Anthropic 让 Claude 自主完成蛋白质设计，命中率 26.8%
Anthropic published one of its most ambitious scientific experiments with Claude…
做 AI for Science 或蛋白质设计的同学必看，这是 Agent 自主跑通完整湿实验闭环的硬核案例，命中率数据扎实，建议细读其协议与算力配置细节。
Anthropic published one of its most ambitious scientific experiments with Claude yet.
Anthropic 发布了其迄今最具雄心的科学实验之一，涉及 Claude。
Anthropic showed that Claude can take a biological target and autonomously run the computational protein-design campaign needed to produce binders that actually work in the lab.
Anthropic 展示了 Claude 能够针对一个生物靶点，自主运行计算蛋白质设计活动，以生成在实验室中实际有效的结合剂。
A lab may no longer need a dedicated protein-design expert to manually run every computational step. This can move the bottleneck from designing and triaging thousands of candidates toward experimentally testing a much smaller, AI-selected set
实验室可能不再需要专门的蛋白质设计专家手动执行每个计算步骤。这可以将瓶颈从设计和筛选数千个候选分子，转向实验测试一个更小、由 AI 选定的集合。
- Given a detailed expert-written protocol, it researched each target, chose where to bind, installed and ran open-source protein-design tools, generated candidates, filtered and improved them, then picked the final proteins for lab testing. Humans did not make the individual design decisions.
- The designs actually worked in the lab. Across 1,320 designs with usable measurements, 354 bound their intended targets, a 26.8% hit rate, and Claude found binders for 14 of 15 targets.
- On several targets, its results were competitive with human/open design competitions. Claude had higher hit rates on 4 of 6 comparable competition targets.
- Giving the agent more attention and compute seems to help. Mythos Preview reached a 35.1% hit rate when each target received its own 24-hour campaign, versus 26.7% when many targets shared a 48-hour campaign.
- The AI still cannot reliably know when a whole campaign has failed. Some unsuccessful targets received computational scores similar to successful ones.
- 根据一份详细的专家撰写的协议，它研究了每个靶点，选择结合位置，安装并运行开源蛋白质设计工具，生成候选分子，过滤并改进它们，然后挑选出最终用于实验室测试的蛋白质。人类没有做出具体的设计决策。
- 这些设计在实验室中确实有效。在 1,320 个具有可用测量的设计中，354 个结合了其预期靶点，命中率为 26.8%，并且 Claude 在 15 个靶点中为 14 个找到了结合剂。
- 在几个靶点上，其结果与人类/开放设计竞赛相当。在 6 个可比较的竞赛靶点中，Claude 在 4 个上具有更高的命中率。
- 给予代理更多关注和计算资源似乎有帮助。当每个靶点获得其自己的 24 小时活动时，Mythos Preview 达到了 35.1% 的命中率，而多个靶点共享 48 小时活动时则为 26.7%。
- AI 仍然无法可靠地判断整个活动是否失败。一些不成功的靶点获得了与成功靶点相似的计算分数。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力