Zhou Jingtian: Biology Has Scaling Laws, but How Far Do They Extend?
In an interview with MIT Technology Review China, Arc Institute lab head Zhou Jingtian said biology certainly has scaling laws, but the open question is how far they extend. He argued that accurately predicting gene expression does not mean a model understands cells or can change their functions. Zhou described moving from large single-cell atlases of the brain and human organs to AI biology models he helped develop: Bolero, which aims to predict gene regulation across cell types, and Devo, which aims to predict cell-state changes during development. He said virtual-cell research must ultimately be validated through real biological experiments. Zhou also said a previous Evo 2 training run at his institute used 2,048 GPUs for about two months and received support from NVIDIA.
The interview sets out an experimental test for virtual-cell models: rather than relying on gene-expression prediction accuracy alone, researchers must show through experiments that models can solve biological problems. It also identifies the Bolero and Devo projects and provides a concrete example of the compute required for large-scale biology-model training.
“肯定有 Scaling Law,我觉得问题在于究竟能够扩展到什么程度。”当越来越多的研究者相信,更大的模型、更多的数据能够让 AI 掌握复杂的生命规律时,美国 Arc Institute 实验室负责人周镜天给出了一个相对审慎的判断。
过去十多年,他亲历了单细胞技术的快速发展,从最初只能分析少量细胞的数据,到如今已经能够绘制覆盖大脑乃至人体多个器官的大规模细胞图谱。但当科学家逐渐看清不同细胞的分子特征,一个新的问题也随之出现:我们能否进一步预测细胞会如何变化,甚至主动改变它们的状态?
周镜天的科研经历,恰好沿着这条路径展开。本科就读于清华大学药学专业的他,后来在美国从事计算生物学与单细胞研究,参与开发过多项实验和计算技术。2023 年,他成为 Arc Institute 首位 Science Fellow,并于次年建立独立实验室。2026 年,他凭借单细胞图谱与疾病相关基因调控研究,入选《麻省理工科技评论》“35 岁以下科技创新 35 人”亚太区名单。如今,他开始将更多精力投入 AI 生物学模型,其参与开发的 Bolero 和 Devo,分别尝试预测不同细胞中的基因调控,以及细胞在发育过程中的状态变化。
#Arc Institute#Bolero#Devo#Evo 2#Zhou Jingtian#NVIDIA#Salk Institute#Stanford University#UC San Diego