige, partly in the blood and partly in the immune cells. The main part of the drug available and the part of the blood that is left is on the cell, and there's nothing we can do about it.
Divergence Labs wants to take a tougher path: to "pull" Ige directly off the immune cells. This requires a sophisticated molecule with two "arms": one to catch Ige and the other to break it out.
Claude splits the target into a detailed design and then operates largely autonomously, producing approximately 35,000 antibody designs for team screening testing in about a month.
As a result, both arms found seedlings. Three were designed to combine Ige and 20 were able to pry down Ige, three of which had a higher detached effect than standard drugs. next step, just put two arms together.
A good drug, not only to combine it with the target, but also to be able to produce in storage and not to block.
BigHat Biosciences connected Claude to the home-based experimental data system and proposed a retrofit programme for two antibody segments, which produced and tested approximately 1010 variants within 10 weeks.
In addition, Enveda's production chemist, who cut off 89 per cent of the audit time to free humanity from the instrument; The Column Group prepared a new company, Pleiades, to use Claude to get a small team to run more sophisticated drug projects; and Baimei Squib is using Claude to assess a clinical trial plan for an early drug and to predict the probability of success.
The majority of past AI pharmaceuticals were found in “single-point breakthroughs”: predicting protein structures, screening compounds, and creating molecules.
AI is a good knife, but it's a man who holds it and decides where to cut it. This time it's different. Claude appears at almost every link in the drug development chain:
- Upstream: it is not worthwhile to assess target values;
- Design: to generate antibodies and to reconstruct CAR-T components;
- Optimization: an iterative candidate, even correcting experimental indicators;
- Production and declaration: review of quality files and drafting of regulatory documents;
- Clinical: planning experiments, estimating success rate;
- Operations: contracting, managing logistics, coordinating 11 laboratories.
It not only answers questions, but also dismantles targets into programmes, autonomous implementation and the identification of methodological gaps.
This is already close to the complex of a junior scientist plus project manager.
And Anthropic is no longer content with selling shovels to gold-diggers, but rather buys them himself and does them himself.
The "Ten Laws" have ruled the pharmaceutical industry for decades, behind which countless new drugs cannot wait, many rare diseases that have been put on hold for "unprofitable" reasons.
Now, a man with an AI is 18 months away from the past years.
Of course, the real exam, in human clinical trials. We'll see about that.
References:
https://x.com/AnthropicAI/status/2115178675050307942
https://x.com/i/trending/2115195261421011324
Edit: Peaches
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.