迷你大脑、跳动的心脏类器官,以及预测它们的AI——Fabian Theis,Helmholtz Munich
- 扰动数据逐渐补齐,虚拟细胞的应用空间随之扩大。Theis提到,单次实验已测得1000万个转录组,药物和CRISPR筛选的覆盖面也越来越广;RJ则提及Biohub宣布了一项18亿美元的合作。在Theis看来,虚拟细胞要回答的问题,是预测细胞在各类扰动下的行为。但把虚拟细胞等同于扰动预测过于简单:真正的目标还包括细胞的输入输出行为、记忆,以及RNA以外的分子层面。药企需要细胞类型特异性的筛选,却还无法大规模开展。
- 类器官将研究拓展到多细胞系统,也带来了尚未解决的建模难题。Theis举例说,培养皿里会出现“跳动的一小团心脏细胞”(“beating little heart cell clumps in a dish”):其中的相互作用,是单一细胞系无法呈现的。免疫成分缺失、数据有限,以及细胞邻域该如何表示仍无定论,都让这一领域仍处于研究前沿。
- 模型复杂度必须先在简单基准和有意义的指标面前证明价值。在器官内部的细胞分类任务上,逻辑回归仍有竞争力;药物反应预测数据噪声很大,有时甚至难以胜过均值。Theis说,ARC挑战也暴露了可被利用的指标;他的要求是“把基线做好”(“Do your baselines properly”),再证明复杂模型能解决更难的问题。
- 学习细胞状态与预测干预效果,是两项不同的建模任务。Theis介绍了一项对比表征工作,覆盖16个物种、3.5亿个细胞,参数量或约1亿。预测扰动效果还要把“状态X加上动作”作为输入,因此需要CellFlow这类生成模型,不能假设只靠表征学习就能解决问题。
- 相比单独的虚拟细胞模型,独特的实验数据更容易构成商业护城河。Theis多次看到平台公司转向“我们来做药”(“We make a drug”),不过主持人也提到,结构生物学和强化学习领域已有公司在探索其他路径。Theis的Cancer Organoid Perturbation Atlas展示了另一种客户价值:药企愿意为独家实验数据的优先获取权付费。
- 在Theis看来,欧洲融资的短板主要出现在公司扩张阶段,而非初创阶段。早期风投和非稀释性资金并不缺,但公司在较大规模的A轮、B轮融资时往往转向美国投资者。Theis所在的Helmholtz研究中心拥有充足的研究资源,约有45位课题组负责人、500至600名成员;他则推动每个课题组在3年内至少孵化1家衍生公司。
- 实验选择正逐渐走向实用,但面向生物学的强化学习仍是远期目标。Lab Compass从目标细胞状态反推可能的干预方案;Theis介绍了与一位剑桥造血研究合作者进行的3至4轮实验反馈。长远目标是“希望所有环节都用强化学习来做”(“we want to RL everything”)。RJ认为目前的方法“可能更接近主动学习”,Theis认同这一描述。
1. 扰动数据让虚拟细胞更具现实意义
- RJ开场提到,Biohub宣布与DOE、NIH、Meta和Google合作,投入18亿美元生成开放的多尺度生物数据。Theis将这项工作与Human Cell Atlas联系起来,同时说明自己的视角“相当学术”。他预计,已经在分子层面推进的生成式建模与实验反馈,未来会逐步扩展到细胞层面。
- 基础测量单位是转录组:每个细胞约有20000个维度,记录信使RNA的计数,其中很多数值为零。潜在表征可以压缩这个空间;聚类结果提示细胞类型,轨迹则可描述分化过程。Brandon用一个直观类比解释两者的区别:DNA提供原料,而“每个细胞都像一道菜”,转录组学试图还原这道菜的组成。
- 变化的关键在于干预覆盖面扩大,而不只是观测到的细胞数量增加。Theis回忆,单次实验已测得1000万个转录组;与此同时,筛选范围扩大到约100种细胞系,或涵盖1000种药物的筛选组合。过去,即使收集了海量细胞,扰动范围也往往很窄——可能只有对照组与疾病组之分。更丰富的干预覆盖,带来了难度显著更高的预测问题。
- 扰动不只是CRISPR基因敲除:邻近细胞的信号、激素、培养基条件、细微的遗传变化及其组合,都属于输入空间。敲除很有用,但它是异常剧烈的干预。Theis希望模型最终能将人类遗传学也纳入扰动;现实需求在于,药企无法大规模开展筛选,同时又保持“细胞类型背景特异性”。
- Theis提醒,虚拟细胞并不等同于扰动预测。他所说的更深层目标,是模拟细胞的输入输出行为、记忆和不同分子层面,而不只看RNA。将虚拟细胞等同于扰动预测模型,在他看来有些过于简单化。
2. 多细胞环境让类器官更具吸引力,也更难建模
- 对Theis而言,类器官说明了研究为何需要超越单一细胞系。细胞可以组织成具有涌现行为的小型系统;他举的例子是“培养皿里跳动的一小团心脏细胞”(“beating little heart cell clumps in a dish”)。这类系统能提供更丰富的实验读数,因为细胞彼此连接、相互交流,而不是只增加某种细胞在另一种处理条件下的观测数量。
- 获得稀缺组织的替代方案,也是类器官的一大吸引力。Theis介绍,可以从皮肤细胞出发,将其分化为类似小型大脑的系统,以期研究神经退行性疾病的遗传因素。但他也指出了局限:类器官可能缺少免疫成分,也缺乏充分的供养机制。它们能复制器官的一些有用部分,但仍存在明显缺口。
- 如何表示这类系统仍无定论。细胞类型是人类便于理解的分类方式——“自然并不在乎”这些标签——但我们还没有同样成熟的词汇来描述局部细胞生态位。因此,空间组学和病理学既带来概念问题,也带来计算问题。当前模型通常把类器官视为细胞分布;图结构或许能捕捉更细微的关系,但Theis说,目前还没有公认的最佳表示方法。
3. 生物学问题不够难时,简单基准依然有竞争力
- 针对主持人提出的归纳偏置挑战,Theis认为,先验约束过强可能牺牲预测能力。不过,扰动数据仍然有限,混合模型“可能仍是值得采用的方案”。以AlphaFold为例,模型如今开始从数据中学习旋转对称性;长期来看,他预计“数据会胜过人为加入的额外概念”。
- Theis在2019年提出的scGen,将非线性编码与潜在空间中的线性位移结合起来;简单方法至今仍难以击败。在器官内部进行细胞类型判定时,CellTypist等逻辑回归工具仍接近最先进水平。跨器官分类则会遇到更细微、受组织影响的差异;他在scTab项目中发现,非线性模型在这类任务上有所提升。究竟复杂模型是否有帮助,取决于具体的生物学问题。
- 谈到ARC的Virtual Cell Challenge,Theis表示,固定指标可能被参赛者利用;在他的印象中,最终被推为冠军的模型本质上基于flow matching。撇开指标利用不谈,他说真正的赢家采用了非线性方法。RJ补充说,Bo Wang在Nature发表的论文指出,一些线性模型看似获胜,部分原因是指标选择所致。Theis的Open Problems方案要求参赛者提交额外指标,让如何定义成功本身也成为科学工作的一部分。
- 药物反应预测提供了一个让人冷静的基准:有时,测量噪声太大,挑战只是要胜过均值。Theis给学生的建议是:“各位——我是说,男生女生都算——我知道你们想用最前沿的深度学习方法,但请先把基线做好。”
4. 有用的虚拟细胞既需要表征,也需要建模干预动态
- Nicheformer展示了一项比给细胞贴标签更难的任务:能否根据一个细胞的表达状态,预测它邻近细胞的表达状态?周围环境留下的影响,会反映在被测细胞中。Theis说,当研究者要提取这类更细微的信息时,更强的非线性模型就能派上用场。因此,现有数据是否还藏着未挖掘的信号,取决于提出的问题,而不能笼统断言基础模型有用或无用。
- 训练目标仍有很大的探索空间。这个领域借鉴了语言建模中的掩码自编码器方法,同时也在探索表格基础模型、JEPA式方法和对比学习。Theis说,早期生物基础模型论文往往没有证明缩放定律。评审应要求研究者证明:随着样本和数据量增加,模型复杂度确实能带来持续提升,而不是把模型规模本身当作价值。
- Theis的对比表征研究曾与Geneformer在约3000万个细胞的规模上比较;其在Hugging Face上的版本已扩展至3.5亿个细胞,参数量或约1亿。更大的潜在空间覆盖16个物种,或许能支持跨物种、跨组织比较,也让研究进化问题成为可能;不过Theis明确表示,这并非其团队的核心专长。
- 干预预测需要在这些表征上增加动态建模:模型接收一个状态和一个动作,再预测下一个状态。谈到扰动建模时,Theis说:“我不会声称纯粹的表征学习就是正确路径。”他的团队增加了一个基于flow matching的生成式推前模型;他表示,已被Nature Methods接收的CellFlow提升了扰动生成能力,但仍高度依赖良好的潜在空间。
5. 德国的科研实力受融资与数据获取限制
- Helmholtz的联邦资金支持长期研究项目,不同于以本科教学为主的大学模式。Theis介绍,他所在的Computational Health Center约有45位课题组负责人和教授,成员约500至600人,研究范围涵盖细胞建模、DNA、蛋白质设计和患者队列。他认为,这是欧洲最大的健康AI学术研究中心。与大学及临床机构的合作则提供了人才培养和转化研究渠道;将这些资源汇集起来,有助于实现更广义的“虚拟人”目标。
- 面对主持人对德国创业官僚主义的担忧,Theis认为,企业初创阶段的竞争力正在提升,尤其是在慕尼黑和柏林。他说,欧盟创立的初创公司数量与美国一样多。他的判断是,欧洲的问题出现在企业扩张阶段:早期风投和非稀释性资金都存在,但愿意承担更大风险的投资者较少。他多次看到公司为规模更大的A轮和B轮融资转向美国投资者,同时承认这样的概括并不全面。
- 健康数据整合仍是另一道阻力。Theis以芬兰的FinnGen和丹麦为例,指出这些地区比其他地方更好地连接了人群抽样与医疗记录;而在德国和美国,病历分散,难以构建患者的连续病程。他提到,欧洲健康数据空间(European Health Data Space)可用于跨境数据整合,“也许未来2年内,这些事情就得开始跨境运作”。他说,愿景正在形成,但实际落地比预期缓慢。
6. 平台公司仍需明确买家,并建立难以复制的数据优势
- Theis对商业化的提醒,始于一些公司承诺“我们解码生物学”,却说不清客户究竟会买什么。平台很有吸引力,但咨询业务并不会自动变成可规模化的生意。他看到不少虚拟细胞公司最终宣布“我们来做药”(“We make a drug”),因为相比开放式建模平台,投资者更容易理解如何为一个明确的治疗资产提供资金。
- 主持人则指出市场正在变化:Chai等结构生物学公司,以及出售强化学习环境的公司,都说明除了持有药物管线,还有其他商业路径。他们希望系统生物学领域也出现类似发展。Theis并不否认这种可能,而是将话题转向这些公司能独特生成并提供什么。
- Theis举例的是Cancer Organoid Perturbation Atlas,这是一个由多家药企合作建立的联盟。公司愿意参与,是因为想优先获取扰动数据集。在他看来,将建模与独特的实验数据生产结合,才是“最佳结合点”:模型附带客户难以从其他渠道获得的信息,药企也就有了明确的参与理由。
- Theis还在追问,市场究竟能容纳多少家独立AI平台。他提出,“Astra 2.0”或许最终能承担其中一部分工作,这也进一步加大了平台在模型之外建立差异化优势的压力。
7. 双导师制将生物学中的模糊问题转化为可计算问题
- Munich School for Data Science采用“Y”形培养结构,为每位学生配备一位方法学导师和一位领域导师。学生同时参加两边的组会,在学习生物学或医学语言的同时开发计算方法。Theis的观点是,重要的领域问题往往一开始并没有明确表述;受过定量训练的学生可以把这种模糊性转化为指标、模型和研究项目。
- 经过约8年,他看到毕业生既能分析数据,也能构建模型,去向包括药企、研究机构、初创公司和纯AI公司。谈话前几个月,NVIDIA在1个月内招走了他的2名团队成员。他认为,这些人的优势不只是展示基准测试的小幅提升,还包括参与定义了一个棘手的实际应用问题。
- 这套方法如今覆盖空间测量、单细胞蛋白质组学和RNA;Theis提到团队前一年发表在Science上的一篇蛋白质组学论文。商业转化也有了明确目标:他要求每位课题组负责人争取在3年内至少孵化1家衍生公司,同时表示学生如今也越来越主动推动创业。
8. Lab Compass让模型参与反复的实验决策
- Theis描述团队的发展路径是“从绘制图谱,到做出预测”,再到指导下一次实验。传统学术项目往往停留在一个模型、一张概念图和有限的验证,然后就转向下一个课题。反复反馈能让学习过程更连续,但他提醒,真正开展多轮计算与实验迭代仍不常见,即使有些机构大力倡导这一理念。
- Lab Compass针对的是逆向问题:给定一个正向模型和目标细胞状态,应该采用什么干预才能实现目标?Theis介绍,团队正在开发一个逆向生成模型来回答这一问题。录制当时,他预计相关工作“今天或明天”会发布在bioRxiv上;他强调,选择下一个最优实验本身就是一道很难的建模题。
- 团队与一位研究造血的剑桥合作者建立了3至4轮实验反馈循环,并通过迭代不断改进模型。关键条件是实验系统的周期相对较快,且双方合作紧密。说服实验人员投入大量时间,去尝试一个不熟悉、由模型推荐的干预仍然不容易;生物学直觉可能让团队选择折中方案,而非直接采用模型提出的实验。
- 机器人技术未来或许能提供帮助,但Theis认为,现在还稍早了一些,因为实验周期仍然很长。“我们想把一切都用强化学习来做”(“we want to RL everything”)代表了目标方向,他称之为“有点超前”。RJ形容目前的方法“也许更接近主动学习”,Theis认同这一说法。
完整逐字稿
本集逐字稿暂无译文,当前显示原文。
We can really start guiding experimentation now. Before, we built these human cell atlases, which had this mapping aspect. Then we talked about the world-model prediction problem. We went from mapping to prediction, but now we can actually steer the next experiment.
Quite frankly, very often the papers ended with some model, maybe a concept figure and a little bit of validation. Then you typically start the next project. Of course, you learn a bit, but it’s not continual learning. A new thing we call Lab Compass is an inverse generative model that allows you, once you have a forward world model or whatever, to figure out what you need to do to reach a certain state. Obviously, the ultimate vision is that we want to use RL for everything, including this loop. It’s a bit out there, but I think we’re slowly getting there.
Welcome to the Latent Space AI for Science podcast. I'm RJ Hanocki, CTO of Mira Omics. I'm joined by my co-host, Brandon Anderson, who develops RNA therapeutics at Atomic AI.
I can’t emphasize enough how much Fabian Theis and his labs and companies have contributed to the development of systems biology and the various iterations of the virtual cell concept. Papers and models come out of his lab faster than I can read them, and the lab is consistently moving the ball forward on our understanding of how to observe and model the behaviors and interactions of cells. Just as importantly, his work often focuses on creating tools that researchers and scientists actually use in their day-to-day work.
Most computational biologists have probably used Scanpy and AnnData at some point. This work has led to the creation of scverse, an amazing collection of open tools that has attracted many contributors. Fabian is one of those academics with more honors than you can possibly list, but suffice it to say that he leads many of the AI and computational biology efforts at Helmholtz Munich and related institutions, and participates in many other industry efforts.
Fabian, it’s really great to meet you. Thanks a lot for joining us.
Thanks for having me. I’m really looking forward to the discussion, man.
1. Why Virtual Cells Are Rising
I’ve been tracking your work for a long time, so it’s a great honor to finally get you on the show and talk to you. Let’s kick things off with some of the hype happening right now. There’s a ton of interest in structural biology. Anthropic and OpenAI’s biopharma initiatives focus heavily on designing molecules.
On the other hand, virtual cell and systems biology approaches really haven’t gotten as much attention recently. Maybe Zera and Noetic are good counterexamples there. But Biohub just announced a $1.8 billion collaboration with the DOE, the NIH, Meta, and Google to collect and open-source multiscale biological data that can be used to develop a new generation of virtual cells. Do you think we’re about to see renewed interest from the pharma industry in virtual cells, based on the activity we’re seeing in academia right now?
I should disclose before I answer that my perspective is quite academic. When I talk to my friends and colleagues who sit in the startups day by day, they have the overview and intuition I have about what different research labs are doing, but for the startups. They know which things are hot and what investments are happening. I don’t really have that kind of overview.
Having said that, you mentioned Biohub’s initiative, which is also heavily coupled to the Human Cell Atlas, an international, worldwide consortium pushing the next level of dataset generation. I do think there are a whole bunch of reasons why we’re going to see these models scale up. They’re working well as generative models, and lab-in-the-loop approaches at the molecule level are now being exported to work one level up, at the cell level.
I think the first reason, and maybe the most important one, is the data generation, which you really hit on. We’ve had the concept of a virtual cell for a while. My lab started perturbation prediction modeling and using machine learning in 2019, with a paper we called SEGen at the time. It was a nonlinear model, but essentially did a linear transformation in latent space, and it was actually quite good. Even nowadays, many of these pseudo-linear approaches are still doing a good job. But I think the datasets and the questions we asked were still on the simplistic side. That’s why it didn’t take off yet.
When people talk about virtual cells, I think what you’re actually asking is: Can you predict how a cell would behave under all kinds of perturbations? That’s a deep question. If you really think about it, it’s maybe even a holy grail in computational biology or systems biology: this input-output system that is a cell, with memory, all its specifications, and all the different levels of molecules—not just RNA.
Being able to map this out raises the question of what understanding even means. That’s a bit of a meta-problem, but I think you understand or model a system very well when you can predict its output behavior given different inputs. This is why people nowadays often equate a virtual cell with perturbation prediction. I think that’s a bit too simplistic, but these perturbation-prediction models—maybe what you mean by a virtual cell—are taking off a lot.
The startup landscape is exploding there. We talk a lot about Europe, but in the Chinese ecosystem, too, there are many additional efforts happening now. I hope this isn’t just a kindling of the flame because everyone is adding AI to their work, but because pharma desperately needs to be cell-type-specific in its screens, which it can’t yet do at that scale.
2. Perturbations At Scale
Maybe you could expound a little bit on what perturbations unlock. We talked to Zera recently, and they’re doing these perturbation screens. What do you think the next generation of data is going to unlock in addition to that?
For me, if I may—stop me if I get too much into these old stories. I’ve been trained as a machine learning scientist for the longest time, doing all kinds of things. I started with single-cell early on, using qPCR. To be honest, initially it wasn’t so much because biology was what I was interested in, but because we had high sample numbers for that. Even early on, it was clear that the number of qPCRs at scale—and then, of course, transcriptomes—was getting really big.
When I started my career, we had a few bulk RNA-seq samples. Now we had thousands of them. But now we’re entering a new regime. One of these eye-opening moments was when Parse Biosciences had the—I think they call it the GigaLab or something like that. There are similar ones from competing companies now. They measured 10 million transcriptomes in a single experiment. I thought that was next-level stuff.
The step from just getting more PBMCs as a control to actually doing something to them at scale gives us the cell space, but then there’s the perturbation space, which was typically very scarce. It was often just control versus disease. The number of points you had in the perturbation space was much smaller than the number of cell observations. But now this is exploding. You might have 100 different cell lines where you do full perturbations, like full CRISPR knockout screens, or batteries of 1,000 drugs added, and suddenly that space gets really rich. Then you can start doing all these fancy perturbation-prediction things. It’s crazy how quickly the scale is increasing.
We’re involved in a whole bunch of things where we go beyond cell lines and do organoid perturbations. Maybe if we have time, we can talk more about that. I’m quite convinced we need to go to multicellular systems to get really interesting readouts. Everyone is scaling these levels up, and at some point, we may even go to human genetics as a perturbation.
Can you just explain what a perturbation is, for people who aren’t familiar?
Actually, would you mind backing up one step more than that? What do we even mean by measuring transcriptomics? You were talking about PCR-based approaches and perturbations.
When I give public talks, I typically say: We have the genome for the body from the Human Genome Project, but it’s pretty much the same everywhere. So why do we have all these different cell types? What’s happening specifically? One of the first levels is that different things are read off the genome in each cell. These messenger RNAs, which encode the various downstream proteins, are expressed differently in different cells.
In a human, maybe there are something like 20,000 different genes, with many, many more transcripts produced through splicing. But that’s the rough level. What you de facto have is a 20,000-dimensional vector for each cell, with a whole bunch of zeros where something isn’t present and a whole bunch of numbers representing how many messenger RNA transcripts you have. These are called transcripts, and this is called the transcriptome.
You then have this 20,000-dimensional space that we often encode in a nicer latent space, with a submanifold where the actual cell cloud lives. Each point in the cell cloud is one cell. If these 20,000-dimensional vectors cluster together, we call those clusters cell types. If they form a sausage-like shape, we call them trajectories, like cell differentiation: a fertilized egg cell giving rise to all the cell types in the body.
DNA is sort of like a recipe book, or maybe it’s a bunch of genes or ingredients. Each cell is its own meal or dish, a combination of different ingredients. Transcriptomics is figuring out what ingredients went into making each meal. Is that a dumb analogy?
I have a bunch of foodies as kids, so I’ll use this metaphor a bit more often. I like it. Thanks, Brandon. I love it.
And perturbations are when you somehow perturb the genome or transcriptome so that it has different transcripts in it. Typically, that’s done with a CRISPR knockout, where you remove a gene from the genome, which causes it not to express that gene. Then you see what phenotypic changes, or actual changes in the physical biology, occur because of the perturbation. Is that more or less right?
Maybe I can take a step back there, because I think you’re fully right, but that’s one of the hardcore types of perturbation. Fully knocking out a gene happens a bit in evolution, but most things are a bit more subtle. Of course, in order to understand the system, it’s a good idea to take out bits and pieces.
In normal development, or when we’re in homeostasis, cells talk to each other all the time. They live in these complex multicellular states, with all kinds of communication happening between cells. They’re always receiving things from outside and responding to them. For me, a perturbation space is really just an input space of things that happen to a cell, maybe because it’s next to another cell and there’s some signaling happening, or some gradient or hormones coming in.
You can’t tune all of that up or down. Even something as basic as the different media conditions used when pharmaceutical companies do drug screens in vitro can have a huge impact. It’s still a bit like bio-driven alchemy, figuring out what cocktails to use to take skin cells and turn them into more or less embryo-like cells, so that you can make new cells to heal different things in your body. These are what are called Yamanaka factors, for example—Nobel laureate-type work.
I think a space of perturbations—things you can do to a cell—can be very rich. It can include changes in a gene, like a full knockout, or something more subtle, like human genetics. We need to put all of that into a big space of things we can do that have an influence on cell state. A virtual cell model should take all of it, including combinations, disease, time, and whatever else impacts the cell.
3. Organoids Reveal Cell Context
That’s a great connection to the idea of organoids. Can you talk a little bit about what an organoid is, how it connects to this concept of perturbation space, and what you do to it in addition to adjusting its context?
When I first learned about this concept—I’m not a trained biologist, so I typically learn about things by reading a bit or talking to colleagues and friends—I was super impressed that people could take cells out of a body, from a cancer or whatever, and let them grow into small multicellular systems that start connecting and doing things.
People can take some cells, put them in a dish, and grow little mini-brains: a set of neurons that start connecting and talking to each other. People have shown that these mini-brains in a dish can compute. Similarly, people can grow heart cells into little heart organoids. We have a bunch of local collaborations with fantastic partners, and those organoids actually start beating. You have little clumps of heart cells beating in a dish.
There are simpler systems, too, like the colon, where this works very well, and there are a whole bunch of cancer systems. Organoids can also help study diseases when material is hard to acquire. If you want to look into the genetics of a neurodegenerative disease, it’s typically very hard to get access to tissue. So you can take cells from the skin, differentiate them, and turn them into these mini-brains, then potentially study some things there.
Replicating little bits of organs in a dish—that’s why they’re called organoids—is fascinating. There are all kinds of limitations, obviously. Often the immune system is missing, and maybe they aren’t being fed properly. There are all kinds of problems, but it’s much more than just a cell line, which is one cell type.
Moving to the modeling side, how do you model that?
That’s a bit of a research frontier. I don’t think people have agreed on the optimal way to model it, particularly because there isn’t as much data as one would like. RJ, you’re involved, if I understand correctly, in a whole bunch of spatial omics and H&E pathology efforts. There are so many samples available for pathology, and now we’re increasingly doing more spatial omics. I think people are less clear about how to represent spatially connected datasets and then break them down at the organoid level, compared with dissociated cells.
Taking this a bit simpler, if we look at the space of dissociated cells and the groups within it, we have this beautiful metaphor of what we humans call a cell type. Nature doesn’t care about that; it’s just us putting categories on top of things to help us understand the complexity. I don’t think we have a similar metaphor for local niches of cells. Maybe a little bit, but in H&E, whenever I talk with pathologists and see what they do, it’s often very qualitative and maybe not at that level yet.
I think we need to develop a categorization language for ourselves. Maybe computationally we don’t need it as much, but we also need to think about how we deal with this. At the moment, we essentially deal with it as distributions of cells. That’s the solution for now, but it could be much more subtle, like graphs or something.
4. Balancing Bias And Data
On the one hand, you have the bitter lesson: you remove inductive bias, and you can use a larger dataset that applies to your problem. That helps you get more accurate models because your distribution is wider, or whatever. But there’s a countervailing tendency, maybe the flip side of the bitter lesson. In data-poor regimes, you have to introduce a lot of inductive bias to get enough statistical power to make good inferences.
There’s also a second side to that, which John Platt brought up in our discussion with him not too long ago: you have predictive models and descriptive models. What you really want as a scientist is to arrive at a simplification, a descriptive model, where you can say, “I’ve tested all the apples, and gravity works in this way on apples.” Then I’m modeling it as a universal concept so we know it works on planets as well.
I want to understand what simplifications, insights, or actual mathematical or scientific models are becoming apparent that will help reduce all this complexity, so we can gain meaningful understanding from these really large datasets.
That’s a very important and, I think, kind of deep point. It breaks the field into maybe two different groups of people with answers. How much inductive bias, or priors, do we need to put into a model to make it more explainable, and potentially more robust when going out of distribution? Or how much are we just going to rely on bigger data to make the model more expressive?
I need to be very clear before we go into the philosophy.
I think the literature has clearly shown over decades, at least for many of our main datasets, that if we build in a priori inductive biases too strongly, we might lose predictive power. So it comes at a cost. If you’re a physician—or even a patient, perhaps, if you might have cancer or not—you may want the best model as a predictor. If you’re getting treated, you might want a model that’s a bit less good but that you understand a bit better. But for prediction, at least, I think you want the best. And I don’t think that’s entirely clear.
In our field, the cell scale is big, I think, but the perturbation-space scale—the things we do to the cells—isn’t as big. I don’t think we’re yet in a super data-rich regime, so in our case, these hybrid models might still be the way to go. But more and more, of course, and this is the honest lesson from large-scale machine learning, is that at some point data beats human-made additional concepts in models. One simple example I often point to is AlphaFold. The first models made an extraordinary effort to make molecules rotationally equivariant. They built that into the learning process. In later versions, rotation symmetry was automatically found by the model with more data, and the model was a bit more flexible. Whatever physical constraint you put in isn’t always as rigid, so you gain something if you let the model learn from the data. I think in the future, the larger share of the more widely used models will be these more flexible, data-based ones, with interpretation added on top afterward.
This also brings up a question for me. You recently wrote or co-authored a paper on the statistical power of current-generation virtual cells, in terms of the lift you get from additional data in those models. I think you said that the objective function or training methodology might be the problem, and not necessarily that the information isn’t in the data. Can you explain that a little, and why you think there might be plenty of untapped information in the data? Where does that get us, and what do we need to do to go further?
I think it’s fair to say that early on, we had very small, limited prediction models, learning latent spaces first and then perturbations on top. When large-scale foundation-model representations came out, everyone bought into the hype, perhaps a bit too much. Then comes the sobering-up phase, or something like that.
Yeah.
There were a whole bunch of papers saying, “Hey, maybe even a more or less linear model does a good job.” Our scGen paper had some of these benchmarks early on, showing that it’s not such an easy task to improve on that. I think there’s some truth to it: with the current state of the data, for some things, like predicting the effect of a CRISPR perturbation in a cell line, you might not need the fanciest model. If you want to predict a cell type, logistic regression is mostly fine.
But if you want to predict something more difficult, for example, in our Nicheformer paper, which is a foundation model for spatial omics, we introduced a new task. You have the gene expression, the state—the latent space—of a single cell. Can you predict the gene-expression state of its neighboring cells? They’re always in communication; they talk to each other, so you can see a remnant of one cell’s state in the other. That’s harder to do. For these harder tasks, you need stronger, more nonlinear models, and you need to look for more subtle information in the data. Whether you can find more information-rich features depends very much on what you ask.
In the self-supervised learning phase, we’re definitely not yet at the best way to represent and learn these things. There’s a lot of copying from NLP and LLMs, like masked autoencoders. People are still debating the typical approach and the types of losses and representations to use, perhaps because this is often tabular data. People are trying tabular foundation models, and there are a lot of JEPA-type approaches. We have one approach where we use contrastive learning, so people are playing around with the losses and the representations.
I think reviewers nowadays are looking much more closely—and I’ve been advocating this for a long time—to make sure you properly show scaling laws. That was missing from the early foundation-model papers. People hadn’t really shown that you need this complexity to scale up. But if you can really demonstrate scaling behavior across examples and samples, then I think you get buy-in for these models. There aren’t as many inductive biases in these approaches, so I do think we’re going to hit this regime quite soon. We’re already seeing it in a few papers. For example, in our contrastive-learning paper, scConcert, we show quite strong scaling behavior. In the paper, we went up to maybe 30 million cells, comparing with Geneformer. On Hugging Face, we now have a 350-million-cell model, with maybe up to 100 million parameters. We’re really going to a large scale, and we’re already seeing nice scaling behavior starting to emerge.
Just to be clear, at 100 million parameters and—did you say 300 million cells?—you get a lift in what kinds of applications? Are these the Nicheformer kinds of questions, where you’re asking, “What is my environment based on my transcriptome?”
Yeah. I mean, the ultimate thing we want to do is predict perturbations better, because that’s what people nowadays understand by a virtual cell. For that, I wouldn’t claim that pure representation learning is the right way to go. Sorry to repeat the subtlety, but typically you learn the state space with these foundation models. The way the model behaves under an action, under a perturbation, is where you need more world-model-like approaches.
World model is the very hip term. I’m sure you’ve had this—
Yeah.
—on your podcast already at some point. It’s been—
Yes.
—all over machine-learning conferences in the last year. But a perturbation-prediction model, like a virtual cell, is effectively a world model in transcriptome or other omics space. It’s not just a model that is F of X, but F of X comma A: the state X plus the action, predicting the next state, X prime. For this, you might need other combinations. What we typically do is learn the representations with, for example, scConcert, and then put a generative model on top to model the push-forward along the action. For this, we use something like flow matching.
Our CellFlow paper has just been accepted for Nature Methods, and we show that these approaches are much better at generating the effects of perturbations. But they depend very much on a good latent space. For this, using the scConcert latent space, we went to 350 million cells and included multiple species. You now have 16 different species where you can potentially look for similarities. Suddenly, you can ask evolutionary questions that you couldn’t before. Not that we can do this particularly well—it’s not our core expertise—but you can start doing these cross-species, cross-tissue latent spaces.
5. Linear Models Set The Baseline
I’m curious about the history of some of these different models. You just gave us a rapid version of it. One thing that was somewhat famous in the community was the point about linear models. Can you explain specifically what the key point was? This is one of those things in classic machine learning: make sure you understand your baselines well, or you can spend a lot of time spinning your wheels and fool yourself. I think there was a famous open competition through the Arc Institute that was a glass of cold water on everyone’s faces.
Can you maybe give a little bit of history about that, why we thought this was exciting, and where this had us pivot a bit?
Yeah. It sounds like such an obvious lesson, right? Do your baselines properly. But in a field rushing toward the gold pot of the virtual cell, people maybe weren’t as careful with the baselines. There are a whole bunch of subtleties with respect to holdouts and so on that weren’t always fully observed, which is why some of these things happen.
One of the first observations was in something as basic as cell type prediction. Nothing super fancy: you give me a new transcriptome, and I tell you which cell type it’s from, maybe even which organ. Even nowadays, some models close to state of the art are really just logistic regression. For such a simple categorization, that’s actually not so bad. There’s CellTypist, for example, which does a good job. I think we had a paper maybe 5 or 6 years ago with the very plain title “Deep Learning Is Not Needed for Cell Type Prediction.”
That’s true if it’s within an organ. If you go across organs, it suddenly gets a bit more subtle because cell types recur in different organs, and immune cells change subtly depending on which tissue they’re in. Then you need to do a bit more. For us, there was a paper called scTab, a tabular foundation model, a predictive model of that type. We showed that this nonlinearity was really adding something beyond the baseline.
For perturbation models, which you were asking about, Brennan, early analyses, particularly of Perturb-seq, were done with classical linear models on pseudobulk. Around 2019, we had scGen, where we started introducing linear shifts in latent space, so they were nonlinearly encoded. We always compared them with simpler linear shifts. But as the field got more complex and datasets started appearing, people began asking more and more perturbation-related questions.
There were a whole bunch of early competitions, for example, one run by MIT. We’ve also been running a regular NeurIPS competition on prediction problems in single-cell data, sometimes for data integration and also for perturbation prediction. That was very well attended. The cherry on top was definitely the Arc competition, which called itself not just a perturbation prediction challenge, but the Virtual Cell Challenge. That just sounds a bit cooler, right? It was very well received, but the metrics were maybe robust and maybe not so easy to work with.
Our solution in the Open Problems consortium, where a lot of labs contributed, was that people should submit their own metrics as well, to quantify subtle things you might not have thought of before. The Arc competition fixed a bunch of metrics that weren’t even very single-cell-specific, and they could just be exploited. I think the winning model, which was eventually crowned the actual winner, was essentially based on flow matching. It was based on this idea of a flow, a CellFlow-type idea. That shows that some nonlinear approaches are needed in this case. If you remove the exploitation, the actual winner was something that incorporated nonlinearities.
Nowadays, we see that linear or additive models aren’t cutting it for more complex things. But even now, if you’re working with a simple setting, with a few cell lines and gene knockouts, then, particularly when the effects are subtle, these more complex models can overfit a bit. That’s why a linear baseline is always very good. I always tell my students, “Guys—I mean, guys and girls—I know you want to do the fanciest deep learning thing, but please have the baseline.”
For drugs, a baseline as simple as predicting the mean can be the hardest baseline—not even a linear model, just doing better than the mean. Sometimes that’s not as hard if there isn’t so much noise in the data, and this stuff is super noisy.
I just wanted to add one point: Bo Wang and some of his collaborators just had a Nature paper on how the metric was actually the problem in some cases where linear baselines beat some of the models as well. So it’s clearly not only metric hacking; asking the wrong question leads to the wrong answer. Or it’s a choice of metric.
Yeah, I think you can discuss that. At some point, we’re going to have agents trying out all kinds of combinations of whatever algorithmic thing you can come up with. In fact, we have some collaborations with AutoML-type labs to try this out. Maybe what’s left sounds a bit too harsh, but maybe our biggest ideas go into this metric formulation, right?
One thing I’ve been really interested in is the different models of funding in research academia. I think Germany has a fundamentally different structure from the US. How is your institution set up, and how does that give you distinct opportunities compared with a US-based research model?
I very much agree that these funding models matter for flexibility, but also for the teams you can set up and the long-term vision you can have. Helmholtz isn’t just in Germany; I would say it’s maybe Europe’s largest research organization in terms of funding. It has different areas of focus, with health being one of the big ones, but also materials, information sciences, physics, and so on. Essentially, it gets federal money to invest in deep research questions.
I’ve been building up what’s called the Computational Health Center at Helmholtz Munich, which I think now has maybe 45 group leaders and professors and 500 to 600 people. I think we’re the largest academic center for health AI research in Europe. We have people working at all scales. I’m focusing on the cellular scale, but there are people doing DNA foundation models, protein design, patient-cohort representation, genetics, and so on. Now you can start bringing those things together, and one of our big visions is to move toward the virtual human.
The key thing is that, in contrast to universities, even though we’re all affiliated with universities—mostly the Technical University or the University of Munich, interestingly, the 2 top-ranked universities in the European Union—we don’t have undergraduate teaching. We essentially just have PhD students and postdocs, but we interact a lot with the universities, and in particular with the clinics, because we want to have that translational component. This may be a bit similar to NIH-type models in the US, but distributed across the country, not just in one place.
One interesting link is that the government and this realization in Germany are increasingly showing that there needs to be a much stronger transfer component—not just translation through universities, but also value creation outside the academic realm, which obviously means more going into startups. The Boston environment has always been one of the key examples, but now, of course, so is the Bay Area. Recently, when I was visiting—maybe Genentech, or some colleagues—we went to a restaurant, and one of the waiters asked us, “Hey, are you guys talking about AI and bio? I’ve been interested, too.” Everyone is thinking about startup ideas, and I think our PhDs are doing that more and more. We’re learning, too. In terms of value creation, Munich in particular is now perhaps the prime place on the continent for VC capital, and deep tech, particularly bio and defense, is really exploding. So this stuff is starting to work.
The thing you were talking about with startups is really fascinating. I’ve heard that Germany has had some issues with its startup ecosystem. It’s difficult to create a startup there compared with even other places in Europe. What’s the competitive landscape for startups in Germany right now, and how does Germany become more competitive broadly?
I have a vested interest in this because I’m probably going to get my German citizenship soon.
Welcome. Let me know if you come by Munich.
Traditionally, Europe—and Germany in particular—has been quite bureaucratic, and processes are far less digital than you’d expect in other places. I’m not saying the U.S. is the best example of fully digitized processes, but things are definitely faster. People have realized that now, and I think startup creation and implementation in Germany are more than competitive compared with the rest of Europe. It’s actually one of the prime places, with hotspots such as Berlin and, in particular, Munich, which I think has the highest number of startups.
Even nowadays, if you look at the European Union, there are just as many startups being created as in the U.S. Where Europe, and Germany in particular, drops off is at the scaling part. There’s early-stage VC and non-dilutive funding, but when you get to scaling, the big VCs that take bigger risks aren’t nearly as present. I’ve seen this again and again: if you want to get into the bigger funding rounds, like Series A and Series B, people start going to the U.S. for that, too. Maybe I’m simplifying a bit, but that’s my current learning, Brennan.
Okay. You’re on the board or, I guess, part of several local startups, right? What does it take to go from an idea to creating a new startup in Germany?
I don’t know if you’ve heard that the European Union tries to regulate a whole bunch of things, and that can sometimes be an advantage. It’s kind of cool that all across Europe we can use a mobile phone, and you may have heard that everyone needs to have USB-C for their mobile phones now. Some good things are coming. But, of course, things like the AI Act make things a bit more complicated.
Luckily, not yet for research, but data access across hospitals can be an issue. We were talking about health deep tech, right? There are some systems in Europe that I think are fantastic. I’m thinking about Finland, FinnGen, Denmark, and so on, where there’s a very detailed description of the population, including sampling, and medical records are connected far better than elsewhere.
The U.S. is also a similar mess to Germany when it comes to getting all the systems to talk to each other. When patients go to different states in the U.S., it’s similar to when they go to different places in Germany: it’s very hard to build patient trajectories. But there are places that know how to do this. The European Union has something called the European Health Data Space, where, maybe in the next 2 years, these things need to start working across borders. You need to start building these combined health records. I do think the visions are starting to happen, but the scale isn’t happening as quickly.
For a tech startup, what I’ve noticed is that the biggest thing is having the best team for it. We have professors and PIs who really think about this. For example, in my department, I said, “Guys, every PI, every group leader should expect to have at least 1 spin-off in the next 3 years. Let’s see how that works.” You need to get people a bit on that track, but I think this is happening from the bottom up nowadays anyway, with students actually pushing for it.
They realize there are a lot of opportunities at the moment, particularly with this computational angle being pushed even more to the forefront, so we can now start guiding experiments. We haven’t really talked about that much, but once you have these perturbation world models, you can go back and drive the experiments. I do think computational startups can be very good for data generation. This won’t just turn into nice papers, which is mostly all our business, but also into this next generation of startups. You see a few of these popping up, where you actually get the lab and the loop implemented.
What do you see as some of the most promising areas for commercialization of virtual-cell-adjacent ideas?
It’s interesting. If you look at earlier examples—and of course, people always refer to what OpenAI, Anthropic, and so on wanted to do—they didn’t really know how to make money, and some people might argue they still don’t know nowadays. There are a few startups that got quite strong investment and said, “We build a foundation model of biology. We decode biology,” without actually saying what they wanted to do with it.
I’ve seen this again and again: these platform-type ideas are super attractive. They fit into what we know. You spoke about some of the tools and platforms we build, like Scanpy and scverse, right? But turning this into something that makes money is not as obvious. There might be some sort of consulting business, but when you say you’re building a virtual cell platform, what investors understand—maybe even more so in Europe’s more conservative VC ecosystem than in the U.S.—is that at some point you develop a drug, and then it’s clear.
I’ve seen so many of these platform companies pivot at some point to, “We make a drug.” On the side, of course, we have our virtual-cell business, but we have 2 drugs that we found, and if you put a lot of money in, maybe you have some other counting sums. A few companies are still sticking with scaling up the AI business, but it’s not as clear how you would turn this directly into a viable, scalable business model, or whether there’s even room for many of these. And, of course, there’s also this anxiety that at some point Astra 2.0 might be doing parts of this already anyway.
Yeah. I think this has actually been a thesis of mine: until recently, every single company that builds a model ends up just building drugs, because that’s the path to monetization over time. Landing sales with pharma is very difficult, or you need some other big-pocket spender. If you’re a tools company and you say, “I’m going to build an asset, and then I can sell that off,” that’s really attractive to an investor.
Yes.
But we’ve seen the opposite happening this year, right? You have a couple in structural biology, you have Boltz, and Brandon will know them better, but there are a couple of those. Chai is one that we talked to, and we’ve interviewed a few that are doing reinforcement-learning environment plays as well. I’d love to see the activity level in structure that we see in Structural also doing things like what Noetic is doing.
I think Noetic is a good example in the sense that they also have a nice scale of data generation, and a big part of that is access to data. We talked before about these organoid perturbation-type things. We’ve actually been setting up a consortium in collaboration with a whole bunch of pharma companies, which we call the Cancer Organoid Perturbation Atlas. Pharma buys in because they want early access to some of these perturbation datasets, right? I think some of these deals go in that direction.
If you couple it with data generation, not just the modeling part, I think that’s where the sweet spot is, where pharma actually buys in. More and more of the companies I see are building their moat not just around a virtual cell model, but in combination with unique, specific data. That’s also more value creation.
I don’t know if this is exactly correct, but I read that you said somewhere that you have your students follow what I think you called a Y model, in which you have to have a methods and a wet-lab advisor. That sounds super interesting and super relevant to our conversation right now about data being so important, and all the attention on lab work and that being a moat for a lot of companies. Can you talk a little bit about that?
Thanks for raising this. I’ve been setting up what’s since become quite a big thing, called the Munich School for Data Science. This is beyond biology; we had a whole bunch of other disciplines. There, we had this supervisor model where 2 PIs always needed to come together to make a proposal and propose it to the students in a graduate-student-type model, which has this Y shape. There’s a methods-oriented one and a domain one, right? And then there’s a supervisor and a student.
Sometimes we had 2 students under the Y model, so Y plus Y equals H. You get it? I think this idea is somewhat trivial nowadays, but I thought it was good to bake it in so strongly. I usually go to 2 lab meetings.
One lab meeting where they really need to learn and understand a deeper biological problem, and maybe understand the language of a medical setup or something like that. Typically, they're all computational, data science-type students, right? But then they go to the computational one. They get pushed on the method aspect of it, and thereby take up a whole bunch of problems that aren't yet precisely formulated in the biology lab, problems you can't directly put into a function.
Because they're method-y and quantitative, they can put this into a metric or whatever that then turns into a model and a paper. I think this is how we've been doing things for maybe 8 years or so, this type of school. A whole bunch of papers came out of it where I learned about systems I had no idea about before. People would approach me and ask, “Could we do something like this?” Then I’d learn something about the lung or the brain that I had no idea about.
I’d say, “Oh, hey, cool data. I think our model types could fit. Let’s build a project around it.” Those students became the experts in my lab on that detailed biological topic, together with a partner, and then we drove it forward. Nowadays, this also extends across data types. Early on, a lot of things were just around RNA, but since then we’ve had a very nice paper in Science last year on single-cell proteomics. We’ve also been extending things to spatial observations. There are all these additional modalities and things you can measure, and interesting questions come up in those domains where these assays work very well.
Yeah, my natural inclination is to gather cross-domain stuff. I think it’s often ripe for the generation of really good and interesting ideas. As a follow-up, you’ve been doing this for 8 years now, I guess. What kinds of longer-term outcome benefits have you seen from this?
I think one thing that got students to join in initially is that, although it sounds like more work—you need to learn two things and go to two lab meetings—students in Germany, and Europe in general, typically do a master’s before they go into a PhD. So, not just a bachelor’s as in the U.S.; you do 1 or 2 more years of a master’s. They know a little bit about science, but they don’t exactly know what interests them. With this kind of combination, they can sample a bit because, in this grad school, they get different proposals. They can explore what they’re interested in domain-wise, and they can also play around with the method side.
I think trying out different things turned the students into very good data analysts, but also modeling people. The areas they went into downstream were equally diverse. Some went into pharma in their biomedical domain, some stayed in research, some went in a startup direction, and some went into very pure AI-type work. NVIDIA hired 2 of my people in 1 month just a few months ago, going very much into the computational aspect and also a bit of the biology part that NVIDIA is now building up. There are similar things happening at other big companies, so I think it gives them an interesting profile.
Nowadays, when you get interviewed, it’s a good idea to say, “I didn’t only do this fancy model that improved maybe 2% in a NeurIPS paper-type of thing. I also had the real, more complex, deeper application that I had to help find as a use case.”
There’s been a movement in agentic AI. Several guests we’ve had, and a blog post we’ve published, have discussed how wet-lab scientists have started basically becoming computational thanks to agentic AI. I don’t know if this is—As I’m curious how this evolves, as agentic AI enables any scientist to become their own computational scientist, what do you think the evolution of this co-advising strategy is? And what do you think the evolution of integration between wet-lab biologists and computational teams will be? I’ve heard several famous CEOs say things like, “Our strength is integration between computational and wet-lab teams.” I’m curious how you think that evolves, and how you set up a culture to take advantage of these new technologies.
Let me keep the culture part out for a moment because I think it’s a hard one to answer, but a very good question. First, let me talk about the very cool point you asked about: where these things happen with deep integration. One of the most exciting frontiers we currently face, and an opportunity we have, is that we can really start guiding the experimentation. Before, we built these human cell atlases and had the mapping aspect. Then we talked about the world-model prediction problem. So from mapping, we went to prediction, but now we can go to actually steering the next experiment.
This has always been the scientific method: putting our hypothesis in the loop and then doing something with it. But quite frankly, very often papers end with some model, maybe a concept figure, maybe a little validation. Then you typically start the next project. Of course, you learn a bit, but it’s not really a continual-learning type of thing. Now I think we can actually loop back more and more. This is happening very much in molecule design or material design, where the lab-in-the-loop is really working out.
Even though some big pharma companies have been putting this out as a general big theme, there aren’t many examples yet where people really do multiple loops of integrating an experiment and a computational model, then going back—whether that’s 2 people or 1. This integration is hard to set up, partly because of modeling. Once you have a model, deciding what the optimal next question is not trivial.
We brought out a new thing, which we call Lab Compass, that should be on bioRxiv today or tomorrow. It’s an inverse generative model that allows you, once you have a forward world model or whatever, to figure out what you need to do to reach a certain state. It’s super exciting for us, and a big part of the work was not just the model but the work with experimental people. In this case, a good friend and colleague from the University of Cambridge was working on hematopoiesis and had a rapid-cycle system. We set up 3 or 4 lab-in-the-loop cycles where we iteratively improved a model.
This is a really hard thing, and it gets into your second question, Brandon: how do you bring up a culture where, once your computational person or model or whatever predicts something, you can convince experimental people to put their lab time in? These experiments often take some time, so they have to buy in and actually measure some of these follow-ups. Sometimes this works a bit better with molecules, because the stakes aren’t as high. But if you say, “You need to do this new cell line; you need to add these very strange things,” people might ask, “Do you really want to do this? Shouldn’t we maybe use a bit of our gut feeling?” Then these intermediate solutions come up.
I think setting this up is far from trivial. That needs culture, and it might be the same person doing it. A lot of people start throwing around robotics and so on as well, but I think that’s a tad too early because our assay cycle times are still a bit on the long side. I do think this will come, and obviously the ultimate vision is that we want to RL everything. We want to do RL in this loop as well. It’s a bit out there, but I think we’re slowly getting there.
At the moment it’s more maybe active learning. Much like we talked about with Lila, that’s their core thesis: basically, the lab is part of the RL environment.
Mm-hmm. Yeah, no, that’s great.
Fabian, it is late at night for you now. I really appreciate you taking the time out of your evening to talk to us. I look forward to watching your research—and when I say you, I mean your lab and colleagues, and you as well. I really look forward to watching this progress and keeping careful track of the things we talked about today.
Yeah, thanks so much, RJ, Brandon—
Pleasure.
—for having me. I think these links are important, looking beyond what you do in your own smaller disciplines, and these discussions help me a lot.
I really enjoyed the discussion. Had a lot of fun. Thanks.
Great. Thanks. Take care.
Thank you.