Summary of Key Points
In June 2026, the American AI company Anthropic launched Claude Science, an AI platform dedicated to life sciences, with the aim of accelerating drug development and uncovering biological principles using AI. To achieve this goal, Anthropic has taken three core approaches: specifically training AI models with biological data, establishing its own experimental laboratories to obtain real-world data, and acquiring startups to enhance its expertise in drug research and development. Training bio-AI faces unique challenges due to the lack of “standard answers.” The industry has moved from concept validation to the early stages of industrialization, but the idea of “creating drugs with a single click” is still not feasible. The long-term vision is for AI to understand biological causal relationships and address issues such as matching patients appropriately for drug trials. Additionally, capital is flowing into this field, but the dual-use risks of AI—both potential for creating beneficial drugs and malicious applications—must be continuously managed.
Anthropic’s Three Key Strategies in Bio-AI
Anthropic isn’t just relying on theoretical AI models; it’s taking concrete steps into the biological domain through these three actions:
- Specifically Training AI Models: Using data from structural biology and clinical documents, Anthropic has been training its Claude model. From Opus 4.6 in February 2026 to 4.8 in May, the model has been continuously improved, with further efforts focused on biological knowledge.
- Establishing Own Experimental Laboratories: These laboratories enable real experiments that generate data. Relying solely on published literature does not provide real-time feedback; by conducting experiments, Anthropic can create a cycle of “experiment → data → model optimization → re-experiment,” making the AI more accurate.
- Acquiring Startups: Anthropic acquired CoefficientBio, a startup founded just eight months earlier, which specializes in identifying drug targets and selecting appropriate types of drugs. This acquisition complements Anthropic’s expertise in drug research and development.
Together, these strategies aim to reduce the drug development cycle by ten times.
Why Is Bio-AI Training So Challenging?
AI learning mathematics or coding is relatively easy because each problem has a clear correct answer, but biology is different:
- Lack of Standard Answers: For example, studying a disease may yield varying results from different laboratories, leaving no absolute “correct conclusion.” AI training requires clear “question-answer” pairs, whereas biological data often represents expert consensus and is not black-and-white.
- Insufficient Literature Data: Published literature is limited and may be outdated or biased, leading to models that are based on incomplete information. Therefore, Anthropic needs to use its own experimental data to refine its models.
In simple terms, training bio-AI is like teaching a child to write essays without standard answers; it requires trial and error to produce reliable results.
AI in Drug Development: From Improving Existing Drugs to Creating New Ones
The development of AI in drug research has progressed through several stages:
- Before 2023: Focus was on optimizing existing drug molecules, such as making them more effective or reducing side effects.
- 2023: David Baker’s laboratory used AI to design a new protein from scratch.
- 2024: Baker’s team designed a new antibody and won the Nobel Prize alongside DeepMind’s AlphaFold team.
- 2026: Companies can now create antibodies with high affinity for most targets, improving drug efficacy.
The industry has moved from demonstrating that AI can create drugs (concept validation) to producing potential molecules on a larger scale (early industrialization), but generating candidate drugs with a single click is still not possible. However, AI’s ability to tackle “undruggable” targets, which are difficult for traditional methods, represents its greatest value.
The Ultimate Goal: For AI to Understand Biological Causal Relationships
Current AI can make predictions (e.g., predicting protein structures), but experts aim for it to understand causal relationships in biology:
- What is causal reasoning? It’s like an engineer identifying which component in a circuit causes a problem; AI should be able to determine why a drug works for some patients and not others.
- Solving Problems: Many drug development failures are due to the inability to find suitable patients. If AI can understand causality and combine it with patient data, it can improve drug success rates.
For example, companies like Xaira train AI by disturbing cells (removing genes or adding chemicals) and measuring the results to help it understand causal relationships.
Capital Rush, but Security Risks Cannot Be Ignored
This field is highly attractive:
- Capital Inflow: Amazon is evaluating the pricing of the Claude model and considering entering the market; other tech companies are transferring engineers to work on bio-AI. Anthropic itself generates annual revenue of $30 billion, providing sufficient funding for investments.
- Security Risks: AI can be used to create therapeutic molecules or harmful ones (e.g., biological weapons). Anthropic addresses this by developing classifiers to detect malicious requests and implementing access control mechanisms. They believe security is a long-term effort that requires as much attention as model development.
In summary, bio-AI represents a significant opportunity, but it comes with security challenges that must be addressed. AI is evolving from a general-purpose tool to a powerful assistant for biological research, though there’s still a long way to go—both technical hurdles and security concerns need to be overcome. The future indeed holds much promise.