拆解Anthropic的生命科学布局：一家AI公司为何下场做实验？
Dissecting Anthropic’s Life Sciences Strategy: Why Is an AI Company Getting Into Experiments?
Summary
An AI company famous for developing large language models is now conducting biological experiments itself. Recently, Reuters revealed that Anthropic has set up its own biology lab in the San Francisco Bay Area. The company’s life‑science …
An AI company famous for developing large language models is now starting to conduct biological experiments itself.
Recently, Reuters disclosed thatAnthropic has set up its own biology lab in the San Francisco Bay Area.Subsequently, the company’s head of life sciences, Eric Kauderer‑Abrams, confirmed the news and said that biological research ultimately still needs to be validated through real experiments. Therefore, Anthropic has begun conducting related work internally while continuing to collaborate with external labs.
A few days later, Anthropic announced an earlyresult:About 950 Claude agents analyzed more than 200,000 reverse transcriptase‑related sequences and discovered a novel enzyme system with features similar to CRISPR repeat sequences.The research team then experimentally confirmed some characteristics of this system, but its exact function and application value remain to be investigated.
This is not Anthropic’s first attempt in the life sciences field. Over the past year, it has successively launched dedicated products for scientists, established collaborations with biological research institutions and pharmaceutical companies, enabled Claude to take part in protein design, invoke specialized models, and even explore controlling laboratory equipment with AI.
From this series of layouts, Anthropic’s goal in the life sciences field appears to go beyond providing AI research tools. It is trying to enable Claude to connect research data, specialized models, and experimental workflows, gradually delving into more aspects of life science research.These actions also suggest a possibility: Anthropic hopes to shape Claude into an AI infrastructure for the life sciences industry.
But why would a company primarily developing general large models extend its business into real biological laboratories?
From Research Tools to Research Platforms
Life science research is not lacking in specialized software, nor in AI models developed for specific problems.
Taking protein design as an example, researchers can use different models to generate protein structures, design amino‑acid sequences, and then predict whether they can bind to a target molecule. However, these tools usually handle only one step each, and scientists still need to decide which models to use, how to set the parameters, and which candidates from the massive computational results are worth further experimental validation.
For the entire development process, even the most powerful single tool does not mean the research can be completed automatically. There remains a great deal of work that must be manually coordinated among data, models, and experiments.
Anthropic is entering the life sciences field, targeting precisely this issue.
In October 2025, the company launched Claude for Life Sciences, enabling Claude to connect with research platforms such as Benchling and 10x Genomics, access experimental records, process sequencing data, and assist in designing research plans. The subsequently released Claude Science further integrates scientific databases, specialized models, and computing resources into a unified workspace, allowing researchers to complete various analysis tasks through Claude.
Compared with building an entire suite of biology‑specific software in‑house, Anthropic chose to have Claude work alongside existing tools. This differs from biological models that are trained specifically for tasks such as protein structure prediction or molecular property calculation:Claude doesn’t necessarily perform every calculation itself; instead, it calls the appropriate tool based on research needs and helps scientists organize the workflow.
This positioning also gives Anthropic the opportunity to enter more stages of development. In addition to early‑stage research, the company has gradually added tool integrations for clinical trial management, regulatory submissions, and other areas, hoping Claude can serve not only laboratory researchers but also other professionals involved in the drug development process.
However, to make Claude a work platform that scientists use regularly, merely being able to connect to databases and software is not enough. What researchers truly care about is whether it can reliably complete real tasks and reduce the complexity of operations in the research process.
Anthropic therefore began collaborating with the Allen Institute, the Howard Hughes Medical Institute (HHMI), and other organizations to explore using Claude for processing complex biological data, assisting experimental design, and generating research hypotheses. The company also launched BioMysteryBench, which uses 99 tasks based on real biological data to test whether the model can autonomously invoke tools, analyze data, and produce verifiable results.
Compared with these evaluations, a protein design experiment more directly demonstrates Anthropic’s approach.
In this study, Claude did not replace specialized models such as RFdiffusion; instead, it coordinated them to perform protein design, parameter tuning, and candidate screening. Ultimately, Anthropic sent 1,320 designs to external laboratories, of which 354 were confirmed to bind the target protein, yielding an overall hit rate of about 26.8%.Among the 15 targets ultimately included in the statistics, 14 had at least one design that successfully bound.
Image | Mythos Preview and Opus 4.8 achieved over 20% hit rate on 15 targets (source: Anthropic)
Although these results are still far from actual drug development, they demonstrate that Claude can already leverage existing specialized models to accomplish a relatively complex research task and produce results that are amenable to experimental validation.
This also reflects a distinction between Anthropic and companies that develop dedicated foundational biology models. It does not need to train its own model for every sub‑field; instead, it tries to make Claude a common entry point for using these tools, helping researchers coordinate different tasks. If this approach can be applied to more research scenarios, Claude’s value will no longer be limited to a single scientific software, but will lie in connecting previously fragmented data, models, and workflows.
However, protein design remains a relatively well‑defined task. Scientists already know what kind of molecule they want, and Claude’s job is to find designs that meet those specifications. In more exploratory basic biology research, scientists sometimes have not even identified a specific target. Anthropic has begun experimenting with having Claude take part in this more open‑ended research.
From Completing Tasks to Contributing to Scientific Discovery
Nature contains a vast number of proteins and gene sequences whose functions remain undefined. As sequencing technologies advance, scientists have accumulated ever more data, but identifying biological systems that may have special functions still requires a great deal of time for comparative analysis.
To address this, Anthropic assembled a dedicated life‑science research team this spring, comprising scientists with experience in computational biology, CRISPR systems, and novel enzyme research. Unlike the team primarily responsible for developing research products, this group conducts fundamental biological research directly, aiming to use Claude to mine massive DNA datasets for unknown proteins and biological systems, and then experimentally validate the findings.
The recently announced ART enzyme system study was one of the early projects undertaken by this team. In this research, the investigators gave Claude a rough direction for searching unknown reverse‑transcriptase systems, then had about 950 agents concurrently analyze more than 200,000 related sequences.
Ultimately, Claude discovered regularly spaced DNA repeat sequences near a class of reverse‑transcriptase genes, whose structure bears some similarity to the CRISPR system. The research team named this system ART, standing for array‑associated reverse transcriptases. Follow‑up experiments confirmed that these repeat sequences can be expressed to produce short RNAs, but the precise biological function of the system remains unclear.
Figure | Discovery of agents in the new RT lineage (Source: Anthropic)
This study does not yet demonstrate that Claude can independently make scientific discoveries. The research direction is set by human scientists, candidate results require manual review, and actual experiments are conducted by researchers.
However, compared with previous protein design work, Claude’s role has changed.In well-defined design tasks, it primarily handles tool selection, generation, and screening of candidates; whereas in ART research, it has begun to participate in identifying problems worth further investigation.
This raises new requirements for Anthropic’s life‑science strategy.
When Claude can analyze large amounts of data in a short time, it may also generate candidate results far beyond what laboratory validation can handle. Anthropic discovered in the ART project that a single research activity can produce hundreds or even thousands of candidate reports, but scientists cannot experimentally test each one.
Therefore, the issue is not only whether AI can discover more clues, but also how to determine which clues are worth validating. Anthropic has begun trying to have Claude learn from researchers’ experience in selecting scientific problems, but the related capabilities are still in an exploratory stage.
A more realistic limitation comes from the experiments themselves. Models can quickly generate candidate molecules or research hypotheses, but whether a protein has the expected function and what role a newly discovered enzyme system actually plays must be confirmed through real experiments.
This is also a key distinction between life‑science research and pure software tasks. Programming tasks can often run code directly, check results, and then modify based on errors; biological research, on the other hand, requires real samples, laboratory equipment, and specialists, and the validation cycle can be much longer.
For Anthropic, which wants to dive deeper into the research process, merely boosting the model’s computational and analytical capabilities will not automatically accelerate every stage of research. If experimental validation can’t keep up, no matter how many candidates Claude finds, it will be hard to judge which results are truly valuable.
And getting the research questions proposed by the model into experiments more quickly is gradually becoming another issue Anthropic needs to address.
Why does Anthropic need its own lab?
In fact, Anthropic can already obtain experimental resources through external partners.
Previous protein design research was validated by external laboratories. The company also partnered with firms such as Adaptyv Bio and Twist Bioscience to host protein design competitions, planning to experimentally test over 5,000 screened designs and publicly release the results.
This approach can leverage the mature experimental capabilities of professional institutions, and it does not require Anthropic to build all facilities itself.
However, as Claude begins to take part in more open research tasks, experimental needs are also changing. Unlike testing a batch of pre‑designed proteins, when exploring unknown biological systems researchers often need to re‑analyze data, adjust hypotheses, and decide the next steps based on each experiment’s results.
This type of research does not have a fixed workflow and requires close coordination between computational analysts and experimental staff.
Anthropic’s head of life sciences, Eric Kauderer‑Abrams, told Reuters that establishing an internal lab is intended, on the one hand, to accelerate certain research projects, and on the other hand, to allow the team to gain hands‑on experience conducting biological experiments.The company will not cease collaborating with external labs; instead, it will choose the most appropriate approach based on the specific task.
Therefore, building its own lab does not mean Anthropic intends to conduct all life‑science research independently; rather, it allows the company to take part in certain experiments directly and understand Claude’s performance and limitations in a real research setting.
To better bridge computational research and actual experiments, Anthropic is also tackling two other challenges:how to make AI more convenient to use laboratory equipment, and how to reduce the computational cost of specialized biological models.
Regarding experimental equipment,Anthropic, in collaboration with Janelia Research Campus of HHMI, developed the Model Hardware Standard (MHS).This specification aims to enable AI agents to connect microscopes, liquid‑handling devices, robotic arms, and other instruments through a unified interface, reducing software integration work between laboratory equipment.
Figure | Comparison of three architectures: traditional academic lab, traditional automation lab, and MHS lab (source: Anthropic)
It shares a similar approach with Claude Science. The latter aims to connect scientific databases, specialized models, and tools within computers, while MHS tries to extend this connection to real experimental equipment.
At the same time, Anthropic also used Claude to optimize more than 30 open‑source biomolecular models. In the company’s published tests, the relevant models ran about four times faster on average. These improvements focus on the computational implementations of existing models, helping reduce the computational resource demands of certain research tasks.
Although these efforts are at different stages of development—MHS is still a research preview, and the actual gains from model optimization depend on the specific task and runtime environment—they are linked to the layout of research platforms and in‑house labs: Anthropic not only wants Claude to understand research tasks, but also hopes it can more conveniently use the resources required to complete those tasks.
This direction is also reflected in talent investment.Anthropic acquired the Coefficient Bio team with a background in computational drug discovery, and John Jumper, who won the 2024 Nobel Chemistry Prize for his AlphaFold research, also announced his joining the company. At the same time, Anthropic is recruiting professionals in computational biology, experimental biology, chemistry, and related fields.
For a general-purpose large model company, these talents can not only help improve Claude’s research capabilities, but also enable it to conduct some biological research on its own, without having to rely entirely on external institutions for research experience.
However, Anthropic does not intend to transform into a traditional pharmaceutical company because of this.
Kauderer‑Abrams told Reuters,the company is especially focused on rare diseases, neglected diseases, and some research challenges that are difficult to drug, but it currently does not plan to conduct commercial clinical trials on its own, nor does it wish to compete directly with biotech firms whose core business is drug commercialization.
Based on its current positioning, Anthropic leans toward providing technical support to research institutions and pharmaceutical companies through Claude, rather than handling the entire drug development process itself. Building its own labs helps the team validate research hypotheses generated by the model and also aids in accumulating practical biological research experience.
This also clarifies the series of moves Anthropic has made over the past year: from linking research data and specialized models, to participating in protein design and exploring unknown biological systems, to building its own experimental capabilities, Claude is entering more aspects of life‑science research.These initiatives may also indicate that Anthropic wants Claude to be more than an AI tool used by scientists, but to gradually become the infrastructure underpinning life‑science R&D.
Reference links:
https://www.reuters.com/world/anthropic-quietly-sets-up-biology-lab-it-ramps-ai-drug-program-2026-09-18/
https://www.anthropic.com/news/claude-for-life-sciences
https://www.anthropic.com/news/claude-science-ai-workbench
https://www.anthropic.com/research/Claude-accelerates-protein-design
https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
Note: Cover/featured image generated with AI assistance
Get MirrorSpeed