Short Wave - How AI Is Speeding Up Scientific Discoveries

Episode Date: October 16, 2023

Artificial intelligence can code computer programs, draw pictures and even take notes for doctors. Now, researchers are excited about the possibility that AI speeds up the scientific process — from quicker drug design to someday developing new hypotheses. Science correspondent Geoff Brumfiel talks about his visit to one protein lab already seeing promising results. Have an AI query? Send us your questions to [e-mail] pcm.adswizz.com for information about our collection and use of personal data for sponsorship and to manage your podcast sponsorship preferences.NPR Privacy Policy

You're listening to Shortwave from NPR. Hey, short wavers, Regina Barbara here. And I'm science correspondent Jeff Brumfield. And today, Jeff, you're back with us to talk about artificial intelligence, which is all over the news. But out in a time of artificial intelligence. And the world of artificial intelligence remains a key issue for Hollywood actors on. Bring artificial intelligence into our daily routine. But out in the real world? Not so much.

I mean, I think that there is AI being used in my kids. Google searches, maybe, maybe helping me with my playlist, but I'm not really using it at work. Yeah, I mean, I've played around with it for fun, but for the most part, that's been about it. But recently, I visited a scientific research laboratory where they're using AI every day, and it's fundamentally changing how they do their jobs. I think it's just a revolutionary that we can come up with a therapeutic in a couple of months now. That's Susanna Vasquez-Torres. She's a graduate student at the University of Water.

And she's on the front end of what could be a big change in science. A lot of researchers are looking at AI right now to solve problems in their fields. And while it might not work for everyone, they're promising science that could help speed discoveries. So today on the show, we look at how artificial intelligence might change science. And what it says about how AI might change other jobs, maybe even ours, Gina. Oh, no. You're listening to Shortwave from NPR. Okay, Jeff, so let's just get started.

At University of Washington, you visited a lab there where AI is already being used. Tell me what you saw. Yeah, I went to a lab where researchers are designing proteins. Okay, so proteins are basically the molecules that do everything in biology, right? The gist I get from all the times it's been explained to me is that they build muscles and organs, they digest food, they fight off viruses. So if DNA is the blueprint, then proteins are the construction workers and the bricks that build us. That's it. And researchers in this laboratory are trying to design new ones from scratch that can help people. Susanna, that graduate student, is trying to work on neglected diseases, you know, diseases that basically have very little funding or research.

And she's particularly interested in snake bites. Snake and benoming still consider an unattended tropical disease. and the current therapeutics are not safe and are very expensive. She told me the right proteins could neutralize different kinds of snake venom. They could be sort of a universal protection against snake bites. Okay, so let's get into this. This is fascinating because I'm sure there are folks already using computers to design their proteins. So how is artificial intelligence helping that specifically?

Yeah, yeah. I mean, computers and protein design are nothing new. So let me walk you through how this all works. proteins are made of simple chemicals called amino acids. These amino acids can be combined in a nearly infinite number of ways to make a nearly infinite number of proteins. In the past, researchers had to systematically test often more than 100,000 possible designs to try and find the right one for a particular job. David Baker is the senior researcher at this lab, and he says that work was really tedious.

So it's kind of like having a whole bunch of sort of predefined sort of keys and just trying them out one at a time to see what fits the best. So rather than dealing with 100,000 keys, AI has really changed the game completely. Rather than having to make a whole bunch of different possible structures on the computer, then try them one by one. We could build something that just fits perfectly from scratch. In Susanna's case, her snake bite project found a few candidates within about six months. And that would have likely taken years before they started using A. at the slab. Wow, that's amazing. Yeah. Now, the particular type of AI being used is known as diffusion modeling. So I don't know if you're

familiar with the popular sort of AI image generators like Mid Journey or Dolly. Dolly, like Wally. I've actually not heard about this. Please tell me more. Well, so Dolly is an image generator from the company that makes ChatGPT. And basically, the way these image generators work is they start with a bunch of random pixels. It's almost almost like a field of white noise, and then slowly tweak each pixel until the image generators made the image that the user asked for. So that might be a picture of like, I don't know, a daisy or pretty much anything you want. Now, in the case of the AI being used in David's lab, the goal is to design a protein with a specific shape. So the image generator starts with kind of a random

assortment of amino acids and kind of builds it. It's almost like building a three-dimensional puzzle. It finds all the pieces. It fits them all together until it's built this perfectly shaped protein. Here you see proteins. We just did this diffusion process, and you can see how they fit perfectly against this structure. The shape of the protein often determines how well it will work. So this kind of AI is actually really well suited for the job, David says. The AI also requires examples to learn from. And luckily, there's this massive database full of proteins. that it can study. It's really a unique situation. Our ability to design new proteins using deep learning rest entirely on the work of 40 or 50 years of graduates and postdocs and scientists, structural

biologists, and it was probably many billions of dollars of investment. And then also the people who collated the database. There really aren't many places in science where you have databases like that. Yeah, it makes me think of the protein folding, like community science project as well, right? Like everyone's been involved in this. There's a long history of computation and protein design, and I think that's why there's so much data available, yeah. Yeah, it sounds like protein design is kind of a unique field because also it's so visual and because it has all this data. Like, can AI help other fields of science? I think that's a really interesting question.

And I, you know, I'm not sure the answer is always so clear. So I spoke to Maria Chan. She's at Argonne National Laboratory in Illinois. and she's working on developing new materials for the renewable economy, things like batteries and solar panels. And she says unlike the field of proteins, there just isn't that much research on the sorts of materials that she studies. There hasn't been enough sort of measurements or calculations,

and also that data is not organized in a way that everybody can use. But Maria is still pretty optimistic about AI, and part of that is just because the old way of trying to do, discover like new battery materials was pretty old-fashioned. The previous, you know, 100-year of science really has to do with a lot of serendipity and a lot of trial and error. So in all these other fields of science, you know, genetics, climate studies, particle physics, pretty much anything you can think of.

Researchers are really racing right now to figure out whether AI is a good fit for them. And I'm not sure it'll work for everyone, but it's bound to work for some. And it may work kind of, you know, at different levels for different fields as well. Okay, Jeff, so that's what's happening with AI in science right now. But we all know this field is developing really rapidly. Where do you think it's going to be heading? Yeah. So a lot of what's happening right now is sort of trying to build these models or make predictions from these large data sets.

But I spoke to some researchers who think that AI may have a more profound effect on science going forward. So one of them is Hanahajee-Sherzzi. She's a researcher at the Allen Institute for Artificial Intelligence, which is also in Seattle. And she wants to develop new AI systems similar to chat, GPT, but for science. And the idea is that they would ingest all the scientific literature in a certain field, maybe even across a couple of fields. And then this AI would develop its own new hypotheses. Wow.

Okay. Let me give you a simple example. What are different relationships between drugs and diseases? Right? As a human scientist, I might go and list all the drugs, all these papers, connect them, see what type of potential effects I have, or I can let my AI system do that, right? So I would argue that at some point, AI would be a really good tool for us to do new scientific discovery. Okay, so AI is making up hypotheses. Yeah, yeah. And you could imagine it could work. There's a lot of scientific literature out there. If it can ingest everything, it might be able to find some new

connection. Getting even more out there, I spoke to another researcher, Yolanda Gill at the University of Southern California, and she wants to develop AI that can do science end to end. Wow. We are working on developing AI scientists that approach questions and hypotheses the way that scientists do. So what Yolanda is imagining are automated systems that can plan and carry out experiments by themselves. She doesn't think a chat GPT type AI could do this, but with some development of new AI technology and using the language models together, there might be a way that they could get there. Okay. I think some of us who struggled doing experiments, this might sound great, but for others who really love like the process of doing the experiments and learning from failure, this might not be

so good. Yeah. And I mean, I think this gets at the bigger question, which is whether AI could someday replace scientists and just be generating new science on its own. And I don't know if that's where this is ultimately heading. It's really dependent on the existing data. Like proteins, big database does great. But in a lot of fields, that may not be the case. But nonetheless, I think there are applications. So, you know, one thing that Yolanda Gill talked a lot about was there are things that need to happen in science right now that just don't. Okay. So you can imagine there's a a paper about Alzheimer's. And maybe there's new data that comes in, but the researcher doesn't want to rewrite the paper, doesn't have time. Well, you could imagine AI could take that data

and actually update the result, update the finding. The amount of questions that we have in science and the amount of updates that we need to do as more data and new methods become available, the amount of work is infinite. There's not enough humans to go around to do all this work. And, you know, this is something I keep coming up against when talking about AI and the the knowledge economy. Often there's jobs that we would like to have done, and AI might be able to do those jobs for us. But that doesn't mean the highly skilled workers that are doing work right now won't have anything to do. It's just, it's going to cover new ground, like updating a research paper from 2012 or gathering a bunch of environmental data and forming a hypothesis in real time

when the researcher has to be doing something else. I don't think that means necessarily the work won't be lost to AI overall. But I do think that there are some hints when looking at science that show AI could allow some things to happen that we just can't do because there aren't enough people working on these problems. We're ending on a slightly light note. Jeff, thank you so much for bringing us this AI story. Thank you. Before we head out, a quick shout out to our Shortwave Plus listeners. We appreciate you and we thank you for being a subscriber. Shortwave Plus helps support our show. And if you're a regular listener, we'd love for you to join so you can enjoy the show without sponsor interruptions. Find out more at plus.mpr.org slash shortwave.

This episode was produced by Burley McCoy and edited by our managing producer Rebecca Ramirez. Jeff checked the facts and Maggie Luthor was the audio engineer. Beth Donovan is our senior director and Anya Grunman is our senior vice president of programming. I'm Regina Barber. Thanks for listening to Shortwave from NPR. Thank you.