Summary of Key Points
Anthropic’s Claude model has recently made a significant breakthrough: it has designed a new protein binder that does not exist in nature from scratch. This binder is more efficient and accurate than human experts or traditional AI tools. For example, it has reduced the time required to design a single protein target from several weeks to just two days, and increased the success rate from 10%-15% to 22%-35%. However, this is just the first step in the drug development process; there is still a long way to go before we have a truly effective medicine (on average, it takes 10-15 years and costs around $1-2 billion to bring a new drug to market). Meanwhile, the AI pharmaceutical industry has become a battleground for giants (Google, OpenAI, and Anthropic are all showing their capabilities), but there are also safety risks associated with AI-designed proteins (such as immune rejection, off-target effects, and potential for abuse). The founder, Amodei’s claim that “most diseases can be cured within 5-10 years” currently sounds like a distant science fiction story.
How Does Claude Design New Proteins?
Claude doesn’t perform magic; instead, it uses a large language model as a “command center” to integrate various scientific tools and create an automated process:
1. Target Selection: Researchers provide Claude with a 16,000-word “protein design manual” and ask it to select 15 targets from 16 candidates (for instance, TNFα, which is the target of powerful drugs like Humira), and identify the “critical regions” on each target that the protein binder needs to bind to.
2. Skeleton Creation: Based on the target, Claude generates the “main chain structure” of the new protein (similar to building a framework for a house).
3. Detail Filling: The generated structure is refined into a specific amino acid sequence (like adding bricks and tiles to the framework).
4. Optimization: The sequences are scored and evaluated to identify the most promising candidates.
5. Experimental Verification: Claude is only responsible for the theoretical design; the final sequences are sent to third-party laboratories for actual testing.
In simple terms, Claude acts like a knowledgeable project manager who breaks down the complex protein design process into manageable steps and uses tools to complete it efficiently, rather than “inventing” new science from scratch.
How Is Claude Better Than Human Experts?
Claude’s performance is impressive:
- Success Rate: The success rate has doubled, and the time required has been significantly reduced.
- Complexity of Protein Design: Designing a protein is like searching for a specific gold particle in a vast amount of sand (10^130 possible sequences for a protein with 10^80 atoms; only a small portion will fold stably and be useful).
- Human Experts’ Past Performance: Traditional methods took months to design a target, with a success rate of only 10%-15% (for example, in 2017, the Baker lab designed 20,000 candidates, and the success rate for high-affinity binders was 11.6%).
- Claude’s Achievements: Out of 15 targets, 14 were successfully designed, with the highest success rate of 40% for a single target (for example, the RBX1 target; human experts achieved only 3.7%, while Claude achieved 40%). Results for each target were obtained in just two days.
- Challenges Overcome: Claude has also managed to design difficult β-fold structures, which have been challenging for humans, but it has not been without failures (for example, all 90 designs for the MBP target were unsuccessful).
This progress is like moving from digging for gold by hand to using a metal detector, which greatly improves efficiency, although we are still far from achieving “turning stones into gold”.
From Protein Design to a Marketable Drug: There’s Still a Long Way to Go
Many people think that designing a protein means it can be used as a medicine, but there are additional challenges to overcome:
- Protein vs. Drug: Protein drugs are just one type of biopharmaceutical (e.g., insulin, PD-1 inhibitors). To develop a new protein into a drug, questions such as scalability, stability in the body, toxicity, and appropriate dosage need to be addressed.
- New Drug Development Process: It takes an average of 10-15 years and costs around $1-2 billion, with clinical trials (Phase III) and FDA approval being essential steps that AI systems have not yet mastered.
- Current Status: No AI-designed protein drug has been successfully launched on the market. Anthropic has also stated that it does not conduct experimental tests or manage drug development pipelines; it only sells its tools to pharmaceutical companies.
Amodei’s claims about curing most diseases within 5-10 years are more like a vision than a reality.
The AI Pharmaceutical Industry: A Battle for Dominance
The AI pharmaceutical industry is the next “gold mine” for giants:
- Google: Has the deepest expertise (with the Nobel Prize-winning AlphaFold), but progress is slow (its spin-off, Isomorphic Labs, has not yet brought a drug to clinical trials, and its key scientist has joined Anthropic).
- OpenAI: Follows a “vertical model + collaboration” approach (using the GPT-Rosalind model in collaboration with Retro Biosciences to design the Shazone factor, increasing efficiency by 50 times).
- Anthropic: Is advancing rapidly by recruiting experts (e.g., John Jumper from Google) and acquiring technologies to enhance its capabilities. Its clients include pharmaceutical companies like Novartis and Eli Lilly, but it does not manage drug development pipelines itself.
- Specialized AI Pharma Companies: Companies like Ingenuity Health’s Rentosertib are in Phase III clinical trials, and CrisprTalents’ drug is in Phase I, which is closer to actual development.
No company has yet emerged as the winner; the industry is still in a competitive phase.
Are AI-Designed Proteins Safe?
New technologies always carry risks, and AI-designed proteins are no exception:
- Immune Rejection: The human body may recognize the new protein as a foreign invader and trigger inflammation (for example, the Abicipar eye drug failed due to an 15.4% inflammation rate).
- Off-Target Effects: The protein may bind to a different target than intended (for example, the AspB10 insulin reduced blood sugar levels too effectively but caused tumors).
- Potential Abuse: AI can generate dangerous protein variants that may evade existing screening methods (a 2025 Science study found that 3% of dangerous variants went undetected), potentially leading to their use as biological weapons.
Anthropic has plans to restrict access to its protein design tool to scientists and is working on a “scientist access program,” but no timeline has been set.
Conclusion
Claude’s achievements in protein design are a significant breakthrough for AI in the biological field and can help pharmaceutical companies save time and costs in the early stages of research and development. However, to achieve the goal of curing most diseases, numerous scientific, clinical, and regulatory challenges must be overcome. For now, Claude is more of a useful “research assistant” than a “cure-all” solution. Amodei’s ambitious claims should be taken with a grain of salt.
(Note: This analysis is based on news content and does not constitute investment or medical advice.)