Protein Structure Analysis (1)
Korea.
The conference room at JH Research.
After returning from Caltech, I held a meeting at JH Research.
In the middle of the whiteboard, written in large letters:
AI × Basic Medicine
Besides the in-house staff, a few more people sat in the conference room.
I turned to the outside guests and greeted them.
"Welcome. Thank you so much for coming."
A physician, a chemist, a pharmaceutical developer—members of the foundational fields needed for new drug development.
I gave the whiteboard a light tap with my finger.
"The reason I asked you all here today is simple. I'm not sure if you're already aware, but JH's AI is receiving global recognition. So I'd like to put our AI to work in basic medicine. Today I want us to talk freely about whether that's possible, and how far we can take it."
"A kind of brainstorming session, then."
"Something like that."
I rolled my wheelchair back a little so the doctor could see me clearly.
"As you can see, my legs are like this."
Looking at my legs, the doctor asked, "Was it a traumatic amputation? Or was the amputation caused by an internal factor like vascular disease or infection?"
"Traumatic amputation."
"Any phantom pain?"
"That's fine. As you can see, I'm like this. It's a personal request, I know—but we have a world-class AI in this company, and I keep wondering whether there's anything it could do for me right now."
The doctor thought for a moment.
"Hmm… In a traumatic amputation, peripheral nerve damage is basically a given. You need to promote nerve regeneration and reconstruct the peripheral nerve tissue. Complete regeneration is difficult, but even at the research stage there have been some cases of partial functional recovery."
I brought up my mother as well.
"My mother suffered a cerebral hemorrhage. Since then her intelligence has declined sharply. Her memory is weak too. Is there any way to address that?"
The doctor spoke calmly.
"Your mother's case is a different kind of problem. A leg involves peripheral nerves, but a cerebral hemorrhage is a central nervous system issue—the brain itself."
"Is the brain that different?"
"You have to evaluate the possibility of reconstructing the neural networks in the damaged brain regions. To be honest, current medicine has its limits."
"I wasn't asking for treatment right this second. The goal of this meeting is to talk about how AI can be used to advance basic medicine. Of course, if that happens to include treatments related to amputations and cerebral hemorrhage, all the better."
One of the outside guests raised a hand.
The chemist—more precisely, a pharmacological biochemist. He had an intellectual air about him, but he also looked worn out.
Pushing his glasses up slightly, the biochemist spoke.
"From what you've said, you'd like AI in medicine to help with peripheral nerve regeneration, tissue regeneration, central neural network recovery, cognitive improvement, and so on."
"That's right."
"Whether it's the brain or the peripheral nerves, it ultimately comes down to recovering from nerve damage. To do that, you have to find candidate compounds that let the damaged nerve cells restore their energy."
"And how do you find compounds like that?"
"How do you find them?"
"Yes."
At my question, the biochemist's expression twisted ever so slightly. Then he answered in a voice thick with resignation.
"…Grunt work."
"Grunt work?"
"Endless manual labor. You run through millions of candidates one by one, check cell toxicity, predict metabolism, tweak the structure… You keep going until something works."
"That must take a lot of time."
"Of course. Which is why people like us—this may sound bad—often end up relying on gut feeling."
"Gut feeling?"
"There are millions of candidates. Which one do you validate first? You guess."
"You guess?"
"Yes, exactly. How do doctors make drugs? With pure logic? No. It's gut. 'That's it—I've decided. Looking at the shape of this molecule, I think you're going to be a new drug.' Then they pick them one by one by feel and run the experiments. Unfortunately, that's the reality of research."
The pharmaceutical developer next to him cut in.
"Everything he said is a hundred percent right. And our side doesn't lose to anyone when it comes to grunt work either."
"What does your side look like?"
"The pharmacology team picks a molecule and validates it at the molecular level. Then they toss it to us and say, 'Turn this into a drug.' And we have to actually make it into one."
The pharmaceutical developer sighed.
"Pill, capsule, injection, patch—how fast it dissolves in the body, how much is absorbed, whether it breaks down in the stomach or gets destroyed in the liver, whether there's toxicity, whether stability holds, whether it can be stored…"
He gave a self-deprecating smile.
"Cell trials, animal trials, human trials phases one through four. The failure rate is ninety-nine percent. And every single time the molecular structure changes even slightly… you do it all over again. Ha ha ha."
"I see."
The pharmaceutical developer continued.
"They say semiconductors double in performance every two years, right?"
"Yes, that was the pattern. These days it feels more like double every year."
"Incredible. But pharma is completely different. The time it takes to develop a new drug keeps getting longer."
"Why?"
"Because the easy ones are already done. What's left are the high-difficulty diseases. What you want, CEO, ultimately comes down to the nervous system—the brain."
"Is the brain that hard to treat?"
The doctor, the biochemist, and the pharmaceutical developer all nodded at once.
"For one thing, with brain-related drugs, ninety-nine percent of them get filtered out by the blood-brain barrier."
"Ninety-nine percent?"
"Yes. The brain is not an ordinary organ. Getting a drug through is extremely picky. Most of the time, it simply doesn't cross. Molecular size, charge, lipid solubility—get even one of those wrong and it doesn't get in."
Somehow the meeting had turned into a session for airing the hardships of pharmaceuticals.
"All right. I understand the difficulties in each field. But this company has a world-class AI. Other places keep asking to borrow it. Even so, I want to allocate it to medicine first. What approach would work best?"
The doctor spoke.
"What I'd want to entrust to AI is… precisely classifying patients' nerve-damage patterns and predicting recovery potential. But even if we used existing treatments and adapted them to each patient more quickly, I don't think it would lead to a revolutionary treatment for you or your mother, CEO."
"I imagine not."
"Yes. If we're going to use AI, it should be applied somewhere more fundamental."
I turned my head toward the biochemist.
"CEO, if it were me, I'd put AI on protein structure prediction."
"Protein structure?"
"Yes. It's been an unsolved problem for the last fifty years. A protein's structure determines how a drug will work."
I didn't fully follow, so I asked again.
"Are drugs proteins?"
"No, not exactly. Most drugs are small molecules. But our bodies are made of proteins. In other words, a drug and a protein are like a key and a lock. The drug has to bind precisely to the protein for any effect to appear."
"Drugs are small keys; our bodies are lock-like proteins. Got it."
"Protein structure is, in the end, a map of the body. 'If I fit this into this groove, that should work. If I block this part, the enzyme should stop. Drugs can't get in here.' With a map, making drugs becomes much easier."
"That makes sense."
"But figuring out a protein's structure is honestly not a job for human beings."
"Is it that hard?"
The biochemist sighed. He sounded like someone who had already prepared the full explanation of why this was impossible.
"A protein is a chain of hundreds of amino acids. Even the rotation angle of a single amino acid has dozens of possibilities. Link about three hundred of them together and the number of possible structures becomes something unimaginable—like ten to the power of three hundred."
"Ten to the three hundred?"
"Yes. One, ten, a hundred, a thousand, ten thousand… three hundred zeros."
"Why is it so complicated?"
"The amino acids in a protein are linked in a line. But the first connection affects the next one. So it isn't a problem you can calculate in order—every part influences every other part."
"But if you want to know the structure, can't you just look at it? A house has a house structure, a desk has a desk structure. You can just see it, right? Is figuring out the structure really that hard?"
"Because they're molecules you can't see with the naked eye. Equipment has advanced, of course, so you can obtain information indirectly. But even that way, determining a protein's structure takes one to two years per molecule. Ah—one moment."
Something seemed to occur to him mid-sentence. He rummaged through his bag and carefully took something out.
A few small plastic tubes—the translucent kind you'd find in a research lab.
"I happened to bring a few."
He set a small vial on his palm and continued.
"These are protein samples. They're diluted in water like this."
I'd expected something more impressive from the word sample. It just looked like water.
"There are protein molecules in here. You can't see them. You can't see them under a microscope either. You put this into analysis equipment like X-ray crystallography or NMR and obtain indirect information. Then you use that information to mathematically reconstruct a three-dimensional structure."
Figuring out the structure of something invisible to the eye. It didn't sound easy.
"Suppose we figured out the structure of one protein in this vial in a single year. But there are tens of thousands of proteins produced by human genes alone, and millions with known amino acid sequences. If one lab takes a year per protein, you're looking at millions of years. Even if many labs divide the work, it could still take thousands or tens of thousands of years."
I asked a question worthy of a semiconductor company CEO.
"Has no one tried approaching this with computers over the years?"
"Of course they have. But even supercomputers couldn't do it. Searching by calculating possible structures was simply too slow—no answer ever came out."
The biochemist gave a quiet smile.
"That's why scientists sometimes call this field Mission Impossible."
The outside guests left.
I stayed where I was, thinking for a moment about how to move the medical business forward.
That was when I noticed Intern Oh beside me, clearing the table—gathering the half-finished drinks and leftover snacks. Just an ordinary intern doing ordinary intern work.
On the surface, an intern. In reality, Intern Oh had come here as the son of the Ohsung Group chairman. A hostage role that didn't even belong in the twenty-first century.
And yet the Ohsung chairman had sent him to me, and all I'd been doing was making him run errands. It felt a little wrong.
Didn't they say an adult's apology comes in money?
Buying DRAM cheap had already cooled some of my anger over Ohsung Electronics' underhanded moves against us.
I turned my wheelchair slightly toward him.
"Intern Oh."
"Yes!"
His voice, at least, was crisp and clear.
"Organizing materials… that is part of an intern's job, right?"
"Yes. It is."
"You heard today's meeting?"
"Yes, I observed as well."
"Then you already know. Gather the materials on protein structures that have already been fully analyzed. We'll feed them into our Nano AI."
Intern Oh's eyes flickered ever so slightly. He seemed to realize this wasn't a simple task.
"You'll probably need to contact research institutes all over the world. Collect the data each lab has researched so far. Then we'll run it through our Nano AI. Can you do that?"
"Yes. I'll try."
A few days later.
I sat in a meeting with Team Leader Gong.
"Team Leader Gong—so even if we try to build molecular structures, at the beginning we still have to teach it directly what they actually look like?"
"That's correct. When we first trained the AI, we labeled things one-to-one—this is an apple, this is a strawberry. It's the same principle."
"So when we teach AI a new field like molecular structure, we have to walk it through everything one by one at first?"
Gong nodded.
"Exactly. And right now there aren't that many labeled molecular structure sets. Once things get rolling, scaling the numbers is what AI does best—but the starting point has to be set by people."
"Intern Oh will bring some."
"Intern Oh? Medical data sources are highly specialized. They won't be easy to obtain."
Fair enough. Apple and strawberry data you could get as much as you wanted if you tried, but molecular structure data was the kind ordinary people could barely obtain—and wouldn't recognize even if it was handed to them.
"Yeah, well. Sounds like he knows people in that world."
Just as we were talking about Intern Oh, the door opened and he walked in.
"CEO!"
A bright, cheerful, confident voice.
"I got the data."
"Oh? Already? Molecular structure data sources are hard information to come by."
Intern Oh came close and whispered quietly.
"…I used the dad card."
Of course he did.
I narrowed my eyes at him.
"Must be nice. Having a dad. I don't."
Intern Oh's eyes went wide with panic.
"I—I'm sorry."
"Heh. I'm joking."
"Ah… thank goodness."
Not having a dad was the true part, though.
I had Intern Oh hand the disk over to Team Leader Gong.
"Team Leader Gong, run the data."
Click-clack, click-clack.
Team Leader Gong began feeding the molecular structure data into Nano AI.