How Is AI Catalyzing Life Science?

**Eric Topol** (0:01)
Patrick Hsu is the first person I've invited back for Ground Truths.
We had our first conversation back in October 2024, and he is my go-to person for digital biology, which is such an exciting field. He's a co-founder of the Arc Institute, and he's also on faculty at Stanford. He is really a leader in this space, so I value his perspective immensely. There's been a lot going on in recent weeks and days on life science, AI, and biomedical models. That's what we're going to be talking about today.
What I also want to make sure you know is that you can send in questions through the message at any time, and I'll try to get to them with Patrice and make it as much interactive as we can. We're going to run through a bunch of new things here. One in particular that's just published in Science Magazine yesterday, a model called Biomni, which is quite extraordinary.

**Patrick Hsu** (1:17)
Hello.

**Eric Topol** (1:17)
Hello.

**Patrick Hsu** (1:19)
Can you hear me?

**Eric Topol** (1:20)
Yes.

**Patrick Hsu** (1:20)
Okay.
I'm so sorry about this, everyone. Thank you for your patience.

**Eric Topol** (1:25)
Yeah. Do you know what happened?

**Patrick Hsu** (1:28)
Well, my alarm for 11.15 went off. It somehow shut off the audio and then I couldn't hear you anymore.
We'll give SubSac some feedback.

**Eric Topol** (1:40)
You know what's so crazy here, Patrick, is these stupid little things. Here we're having this accelerated biomedical discovery at warp speed. I mean, geez. Anyway, we're going to get started here.
I did introduce you the first time we've had one of our guests back on Ground Truths. It's really exciting for me because this is such a hot space and you are my go-to digital biology force and what you're doing at Arc Institute has really been so impressive. So back in March in Nature Biotech, Patrick, one of his colleagues, Brian Plotsky and many others and I published a review on this language of life story. You'd think in March, it wouldn't be that old, but it's like, collected dust since then because that was about all the fact that we're learning about the complexity, deconvoluting the language of life. But what's happened since then is just, I've never seen anything like it.
Back in May, there were two papers published side by side. One was by Google called Co-Scientists, and the other was by this non-profit in San Francisco called Full House, and its model was Robin. They were basically discovering repurposing drugs. For Robin, it was about macular degeneration, and for the Co-Scientists, Google, it was for leukemia and liver fibrosis and antimicrobial resistance. These were very impressive papers that were starting to show us that you could take raw experimental data and get an agentic AI to really make that move forward.
What did you think of those two papers when they came out? You know, just weeks ago.

**Patrick Hsu** (3:40)
I think they're very exciting applications of models that can reason over the entire scientific literature, right? I think the idea that the human context window is limited and we can only read so many papers and think about so many ideas at a given time, but an AI that is able to create hundreds or thousands of its instances and become an expert in different scientific fields and then coordinate discussions across all those different sub-fields is a really exciting one. So you could have someone who's an expert in drug target ID, someone who's an expert in genetics, another who's an expert in hepatocyte biology, and then you can essentially simulate in silico conversations and scientific debates and hypothesis generation across all these different fine-tuned AI models, right? And so one of the things that these people could do is design new scientific projects.
They could be a data scientist and analyze data. They can make figures the way that you do in biorender or RISM or Excel or RR today.
And the question is really, how do you orchestrate all of these into a workflow to actually do something useful?
And so I think these, the sort of the initial work from Google and Future House, you know, has really focused around the idea of, can we do a proof of concept around end-to-end target discovery and then discovering or repurposing a molecule that already exists, maybe has even been tested in humans, that will work for this net new target. And I think the core idea is there's just a lot of disparate knowledge that you have to figure out how to put together in the right ways. And I think a lot of the questions about AI models making mistakes or hallucinating things are all very real today and the models forget things, they hallucinate values when you're trying to have them do data science, they'll plot the wrong thing or something that didn't actually exist. And, you know, I think the sort of long AGI bet is that all of these things get better over time, right? If you think about where the models were a year ago and where they are today, if we can project similar improvements over the next 12, 18 months, right, hopefully they'll continue to become more accurate and more interesting. Whereas I think today top scientists would tell you these are interesting, but don't really satisfy my taste function, right? If you ask a domain expert, right, did this specific drug target for macular degeneration, is it going to cure the disease? Maybe, maybe not. But I think the lucky thing is, in our domain of biology, there is a ground truth, right? You actually have to test it in real cells, in tissues, in animals, in patients, and actually make sure. And that's sort of, I think, the saving grace of this field is we can actually make sure that they actually work. So, I think very exciting.