AlphaFold 3 AI Just Won The Nobel Prize!

We have partnered with Google DeepMind for this video to celebrate the launch of a follow up to one of the best papers ever written, AlphaFold. Now version three is here and I am out of words. I'll try my best to explain why. What you are seeing here is history in the making. So what is all this about? Well, this work is about protein folding and more. You'll see in a moment. A protein is a string of amino acids. These are the building blocks of life. This is what goes in, which in reality has a 3D structure. And guessing that structure is protein folding. Letters go in and a 3D structure comes out. That is brutally hard. And no one could produce a good enough solution before until AlphaFold. However, wait. A protein goes in and this structure comes out. But what is all this good for? Well, look at this follow up paper. This project was enabled by Alphafold and is about evolving enzymes. What kind of enzymes? Well, these are enzymes that are made to be able to digest these plastics. Yes, that's right. It is breaking these plastics down. But it gets better. It even helps recycling them so they can be used over and over again infinitely like glass or tin cans. We don't need to produce more of these, which requires fossil fuels. And we can clean up these landfills with it. This sounds like a science fiction movie. And it came right out of a paper that was published in 2021. And this was just one citation, one follow up research paper that used this work. And hold on to your papers fellow scholars, because it now has 21,000 citations. Holy Mother of papers. Just imagine all the ways this work already changes the world and saves lives. And it gets better. That was only AlphaFold2 and now and AlphaFold3 is here and it just got even better. So how could it possibly get better? Dear fellow scholars, this is two minute papers with Dr. Zsolnai. So this just came out today and I had the honor of spending a little quality time with this paper earlier than its release. And here is what I have found. By the way, this might be the first video you see on this anywhere. So is it better on proteins? Yes it is. It is marginally better in most categories. However, on protein antibodies, its accuracy has more than doubled. That is an absolutely amazing improvement worthy of a legendary paper all by itself. However, that is not all. Not even close. This chart shows perhaps something even more insane. Ligands are small molecules that interact with proteins. They can have complex chemistry and a huge amount of medical and biological research involves these. In fact, the majority of drugs are small molecules like these. There are specialist physics based systems that try and predict the interactions of proteins and ligands. And there are machine learning based systems that try to do the same thing too. Previously, AlphaFold was limited to only proteins and could not say anything about other molecular structures such as ligands. And oh my goodness, what happened here? AlphaFold3 can now not only deal with these molecules too, but it also beats these specialist systems at their own game. And not by a little. For the first time, an AI based system is outperforming the physics based systems that are widely used in the industry. This is crazy and it will change everything. So this game is not just about proteins anymore. Today the 3D structure of ligands, ions and even DNA and RNA can also be predicted with the accuracy that is way beyond the capabilities of previous methods. This new work can thus help develop biorenewable materials, more resilient crops, accelerate drug design and genomics research. And once again, as we have seen so many times recently, this is a multitasking AI that outperforms specialists. Like just recently we saw sima, a technique that can play loads of games and outperforms specialist individual game playing agents for each game. The way it does this is different from AlphaFold 3 and it is playing across different environments. But my takeaway is the same generalist AI models are going to be better than specialist models at their own game. I am completely stunned by this. So what changes were necessary to pull this off? Well, a great deal has changed. For instance, the previous version, AlphaFold 2 had an Evoformer module that imagines this protein folding problem problem as a graph. You can imagine such a graph as a social media website where each person is a node in the graph and they are connected if they know each other. However, here nodes represented the amino acid residues in the protein. These are the people where the connections represented their relationships, much like in a social media graph. Now this evil former has been switched out to a new and simpler module called pair former. If we pop the hood and look inside, we see this. We will talk a little more about this in a moment. There is also a new diffusion module, somewhat similar to the ones that you see in text to image models that start out from a bunch of noise and over time reorganize this noise to create a beautiful image. However, here Instead it creates 3D molecular structures. So cool. So these are the pieces that enable all this AI magic to happen. And this is the corner of the Internet where we look at bar charts and flip out together. Welcome to two Minute Papers. And this work is a start of a revolution. Wow. We now see that one single learning based system over time can do more and more and perhaps one day we will even have just one AI that can perform every task for full drug discovery. This is a paper that is likely already saving lives and will save many more lives in the future. Now for this video we have partnered with Google DeepMind and of course we will still talk about the limitations of this technique, even more so than usual. So what do I see here from the previous paper? What I see is not just a transformer neural network being thrown at a problem. No, no, what I see is plenty of hand engineering, lots of little additions to the algorithm, each of which adds only a little more accuracy. No silver bullet here, but lots of small moving parts. And the new version of AlphaFold 3 also inherits many of these moving parts. The new paper is a step in the right direction. It is a little simpler, but I am hoping that over time we will find a simpler and more elegant technique to perform these tasks. Now, two more limitations. One, this can only predict static structures. Unfortunately it cannot capture more dynamical behaviors. 2. The diffusion module starts out from a bunch of noise and this noise is generated randomly. And the technique shows a little sensitivity to that. This means that different runs might provide different solutions, some of which will be less accurate. This can be mitigated by running it a few more times from different starting points to improve accuracy. And one more thing, this is AlphaFold Server, a way for you to play with AlphaFold 3 for free. If you are not a biologist. Don't worry for a second. There are plenty of presets to choose from to have a look at and it has everyone's favorite dark mode. The link is available in the video description. So now I hope you fellow scholars also feel what I feel when reading this incredible paper. And that is why I have a framed copy of the earlier AlphaFold paper in the 2 Minute Papers lab to remind me that we are living in a world where all this is possible. What a time to be alive. And don't forget they are giving away all of this for free. What an amazing gift for humanity. Also, what applications of AlphaFold would you fellow scholars like to see in in the world? Let me know in the comments below. I will show the best ones to the Authors and Google DeepMind. They'd love to hear your thoughts.