# A virtual cell must predict unseen interventions

Markdown copy of https://realalpha.app/s/a-virtual-cell-must-predict-unseen-interventions/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Alex Rives](https://realalpha.app/builders/alex-rives/) · Head of Science, Biohub · Created the ESM protein language models at Meta and co-founded EvolutionaryScale, now leads Biohub's AI-for-biology science.
- Host: [Elad Gil](https://realalpha.app/people/elad-gil/) · CEO, Gil & Co · Investor in Airbnb, Stripe, Coinbase and Anthropic; co-founded Color.
- Host: [Sarah Guo](https://realalpha.app/people/sarah-guo/) · Founder, Conviction · Ex-Greylock partner; early investor in Cognition, Harvey and Mistral.
- Show: No Priors
- Episode: [“Curing All Disease by next century is too conservative" - Mark Zuckerberg](https://realalpha.app/v/u8cYwaLzN-k/), Jun 10, 2026
- Moment: 46:51 to 49:37 of the episode (3 min), playing on https://realalpha.app/s/a-virtual-cell-must-predict-unseen-interventions/; the episode from there: https://realalpha.app/v/u8cYwaLzN-k/?t=2811

## Quote

> “you need enough generality so that you can ask the model questions about **a new intervention** in a context that it hasn't been trained on”
> — [Alex Rives](https://realalpha.app/builders/alex-rives/) at [47:56](https://realalpha.app/v/u8cYwaLzN-k/?t=2876) · No Priors

## Takeaways

1. Connect protein, genetic and transcriptomic layers to phenotype.
2. Predictions should generalize to new contexts.
3. Generating enough relevant data is a major challenge.

## About this segment

Alex Rives defines a virtual cell as a model connecting molecular layers to phenotype and generalizing to new interventions. Making those predictions requires substantial new data.

From ““Curing All Disease by next century is too conservative" - Mark Zuckerberg”, Jun 10, 2026. [Watch the full episode](https://realalpha.app/v/u8cYwaLzN-k/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[46:56](https://realalpha.app/v/u8cYwaLzN-k/?t=2816) what the next step in the research agenda is? Um, it's like world model for biology. And then I could, I'm just gonna be very coarse here. Like I could scale it up. I could add more data. I could like adding data is a non-trivial thing in terms of new methods and domains. Like what is, do you take input from the, um, the larger ecosystem about, you know, how people are using it and what would make it more useful?

[47:19](https://realalpha.app/v/u8cYwaLzN-k/?t=2839) Or is it really like we understand like the next step of structures or coverage that we're looking for? I mean, I think there's two things. So like we have a view on kind of the next big challenge, which I think is... you know, the virtual cell and, you know, really being able to kind of ladder up the hierarchy of biological complexity to the cell. Sorry, very basic question. This virtual cell model, like what is the input and output I should expect? Yeah, I mean, I think there's different views on that, but I think kind of what you ultimately want is a system that can really model each of the levels of complexity. So, you know, the proteomic layer, the genetic layer, the transcriptomic layer and connect that to the phenotype. And you need enough generality so that you can ask the model questions about a new intervention in a context that it hasn't been trained on and

[48:10](https://realalpha.app/v/u8cYwaLzN-k/?t=2890) kind of get an answer from it. And, you know, the gap that we need to close as a field is being able to really make those predictions that can generalize. So that's going to require, you know, enormous effort to generate data. Yeah. And then, I mean, in terms of what you decide to do next, I think this is like, you know, a pretty normal process of constraint management, right? I mean, it's like, like, I think every lab in every field across the world probably feels compute constrained. I think that that's probably true here too, right? It's like, so, I mean, I know, like, you know, there's always questions like, okay, should we double down more on advancing the protein piece? Should we do more of the cellular stuff? I think that those are kind of ongoing debates in terms of how you sequence that. Um, and then, yeah, within that, there's

[48:58](https://realalpha.app/v/u8cYwaLzN-k/?t=2938) kind of being at the Pareto frontier about how much you want to train the different models in order to like, and the size of the models is also dependent on the scale of the data that you have because you have, yeah, for obvious reasons. So yeah, I mean, I think it's, there's some of that is just where you want to be on the curves and the normal constraints. But I think that this is like probably the same process that like any research organization goes through of like, You want to go in all these different directions, and you're just trying to constraint, optimize, and make enough progress to do world-class work at one thing at a time while planting some seeds that can blossom over the next couple years as well.