Here’s the link to the article: https://www.wired.com/story/how-the-ai-nobel-prizes-could-change-the-focus-of-research/ It’s well-written. However, the author doesn’t have the forty years of perspective that I do. I can do better; that’s today’s blog post. I am drawing on my teaching abilities, honed as a TA at Stanford, guest spots at Berkeley, seven years at U. of Delaware as an Assistant Professor, two years at FGCU, and both adjuncting and teaching dual enrollment at a place that used to be called Edison Community college that is now FSW.
My knowledge of this field includes a course I took at Stanford from Buzz Baldwin. Buzz was in his 60s at the time. He was the world’s acknowledged expert in the field of protein folding. It also includes a different course from another Stanford professor—a new one; I took it within weeks of his arrival. Michael Levitt, fresh from Oxford and the Weizmann. It was on Computational Biology; he was more qualified than anyone in the world to teach that course. It also includes conversations with my one of my housemate Britt Park, who entered grad school in the Stanford Department of Biochemistry in the same class as me.
Michael and Roger Kornberg were his co-advisors. Roger won the Nobel Prize for what he discovered about RNA transcription. I was on a first-name basis with both Roger and his dad, Arthur Kornberg, who won the Nobel Prize for the discovery and mechanism of DNA synthesis by DNA polymerase. And with Michael. My best friend and another of my housemates, Brad Cairns, was also a graduate student for Roger Kornberg.
I kept up on the literature throughout my scientist years, and in the latter part, when I was collaborating with computer scientists and chemical engineers, I was aware of all the failures and limitations of the artificial intelligence-based approaches to this problem. And yes, I’ve played around with Chat-GPT, just for fun, to be goofy.
Listening? You are going to learn the difference between weak AI, generative AI, and strong AI, but we need to start with protein folding, a problem for which weak AI (also called machine learning) made a major advance that won the Nobel Prize in Chemistry (two of the recipients doing the computer science; one doing the computational biochemistry). This is a real advance; the criticisms are garbage. These guys deserved their award.
To get there, you need to learn about proteins and protein folding. I am going to start where I would if teaching a high school course, but I’m going to proceed at the level an educated, non-scientist adult would understand.
Proteins are chains of amino acids. An amino acid is a carbon atom with an amino group (NH3+) and a carboxy group (COO-) attached. There is also a hydrogen atom, and the fourth position in an arrangement that can be written on a piece of paper as four things attached to a carbon in a cross, but is really a tetrahedron in 3-D space (Wikipedia if you’ve never seen a tetrahedron). The fourth position can be a lot of things—it is called the R group.
When a peptide bond is made, it is what is called a nucleophilic attack in organic chemistry terms. The amino group is positively charged. If it loses a hydrogen atom, it is neutral, but electrons on the nitrogen are negatively charged. Negative charge is attracted to positive charge. The carbon of the carboxy group can be electropositive. We are talking about quantum mechanical slight shifts in positions of electrons, not absolute movement of electrons and protons—how it is presented in badly taught high school classes where the teacher is just reading the textbook.
A general base (again, organic chemistry) can pull a hydrogen from an amino group. Attack of a N-group (an amino group that has lost a hydrogen) on a carboxy group (where the carbon is more electropositive than the oxygen; oxygen likes its electrons. I can teach any high school kid this at a deeper level, but this is fine for now.)
This can occur in a chamber with conditions that occurred in primitive Earth. It can occur in a peptide synthesizer—a machine that has existed since the 1970s. It occurs in your body, but the reactions are all catalyzed. The enzyme is the protein synthetase part of your ribosomes. Unlike most other enzymes in your body, this one is built from RNA, not protein. It is one of the only enzymes so critical to life that evolution never replaced it with a protein. There are proteins that stabilize it and help it, but the important part is made from RNA.
Catalysis is making a reaction go fast. Give it a billion years, and it will happen anyway because of thermodynamics. The energy of the products of the reaction is less than the reactants. The loss is heat.
But it won’t happen fast enough in your body. That’s kinetics. Things will sit around and do nothing for a billion years.
Your body makes proteins every day that are complex and interesting. There are 20 building block amino acids plus some specialized ones. There are other things attached like carbohydrates and lipids. There are disulfide cross-links in which a sulfur from a cysteine attached to another sulfur in a cysteine.
The linear structure of which amino acid is attached in order to which one is called the primary structure. You can think of it as NCC NCC NCC. The N is the amino group. The first C is the central carbon to which the hydrogen and the variable (R) group are also attached (in tetrahedral arrangement n 3-D space). The last C is the carboxy group. The far left is called the amino terminus. The far right is called the carboxy terminus.
Lots of stuff happens on both end—too much to get into here.
This is primary structure—we need to get to secondary, tertiary, and quaternary structure for you to understand today’s news. You may need to have a bathroom break or eat something right now.
Secondary structure is the very local ways in which the amino acid chains fold. “Fold” is the key word. I do an active learning exercise when I teach that I stole from Jim Spudich, a Stanford Cell Biology professor, that I learned in his class. I also rotated in his lab. We got along well, he’d have taken me as a grad student, but I was totally incompetent at things most entering graduate students would have found easy at a time when I could converse with him at a level that most beginning grad students could not.
Break for personal stuff, because after all this hardcore science, you deserve light reading.
It was MIT Biology that got me to this place with Jim Spudich. They purposely avoided most Chemistry classes because except for my initial faculty advisor’s; chemistry department teaching was awful. The worst teaching at MIT. Both Biology and Chemical Engineering prided themselves on teaching. This one guy taught everything a Biology Major or Chemical Engineering major needed to know about Physical Chemistry. My Stanford Ph.D. advisor was smart enough to not complain when I wanted to sit in on undergraduate Quantum Mechanics and undergraduate Statistical Mechanics and Kinetics at Stanford. Stanford’s Chemistry Department prided themselves on their teaching. My AP Chemistry classes have benefitted from what I learned at Stanford. With only Algebra II as background, the bare minimum of calculus ties everything together. The learning objective is understanding my direct lecture. If you can follow me, you’re okay. They all could—all smart cookies. It tied all of this together in a way a textbook just says is kind of magic that maybe you’ll understand someday. My approach works better.
As for MIT, when I had four hours of one-hour classes in a row, every day, ten am to two pm, a cute girl and I ditched organic chemistry to go for pizza just about every day. I had no idea that her father was a famous physicist and running my recitation section in physics just for an easy teaching load. French guy with a great Parisian accent. I got an A. He also probably didn’t know his daughter was ditching organic chemistry class. It didn’t matter, with the final being based on average and standard deviation, 40 +/- 50 points was a B. I got a B. Anyone who could understand these guys (they co-wrote the textbook) well enough to get 90 or above got an A. I think they may have called zero is a B ridiculous and given some C grades, violating MIT policy. Nobody failed. So, I got a B for hanging out with a cute girl and eating pizza. And her dad liked me, even though it took me months to make the connection.
And this was freshman year, where every course at MIT was pass/fail. You got real grades; you saw the real grades. A few very obstinate medical school programs wouldn’t admit you unless you showed your freshman grades. Nobody else saw unless you wanted them to.
I had my junior year also pass-fail because I was overseas. My GPA was 3.7/4.0. If you included either or both freshman or junior year, it was still 3.7.
I went to a very good public high school. Half Black, half white, no other real ethnic diversity. Students from dirt poor to phenomenally wealthy among both the Black and white population. I once was out once socially with a white guy whose mom was on welfare and he was bussed from the area of Cleveland that busses in and a Black guy whose Dad was an Undersecretary in the Carter administration. The rich people kept their kids in the public schools. They were better than the private schools. Very few students went to MIT, but there was a year 24 were accepted to Cornell. Number one in the class went to Harvard, as did number two and number four. (All of this is when it counted for numbers, after junior year). Number five turned down Harvard to go to Brown. I was number three. Far behind number two—I didn’t care. Number one is a tenured law school professor. Number two is a tenured economics professor. Number four (who I think became number three by the time of graduation—I didn’t care) is the head of Public Citizen. Number five is one of the country’s experts on doing business with China.
I knocked him out of what he likely assumed was a championship in debate at the Northeast Ohio level, where our school took the top three places to go further and he was third. I beat both of my team members: him with me taking the easy side; and the other person, with me taking the tough side. He lost taking the tough side to both of us. The third person who lost to me taking the tough side, beating her taking the easy side. Sorry, Jimmy. You got a raw deal.
Anyway, I came into MIT on easy street. I could write well. I could do math well. For me the first quarter was me finishing all work, ready to party on Wednesday night. For most people it was the hardest place they’d ever been. Only pass-fail grading (and making all three-day weekends into four-day weekends) prevented some suicides.
I was having the time of my life. MIT said it was a liberal arts school; I said you’re living up to that. My Middle East studies concentration, including becoming fluent in Hebrew and taking conversational Arabic, and learning a hell of a lot of politics, history and sociology—all in Jerusalem except for travel that included a month in Egypt—got me to the point where I could co-edit an anthology set in the Middle East.
My conclusion after a year where I could talk to everyone from the head of a Palestinian militant group by invitation in his office on the West Bank, and the synagogue of a leading Jewish settler ended with the conclusion that things were about to explode. The first intifada was three months later. Check out my story in Dark Matter Magazine which is really a comment on the role of racism in the conflict. The Jewish protagonist is a racist; no other character, Jewish or Palestinian, is. It is simultaneously hard SF on pharmaceutical side effects and very soft SF on quantum mechanics. Free click: https://darkmattermagazine.shop/blogs/issue-014/side-effects
I took the music course where the next stage was professional. It was a catch-all where I sang in a sight-singing chorus, went over every note of three pieces (Mozart and Beethoven’s were two), learned about where everyone in an orchestra sat, took piano lessons, etc. 20 hours per week with 40 hours of homework expected.
I took a rigorous course on impressionist and post-impressionist art. I took macro and micro economics. Only at MIT would freshman macroeconomics involve partial differential equations. I took a poetry reading and writing workshop. The readings were wonderful; my poetry was juvenilia.
My MIT degree was ¼ biology classes; ¼ other math/science courses; ¼ foreign languages; and ¼ other arts, humanities and social sciences.
Yep, MIT is a liberal arts school if you want it to be. One course was at Harvard: Evolution with Stephen Jay Gould and Richard Lewontin. Out of 98 Harvard students and 2 MIT students, when Gould told the TAs, he wanted to see the molecular biology, he chose one student’s papers to grade. Mine. The TAs graded the other 99.
This was the undergraduate education that I entered Stanford with: able to converse at a high level but unable to avoid screwing up a simple reagent preparation. Sorry for wasting money, Jim. My other rotation was with Paul Berg—another Nobel Laureate. I proved definitively that two of his postdocs were wasting time on a completely useless project. And spending ridiculous amounts of his money on it. I think that redeems me overall.
Rotations were picked out of a hat: I got Jim Spudich and Paul Berg. I wasn’t considering either of them. My second choice who would have taken me was Pat Brown. A Nobel Laureate who created Impossible Foods as the best thing he could do for the planet. I cooked Impossible Chili on Monday. It was delicious. Thanks, Pat. The project he proposed would have failed. It did for the grad student who took it (Ray). The viral genetics you proposed as a side project with a mouse virus might have worked and been important. I would have focused on that, probably, and had a nice Ph.D. I would have been your first grad student, but I wouldn’t have been the stars you had later.
Are you all ready to get back to proteins?
Secondary structure: The most brilliant teaching demonstration I ever saw, which I have stolen from Jim Spudich, goes this way. I take an object, usually a manilla folder, hand it to a student and say place it on the table. I then give them another, and I say place it on top with an angle offset anything other than 90 degrees. Literally, any other angle works. I then tell them to look at the angle between the manilla folders and put the next folder on top, the same angle offset. We continue with a few more folder. I say, “you just made a helix.”
This is why DNA and RNA are helices. It explains the collagen helix and every other helix in biology as pure geometry. For our purposes, it explains the protein fold called an alpha helix.
There are 90-degree angles in proteins. They are called beta sheets. It’s another type of fold.
At U. of Delaware, once, teaching, I saw the biggest smile I’d ever seen from a student. He was a truly top DuPont technician who DuPont was making part of his job to take my course. Industry is lousy at teaching, and they know it. Universities teach. DuPont paid top-dollar tuition for any technician who wanted to take my course to take it and include the driving time as part of their full-time job. The university wasn’t displeased with taking DuPont’s money.
But dual enrollment 17-year-olds have given me the same understanding smile.
In an alpha helix, hydrogen bonds line up precisely. Remember that hydrogen on the central carbon of an amino acid. It’s small. Remember those electrons on the carboxy oxygen that can do a nuclear attack. They are also tiny. Coulomb’s law says charge attractions are strongest at short distances. This is very short. It’s called a hydrogen bond.
Hydrogen bond energies can be measured using physical chemistry. One it weak. But with a billion, you just explained most of the energy putting the DNA double helix together.
This is modeling like a physicist. You will also hear about modeling like a chemical engineer and modeling like a computer scientist later.
There are lots of other interactions in an alpha helix. Most of them involve R groups. Some are electrical. Some involve types of chemistry. The point is that physicists have math that describes them all. You can make a total of all energy holding that protein together using just physics.
Michael Levitt was a physicist working in a cell biology department. He came to the conclusion that one can figure out the minimum energy needed to hold a protein together. It’s first-year calculus. You minimize a function by taking the first derivative and setting it equal to zero. This is called energy minimization: Levitt has programmed it on a Mac computer.
Then, you allow movement. Random numbers. You run the program again. Keep going until the energy doesn’t change.
This is critical to getting pictures of proteins. Crystal structure data is a mess. That plus energy minimization software, and we know what a protein looks like.
First day of class, he hands us all the first floppy disks we’ve seen that aren’t floppy. Master at computers who I am, I manage to erase the entire disk and have to go back for another. Not my most impressive moment with Michael Levitt. But I sit in on every group meeting he has and learn what he is doing. I am his student in his course; I am welcome. I am the only one who takes him up on the offer. My friend, Britt, is there anyway—he’s his student.
The project for the class is doing the computer science, using Levitt’s program, to improve a protein. I go for heat stability. Mundane. I get an A on the project, largely because I can write about what I did well; a B in the class.
Eventually the class includes molecular biology—go in the lab and make and purify that protein—and biochemistry. Prove you made what you said you did. My proof would involve physical chemistry—a differential scanning calorimeter. But that first year, it was just the computer science.
Britt continues every possible variation on the physics approach. If you chemically cross-link two parts of the protein and model that, do things work better? If we leave out some of the data, how much worse does it make it? Britt publishes a nice thesis, goes on to work for a company in San Francisco. He gets seriously rich (he used to duck-tape the bottom of his sneakers). He is now a millionaire, living with his wife, who was also a housemate, in Noe Valley. His part of the story is over.
The important point here is physicist thinking was logical, useful in other ways, valuable, understandable, but it completely and miserably failed to solve the protein folding problem. Think like a physicist meant fail to solve anything.
It has solved other things. Physicists do a lot of good work; they couldn’t solve this.
Tertiary structure is the same sort of interactions, but they are big pieces of proteins that move around in 3-D space. Most of the action of enzyme is making or breaking a secondary or tertiary structure interaction. Most of the big things that happen are tertiary interactions flopping around until they minimize energy.
The 1962 Anfinsen experiment makes a simple protein and sees it all fold correctly with only letting it sit in water for a while. This works every time with simple proteins. In a cell, where things are crowded, with stuff in the way everywhere, you need what Anfinsen called “folding helpers.” They are proteins. There are lots of them.
Quaternary structure is completely separate proteins using these same interactions (that can be measured and modeled using physics thinking) to stick to each other. Think antibodies, which are multiple protein chains.
Think also proteins sitting in a phospholipid bilayer membrane. Their inside could be in the cytoplasm; their outside sticking out of the cell, and their middle in the cell membrane. Lots of important biology here.
Anyway, that’s proteins and protein folding. But if all I have is data, and I’m a computer scientist, can I predict protein folding? If you give me the primary sequence—the order in which amino acids come, can I write a program that shows you the 3-D structure? When I use crystallization and physical chemistry techniques (or NMR) to determine the structure, is it identical to what the computer predicted?
The answer up to now had been not quite, but getting better and better. The computer scientist’s type of modeling almost got us there. And weak AI got us there and deserved its Nobel Prize today.
That’s the next part. This is hard work. It’s all computer science. It’s assigning values. It’s systems of math. It’s ideas for how to model. And it’s just not getting there before weak AI.
Weak AI is a repetitive algorithm. Try this math, make random changes, keep running the math. Look at the prediction. It’s a sort this with a criteria: looks like a protein versus doesn’t.
AI is running blind. It’s not really “looks like the protein”; instead, it’s same versus not same. In one famous experiment, it was a watermark of a horse in the back of photos that nobody saw that the program glommed onto. In other cases, the program could sort “dog” versus “cat.”
But the key is that nobody understood what the program was doing. It was a black box. A challenge today is to understand what weak AI is doing. It’s looking at every step taken and trying to make sense. This is good work and making progress, but most is still black box.
Black box AI solved protein folding. Feed it the primary sequence, black box AI gives you the 3-D structure, and it looks precisely how what came from very hard and sometimes near-impossible chemistry got. Not yet for the toughest problems—multi-protein things in membranes—but it will get there.
That’s today’s Nobel.
But what of the third type of modeling I mentioned: “chemical engineering thinking.” That was a lot of what I did at U. of Delaware, paired off with the world’s top chemical engineers. A team of four was critical, Professor Prasad S. Dhurjati, our co-advised grad student Vikas Agarwal (now Dr. Vikas Agrawal, Sr. Principal Data Scientist, Oracle, India, up to recently with Intel, India, and considered one of the top ten data scientist in India by their Indian Institute of Technologies), Zhang Chu, my other Ph.D. student, now the head of the Center for Translational Cancer Research at the University of Delaware, and me.
In a talk honoring Vikas, which he recently sent me, he started with a figure from his Ph.D. thesis in Plant Biology with Prasad and me as co-advisors. Literally everything underlying this approach to modeling came from one meeting between the four of us. I put everything together, Prasad realized he’d been the one who developed the math I’d just said must exist, Chu and I had all the parameters from work we’d done with our own hands, and Vikas did the MatLab coding that made it all real. Prasad’s and my names on his thesis (and Chu’s name on the papers going into his thesis) says to academia the four of us are responsible for one of the most important things powering data science right now. Nobody in real world space has heard our four names together. No biologist will read the paper because of one mistake I made with the genetics and my failing to appreciate one other scientist’s papers. They obsess on these details. They were all contained within one circle on a diagram I wrote that night. None of what they cared about mattered for getting our approach to correctly model one system of biological data or Vikas’ later generalization into a generic approach in data science that truly works.
Here's what I did on a piece of paper that night: I drew circles and said this math in here is this real biology thing we measured and is a variable. All circles are the variables. The math inside can be played with. What matters are the arrows connecting the circles. All will be positive (a causes b) or suppressive (a prevents b). They will also have timing: fast, medium or slow. Prasad had chemical plants in Texas running on that math in the ‘90s based on papers he wrote in the ‘70s. He recognized his own math being described by me. Vikas is famous now for taking his Ph.D. thesis and generalizing it as a data science approach.
This “chemical engineering” thinking depends on factors that would horrify physicists or computer scientists: 1) Lots of things are data. Prasad once solved problems based on a worker telling him what times a day that things smelled bad; 2) Parts of what you think you know can be wrong and the model will still give the correct answer for other things. The stuff that got the biologists into the National Academy—very important stuff—didn’t matter in our model. 3) Being wrong is great! It first means your data is bad, your assumptions are bad, or your math is bad. Make them all the best you can, and life is still great! You have just discovered you don’t understand your own system. Look for the missing answers. We found some. Others with other models did, too. You just did “systems biology” using chemical engineering and math. 4) Only some variables are important. A dozen-equation model with 12 variables you understand can say a lot when a 100,000-equation model says nothing because your brain can’t follow it all.
Nobody in biology knows we did this. Indian data scientists do. We have been given formal academic credit in the language they use.
My friend, Prasad, is now deceased. Very brilliant man. Chu, Vikas, and I are still around and occasionally say “hi.”
So, what does this have to do with Chat-GPT. Nothing. That’s generative AI. Take every word ever spoken and model what should be said next. It’s just pattern matching. It’s useless in many cases. It can be racist and ridiculous.
Weak AI is a very valuable scientific tool. Generative AI—the jury is still out.
Hey, Ted Chiang, when you plopped down on the chair next to me at READERCON, and we talked for two hours, and the first thing you did was agree with me that biologists were using it right and everyone else is dangerous—this was too much to say—but you told me we agree on these issues. You of all people in the world will like my blog post most. I loved our discussion of what you are doing on AI issues. I 100% buy what you say about it being great for billionaires and dangerous for labor.
Folks, Ted is doing some very important nonfiction stuff on AI these days. Read it.
And Wired folks—you are correct, funding is everything. I am not a professor right now because I couldn’t keep grants funded. I once hit top grant in the country in a joint NSF/NIH program that NSF put in 90% of the money, and then petty politics killed it.
The big problem is capitalism. A professor shouldn’t be a small businessman, they should be a thinker, a teacher, and an advisor. Stop making them be capitalists. You have already killed American science with this approach.
The Wired article is coming two decades too late for me. But if they’re correct, maybe it’s not too late for American science to recover. The key issue: capitalism needs to end.
And I’ve sold short stories with weak AI in them (hard SF), and written but not yet sold fun stories with strong AI (Is that you, Dave?) type characters. You can read my fiction under the tabs “Stories I’ve Sold” and “Freebies.”
Comments