Xaira：虚拟细胞与药物发现要因果模型，就需要因果扰动数据；当前测序常杀死细胞，活细胞时间分辨测序才是真正「魔杖」。
- 因果模型←因果数据
- 扰动实验设计中心
- 相关≠药效
嘉宾：Bo Wang & Ci Chu（Xaira） · 日期：2026-07-21（PT）
- 原集网页: https://www.latent.space/p/xaira
- YouTube: https://www.youtube.com/watch?v=2AdS-2uuH80
- 音频: https://api.substack.com/feed/podcast/207941607/06ca3e4bd112d187109d3b25b2c9d8fc.mp3
- EN 转写（分析源）: https://www.youtube.com/watch?v=2AdS-2uuH80
播出信息
- 日期：2026-07-21（PT）
- 时长：约 89 分钟（RSS=5387s）
- 章节末段：1:28:43（约 99%）
节目结构（YouTube 章节）
- Intro — 00:00 🎧 · ▶️
- Guest intros & Xaira's three-platform overview — 02:07 🎧 · ▶️
- Why biology lags behind protein design — the data bottleneck — 07:12 🎧 · ▶️
- What X-Cell does — perturbation prediction explained — 13:07 🎧 · ▶️
- History of virtual cell modeling — from differential equations to scGPT — 16:05 🎧 · ▶️
- Perturb-seq at scale — building the PISCES dataset — 22:42 🎧 · ▶️
- Spatial transcriptomics and future modalities — 36:32 🎧 · ▶️
- X-Cell architecture — why diffusion beats autoregression for gene expression — 44:00 🎧 · ▶️
- Ablation results — what actually moves the needle — 54:25 🎧 · ▶️
- X-Cell generalization — beating linear baselines in unseen cell types — 57:29 🎧 · ▶️
- Single vs. combinatorial perturbations and platform expansion — 1:07:10 🎧 · ▶️
- Academia vs. industry in the agentic AI era — 1:16:07 🎧 · ▶️
- Open science and what academic labs should focus on — 1:25:14 🎧 · ▶️
- You can't have the cake and eat it—currently, to sequence a cell, you have to kill it. The real 'magic wand' for virtual cells would be temporal sequencing of live cells. — 1:28:43 🎧 · ▶️
分节详解
1) Intro — 00:00 🎧 · ▶️
💬 他们说了什么： 开场引入核心命题与嘉宾背景，定下本集问题意识。
💡 为什么重要： 该段推进本集核心主张与证据链。
⚡ 争议点： 自动字幕非人工精校；技术细节以论文/官方材料为准。
2) Guest intros & Xaira's three-platform overview — 02:07 🎧 · ▶️
💬 开场引入核心命题与嘉宾背景，定下本集问题意识。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
3) Why biology lags behind protein design — the data bottleneck — 07:12 🎧 · ▶️
💬 指出生物相对蛋白设计更缺高质量、可对齐的数据，这是滞后主因。 依据字幕：models so that we have something to show uh clinical utilities. So I think what uh clinical utilitie…
💡 数据与扰动设计决定因果/预测模型能否可信落地。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
4) What X-Cell does — perturbation prediction explained — 13:07 🎧 · ▶️
💬 解释扰动预测：打断通路后基因/蛋白连锁反应，模型要预测表达变化。 依据字幕：so forth and so on. There's this so forth and so on. There's this longchain reaction of of genes and…
💡 数据与扰动设计决定因果/预测模型能否可信落地。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
5) History of virtual cell modeling — from differential equations to scGPT — 16:05 🎧 · ▶️
💬 回顾虚拟细胞从微分方程到 scGPT 等路线，强调数据不可比与评估难题。 依据字幕：the data that you collect. And so that there's a big problem of how do I even there's a big problem …
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
6) Perturb-seq at scale — building the PISCES dataset — 22:42 🎧 · ▶️
💬 讲述大规模 Perturb-seq / PISCES 类数据如何支撑训练与泛化。 依据字幕：there. And there's n number of way to fit a causal regulatory network into fit a causal regulatory n…
💡 数据与扰动设计决定因果/预测模型能否可信落地。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
7) Spatial transcriptomics and future modalities — 36:32 🎧 · ▶️
💬 讨论空间转录组等未来模态，扩展超出解离单细胞的信息。 依据字幕：quite in depth there. So if you want to quite in depth there. So if you want to background you can g…
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
8) X-Cell architecture — why diffusion beats autoregression for gene expression — 44:00 🎧 · ▶️
💬 说明为何在基因表达上偏好扩散而非自回归等架构选择。 依据字幕：they're fundamentally objects which operate on sets. Uh the community spends operate on sets. Uh the…
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
9) Ablation results — what actually moves the needle — 54:25 🎧 · ▶️
💬 用消融实验指出真正拉动指标的模块，而不是叙事。 依据字幕：color prediction in human know what drugs will work in uh which patients but drugs will work in uh w…
💡 评估与消融是判断「模型是否真有用」的关键。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
10) X-Cell generalization — beating linear baselines in unseen cell types — 57:29 🎧 · ▶️
💬 展示在未见细胞类型上相对线性基线的泛化表现。 依据字幕：be a linear baseline and second we apply other models um from the field and last other models um fro…
💡 评估与消融是判断「模型是否真有用」的关键。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
11) Single vs. combinatorial perturbations and platform expansion — 1:07:10 🎧 · ▶️
💬 解释扰动预测：打断通路后基因/蛋白连锁反应，模型要预测表达变化。 依据字幕：always single gene pertivations or do you have more? Because my understanding you have more? Because…
💡 数据与扰动设计决定因果/预测模型能否可信落地。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
12) Academia vs. industry in the agentic AI era — 1:16:07 🎧 · ▶️
💬 讨论智能体时代学界与工业界分工、资源与目标差异。 依据字幕：myself just update the models and by myself just update the models and by talking to different stude…
💡 开源与学术分工影响生态与长期信任。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
13) Open science and what academic labs should focus on — 1:25:14 🎧 · ▶️
💬 谈开源与学术实验室应聚焦的问题与贡献方式。 依据字幕：I think we need the scale the I think we need the scale the industrialization the innovation from in…
💡 开源与学术分工影响生态与长期信任。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
14) You can't have the cake and eat it—currently, to sequence a cell, you have to kill it. The real 'magic wand' for virtual cells would be temporal sequencing of live cells. — 1:28:43 🎧 · ▶️
💬 收束：测序常需杀死细胞；活细胞时间分辨测序才是虚拟细胞「魔杖」。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
金句
- 因果模型需要因果数据。
- 你不能既要蛋糕又吃掉它——目前测序细胞往往意味着杀死它。
来源
- 原集：https://www.latent.space/p/xaira
- YouTube：https://www.youtube.com/watch?v=2AdS-2uuH80
- 音频：https://api.substack.com/feed/podcast/207941607/06ca3e4bd112d187109d3b25b2c9d8fc.mp3
- 分析源：YouTube 官方章节 + 英文自动字幕（未用 Whisper）。
- 未创建中文全文转写页；本页仅为深度中文详解。
