The Information's TITV
00:31:36 · Stephanie Palazzolo, Lori Goler
[00:31:36] S0: And when it comes to building these datasets, are there milestones that you're looking forward to, to kind of show the progress you're making? Can you walk me through some of those milestones? [00:31:46] S6: Sure. Yeah. I mean, the first thing will be just that we're creating novel datasets. So, you know, we'll know when we have those. I And hope we'll have those in the next twelve months. I hope they'll start becoming more there an example [00:31:56] S0: of one type of data set? [00:31:58] S6: Yeah, areas that us and other people are very interested in are areas like spatial biology and perturbation data, because a cell exists not by itself at a moment in time, but in context. In context means both over time, but also in context with other cells in your body, and things that are happening, and changes. To address something that's happening in your body, need some intervention. You want to have enough different types of data that the AI can really come up with new insights or observations or discoveries, not just tell us what we already know. The idea is to create this almost an oracle that's a generalizable reasoning model for the cell and for biology. How will we know when we get there? There are a lot of players now who are working on models. They'll be using this data. You know, we'll take a step back together and think about, like, okay. Well, this model did this. This one did this. Like, what is missing? What data is missing? And then together, we can go and create the datasets.
00:31:58 · Lori Goler, Stephanie Palazzolo
[00:31:58] S6: Yeah, areas that us and other people are very interested in are areas like spatial biology and perturbation data, because a cell exists not by itself at a moment in time, but in context. In context means both over time, but also in context with other cells in your body, and things that are happening, and changes. To address something that's happening in your body, need some intervention. You want to have enough different types of data that the AI can really come up with new insights or observations or discoveries, not just tell us what we already know. The idea is to create this almost an oracle that's a generalizable reasoning model for the cell and for biology. How will we know when we get there? There are a lot of players now who are working on models. They'll be using this data. You know, we'll take a step back together and think about, like, okay. Well, this model did this. This one did this. Like, what is missing? What data is missing? And then together, we can go and create the datasets. [00:32:58] S0: So it sounds like some milestones are, you know, getting the datasets in the first place and then this reasoning model you're talking about that will be some sort of model of the way cells Yes. Yeah. [00:33:07] S6: And you know, the model that we'll build will be a little bit different from the model that other folks will build. And so together, we'll be able to see how well it's working. And eventually, we're actually going have to validate it in the, you know, in the lab. So that's another step and another milestone that will happen down the road.
00:33:07 · Lori Goler, Stephanie Palazzolo, Nicole Su
[00:33:07] S6: And you know, the model that we'll build will be a little bit different from the model that other folks will build. And so together, we'll be able to see how well it's working. And eventually, we're actually going have to validate it in the, you know, in the lab. So that's another step and another milestone that will happen down the road. [00:33:21] S0: Are there any certain diseases that you think, through this type of work, general, like AI for biology work, any diseases you think can be cured in the next five [00:33:31] S6: to ten years? I think we'll see. We're more disease agnostic. We're sort of trying to bring the tools to everyone. I think what we're going to find out, though, is that the big disease categories are actually multiple diseases, and in some cases, maybe N of one diseases. It's like big areas like Parkinson's or cancer or even autism. These might be a variety of different diseases, so we may decide that we have more diseases before we start to actually cure them because we are understanding them better. And I hope eventually they'll be kind ofthey really will be N of one. You know, it'll bring to life the promise of personalized medicine. Yeah. [00:34:06] S0: I mean, again, the idea of personalized medicine I think [00:34:09] S3: is one that people have been
00:34:09 · Nicole Su, Stephanie Palazzolo, Lori Goler
[00:34:09] S3: is one that people have been [00:34:09] S0: very excited about for a long time. And so I think that's really exciting work. I mean, I am curious, like I kind of hinted at earlier, a lot of the AI labs, especially Anthropic, it feels like, and obviously Google DeepMind has been doing work here too. They have been focused a lot on building AI for science, AI for healthcare, AI for bio. I think Anthropic even has its own wet labs, for instance. What about your approach do you think will work better than the approach Anthropic is taking, for instance? [00:34:42] S6: I I hope that we all have success in this area because the more of us that are successful, the better for humanity. But I think one of the things that really makes Biohub different is that we have all the pieces really under one roof. So, for ten years, and Priscilla built frontier biology capabilities. We've added in the last year frontier AI capabilities. We have a meaningful amount of compute. We have, and then we're building all the data sets as we go. We have all of the things we need under one roof, and we're really taking advantage of that. It's really a cohesive system where it's not that we're doing something in one spot, and then we sort of have this other effort over here. It's like we really have an integrated strategy around it.
00:34:42 · Lori Goler, Stephanie Palazzolo
[00:34:42] S6: I I hope that we all have success in this area because the more of us that are successful, the better for humanity. But I think one of the things that really makes Biohub different is that we have all the pieces really under one roof. So, for ten years, and Priscilla built frontier biology capabilities. We've added in the last year frontier AI capabilities. We have a meaningful amount of compute. We have, and then we're building all the data sets as we go. We have all of the things we need under one roof, and we're really taking advantage of that. It's really a cohesive system where it's not that we're doing something in one spot, and then we sort of have this other effort over here. It's like we really have an integrated strategy around it. [00:35:20] S0: And, you know, obviously through this collaboration, you are working with the U. S. Government, Google DeepMind, Isomorphic Labs, Meta as well. Do you see yourself adding other, you know, companies or labs like Anthropic, for instance, moving forward? [00:35:35] S6: Yeah, I hope so. I hope that this will really be an effort. Since the announcement, actually, we've had great interest from a lot of players who've reached out to learn more and to find out what it's really about and to consider kind of joining the initiative. Again, think it's bigger than any one of us. It really is an industry wide cross sector initiative, and I really hope that we'll have lots of different players who are involved in contributing over time.
00:35:35 · Lori Goler, Stephanie Palazzolo
[00:35:35] S6: Yeah, I hope so. I hope that this will really be an effort. Since the announcement, actually, we've had great interest from a lot of players who've reached out to learn more and to find out what it's really about and to consider kind of joining the initiative. Again, think it's bigger than any one of us. It really is an industry wide cross sector initiative, and I really hope that we'll have lots of different players who are involved in contributing over time. [00:35:59] S0: Why does right now feel like the right time to be doing this sort of thing? Right? You know, I've had people tell me that one reason why AI is really good at doing science is because, you know, the average scientist, maybe you're an expert in bio, but maybe you're not an expert in physics or all these other areas. And AI models are really good at connecting these disparate dots across different fields. That's like one reason I've heard. Like, why do you think LLMs, today's frontier AI models, are making it so that now is the right time to go after this?