MALCOLM BURNLEY, HOST: You’ve probably heard about COVID-19 spike proteins, like this explanation on CNN.
CNN CLIP: Why the spike protein? Because it’s the key the virus uses to enter the human cell.
MB: Spike proteins look like little pieces of popcorn stuck to the outer edges of the coronavirus. They’re usually colored red in the diagrams, and rightfully so, because they’re what made COVID-19 so transmittable and deadly. They’re also what the vaccines are trying to stop.
CNN CLIP: But if you create antibodies to the spike protein, it’s then blocked.
MB: Shape is everything in the world of proteins, which are like tiny 3-D puzzle pieces. For example, if you want to stop an RNA-based virus like Ebola, you need to know the shapes of its harmful proteins so you can block them.
There are a couple reasons why protein shapes are so hard to figure out. One is that they’re tiny. Their grooves and ridges consist of atoms, meaning their structure can vary by one millionth of an inch or less. But on top of that, they twist into compact 3D shapes. They fold up like origami, which makes proteins all the harder to decipher.
The tools scientists have used to figure out protein shape — things like X-ray crystallography and electron microscopy — are complicated, time consuming, and often hit or miss. Which is why scientists started asking: Is there another way to solve this problem? What if we could do this way faster — with the help of AI?
From WHYY in Philadelphia and Princeton University, this is A.I. Nation, a podcast exploring how humans are handing over more and more control to machines and what that means for us. I’m Malcolm Burnley. My co-host is Ed Felten.
ED FELTEN, HOST: Yeah. Hey, Malcolm.
MB: Ed’s a computer science professor at Princeton University. He also advised the Obama administration on AI. Ed, I know age is a major force in health care. Did we see it use much in the pandemic?
EF: Absolutely we did. It was used for all kinds of things. It played a role in many of the medical studies that helped us understand what was going on with the pandemic, with developing drugs and vaccines, and helping hospitals and others deal with limited resources and changing schedules, managing patients, telemedicine. Really AI went through a lot of what happened.
MB: Today, we’re going to take a look at how AI has improved our response to the pandemic, and how it could supercharge that response in the future. We’ll also take a look at the sacrifices we might have to make to let AI into our lives like that.
EF: AI is built on data and our health data is some of the most sensitive information we have.
MB: But first, let’s get back to those protein shapes. Folding is how scientists describe protein shape. And within our bodies, the way a protein is shaped — folded, or misfolded — can determine if we are immune to a certain disease — or prone to it.
ANDREI LUPAS: In fact, most degenerative diseases in humans like Parkinson’s disease, Alzheimer’s, cystic fibrosis, they’re all protein misfolding diseases.
MB: Andrei Lupas is a molecular biologist who studied at Princeton and is now the director of the Max Planck Institute in Germany. Many scientists have spent their entire careers studying the structure of a single protein. Andrei has been looking at one of them for more than 10 years.
AL: You know how asteroids are called NK7-7-5-0? Our protein is called AF-15-03.
MB: The COVID vaccines were developed in record time because we weren’t starting from scratch with the spike protein. Researchers already had a good indication of its shape based on previous research into SARS. But we might not be so lucky, getting the same head start in the next pandemic. Scientists have been looking for a shortcut to determine protein shapes, one that will let us better treat disease and respond to pandemics, for decades.
That’s why John Moult founded CASP.
JOHN MOULT: I’m John Moult, I’m a professor at the University of Maryland and I am co-founder of an organization called CASP, Critical Assessment of Structure Prediction.
CASP has been called the Oscars of the Protein Folding World.
[music, cameras clicking, voices saying “oh”]
MB: Since 1994, it’s been the international conference dedicated to this shortcut.
JM: So every two years we’ve been doing this. We’ve just completed the 14th round.
MB: CASP is based on something Christian Anfinsen, a Nobel laureate, said in 1972: that you didn’t need to look under a microscope to figure out the shape of a protein. In theory, all you needed was its genetic code, the amino acid sequence that’s embedded in the RNA.
[music ends, applause]
MB: Artificial intelligence should be able to use that sequence to predict a protein’s three-dimensional shape.
JM: You shouldn’t need all that expensive and risky experimental stuff. You should be able to directly compute the shape, the folded structure.
MB: In other words, you can unfold the origami, and by simply staring at the piece of paper, calculate exactly how it folds up. You never have to see the origami in the first place.
JM: The problem is that’s a very difficult computational problem. It’s been recognized for about 40 years. It’s been a grand challenge in computing, to try and solve this.
MB: CASP brings together leading minds who are independently searching for this shortcut, the answer to what’s been dubbed “the protein folding problem.”
CASP sends them a problem to solve, an amino acid sequence. This is a shape that’s already been figured out, but the results haven’t been published yet. The people who run CASP know the answer to the problem, but the participants don’t. So all of the teams run the amino acid sequence through their algorithms. They come up with a 3D model of what they think the protein looks like and they send it in. CASP compares the predictive models with the real shape. Then, everybody meets up at a big hotel and the judges announce who came the closest.
Andrei Lupus has been participating since the third round of CASP.
AL: From about fifteen years ago, I could see there was no progress being made. And so I stopped believing that I would see the solution to the protein folding in my lifetime.
MB: In 2018, things started to shift. There had been some massive leaps in machine learning, and it seemed like AI was starting to nip at the heels of the protein folding problem.
JM: So this was extremely exciting. And then, of course, you know, we had to wait two years to see what would happen. And what happened was that these same successful groups in the previous round did push things forward quite a lot.
MB: At the 2020 conference, a Google company called Deep Mind used a neural network-based AI called AlphaFold II. If CASP is the Oscars of the protein folding world, Deep Mind won best actor. What they were able to do with AlphaFold II blew the CASP judges away.
JM: They were suitably astonished. And then finally we had the virtual international meeting and it was very interesting to see, when I announced the results at the beginning of the meeting, the sort of explosion on the chat of people reacting to this.
MB: AlphaFold II could predict protein shapes pretty much as accurately as lab experiments.
JM: For most of the models. When you compare them with the experiment, you can’t tell when there are differences, whether it’s the experiment or it’s the model.
MB: Remember that protein shape that Andrei had been trying to figure out for a decade?
AL: [distant, echoing] AF-15-03.
MB: He gave AlphaFold II the amino acid sequence and it solved it in half an hour. In November, John and the AlphaFold II team made a stunning announcement: the protein folding problem was solved
JM: To see something as dramatic as this. Such an almost unimaginably accurate solution to this problem that’s been around for 50 years was just mind blowing.
MB: Which brings us back to that spike protein. Imagine if we’d had this AI power at the start of the COVID-19 pandemic?
JM: You can imagine in a future pandemic very quickly being able to design proteins that bind to the virus and knock it out as a drug. You can also imagine that these might be adapted to stimulate the immune system more to get a fast response in a vaccine-like situation.
MB: Our understanding of proteins is about to change forever. The AI revolution is allowing scientists to dream big in healthcare. It’s opened up a future where the next pandemic might not be that big of a deal.
Ed, what was your reaction to hearing about AlphaFold II?
EF: It was an impressive piece of work. This is one of these things where we expected that AI would do it eventually, but we didn’t know when. And it was a surprise that it happened as quickly as it did.
MB: I was really blown away by how far we’ve come and just the last five years. I know we can’t predict the future, but I’m curious — what do you think a pandemic will look like, say, a hundred years from now?
EF: I don’t mind speculating as long as, you know, we’re clear that it’s rank speculation. And that’s fine, right? We’re sort of in science fiction territory here if we’re trying to go a hundred years.
I mean, a hundred years from now, certainly, I think we could respond to and move past a pandemic in a matter of months, and possibly even in weeks.
MB: Weeks?! This thing has been with us for more than a year!
But that would involve the government having some pretty advanced capabilities.
EF: The government would be collecting data about who’s going to the doctor, who’s going to emergency rooms, how many people are going there, what symptoms they’re having, what the diagnosis is. They may be getting data from health tracking apps that people have or different appliances, your smartwatch or your smart earrings or whatever it might be that collects all kinds of information. Sniffing DNA that comes out in your breath. And it’s sequencing that DNA and categorizing it. So you can imagine flows of data that go to the government or to your local health department or something, which they’re constantly analyzing.
MB: The future Ed’s imagining assumes a government, however benevolent, that’s harvesting a lot of our data. It’d be keeping track of hospital admission rates, drugstore purchases and Internet searches — all to detect some blip that could mean a new pandemic is brewing.
EF: You know, a benign government is trying to figure out what is the minimum we can do to prevent this pandemic in a way that minimizes the impact on people’s lives and that relies as much as possible on voluntary measures. But we’ll do mandatory restrictions if we need to.
MB: With all that data, you can make your restrictions really targeted. The government could say some people need to be stuck inside quarantining, but others are free to go about their business.
EF: But we need to have a conversation about how to think about this because even if there’s a very good reason to single out some individual or to single out some group for restriction, that causes a lot of pain, and it falls disproportionately, arguably unfairly, on that group. And so we need to talk about, What are we going to do to make sure that the people who are disadvantaged in this way are, first of all, are only disadvantaged if it’s really necessary. And second, that their interests are protected and taken care of in the right way.
MB: How we’ll use AI in a future pandemic is not just a technological question. It’s a question of values and priorities. Ed thinks we’ll be more willing to make those sacrifices and work through problems like this because, unfortunately, this could happen again.
EF: You know, a hundred years in the future, we will have seen multiple pandemics, you presume, not just COVID, but more. So this, I think, will be something that will be seen as really serious, and an important part of the operation of of a well-functioning society at that time.
MB: While you might think this future sounds far off — a hundred years down the road, the truth is — some of it’s not. We don’t have smart earrings, but some of what Ed laid out is already here in some places. When we come back, we’ll hear about one of those places that used AI and big data to virtually squash the COVID-19 pandemic. I’m Malcolm Burnley.
EF: And I’m Ed Felten.
MB: This is A.I. Nation.
MALCOLM BUNRLEY, HOST: Welcome back to A.I. Nation. I’m Malcolm Burnley.
ED FELTEN, HOST: And I’m Ed Felten.
MB: South Korea is a three hour flight from Wuhan, China. But the country was very well prepared for an outbreak. SARS swept through the country in the early 2000s. Bird flu before that. And in 2015, MERS killed dozens of people. So the COVID-19 pandemic wasn’t the first time South Korea went through something like this.
And in the spring and summer of 2020, it showed. While Americans were hoarding toilet paper, people in South Korea, like John Lim, were allowed to go out clubbing.
JOHN LIM: The weekend before my birthday, I went out with a couple of friends in Itaewon. We went to a restaurant first. A bar, another bar, and another bar, another bar. We were just bar-hopping.
MB: The US and South Korea reported their first COVID-19 cases on the same day, January 20. The US was more than a month away from imposing any restrictions. But the South Korean government sprung into action with big data and AI.
I spoke to Lee Rainee about this. He’s the director of internet and technology research at the Pew Research Center.
LEE RAINEE: Well, they use contact tracing apps in smartphones to watch where people were going. They also combined it with data from their credit card use and ATM machines they were using and things like that. So they had a really robust data collection process.
MB: A Korean supercomputer fast-tracked the development of COVID test kits. Algorithms developed by the Korean CDC classified and sorted positive cases, so people were sent to different hospitals based on their level of need. There were also AI medical tools, like a machine that diagnoses X-rayed lungs in less than 10 seconds, which can take doctors an hour or more.
But contact tracing was the most visible way AI and big data made their way into South Korea’s pandemic response.
LR: There were ways in which they could sort through databases and see who had been in close proximity to people who were known to have the coronavirus, or were known to have been in contact with others or in other places where the coronavirus had popped up.
And so there was a massive sort of networking effect where a lot of the population was very much included in all these databases, and willing to be sharing their information, and willing to listen to public health officials once they were contacted that they might have something to worry about.
MB: Before the pandemic, South Korea already had a law in place, the Infectious Disease Control and Prevention Act, that allowed the government to access incredible amounts of private data during a crisis.
When someone tested positive for COVID, this system allowed public health officials within minutes to pull up that person’s recent credit card transactions, their cell phone location data, and even closed-circuit camera footage of where they may have travelled.
This is a pretty intense response, but what it meant is that South Korea never had to shut down its economy. Business people, including foreigners like John Lim, were able to keep on doing business in the country. Along with other things.
JL: And then we went to a bar and another bar and another bar. We were just bar hopping.
MB: John is a Korean-American who runs a pet transportation company. At the start of 2020, he was able to make four round trips between New Jersey and South Korea.
He saw up close the difference in how serious the countries were treating this.
JL: The US has done nothing. Since last year, even during the pandemic started, everything is closed except Home Depot and Lowe’s. What was that all about? Why do you need to go to Home Depot and Lowe’s? People were just hanging out there.
MB: John was frustrated with the US response. Compared to Korea, it was like night and day. Over there, he was bombarded with updates about what the government was doing. And, he was updating them.
JL: In Korea, they announced to everyone that due to the national pandemic, all your data for the next 30 days will be saved and shared with the national health department.
MB: The South Korean government already stores mass troves of data about its citizens’ activities. To stop fraud and other crimes, the government receives data from cell phone carriers and banks.
But during the pandemic, it monitored locations in real time. Anyone picked up through contact tracing or who’d travelled internationally, had to go one step further.
IRENE YOON: As soon as I arrived at the airport, they had stations set up. It was like Koreans vs foreigners.
MB: Irene Yoon is an English teacher in South Korea, who returned to the country in April.
IY: And they have you download this app on your phone, which you have to allow permission to always know your location basically 24/7.
MB: She did this before a mandatory two-week quarantine at a place approved by the government.
IY: I have a valid long term visa and because of that I had the option of staying at Airbnb or if I had a house here, I could stay at my house or I think even a relative’s house.
MB: Irene chose the AirBnB. While she was there, she had to update her temperature on the app twice a day and write down any symptoms she was feeling. She also communicated with a public health worker assigned to her case.
Violating quarantine had consequences.
JL: Several people got caught by turning their phone off or leaving their phone at home and then and walking outside.
MB: These kinds of stories were all over John Lim’s facebook feed.
JL: Normally the process was, you do it once, they’ll give you a fair warning. You do it twice or three times, you’ll get fined. 10,000 dollars, I think it is?
MB: If your phone left the quarantine location, you’d immediately get a call. But you’d also get contacted if your phone wasn’t moving at all. That happened to someone staying down the hall from Irene at her AirBnB.
IY: And I woke up one morning because someone was banging on their door, like really loudly, for like over ten minutes. And I heard him get out his phone and call someone and say, this guy’s phone is turned off and they’re not answering the door. And I guess the person on the other line said, like, call the police.
MB: The app also allowed the government to send location-specific text alerts, similar to Amber Alerts, with information about every new positive case in an area. The alerts did not include the name of the person who had the virus, but where they had been in the days before they tested positive.
John found out from these alerts that he might have crossed paths with someone infected, that night he went bar-hopping.
JL: They said, if you were on these bars or these restaurants on this day, please get tested. So I went and got tested.
MB: They found him through a credit card receipt.
JL: But then they called me about 4 days later and said you need to come in and get tested again. And stay home.
MB: I expected Irene and John, two Americans, to say the restrictions were at least a little bit overboard. But they didn’t. They thought the government was pretty transparent, and, besides, what’s digital privacy anyway these days?
JL: We already give away so much of our privacy through Facebook, and Instagram, and our location and all of that. So yeah, I didn’t mind.
MB: They saw what the U.S. didn’t do — and preferred the alternative.
LR: There are a bunch of societies that are a lot less focused on privacy and a lot more willing to have surveillance and data capture systems applied to their lives, partly because it makes their life more convenient and partly because they think that this is a way to be a good citizen.
MB: That’s Lee Rainee of the Pew Research Center again. He says in the US, there’s much more of a ‘not in my backyard’ attitude. when it comes to government and data.
LR: Contact tracing is a perfect example of how complicated this environment is.
MB: According to a Pew poll conducted last year, six out of 10 Americans believe the government tracking people’s locations through their cellphones wouldn’t even make a difference in flattening the curve.
LR: And Americans were deeply skeptical, both at the level of thinking that a system like that would work and be an effective tool in fighting the coronavirus, but also in just being comfortable with the whole idea that smartphones would become sort of a surveillance tool for the folks who are trying to track the pandemic and its spread.
MB: There was some significant variance among different demographics, including young people being more willing to hand over their data than older generations. But the thing that’s true across the board with Americans, Lee says, is that we’re more willing to support the use of AI when it’s out of sight, out of mind.
LR: The distinction that we see, and it shows up in a lot of other ways that we ask about privacy issues, is at the population level, Americans are more or less comfortable with data being collected and anonymized and used to understand big phenomenon — whether schools are working or not, how fast a pandemic is spreading or not. Once you get down to the idea of specifically enforcing specific things about specific people, that’s when Americans really part company with it
EF: There’s kind of an irony here that sometimes it feels like we don’t object to being classified, categorized, and surveilled. We just object to finding out about it.
MB: That’s really interesting to me because I think for a lot of people, that’s how government functions in their lives, or that’s how they feel about government in general. If they don’t know it’s working, then they’re OK with it. But if it’s suddenly knocking at their door, being intrusive, or even just asking someone to do something, people suddenly don’t like the government then.
EF: Yeah, and here it’s really about data collection because, you know, in order to be able to detect a pandemic, a possible pandemic really early, the government needs to know what’s happening. It needs to be able to separate that very faint signal from all the noise of people showing up in emergency rooms with weird symptoms that happens all the time. And that means they have to have eyes and ears out. The question of how much of that we are willing to accept in exchange for maybe the next pandemic being noticed quicker, getting over more quickly. That’s something we have to decide.
MB: It’s also a choice of priorities, I think, where in a vacuum people may not want to give up their data for a government pandemic response. But if you told them it will save a thousand lives, they’d probably be more willing to give that up.
EF: They probably would, but it’s hard to tell in practice, right? It’s hard to tell. We don’t always know how much data is being collected or what’s being done with it. And of course, we don’t really know what would have happened if government had collected less data. The better a system like this is at preventing pandemics, the more it seems like we don’t need it.
MB: Right. We never know what the alternative could have been.
EF: Sure. Right. So we don’t know how many lives, say, airplane safety inspections are saving if airplanes don’t crash. And we might say, well, you know, I didn’t get wet, so therefore I didn’t really need an umbrella. But of course, that’s not how it works.
MB: I do want to spend a minute on some other ways AI was used in the pandemic because there are cases where I being a part of health care isn’t a big, fraught, moral issue. And when it comes to the pandemic, we could all use more straightforward good news. For that, I called up Tayab Waseem.
TW: I am the chief scientific officer at Stability AI.
MB: Tayab is a computational biologist with a PhD in immunology, which means that he lives at this intersection of medicine and AI, and he and his colleagues used AI to do some really interesting problem solving at the beginning of the pandemic.
Think back to March 2020. There was an overwhelming amount of information out there about the virus: theories, studies, statistics. I know I felt lost. You probably felt it. But imagine being a frontline health care worker at that moment.
TW: At one point, there were thousands of papers being published a week. And if you’re a clinician seeing patients in the hospital, you know, it’s your best guess on whether or not you’re up to date with the latest literature. It’s just coming out way too fast.
MB: Medical journals tried to help. They took down their paywalls for COVID-related papers. So there was a lot of information on how to fight the virus. But who had time to sift through it?
Tayab and his colleagues went to work building a platform that could help process this dense universe of medical text. The goal was to aggregate all these papers, then, using an algorithm, have the platform highlight for doctors what’s most relevant and what’s redundant, like having an AI research assistant.
TW: And the papers that you actually want to read, you can, you know, maybe sort them by type of study, or how many patients actually had enrolled in it. You can click those and and delve in further to to learn some of the nuances behind how they came up with their decision.
MB: Tayab and his team called their platform Collective and Augmented Intelligence Against COVID-19, or CAIAC. When the whole world was talking about the importance of masks last year, the platform showed how little research was out there on the topic.
TW: Everywhere you turn to its, mask this, mask that. It’s all over the news. It’s everywhere. When we actually did the literature search on it, we found there is actually almost no primary literature on it.
MB: It didn’t mean the advice about masks was wrong. Rather, it just required more attention.
TW: And I think another important part of what we were doing was, because you’re combing through all the literature on a specific topic, you end up finding gaps in the literature. So we have thousands of scientists, hundreds of thousands of scientists, across the world, doing research, and most of them are doing the exact same experiments. There’s no real coordinated effort.
MB: Organizations like the U.N. and World Bank have taken an interest in Tayab’s platform, which is still being improved. This isn’t the same kind of revolution as AlphaFold, but it has the power to help us with information so medical professionals can make better choices.
EF: It’s not going to make us better, more moral people, or make us better at agreeing with each other about what we should do. But it can certainly give us better data ,so that as we have the difficult conversations about what we should do, we understand what we’re trading off and we understand really what our options are and have a better understanding of what happens if we choose A versus B.
We’re not going to get to the point where scientists can just all retire and AI is going to decide or figure out everything. But the idea of AI as being an indispensable part of the toolkit of science, I think, is definitely something that’s out there and growing.
MB: Slowly, we’re getting closer to that future when maybe a pandemic will be a lot less painful.
EF: A lot of things we’ll be able to do more efficiently, more comfortably in the future, because AI is helping in little ways. It’s built into a lot of the things that we use. And so one of the big challenges with AI is how do we make it a tool, and how do we harness it so that it’s not an exotic technology that we turn to sort of in a lab, but it’s just helping to make a lot of everyday things better.
MB: On our next episode, when AI puts you behind bars, has it gone too far? We dive into the world of predictive policing.
A.I. Nation is produced by me, Malcolm Burnley, and Alex Stern. My co-host is Ed Felten. We have editing help from Katie Colaneri and John Sheehan. You can find A.I. Nation wherever you get your podcasts — and please don’t forget to subscribe, and leave us a review. This podcast is a production of WHYY and Princeton University.