The WATTBA Series: AlphaFold3
I thought I had coined a new word, which only goes to show that old fuddy-duddies should never try to do cool things on the internet.
WATTBA stands for What A Time To Be Alive, but there is (of course there is) already a Wikipedia page on it. It gets worse, because there is a Wikipedia disambiguation page.
Pah.
But anyways, I'm starting a new series called WATTBA, because the present time really and truly qualifies. It might not seem like it given our awesome ability to doomscroll ourselves to death at every given opportunity... but that's just us choosing to focus exclusively on the bad stuff. I've been very, very good at focusing on the depressing stuff recently, and am trying to get worse at it.
The WATTBA series will help.
And we kick things off with AlphaFold 3
What is AlphaFold 3?
Inside every plant, animal and human cell are billions of molecular machines. They’re made up of proteins, DNA and other molecules, but no single piece works on its own. Only by seeing how they interact together, across millions of types of combinations, can we start to truly understand life’s processes.
In a paper published in Nature, we introduce AlphaFold 3, a revolutionary model that can predict the structure and interactions of all life’s molecules with unprecedented accuracy. For the interactions of proteins with other molecule types we see at least a 50% improvement compared with existing prediction methods, and for some important categories of interaction we have doubled prediction accuracy.
Sounds big, "all of life's molecules". But what exactly does it mean?
(Note that while I have written this blogpost myself, it is based on extensive conversations I have had with Google Gemini, Claude and ChatGPT)
Imagine a human body as being constructed out of Lego blocks. Turns out that this is actually not all that bad an analogy, because that's helpful at two different levels. The cells in our body can indeed be thought of as Lego blocks, but each of those Lego blocks are themselves composed of smaller Lego blocks. What kind of Lego blocks are these smaller Lego blocks?
Well, there's proteins, carbohydrates, lipids and nucleic acids (and tons of other stuff besides), and not all of them work the way they are "supposed" to. This inability to work the way they are supposed to can be because of many reasons - disease, injury, old age, and many others besides.
Medicine works by "repairing" these miniature Lego blocks. This needs a deep understanding of the structure of these Lego blocks, and AlphaFold is a machine learning model that is awesome at this game - and we're currently on v3.
Say hello to the proteins
But repairing them isn't quite so simple. Medicines work (some of the time) by "fitting into" these miniature Lego blocks. But fitting into them is only possible if you know the shape and structure of these Lego blocks. Which is why we need to "predict the structure and interactions of all life’s molecules".
Now, here is where the Lego blocks analogy breaks down, because if you're thinking "Ah, Lego blocks, nice and simple structure"...
... boy are you in for a treat:
And that's just one protein. There are, perhaps, hundreds of thousands of such proteins in our body, and that may well be an underestimate. Figuring out the structure of each of them used to take, well, years and years.
Why did it take years and years?
Because before AlphaFold came along, the way to find out the structure of a protein was by "growing a crystal" of the protein. Growing a crystal of the protein, best as I can tell, involves taking a pure solution of the protein, and adding some specific chemicals to it to help these proteins start linking up with each other, which is what crystallization is. This process alone can take weeks, if not months.
Once you have these crystals, you shine x-rays through these crystals, and analyze the patterns created by the manner in which these X-rays are scattered by the crystallized protein. (Google Gemini describes this as kinda sorta trying to figure out what the pebble must have looked like by analyzing the ripples it makes on the surface of a pond. Huh. Good luck with that.) Those patterns give you a general sense of what the shape of the protein is like.
This is obviously a simplified explanation, and that is a massive understatement. But it is good enough for our purposes for the moment.
This process, by the way, is called X-Ray Crystallography. A typical process for a reasonably easy protein can take multiple years, and can cost a fair bit (fifteen to twenty million dollars per protein for some of the more complicated ones).
So What Did AlphaFold (AF) do?
A better question is what did AF1, AF2 and Af3 do - and each matter. And to understand what AF1, Af2 and Af3 did, we need to talk about CASP13 (and 14 and 15).
CASP stands for Critical Assessment of Structure Prediction. It is the T20 World Cup of protein prediction, and this is a reasonable analogy in two different ways. One, it is a global competition, and two, it runs every two years. It's been going on since 1994, and we are currently at CASP15, which was conducted in 2022. CASP16 is scheduled for December of this year, if you feel like taking a swing at it.
So anyways, back in 2018, AF1 shot to global prominence by winning CASP13. Even more impressively, in 2020 AF2 won again - CASP14 this time. And not just won it - it won it by beating expectations that were guided by its outstanding performance in CASP13:
So when I learned that DeepMind will declare the problem solved I assumed they had done what they did at CASP13 again, achieving in two years what I thought they would do in four, getting a median GDT_TS of ~80. I was impressed and eager to hear the details. Imagine my surprise then when I was informed a few days later of the final number, a median GDT_TS of 92.4. Never in my life had I expected to see a scientific advance so rapid.
This blogpost, and the section within it titled "The Method" is recommended reading if you'd like to really get into the weeds.
And now, in 2024, AF3 is getting the same kind of mind-blowing results, but not just for proteins - but for all of life's molecules.
Imagine this:
What if Boston Dynamics came up with a robot athlete that was the player of the tournament at the 2018 T20 World Cup. By 2022, this robot athlete was the best T20 player, the best one day international player and the best Test match player - of all time. That is, it was ABD, Kohli and Don Bradman all rolled into one - in fact, better than all three rolled into one.
And in 2024, Boston Dynamics tells us that this robot athlete is the GOAT not just in cricket, but also in football, all Olympic sports, and every other sport you can think of.
That's AlphaFold3.
TMKK?
What is an EFE post without a Toh Main Kya Karoon?
Three major advancements have been made much more likely as a direct consequence of AF2.
First, we now have the ability to think about designing drugs that target getting the whole body to get better, not just the specific part that has been negatively impacted by the disease in question. As Gemini puts it, it won't do to solve a traffic jam only in the immediate vicinity of a malfunctioning traffic signal - you need to clear traffic that has been clogged up in the entire neighborhood. This is "systems pharmacology", and this is more likely now as a consequence of AF2.
Second, we now have the ability to make drugs that can target multiple diseases at the same time. Imagine a superhero that is Iron Man, Superman, Batman and any other superhero you can think of, all rolled into one. This is "polypharmacology".
And that was the good news from AF2!
AF3, remember, is about all of life's molecules, not just proteins. Or, to go back to our Lego analogy, we now have the ability to not just understand Lego blocks and how they work, but also everything else that comes in a Lego set - the accessories, the instruction manual (DNA/RNA) and Lego blocks that we haven't even imagined so far.
So Life Becomes Awesome Starting Tomorrow?
Ah, no, not really, alas.
Protein and all of life's molecules is all well and good, but drug development, drug testing and drug roll-outs will still take time. So things will take a bit of time, still.
But if they were going to take, say ten years until yesterday, they're likely to take maybe seven years now. And saving thirty percent of the time it takes to roll out unimaginably awesome drugs is nothing to sneeze at.
Sneeze at it all you want, and for long as you want. We have a drug to cure that sneeze, and it is just around the corner.
WATTBA!