Our Walter Bradley Center director Robert J. Marks is back with the second instalment of 2020 smash hits in AI—and now, the #1 Smash Hit! Readers may recall that we offered a fun series during the holidays about the oopses and ums and ers in the discipline (typically hyped by uncritical sources). This time, Dr. Marks talks with Jonathan Bartlett and Eric Holloway about AI programs that can beat humans at understanding the complexities of protein folding, the immensely complex ways that proteins in our bodies actually work.
Robert J. Marks: AI has cracked a problem that stumped biologists for 50 years. And it’s a huge deal. Jonathan and Eric, elaborate on this a little bit.
Jonathan Bartlett: Well, protein folding has been a tough problem for biology for a long time, just because there is all the interactions. It is hard to predict exactly how a protein is going to actually fold. So typically, or historically they’ve done it by x-ray crystallography where they basically shoot x-rays at proteins and watch them bounce off. And then guess what the protein looks like based off of these x-rays. But what they really always wanted to do is be able to guess what the structure is just from the sequence. So for those who don’t know, you have DNA and DNA is basically this long strand of what are called bases, which are basically the letters of DNA. And so those letters of DNA get translated into proteins, which again, there’s just a long strand of proteins and there’s again, basically letters of proteins that are just connected all along.
But unlike the DNA, the proteins actually do things. They connect with each other. They have interactions between the individual amino acids. The problem is, is that, there’s all sorts of interactions that might happen between these different amino acids that would cause the protein to go into different shapes. And so the question is, is which way will it actually fold? And being able to suss that out has been a difficult problem. And the biologists always want to just be able to see the sequence and infer what the final shape is going to be. So that’s what they’ve been… That’s been the problem and humans have generally been bad at coming up with rules for this. And so this is why they put it to AI is to try to get the AI to develop a system that can take a sequence and predict what the final structure is going to be.
Robert J. Marks: I think I learned this from you, Eric, was that there was a game called Foldit. Now, again, this is way before this artificial intelligence, but Foldit was a demonstration of human creativity. Could you walk us through that?
Eric Holloway: Yeah. The basic protein folding problem is what’s known as an NP hard problem. Just looking at the basic structure of the DNA, when you have a string of 10 DNA, you get two to 10 different possible ways that could fold. And when you have computers just trying to go through all their permutations, it takes way too long to do it by brute force.
So they found that complete amateurs, people who have no understanding of biology whatsoever, when they saw these folding algorithms doing their thing, they could easily spot optimizations that the algorithms were missing. And so the researchers just turned it into a game and they started making breakthroughs.
Robert J. Marks: Foldit used to ask you, “Hey, you’re not using your computer today at 2:00 AM to 5:00 AM. Let me use your computer.”
Eric Holloway: Yeah, that was the original approach. That’s where they got this insight from. First they just were trying to use people’s spare CPU’s. And they were very successful at that. They had thousands and thousands of people donating loads and loads of CPU time, but still even with all the free CPU time, they made very, very little progress. But part of the software was a screensaver that showed the computer owner what the software was doing with the protein folding. And that’s where people started contacting the researchers saying, “Hey, I could do a much better job than this algorithm is doing.” And so they just turned it into a game.
Robert J. Marks: Indeed. This is, I think, a very, very interesting insight into the Foldit program. This breakthrough was made by, I believe DeepMind and DeepMind is famous for reinforcement learning, which was used for example, to beat the world champion, Lee Sedol in the game of Go. And it’s a game. Protein folding is a game. And the game was proposed to users and the users before the AI were able to solve it much more quickly than the artificial intelligence.
But that being said, I do believe that AlphaFold does even better than the human players do.
Eric Holloway: Although that being said, I don’t know what exactly it means when it says AlphaFold has solved protein folding. Are they testing this on entirely new DNA sequences? And if so, how do they actually know it’s telling them the truth? I think they’re actually still using a known data set as a reference data set. And so they’re… Yeah. I don’t know if it’s actually completely solved protein folding.
Robert J. Marks: Biologists really are excited. One of them, Andrei Lupas said this is going to change medicine.
Eric Holloway: They try to hold Andrei Lupas up as an example of how incredible AlphaFold is. And they say, Oh, he spent a decade trying to figure out the shape of one protein and AlphaFold does it in half an hour. But why was AlphaFold able to do it in half an hour? Because it depends on decades and decades and decades of researchers just like Andrei Lupas trying to figure out the shape of one single protein. And so really Andrei Lupas, his research has accelerated because he has better access to all these other scientists research through AlphaFold.
Robert J. Marks: Let me ask you, Jonathan, Eric, what is going to be the impact if AI is able to solve the protein folding problem? Where is it going to be used? Will it affect me?
Jonathan Bartlett: Basically, what it’s going to allow you to do is model drugs before we actually test them. So for example, if some drug company has a drug that they want to put out, they’re going to be much more able to test its effects against various proteins, because it’s going to be able to have a model for them. So it can estimate what it thinks is going to happen and run a lot of those tests in silico rather than in life.
Eric Holloway: Although also there, I would be a bit cautious too, because this again is one of those deep learning models. And I’m not quite sure how they’re verifying the results. So you could have a lot of corner cases that deep mind is just totally off on with the folding. So I think it can probably accelerate them by showing them where to look. But I think they probably can’t completely replace real experiments, too.
We are all still learning.
Here are the AI 2020 Smash Hits for 2020 to date:
2 To win against AI poker programs, humans need to identify their blind spots. Each pro separately played 5,000 hands of poker against five copies of Pluribus and Pluribus won. Now this, in itself, was an astonishing result. Eric Holloway: I’d be curious if down the road with all these game-playing AIs, people start finding out these blind spots and figuring out how to control the game AIs.
3 AI Smash Hits 2020: AI can help paralyzed people move again. The human brain can interface directly with electronics. An “exoskeleton walking device could get many paralyzed people out of their wheelchairs.
4 AI Smash Hits 2020 AI helps detect dreaded White Eye disease. The first step in treatment is correct diagnosis. Baylor University profs developed an app that enables eyeball disease in small children to be detected easily.
5 AI 2020 Smash hit: Deepfakes—What they can and can’t do. Deepfakes? Our minds often actually fill in a lot of our background for us when we are not even aware of it. One way of thinking about deepfakes is that they are liked instant mashed potatoes. The water is removed and re-added later.
6 AI Smash Hit: AI defeats fighter pilot hands down. The future of warfare may involve more machine waste but less human carnage. Eric Holloway: It’s going to be a more a hybrid approach where you have the fighter pilot and then a bunch of robot wingman that he can control.
7 AI Smash Hit: Why AI can’t do your thinking for you. Robert J. Marks: you change a pixel or two in an image and the deep convolutional neural network is totally wrong. Eric Holloway: The machine’s confidence in its result is complete certainty and it’s absolutely certain about the wrong result.
8 AI 2020 Smash Hit: Big gains in practical self-driving cars. The people who have been pursuing Level Five self-driving are nowhere but Level Four is working well. Jonathan Bartlett: You can think of Level Four self-driving as an engineering project and Level Five as a philosophy project
9 AI Success: Smarter cars for non-millionaires If your car is a recent model, an affordable aftermarket kit might transform it into a much smarter car. One possible risk is that a hacker could take over your car but, no matter what we do with AI, we must deal with security issues.
10 Smash Hit: #10 AI Success!: Translation gets faster and better. Machine translation, properly used, can help us communicate better. What’s made AI tech translation work so well is not that it’s perfect, but we’re going to have a second pass.
- 00:31 | Introducing Jonathan Bartlett
- 00:40 | Introducing Dr. Eric Holloway
- 02:47 | #5: Deepfaking for Entertainment
- 10:12 | #3: Paralyzed Man Moves in Mind-Reading Exoskeleton
- 14:32 | #4: Deep Learning for leukocoria, or “white eye”
- 16:36 | #2: AI Beats Professionals in Six Player Poker
- 20:23 | #1: AI Cracks Protein Folding
- Jonathan Bartlett at Discovery.org
- Eric Holloway at Discovery.org
- #5: “Disney’s deepfakes are getting closer to a big-screen debut” (The VERGE)
- #4: “An App That Can Catch Early Signs of Eye Disease In A Flash” (NPR), “Eye-catching tech” (Waco Trib)
- #3: “Paralyzed Man Moves in Mind-Reading Exoskeleton” (BBC News)
- The Brain That Changes Itself by Norman Doidge, M.D.
- #2: “Carnegie Mellon and Facebook AI beats professionals in six-player poker” (Carnegie Mellon)
- #1: “Protein Folding: AI has cracked a problem that stumped biologists for 50 years. It’s a huge deal.” (VOX), “AlphaFold Scores Huge Breakthrough in Analyzing Causes of Disease” (Mind Matters News)