Alphafold 3.0: the AI protein predictor gets an upgrade

Benjamin Thompson

Welcome back to the Nature Podcast, this week, a clock that could probe the laws of physics…

Nick Petrić Howe

…and AlphaFold gets an upgrade. I’m Nick Petrić Howe.

Benjamin Thompson

And I’m Benjamin Thompson.

<music>

Benjamin Thompson

As we've heard in recent podcasts, timekeeping is super important to science. And the precision of clocks has gotten pretty good — to put it mildly — with the development of optical clocks, which keep time so well they only lose a second every 30 billion years or so. But scientists want to do better than that. And a paper came out last week, which describes a big step towards making a new type of clock that shows potential to be even more precise, but perhaps more excitingly could allow researchers to even test the very laws of physics. Our regular clock correspondent Lizzie Gibney has been reporting on the story and as sure as tick-follows-tock she joins me now. Lizzie, hi.

Lizzy Gibney

Hi Ben.

Benjamin Thompson

Now, Lizzie the most precise clocks are all atomic clocks, but researchers have wanted to make something else a nuclear clock. Now, they kind of sound similar, but they're actually quite different things, right?

Lizzy Gibney

Yeah, there's a quite similar fundamental principle there. But in an atomic clock, what we're talking about is pumping an electron up an energy state, you’ve got electrons whizzing around an atom, you pipe that up or down an energy state — you might remember the shells from GCSE or high school level chemistry — and when you do that, that's the same energy difference. And that means you've got a kind of fundamental number that will apply to any atom like that, in the world, in the universe. So that's the kind of thing that makes for good clock, we use as the actual ticking the frequency of the light that you need to shift that electron up or down. So the frequency of the light is that tick. Now, what a nuclear clock is, is not looking at the electrons in the shells around an atom, it's actually the nucleus itself, so the neutrons, the protons that are inside the nucleus of the atom. And you can imagine they're much more strongly bound. And you know, you've got this teeny, tiny, dense little hub, rather than these big whizzing round electrons, we're using the energy state in that instead. So it's the same principle in that you're gonna have kind of low-energy state, and then one slightly higher energy state. And it's the frequency of light that it takes to do that, which becomes your clock.

Benjamin Thompson

Now these have been quite hard to make, though, as I understand. Like, in principle, that was a fantastic idea but the actual gap between in principle and reality has proven quite difficult to get across.

Lizzy Gibney

That's right. So there are two big reasons why it's tricky. One is you need to find a nucleus where this actually applies. So although we have these energy states that you can raise a nucleus to, most of them are going to be a huge energy difference. And what we want to be able to do to nudge it with a laser that you'd have in a lab that we can practically use to make a clock, right. So, the most nuclei is going to be really, really big energy difference. There’s one which is thorium-229, which is an isomer for type of the radioactive metal, thorium, where it's a real anomaly, and it has a teeny, teeny tiny transition, this energy to go from its lowest state to its first energy state is very, very small. So it's just about accessible with a laser. So first of all, we found our nucleus, but then the second thing you need to do is know exactly what energy to trigger it at. So the way in which these clocks work is you fire a laser at it and when you get it right, the photons of light get absorbed by the nucleus, and it goes up to the state, and then it releases them and emits them when it falls back down again. So you need to fire the right energy at it. And actually, it's really, really hard to know exactly what frequency of light you need. Theory actually doesn't help us here because the most accurate, they can figure out using theory and making calculations is many, many times greater than the actual number that we need. So it gives us a ballpark much bigger then is useful. So instead, we have to rely on experiment. So it's been looking for a needle in a haystack.

Benjamin Thompson

So we know that thorium-229 doesn't need much energy to nudge it up to this next state. But exactly what that frequency is then, there's quite a wide spectrum, I suppose. And that's been slowly narrowed down but now the researchers have been trying to pinpoint the exact one then that is required.

Lizzy Gibney

Exactly. So there have been lots of different methods of trying to figure out exactly what this frequency is. There have been decades of work and lots of different groups they've been getting closer and closer, but this is really a milestone paper because they've narrowed it down the precision is about 800 times greater than next best effort. And it was a little bit in the end of just sweeping a laser up and down this region, that we know this frequency is hiding in, until they found it. And they found it by this crystal which had the thorium studied in it, until that crystal glow fluoresced when the thorium was nudged into this high-energy level and then released the photon again.

Benjamin Thompson

Right so a little bit like tuning an FM radio for the younger listeners. And so they found it then at the right frequency and what does that mean then? Because this actually isn't a clock itself, this is just a step towards it, right?

Lizzy Gibney

Exactly. They've found their ticking mechanism, the frequency of that light that becomes your tick. So now they've got that then got the basis of the clock. What they need is a few things, they need to check that the system they've got it made in doesn't affect that frequency somehow, the fact that they've done it in this crystal that isn't messing with the frequency somehow. And they also need to narrow down the number of frequencies that are contained within the laser that they're using to do this nudging up an energy state. It's still quite a broad band at the moment, this was one year, the laser was made by the lab themselves, you know, you can't buy a laser like this off the shelf. At the moment essentially, they're sending in loads of photons that are doing absolutely nothing, because some of them are hitting this exact frequency, but many of them aren't. So that laser needs to get much, much, much better in resolution. Now, that's going to be a lot of work. But it's also the kind of thing that people I spoke to for the story said, is doable, like there's not a huge amount of doubt that we should be able to make that laser, which means this kind of clock should be possible we're just not there yet.

Benjamin Thompson

Let's sort of fast forward to when someone does crack this, researchers have wanted to make one of these nuclear clocks for quite a while, why? Like how much better are these than the sort of the optical clocks, which we've discussed on the show fairly recently?

Lizzy Gibney

So there are a few advantages, you could get greater precision, because as we discussed with optical clocks, there you're using optical visible light, here, it's UV, that's a little bit higher frequency, that's kind of getting more ticks in, you're more finely slicing time. So you can increase the precision a bit. But optical clocks are getting better all the time as well. So that's not the main selling point, I would say. The best part about this clock is that it's really, really stable. It's really robust, because the ticking mechanisms come from within the nucleus, which is this really tightly bound little nugget inside the atom, it's much harder for fluctuations in electric or magnetic fields or temperature or anything like that to affect your clock. And that can happen when what you're looking at is these electron levels as you are in an optical clock. So that fundamental difference makes it more stable so your lock is likely to keep time better over time. And then also this system that it’s in, this kind of solid state, the fact that its inside a crystal maybe could mean it's more portable, maybe again that's more amenable to use than one of these massive vacuum systems that you use in the lab for an atomic clock. And there's one more thing that's really, really cool that this enables. So the nuclear clock, this frequency that we need to nudge the nucleus up that energy, depends on what's going on within the nucleus. So it depends on the strength of the fundamental forces, the strong nuclear force, the electromagnetic force. And so we can actually use this ticking of the clock to figure out if anything strange is going on at a very, very, very fine level within physics. So if, for instance, some kinds of dark matter are hypothesised to affect these fundamental constants, the strength of the strong nuclear force for instance, then we'll be able to see that that will be revealed in a slight wavering in the ticking of this clock. So that has got physicists quite excited.

Benjamin Thompson

So it's only in the nucleus then that you can see these potential slight changes between how neutrons and protons and so forth are interacting then. And this wouldn't work just using electrons as are in current atomic clocks right?

Lizzy Gibney

Right. So electrons have some sensitivity to the strength of these forces. But this is about 10,000 times more sensitive than you could get with an atomic clock, which you could get from the dynamics of what happens to electrons. The nucleus itself, because we're talking about these more massive and more strongly bound entities, they are much more affected by any changes in the strength of these forces. So, it gets revealed it gets amplified in a nucleus.

Benjamin Thompson

And you said researchers are kind of excited about this, what have they been telling you?

Lizzy Gibney

Well, even based on this paper, which as we said was, you know, a very broad bandwidth of laser and they found this frequency, but it's very preliminary. They're already using that to put some limits on the kinds of dark matter that can be out there. So different kinds of dark matter would affect these fundamental constants in different ways. And so there's going to be a lot of different studies and testing a lot of different hypotheses about dark matter, just through using these clocks. So people are quite excited. And actually, if I can, quote also one of my colleagues, so when we were talking about this story, Davide Castelvecchi another physics reporter here, said something nice, which I wish one of our interviewees had said, that whenever we use a new system like this here we are tapping right into the nucleus of an atom, that we're going to open ourselves up to lots more scientific discoveries. This is really finely tuned peering into the nucleus in a very, very exact way. And we don't really know yet maybe what all of the uses are going to be. But it's going to be, I think, really exciting.

Benjamin Thompson

We'll link to your story in the show notes. But until next time, Lizzie, thank you so much for joining us.

Lizzy Gibney

Thank you. Maybe more clocks again, we'll see.

Nick Petrić Howe

Coming up, the powerful protein prediction tool, AlphaFold, can now accurately model complex combinations of proteins and other molecules. Right now, though, it’s the Research Highlights, with Dan Fox.

<music>

Dan Fox

If you look around our plant-covered planet, you may think that green is a good indicator for life. And indeed, visible light colours are one of the variables used by researchers when searching for life on other planets. But not all life is green. And now a new guidebook will help astronomers look for purple, brown, orange and yellow hues that could be reflected from worlds dominated by certain kinds of bacteria. Plants are the dominant life form on Earth and the chlorophyll they use to make food absorbs the blue and red wavelengths from sunlight and reflects back green. But the colour of life could be different on other planets. Even on earth, there are ‘purple bacteria’ found from shallow waters to deep-sea hydrothermal vents. These actually come in a range of colours because they rely on other compounds that absorb red and infrared light to produce energy. Microorganisms like these, could thrive on worlds orbiting stars redder than our own Sun. And so a team of astrobiologists have analysed the light reflected by various kinds of purple bacteria recording the spectrum for each species and modelling what spectra from exoplanets covered by such bacteria would look like. From this, they created a database of spectra, which will help astronomers to broaden their search for life on newly discovered exoplanets. You can read that in full in Monthly Notices of the Royal Astronomical Society.

<music>

Dan Fox

A fluid that retains its magnetic structure even after an external magnetic field is removed could form the basis of new liquid bioelectronics. A ferrofluid is a substance consisting of nanometre-sized magnetic particles dispersed throughout a fluid. In the presence of an external magnetic field, the particles align and the ferrofluid exhibits useful magnetic properties. But if you turn off the external field, the particles become randomly oriented, and the properties are lost. Now researchers have produced a new ferrofluid-like material that lacks this drawback. The substance comprises magnetic nanoparticles suspended in a viscous fluid. When exposed to an external magnetic field these particles self-assemble into an intricate 3D network that remains intact even after the field is removed. The team used their substance to make liquid bioelectronics that could be injected into the body and convert biomechanical motion into detectable electrical signals. These kinds of devices could be used for self-powered and wireless monitoring of the cardiovascular system. If you are magnetically attracted to that research, you can read it in Nature Materials.

<music>

Nick Petrić Howe

AlphaFold and its successor AlphaFold 2 have been game changers in biology, as the AIs have made it easier than ever to predict the structures of proteins – the molecules that make up so much of life. And now there's a new version, AlphaFold 3, which promises even greater predictive abilities. To dive into the details of this upgrade, I'm joined by Ewen Callaway who has been writing about AlphaFold, well, basically since the beginning — for years now — and has been working on a new story about the latest iteration AlphaFold3. Ewen hi, how's it going?

Ewen Callaway

It's good. Yeah, I like to see AlphaFold 3 as “The Return of the Jedi” of AlphaFold kind of completing the trilogy. Or “The Godfather III”. I don’t know which was deemed better.

Nick Petrić Howe

Hopefully it's not Godfather III because that one was not the best one in my opinion. But let's start off by talking about what this version can do. So, what is it that AlphaFold 3 promises that maybe the other versions lacked?

Ewen Callaway

Yeah. So the real revolution was AlphaFold 2. Mark one, it did a lot and it won this competition of protein protection but AlphaFold 2 was the game changer. And what it did was you know, you'd input a sequence of amino acids, which are the building blocks of proteins, and it would give you a pretty darn good prediction of what that protein looked like in three dimensions. The only problem — I wouldn't say a problem — was that it just predicted protein structures, it didn't predict proteins alongside all the other, you know, players of the cellular ecosystem. That was just not part of AlphaFold 2’s language, part of its abilities. And so AlphaFold 3, this latest update, is exactly that it brings in the rest of the ecosystem, all these other players and you know, predicts proteins alongside of them.

Nick Petrić Howe

So, when you say like these other players, this is things like DNA, RNA…

Ewen Callaway

Yeah, all of these things. So say you've got a protein that's involved in copying our DNA, which is something that's one of the most basic features of life, well, that protein needs to attach to DNA, and AlphaFold 3 can do that. So you've got proteins that are helping to turn, you know, DNA into proteins via an RNA intermediate, you got some proteins that recognise RNA well, Fold 3 can do that. There are lots of modifications, you know, you plonk on something called a phosphate group. It's called phosphorylation and that activates many proteins, and they propagate signals throughout cells. And so it's really kind of embedding proteins in their ecosystem in their environment, you know, what they do with all these like complex roles, you really need to know about these other players.

Nick Petrić Howe

And I mean, given how many different kinds of molecules and how many different proteins there are, like, it seems like this would be a huge challenge to actually bring these things together in all the possible, you know, iterations they could be. How have they managed it?

Ewen Callaway

I mean, it's a very sophisticated neural network. But I think the principle that, you know, I understood from talking with John Jumper — who led the development of AlphaFold 3 — was that all this information, you know, all these modifications, all these accessory molecules, they're experimental examples, real world examples of them with their protein partners sitting in this database called the protein databank that AlphaFold 1, 2 and now 3 was trained on. And so you've got lots of examples, lots of good data for a machine-learning model and artificial intelligence to learn from. And you know, with a bunch of bells and whistles that I won't bore you with, like transformer and diffusion and embeddings, they've, you know, created a model that can represent all this additional data, not just the protein sequence, but all these other atoms that are sitting there in this database.

Nick Petrić Howe

And so this has been developed, as you said, by Google DeepMind. Have other researchers got to try out this new tool?

Ewen Callaway

Yeah, it seems like a fair number of researchers have got a sneak preview of it. I actually spoke with somebody this morning, who didn't get an official sneak preview, but reviewed the paper for me, and he was able to get onto the server. And people like it, you know, they say it's really fast, it's really convenient. But it lowers the barrier to entry I think, compared to AlphaFold 2 which to use it you almost sometimes had to download your own version and run it or run it on a server. This is, as far as I can tell, kind of a web form where you input your sequence, pick some boxes of sorts of modifications that you want to see, and bada bing bada boom, 10 minutes later, you've got a prediction that can help you do some experiments. So the limited feedback I've gotten so far is that it's really helpful.

Nick Petrić Howe

What are researchers hoping that this tool could be used for?

Ewen Callaway

I think it's part of maybe getting a better approximation of how your protein of interest is doing its job, how it's playing its part in the cellular ballet, whatever we want to call it. So the one example that, you know, I talked with the scientists who used it, he's a scientist at the Crick Institute, across from, you know, Nature HQ in London. He studies DNA replication and a lot of the proteins that his lab is interested in, directly bind DNA. And they'll be, you know, portions of these proteins that binds DNA. And so with the predictions that he got from AlphaFold 3 his lab started making mutations to their protein to try and alter, you know, how it bound DNA and found that some of these really panned out that the predictions were kind of on the spot. So it gives him some insight into how this protein he's interested in does its job.

Nick Petrić Howe

So one thing about AlphaFold’s previous iterations that researchers were a bit sceptical of was its ability to help with drug discovery. Do you think this new tool will help bridge that gap?

Ewen Callaway

AlphaFold has been– it’s been revolutionary, but I think, you know, with drug discovery, you know it's been maybe a mixed bag. There have been you know, a lot of a lot of scepticism whether it's really a game changer. There have been some studies suggesting that it's structures can be useful for drug discovery, but drug discovery is a complex process with lots and lots of steps. And you know, one AI is not going to disrupt this. AlphaFold 3 I think, you know, I've talked with people because it can model because it predict potentially how other molecules interact with proteins, it could potentially be very useful for drug discovery. And in fact, Google DeepMind has a spin off called Isomorphic Labs and they're using AlphaFold 3 to do just that. The hitch is that, you know, the way that Google DeepMind is making AlphaFold 3 accessible to the scientific community, they're not going to allow researchers to straightforwardly look for how their protein of choice binds to a new drug, that's just not possible for people to do. You know, and that was a decision that Google DeepMind made, you know, they put a lot of resources into developing it, and they're going to reserve the commercial pursuits for their partners. But I did speak with, you know, some scientists who said that, you know, the paper that they published in Nature, you know, releases enough information about how this model was developed, that within a year or so, other researchers can develop open-source versions that you can plug any potential drug into it. So it could potentially have, you know, a really significant impact on drug discovery, not just for, for Google and its partners, but for the field as a whole. I think that remains to be seen. But you know, that's a possibility.

Nick Petrić Howe

Well, Ewen thank you so much for joining me.

Ewen Callaway

Yeah, thank you.

Nick Petrić Howe

And listeners, for more on that, check out our show notes for a link to Ewen’s News article.

Benjamin Thompson

That's all for this week, as always you can keep in touch with us on X, we're @NaturePodcast, or send an email to podcast@nature.com. I'm Benjamin Thompson.

Nick Petrić Howe

And I'm Nick Petrić Howe.