STELLA Reportedly Identifies 40 AML Target Candidates
In a DeepTech interview, Zhang Zaixi said the research agent STELLA generated 40 previously unreported candidate targets for acute myeloid leukemia for a national key laboratory. According to Zhang, researchers tested the top-ranked target through gene knockouts in four cell lines and saw what he described as clear, highly consistent effects. The other 39 candidates had not mostly been experimentally validated. These results were reported by Zhang, a participant in STELLA’s development, and were not independently documented in the article.
The account highlights a central AI-for-science challenge: AI systems can generate and rank candidate targets quickly, while wet-lab validation remains much slower. It also illustrates the gap between computational hypotheses and experimentally verified results.
过去两年，AI for Science 正在经历一场肉眼可见的转向。自 AlphaFold 拿下诺贝尔奖、AlphaFold 3 和 ESM3 等模型相继问世后，AI 在生命科学里的角色，已经不再局限于预测一个蛋白结构或生成一段序列，而是开始直接接管“提出科学假设”这一核心环节。
2025 年，斯坦福大学加里·佩尔茨（Gary Peltz）团队曾做过一次对比测试：他们请谷歌研发的科研智能体 AI co-scientist 为肝纤维化筛选可以“老药新用”的候选药物。AI 推荐了三种，佩尔茨凭经验另选了两种。这五种药物在实验室培养的肝脏类器官中一同接受测试，结果 AI 推荐的两种表现出显著的抗纤维化活性，而佩尔茨选的两种均未见效。