AI 模型和平台
Anthropic 早期药物开发进展报告
在 Google 上的首选来源中添加 Unite.AI
Anthropic于2026年10月27日发布了其生命科学工作的最新进展，报告与六个药物开发组织——BigHat Biosciences,Bristol Myers Squibb, Divergence Labs,Enveda，生成生物医学，以及The Colonor Group——的研究伙伴关系的早期成果，以及它的第一个内部药物方案。
称公司的战略是让Claude参与药物开发过程，从目标鉴定和分子设计到制造文件和临床试验规划，因为瓶颈在整个过程中分布，而不是集中在一步。 Anthropic 称的工作旨在绘制其模型的优点和局限性，为更广泛的生命科学界提供案例研究，如果目前的轨迹继续下去，则生产食品过敏症、罕见儿童疾病和几种癌症的药物。 公司警告说，研究时间仍然很早，而且大多是临床前的，这意味着所涉药物大多尚未在人类身上进行测试。
重新设计细胞治疗和抗体
在"生成生物医学"(Generate BioMedicines)中，六位科学家使用Claude Code来构建AI agents系统，该系统为CAR T细胞疗法提出了重新设计的信号域，癌症治疗生成者旨在通过使受体的遗传指令更加紧凑来从血液癌适应固体肿瘤。 代理系统于2026年2月开始建设，团队选择了186项设计进行第一轮实验室测试。 Generate报告说，一项设计比标准版本短约70%，同时在实验室测试中，在细胞表面产生大约十倍多的受体，更高效地杀死肿瘤细胞，使免疫细胞活动时间更长。 生成和Anthropic 称都是早期实验室结果，尚未对病人的安全或有效性进行测试，以鼠标模型测试和固体肿瘤目标战略作为下一步。
Divergence Labs是一家成立于2025年的生物技术公司，它正在设计一种食品过敏药物，目的是既中和自由漂浮的IgE，又采用双臂抗体设计，将已经捆绑在免疫细胞上的IgE驱离。 Anthropic报告说，Claude在22个自主工作上产生了大约35,000个抗体设计，在Nvidia H100 GPU上共延伸了约2,750小时的计算时间。 两种武器都报告早期的达到差异：在91个经测试的干扰器设计中，有20个可以测量地驱散IgE，其中3个似乎在Xolair的浓度下超过了核准的药物，病人可以实际得到；在57个经测试的锚设计中，有3个附在IgE上。 根据报告，从想法到实验室，干扰臂大约需要三周时间，锚臂大约需要四周半时间，而根据报告，前一次的类似工作需要九个月时间。 差异调查小组还注意到存在一些局限性，发现Claude在决定进一步调查某项结果和继续调查与称之间结果与新药相比不够务实。
实验室优化和管理文件工作
在与BigHat Biosciences的飞行员中，Anthropic研究人员与Claude合作，与BigHat的Reccy实验室数据系统相连，优化了两个抗体碎片：一个抗甲状腺素VHH和一个针对癌症的抗美食素ScFv. BigHat的自动化实验室在7个回合的大约10周内，在1,232个样本中制造并测试了大约1,010个变体，每轮有176个变体。 BigHat报告说，最终分子的结合紧密度是起始分子的六至九倍，在温度8至18°C时被粘合在一起，并更能耐结聚；最终VHH熔融度约为71°C，而起始分子的熔融度约为62°C，最终ScFv在大约74°C时从溶液中溶解，约56°C时溶解。 部分经过，研究人员和Claude对原始仪器数据进行了重新分析，得出标准熔化和熔化的分数有误导性，并提出了修改的度量法外加一个协议变化测试样本，分两个浓度和两个加热率；Claude对哪些数字将保持稳定的预先预测证明是正确的。 Anthropic向BigHat支付了实验室工作费。
Enveda在Claude周围建造了一个CMC文档处理器，以审查制造批量记录。 在对后来提交给林业发展局的第二阶段制造运动的总共1 061页的4个记录进行评价时，该系统的融资所有11个问题都发现，还有7个问题没有被人类标注，将审查时间从25.9小时缩短到2.8小时，其中89%的Enveda报告了使用情况。 Enveda还用Claude在4小时内起草调查新药应用单元2，而用手动大约3个月，并起草调查员手册，其中审查人员在文件提交林业发展局之前发现一个小单位转换错误。 Enveda 称的团队认为IND草案与专业医学作家的作品相当。
公司建筑和住宅方案
自2007年以来投资生物技术的一家风险投资公司Colonel Group正在建立Pleiades，这是一家新的生物技术公司，由Anthropic工程师建造，从第一天起就全面启用AI。 TCG科学家和Anthropic开发了Claude技能进行初始目标评估，对45对目标疾病进行了评价，TCG发现在结合专家判断时将相关科学文献的查找和解释过程从几周缩短到几天。 自8月以来，第二个临床战略技能正在开发之中。
Anthropic也开始与布里斯托尔·迈尔斯·斯基布合作，找出临床试验规划中的瓶颈，BMS使用Claude来评估早期药物的计划，并估计试验成功的可能性。 “Claude从一个瞬间产生了一个强有力的初步分析，帮助我们评估不同的发展选择，否则需要几天时间才能完成，”BMS的高级副总裁兼早期临床发展、血液学、肿瘤学和细胞治疗主任称 Michael Pourdehnad。 公司计划在2027年扩建工程。
Anthropic单独拥有收购 Mithril Medical，是科迪·特兰巴格(Cody Tranbarger)创立的一家单人公司，正在开发治疗LAMA2-相关先天肌萎缩症和α-脱氧核糖核酸，两种罕见的肌萎缩症。 在7月收购之前，Tranbarger利用Claude组织与11个外部实验室的合同，后勤，以及实验，在不到18个月的时间里将一个双特异性的抗体从思想到老鼠证明原理。 Anthropic在对严重形式的LAMA2-CMD的鼠标进行短期研究时报告说，在最高剂量上被处理的小鼠的握力和三重力大约是未经处理的小鼠的两倍，其中8只被处理的小鼠中有1只在2至6周之间死亡，而8只未经处理的小鼠中有4只死亡。 Anthropic指出，这些是鼠标早期的结果，对病人没有安全或功效。
Mithril在细胞基测试中已经新建了100多个双体特异性抗体，并打算在2027年为这两种疾病命名正式的药物候选者，目标是在2028年末尽快在LAMA2-CMD开始临床试验，并在2029年尽快在α-dystroglycanopy开始临床试验。 Anthropic还指出了它的Claude科学工作台，用于生命科学家的Claude学分，生命科学验证方案，以及公开呼吁客座调查员为扩展生命科学工作做出相关努力。
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。
AI Models & Platforms
Anthropic Reports Early Drug Development Progress With Claude
Add Unite.AI to your preferred sources on Google
Anthropic on October 27, 2026 published a progress update on its life sciences work, reporting early results from research partnerships with six drug development organizations — BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group — alongside its first in-house drug programs.
The company said its strategy is to involve Claude across the drug development process, from target identification and molecule design to manufacturing documentation and clinical trial planning, because bottlenecks are distributed throughout that process rather than concentrated in one step. Anthropic said the work serves to map the strengths and limits of its models, provide case studies for the wider life sciences community, and, if current trajectories continue, produce medicines for food allergies, rare childhood diseases, and several forms of cancer. The company cautioned that the research remains early and mostly preclinical, meaning the drugs involved are for the most part not yet being tested in humans.
Redesigning Cell Therapy and Antibodies
At Generate Biomedicines, six scientists used Claude Code to build a system of AI agents that proposed redesigned signaling domains for CAR T-cell therapy, a cancer treatment Generate aims to adapt from blood cancers to solid tumors by making the receptor’s genetic instructions more compact. The agentic system was built starting in February 2026, and the team selected 186 designs for first-round laboratory testing. Generate reported that one design was about 70% shorter than the standard version while producing roughly 10 times as many receptors on cell surfaces in lab tests, killing tumor cells more efficiently and keeping immune cells active longer. Generate and Anthropic said these are early laboratory results not yet tested for safety or effectiveness in patients, with mouse-model tests and a solid-tumor target strategy as next steps.
Divergence Labs, a biotechnology company founded in 2025, is designing a food-allergy drug intended to both neutralize free-floating IgE and dislodge IgE already bound to immune cells, using a two-armed antibody design. Anthropic reported that Claude generated roughly 35,000 antibody designs over 22 autonomous work stretches totaling around 2,750 hours of computing time on Nvidia H100 GPUs. Divergence reported early hits for both arms: 20 of 91 tested disruptor designs measurably dislodged bound IgE, with three appearing to outperform the approved medication Xolair at concentrations patients could realistically receive, and three of 57 tested anchor designs attached to IgE. Going from idea to lab results took roughly three weeks for the disruptor arm and four and a half weeks for the anchor arm, versus up to nine months for comparable prior work, according to the report. The Divergence team also noted limitations, finding Claude less pragmatic than a human operator in deciding between further investigating a result and moving on, and said the results remain far from a new medicine.
Lab Optimization and Regulatory Paperwork
In a pilot with BigHat Biosciences, an Anthropic researcher working with Claude, connected to BigHat’s Reccy lab data system, optimized two antibody fragments: an anti-albumin VHH and a cancer-targeting anti-mesothelin scFv. BigHat’s automated lab made and tested about 1,010 variants across 1,232 samples in about 10 weeks over seven rounds, with 176 variants per round. BigHat reported that the final molecules bound six to nine times more tightly than the starting molecules, held together at temperatures 8 to 18°C higher, and resisted clumping better; the final VHH melted at about 71°C versus about 62°C for the starting molecule, and the final scFv collapsed from solution at about 74°C versus about 56°C. Partway through, the researcher and Claude reanalyzed raw instrument data, concluded the standard melting and clumping scores were misleading, and proposed revised metrics plus a protocol change testing samples at two concentrations and two heating rates; Claude’s advance predictions about which numbers would hold steady proved correct. Anthropic paid BigHat for the laboratory work.
Enveda built a CMC Document Processor around Claude to review manufacturing batch records. In an evaluation on four records totaling 1,061 pages from a Phase 2 manufacturing campaign later submitted to the FDA, the system raised all 11 issues human reviewers had found plus seven more that no human had flagged, cutting review time from 25.9 hours to 2.8 hours, an 89% reduction Enveda reported for that use case. Enveda also used Claude to draft Module 2 of an Investigational New Drug application in four hours, versus approximately three months manually, and to draft an Investigator’s Brochure in which reviewers found one minor unit-conversion error before the document was submitted to the FDA. Enveda said its teams deemed the IND draft comparable to work produced by professional medical writers.
Company Building and In-House Programs
The Column Group, a venture capital firm investing in biotechnology since 2007, is establishing Pleiades, a new biotech company built with Anthropic engineers to be fully AI-enabled from its first day. TCG scientists and Anthropic developed a Claude Skill for initial target assessment, evaluated on 45 target-disease pairs, which TCG found cut the process of finding and interpreting relevant scientific literature from weeks to days when combined with expert judgment. A second Skill for clinical strategy has been in development since August.
Anthropic has also begun working with Bristol Myers Squibb to identify bottlenecks in clinical trial planning, with BMS using Claude to assess plans for an early-stage drug and estimate the likelihood of trials succeeding. “From a single prompt, Claude generated a robust initial analysis that helped inform our assessment of different development options, which would otherwise take days to complete,” said Michael Pourdehnad, BMS’s Senior Vice President and Head of Early Clinical Development, Hematology, Oncology and Cell Therapy. The companies plan to expand the work in 2027.
Anthropic has separately acquired Mithril Medicine, a one-person company founded by Cody Tranbarger that is developing treatments for LAMA2-related congenital muscular dystrophy and alpha-dystroglycanopathy, two rare forms of muscular dystrophy. Before the July acquisition, Tranbarger used Claude to organize contracts, logistics, and experiments with 11 external labs, taking a bispecific antibody from idea to mouse proof-of-principle in less than 18 months. In a short-term mouse study of a severe form of LAMA2-CMD, Anthropic reported that treated mice on the highest dose had roughly twice the grip strength and tricep force of untreated mice, with one of eight treated mice dying between two and six weeks of age versus four of eight untreated mice. Anthropic noted these are early mouse results that do not show safety or efficacy in patients.
Mithril has built more than 100 new bispecific antibodies now in cell-based testing and aims to name formal drug candidates for both diseases in 2027, targeting the start of clinical trials in LAMA2-CMD as soon as late 2028 and in alpha-dystroglycanopathy as soon as 2029. Anthropic also pointed to its Claude Science workbench, Claude credits for life scientists, a Life Sciences Verification Program, and an open call for guest investigators as related efforts to extend its life sciences work.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.