What if, all of the world's biggest problems from climate change, to curing diseases, to disposal of plastic waste, what if they all had the same solution?

A solution so tiny it would be invisible.

I'm inclined to believe this is possible, thanks to a recent breakthrough that solved one of the biggest problems of the last century. How to determine the structure of a protein?

- It's been described to me as equivalent to Fermat's last theorem, but for biology.

- Over six decades, tens of thousands of biologists painstakingly worked out the structure of 150,000 proteins.

Then in just a few years, a team of around 15 determined the structure of 200 million.

That's basically every protein known to exist in nature.

So how did they do it and why does this have the potential to solve problems way outside the realm of biology?

A protein starts simply as a string of amino acids.

Each amino acid has a carbon atom at the center.

Then on one side is an amine group, and on the other side is a carboxyl group.

And the last thing it's bonded to could be one of 20 different side chains, and which one determines which of the 20 different amino acids this molecule is.

The amine group from one amino acid can react with the carboxyl group of another to form a peptide bond.

So a series of amino acids can bond to form a string and pushing and pulling between countless molecules, electrostatic forces, hydrogen bonds, solvent interactions can cause this string to coil up and fold onto itself.

This ultimately determines the 3D structure of the protein.

And this shape is the thing that really matters about the protein.

It's built for a specific purpose, like how hemoglobin has the perfect binding site to carry around oxygen in your blood.

- These are machines, they need to be in their correct orientation in order to work together to move, for example, the proteins in your muscles.

They change their shape a little bit in order to pull and contract.

- But it would take people a long time to get the structure of just one protein.

- Absolutely.

So what should proteins look like? Was only started to answer really with experimental techniques.

- [Derek] The first way protein structure was determined was by creating a crystal out of that protein.

This was then exposed to x-rays to get a diffraction pattern, and then scientists would work backwards to try to figure out what shape of molecules would create such a pattern.

It took British biochemist, John Kendrew, 12 years to get the first protein structure.

His target was an oxygen storing protein called myoglobin, an important protein in our hearts.

He first tried a horse heart, but this produced rather small crystals because it didn't have enough myoglobin.

He knew diving mammals would have lots of myoglobin in their muscles since they're the best at conserving oxygen.

So he obtained a huge chunk of whale meat from Peru.

This finally gave Kendrew large enough crystals to create an x-ray diffraction image.

- And when it came out, it looked really weird.

People expected something kind of logical, mathematical, understandable, and it almost looked, I wouldn't say ugly, but intricate and complex and kind of like if you see a rocket motor, and all the parts hanging off.

- [Derek] This structure, which has been called "Turd of the century," won Kendrew, the 1962 Nobel Prize in chemistry.

Over the next two decades, only around a hundred more structures were resolved.

Even today, protein crystallization remains a big challenge.

- Frankly it is not uncommon that just a couple protein structures can be someone's entire PhD.

Sometimes just one, sometimes even just progress toward one,

- And it's expensive.

X-ray crystallography can cost 10s of thousands of dollars per protein.

So scientists sought another way to work out protein structure.

It only costs around a hundred dollars to find a protein sequence of amino acids.

So if you could use this to figure out how the protein would fold, that would save a lot of time, effort, and money.

I kind of know how carbon behaves and I know how carbon sticks to a sulfur and how that might stick next to a nitrogen.

And if these ones are here, then I can imagine this one folding, making that bond there.

So it seems like if you have some sense of basic molecular dynamics, you might be able to figure out how this protein's gonna fold.

- One of the few true predictions in biology was actually Linus Pauling looking at just the geometry of the building blocks of proteins and saying, actually they should make helices and sheets.

That's what we call secondary structure, the very local kind of twists and turns of the protein.

- But beyond helices and sheets, biochemists could not figure out any reliable patterns that would lead to the final structure of all proteins.

One reason for this is that evolution didn't design proteins from the ground up.

- It's kind of like a programmer that doesn't know what they're doing, and whenever it looked good, they just kept adding that kind of thing.

And that's how you end up with these both amazing objects and incredibly complex and hard to describe.

They don't have purpose underneath them in the same way as like a human designed machine would.

- [Derek] To illustrate just how complicated this process can get, MIT biologist Cyrus Levinthal did a back-of-the-envelope calculation,

and he showed that even a short protein chain with 35 amino acids can fold in an astronomical number of ways.

So even if a computer checked the energy instability of 30,000 configurations every nanosecond, it would take 200 times the age of the universe to find the correct structure.

Refusing to give up, the University of Maryland professor John Moult started a competition called CASP in 1994.

The challenge was simple, to design a computer model that could take an amino acid sequence and output its structure.

The modelers would not know the correct structure beforehand, but the output from each model would be compared to the experimentally determined structure.

A perfect match would get a score of a hundred, but anything over 90 was considered close enough that the structure was solved.

CASP competitors gathered at an old wooden chapel turned conference center in Monterey, California, and at any point where a prediction didn't make sense, they were encouraged to tap their feet as friendly banter.

There was a lot of foot tapping.

(foot tapping)

In the first year, teams could not achieve scores higher than 40.

The early front runner was an algorithm called Rosetta, created by University of Washington biologist David Baker.

One of his innovations was to boost computation by pooling together processing power from idle computers in homes, schools, and libraries that volunteered to install his software called Rosetta at Home.

- As part of it, there was a screensaver that showed basically the course of the protein folding calculation.

And then we started getting people writing in saying that they were watching the screensaver and they thought they could do better than the computer.

- So Baker had an idea. He created a video game.

(upbeat music)

The game called Fold It, set up a protein chain capable of twisting and turning into different arrangements.

- But now instead of the computer making the moves, the game players, the humans could make the moves.

Within three weeks, more than 50,000 gamers pooled their efforts to decipher an enzyme that plays a key role in HIV.

X-Ray crystallography showed their result was correct.

The gamers even got credited as co-authors on the research paper.

Now, one man who played Fold It was a former child chess prodigy named Demis Hassabis.

Hassabis had recently started an AI company called DeepMind.

Their AI algorithm, AlphaGo made headlines for beating world champion Lee Sedol at the game of Go.

One of AlphaGo's moves, move 37, shook Sedol to his core.

But Hassabis never forgot about his time as a Fold It gamer.

- So of course I was fascinated this just from games design perspective.

You know, wouldn't it be amazing if we could mimic the intuition of these gamers who were only, by the way, of course, amateur biologists.

- After returning from Korea, DeepMind researchers had a week-long hackathon where they tried to train AI to play Fold It.

This was the beginning of Hassabis' longstanding goal of using AI to advance science.

He initiated a new project called Alpha Fold to solve the protein folding problem.

Meanwhile at CASP, the quality of prediction from the best performers, including Rosetta had plateaued.

In fact, the performance went downhill after CASP eight.

The predictions weren't good enough, even with faster computers and a growing number of structures in the protein data bank to train on.

DeepMind hoped to change this with AlphaFold.

Its first iteration, AlphaFold 1, was a standard off-the-shelf deep neural network like the ones used for computer vision at that time.

The researchers trained it on lots and lots of protein structures from the protein data bank.

As input, AlphaFold took the protein's amino acid sequence and an important set of clues given by evolution.

Evolution is driven by mutations, changes in the genetic code, which in turn change the amino acids within a given protein sequence.

But as species evolve, proteins need to retain the shape that allows them to perform their specific function.

For instance, hemoglobin looks the same in humans, cats, horses, and basically any mammal.

Evolution says, if it ain't broke, don't fix it.

So we can compare sequences of the same protein across different species in this evolutionary table.

Where sequences are similar, it's likely they are important in the protein structure and function.

But even where the sequences are different, it's helpful to look at where mutations happen in pairs because they can identify which amino acids are close to each other in the final structure.

Say two amino acids, a positively charged lysine and a negatively charged glutamic acid attract and hold each other in the folded protein.

Now, if a mutation changes lysine to a negatively charged amino acid, it would repel glutamic acid and destabilize the whole protein.

Therefore, another mutation must replace glutamic acid with a positively charged amino acid.

This is known as co-evolution.

These evolutionary tables were an important input for AlphaFold.

As output, instead of directly producing a 3D structure, AlphaFold predicted a simpler 2D pair representation of that structure.

The amino acid sequence is laid out horizontally and vertically.

Whenever two amino acids are close to each other in the final structure, their corresponding row column intersection is bright.

Distant amino acid pairs are dim.

In addition to distances, the pair representation can also hold information on how amino acid molecules are twisted within the structure.

AlphaFold 1 fed the protein sequence and its evolutionary table into its deep neural network, which it had trained to predict the pair representation.

Once it had this, a separate algorithm folded the amino acid string based on the distance and torsion constraints.

And this was the final protein structure prediction.

With this framework, AlphaFold entered CASP 13 and it immediately turned heads.

It was the clear winner after many additions, but it wasn't perfect.

Its score of 70 was not enough to clear the CASP threshold of 90.

DeepMind needed to get back to the drawing board to get better results.

So Hassabis recruited John Jumper to lead AlphaFold.

- AlphaFold 2 was really a system about designing our deep learning.

The individual blocks to be good at learning about proteins, have the types of geometric physical, evolutionary concepts that were needed and put it into the middle of the network instead of a process around it.

And that was a tremendous accuracy boost.

- [Derek] There were three key steps to get better results with AI.

First, maximum compute power.

Here, DeepMind was already better positioned than anybody in the world.

It had access to the enormous computing power of Google, including their tensor processing units.

Second, they needed a large and diverse data set.

Is data the biggest roadblock and why?

- I think it's too easy to say data's the roadblock and we should be careful about it.

AlphaFold 2 was trained on the exact same data with much, much better machine learning as AlphaFold 1.

So everyone overestimates the data blockage because it gets less severe with better machine learning.

- [Derek] And that was the third key element, better AI algorithms.

Now AI is not just good at protein folding.

It can do all kinds of tasks that no one likes from writing emails to answering phone calls.

Something I hate is building and maintaining a website.

It's so much work from optimizing the website for different platforms, finding a good design so it looks professional to constantly updating it with new information about the business as it grows.

That's why we partnered with Hostinger, the sponsor of today's video.

Hostinger makes it super easy to build a website for yourself or your business.

And with their advanced AI tools, you can simply describe what you want your website to look like. And in just a few seconds, your personalized website is up and running.

Hostinger is designed to be as easy as possible for beginners and professionals.

So any tweaks you need to make after that are super easy too.

Just drag and drop any pictures or videos you want, where you want them, or just type what you want to say or have the AI help you here too, if writing isn't your thing either.

And if you still want that human touch, Hostinger is always available with 24/7 support if you ever run into any issues.

But when you're done building in just a few clicks, your website is live.

It's all incredibly affordable too, with a domain and business email included for free.

So to take your big idea online today, visit hostinger.com/ve or scan this QR code right here.

And when you sign up, remember to use code VE at checkout to get 10% off your plan.

I wanna thank Hostinger for sponsoring this part of the video.

And now back to protein folding.

As the AlphaFold 2 team searched for better algorithms, they turned to the transformer.

That's the T in ChatGPT.

And it relies on a concept called attention.

In the sentence, the animal didn't cross the street because it was too tired. Attention recognizes that it refers to animal and not street based on the word tired.

Attention adds context to any kind of sequential information by breaking it down into chunks, converting these into numerical representations or embeddings and making connections between them. In this case, the word it and animal.

3Blue1Brown has a great series of videos specifically about transformers and attention.

Large language models use attention to predict the most appropriate word to add to a sentence,

but AlphaFold also has sequential information, not sentences, but amino acid sequences.

And to analyze them, the AlphaFold team built their own version of the transformer called an EVO Former.

The EVO Former contained two towers, evolutionary information in the biology tower and pair representations in the geometry tower.

Gone was AlphaFold 1's deep neural network that started with one tower and predicted the other.

Instead, AlphaFold 2's EVO Former builds each tower separately.

It starts with some initial guesses, evolutionary tables taken from known data sets as before, and the pair representations based on similar known proteins.

And this time there's a bridge connecting the two towers that conveys newly found biological and geometry clues back and forth.

In the biology tower, attention applied on a column identifies amino acid sequences that have been conserved.

While along a row, it finds amino acid mutations that have occurred together.

Whenever the EVO Former finds too closely linked amino acids in the evolutionary table. It means they are important to structure and it sends this information to the geometry tower.

Here attention is applied to help calculate distances between amino acids.

- There's also this thing called triangular attention that got introduced, which is essentially about letting triplets attend to each other.

- [Derek] For each triplet of amino acids, AlphaFold applies the triangle inequality.

The sum of two sides must be greater than the third.

This constrains how far apart these three amino acids can be.

This information is used to update the pair representation,

- And that helps the model produce like a self-consistent picture of the structure.

- [Derek] If the geometry tower finds it's impossible for two amino acids to be close to each other, then it tells the first tower to ignore their relationship in the evolutionary table.

This exchange of information within the EVO Former goes on for 48 times, until information within both towers is refined.

The geometrical features learned by this network are passed onto AlphaFold 2's second main innovation, the structure module.

- For each amino acid, we pick three special atoms in the amino acid and say that those define a frame.

And what the network does is it imagines that all the amino acids start out with the origin and it has to predict the appropriate translation and rotation to move these frames to where they sit in the real structure.

So that's essentially what the structure module does.

- But the thing that sets the structure module apart is what it doesn't do.

Previously, people might have imagined that you would like to encode the fact that this is a chain, you know, and that certain residue should sit next to each other.

We don't really explicitly tell AlphaFold that.

It's more like we give it a bag of amino acids and it's allowed to position each of them separately.

And some people have thought that that helps it to not get stuck in terms of where things should be placed.

It doesn't have to always be thinking about the constraint of these things forming a chain, that's something that emerges naturally later.

- [Derek] That's why live AlphaFold folding videos can show it doing some weirdly non-physical stuff.

The structure module outputs a 3D protein, but it still isn't ready.

It's recycled at least three more times through the Evo Former to gain a deeper understanding of the protein only then the final prediction is made.

In December, 2020, DeepMind returned to a virtual CASP with AlphaFold 2, and this time they did it.

- I'm going to read an email from John Moult.

"Your group has performed amazingly well in CASP 14, both relative to other groups and in absolute model accuracy. Congratulations on this work."

- [Derek] For many proteins, AlphaFold 2 predictions were virtually indistinguishable from the actual structures and they finally beat the gold standard score of 90.

- For me, having worked on this problem so long, after many, many stops and starts, and suddenly this is a solution.

We'd solved the problem.

This gives you such excitement about the way science works.

- [Derek] Over six decades, all of the scientists working around the world on proteins painstakingly found about 150,000 protein structures.

Then in one fell swoop, AlphaFold came in and unveiled over 200 million of them.

Nearly all proteins known to exist in nature.

In just a few months, AlphaFold advanced the work of research labs worldwide by several decades.

It has directly helped us develop a vaccine for malaria.

It's made possible the breaking down of antibiotic resistance enzymes, which make many life-saving drugs effective again.

It's even helped us understand how protein mutations lead to various diseases from schizophrenia to cancer,

and biologists studying little known and endangered species suddenly had access to proteins and their life mechanism.

The AlphaFold 2 paper has been cited over 30,000 times.

It has truly made a step function leap in our understanding of life.

John Jumper and Demis Hassabis were awarded one half of the 2024 Nobel Prize in chemistry for this breakthrough.

The other half went to David Baker, but not for predicting structures using Rosetta.

Instead, it was for designing completely new proteins from scratch.

- It was really hard to make brand new proteins that would do things.

And so that's kind of the problem that we solved.

- To do so, he uses the same kind of generative AI that makes art in programs like Dall-E.

You can say draw a picture of a kangaroo riding on a rabbit or something, and it will do that.

And so it's exactly what we did with proteins.

His technique called "RF Diffusion" is trained by adding random noise to a known protein structure, and then the AI has to remove this noise.

Once trained in this way, the AI can be asked to produce proteins for various functions.

It's given a random noise input, and the AI figures out a brand new protein that does what you asked it to do.

This work has huge implications.

I mean, imagine you got bitten by a venomous snake.

If you're lucky, you'll have access to anti-venom prepared by milking venom from the exact kind of snake,

which is then injected into live animals, and the antibodies from that animal are extracted and refined and then given to you as an anti-venom.

The trouble is often people have allergic reactions to these antibodies from other organisms.

But your odds of survival can be a lot better with the latest synthetic proteins designed in Baker's lab.

They've created human compatible antibodies that can neutralize lethal snake venom.

This anti-venom could be manufactured in large quantities and easily transported to the places where it's needed.

With these tiny molecular machines, the possibilities are endless.

What are the applications you're most excited about?

- So I think vaccines are gonna be really powerful.

We have a number of proteins that are in human clinical trials for cancer, and we're working on autoimmune disease now.

We're really excited about problems like capturing greenhouse gases.

So we're designing enzymes that can fix methane, break down plastic.

- What makes this approach so effective is how fast they can create and iterate the proteins.

- It's really quite miraculous for anyone who's a conventional school biochemist or protein scientist.

We can now have designs on the computer, get the amino acid sequence of the design proteins, and then in just a couple days we can get the protein out.

Yeah. We've given a name to this, which is "Cowboy Biochemistry" because we just like, you just got kind of go for it as fast as you can, and it turns out to work pretty well.

- What AI has done for proteins is just a hint of what it can do in other fields and on larger scales.

In materials science, for example, DeepMind's GNoME program has found 2.2 million new crystals, including over 400,000 stable materials that could power future technologies from superconductors to batteries.

AI is creating transformative leaps in science by helping to solve some of the fundamental problems that have blocked human progress.

- If we think of the whole tree of knowledge, you know there are certain problems where you know if their root, no problems.

If you unlock them, if you discover a solution to them, it would unlock a whole new branch or avenue of discovery.

- And with this, AI is pushing forward the boundaries of human knowledge at a rate never seen before.

- Speed ups of 2x are nice, they're great, we love them.

Speed ups of a 100,000x, change what you do.

You do fundamentally different stuff and you start to rebuild your science around the things that got easy.

- And that's what I'm excited about.

These discoveries represent real step function changes in science.

Even if AI doesn't advance beyond where it is today, we will be reaping the benefits of these breakthroughs for decades.

And assuming AI does continue to develop, well, it will open up opportunities that were previously thought impossible.

Whether that's curing all diseases, creating novel materials, or restoring the environment to a pristine state.

This sounds like an amazing future as long as the AI doesn't take over and destroy us all first.

(slow cosmic music)