从映射到操控:虚拟细胞与实验闭环From Mapping to Steering: Virtual Cells and the Lab-in-the-Loop
The Takeaway:AI 正在从绘制和预测细胞行为,转向用 Lab Compass 这类工具主动引导下一轮实验,真正闭环虚拟细胞研究。
Helmholtz Munich 计算健康中心负责人、Scanpy 与 Anndata 等开源工具的核心推动者 Fabian Theis,分享了系统生物学的最新前沿。虚拟细胞的核心已逐渐等同于扰动预测:给定任何输入(基因敲除、药物、培养基条件、邻细胞信号),模型能否准确输出细胞下一状态。早期 SEGen 等伪线性模型在简单任务上仍有竞争力,但随着 GigaLab 级千万转录组、全 CRISPR 扰动和类器官数据的爆发,非线性世界模型开始显现优势。
Theis 实验室最新推出的 Lab Compass 是一个逆生成模型:在已有前向世界模型的基础上,直接告诉你需要施加什么扰动才能到达目标细胞状态。他们已在造血系统中完成了 3-4 轮真实实验闭环。他强调:“我们现在可以真正开始引导实验。从映射,到预测,再到实际操控下一轮实验。”最终愿景是把强化学习完整嵌入这个闭环,虽然还早,但已经在路上。
数据规模与任务难度决定了模型选择:简单细胞类型分类逻辑回归就够,预测邻细胞状态或复杂扰动则需要更强的非线性与更好的表征学习。scConcert 等对比学习模型已在 3.5 亿细胞、跨 16 物种上展现清晰的 scaling law。类器官与空间组学是下一数据前沿,因为真正有趣的生物学发生在多细胞系统里。
Helmholtz Munich 计算健康中心负责人、Scanpy 与 Anndata 等开源工具的核心推动者 Fabian Theis,分享了系统生物学的最新前沿。虚拟细胞的核心已逐渐等同于扰动预测:给定任何输入(基因敲除、药物、培养基条件、邻细胞信号),模型能否准确输出细胞下一状态。早期 SEGen 等伪线性模型在简单任务上仍有竞争力,但随着 GigaLab 级千万转录组、全 CRISPR 扰动和类器官数据的爆发,非线性世界模型开始显现优势。
Theis 实验室最新推出的 Lab Compass 是一个逆生成模型:在已有前向世界模型的基础上,直接告诉你需要施加什么扰动才能到达目标细胞状态。他们已在造血系统中完成了 3-4 轮真实实验闭环。他强调:“我们现在可以真正开始引导实验。从映射,到预测,再到实际操控下一轮实验。”最终愿景是把强化学习完整嵌入这个闭环,虽然还早,但已经在路上。
数据规模与任务难度决定了模型选择:简单细胞类型分类逻辑回归就够,预测邻细胞状态或复杂扰动则需要更强的非线性与更好的表征学习。scConcert 等对比学习模型已在 3.5 亿细胞、跨 16 物种上展现清晰的 scaling law。类器官与空间组学是下一数据前沿,因为真正有趣的生物学发生在多细胞系统里。
The Takeaway: AI is moving from mapping and predicting cell behavior to actively steering the next experiment with tools like Lab Compass, closing the real lab-in-the-loop for virtual cells.
Fabian Theis, who leads computational health efforts at Helmholtz Munich and whose lab created the widely used Scanpy and Anndata tools, lays out the current frontier of systems biology. The virtual cell is increasingly defined as perturbation prediction: given any input (CRISPR knockout, drug, media condition, or neighboring-cell signal), can the model accurately output the cell’s next state. Early models such as SEGen already showed that pseudo-linear shifts in latent space worked surprisingly well on simple tasks, but the explosion of GigaLab-scale data (tens of millions of transcriptomes), full CRISPR screens, and organoid perturbations is finally rewarding more expressive nonlinear world models.
Theis’s newest tool, Lab Compass, is an inverse generative model: once you have a forward world model, it tells you exactly which perturbation to apply to reach a desired cellular state. They have already run three-to-four real experimental loops on hematopoiesis. As he puts it, “We can really now start guiding the experimentation. From mapping, we went to prediction. But now we can go to actually steering the next experiment.” The ultimate vision is full reinforcement learning inside that loop.
Model choice still depends on task difficulty and data regime. Cell-type classification is still often solved by logistic regression; predicting neighboring-cell states or complex multi-perturbation outcomes needs stronger nonlinearity and better representation learning. Contrastive models such as scConcert already show clear scaling laws at 350 million cells across 16 species. Organoids and spatial omics are the next data frontier because the biology that matters happens in multicellular systems.
查看原文 →
Fabian Theis, who leads computational health efforts at Helmholtz Munich and whose lab created the widely used Scanpy and Anndata tools, lays out the current frontier of systems biology. The virtual cell is increasingly defined as perturbation prediction: given any input (CRISPR knockout, drug, media condition, or neighboring-cell signal), can the model accurately output the cell’s next state. Early models such as SEGen already showed that pseudo-linear shifts in latent space worked surprisingly well on simple tasks, but the explosion of GigaLab-scale data (tens of millions of transcriptomes), full CRISPR screens, and organoid perturbations is finally rewarding more expressive nonlinear world models.
Theis’s newest tool, Lab Compass, is an inverse generative model: once you have a forward world model, it tells you exactly which perturbation to apply to reach a desired cellular state. They have already run three-to-four real experimental loops on hematopoiesis. As he puts it, “We can really now start guiding the experimentation. From mapping, we went to prediction. But now we can go to actually steering the next experiment.” The ultimate vision is full reinforcement learning inside that loop.
Model choice still depends on task difficulty and data regime. Cell-type classification is still often solved by logistic regression; predicting neighboring-cell states or complex multi-perturbation outcomes needs stronger nonlinearity and better representation learning. Contrastive models such as scConcert already show clear scaling laws at 350 million cells across 16 species. Organoids and spatial omics are the next data frontier because the biology that matters happens in multicellular systems.