Your DNA contains roughly three billion letters, but reading them doesn't tell us everything about you. Hidden among them are millions of little molecular switches to control when and where your genes become active. And many diseases begin where the switches go wrong. And DeepMind's new AI technique, Alpha Genome 2, is finally able to predict what these little switches do, but with a twist. Okay, so what is this exactly? With Alpha Genome 1, you give it a piece of DNA and you change a letter. Is nothing going to happen, or could this tiny mutation switch on a gene that drives terrible diseases like leukemia, or could it alter the genes linked to dangerously high cholesterol? The previous Alpha Genome 1 AI can predict exactly this kind of effect, which is basically insane. I can't believe that I'm saying this. Meanwhile, my Bluetooth headphones still can't connect reliably. So there's that. However, there is a problem. The same mutation can have a completely different effect based on where it is, which cell it is in. For instance, your brain and skin contain the same DNA, but they use it completely differently. And Alpha Genome 1 struggles to predict these finer differences. Of course, because it's kind of impossible. There are billions of single letter mutations and hundreds of different cell types. Unfortunately, that is trillions of possible combinations. Testing all of these in a lab is completely hopeless. So how on earth would you predict all this from just a sequence of DNA? All this knowledge from a bunch of letters? Well, for example, they tested mutations known to cause dangerously excessive insulin production. And finally, Alpha Genome 2 predicted that these mutations on the HK1 gene in insulin producing beta cells, but have almost no effect in other pancreatic cells. But wait a second. This problem might be exceptionally well suited for a powerful modern AI. Why? Because we can create input output pairs to learn on. The input could be a piece of DNA from the pancreas and a pancreatic beta cell as a cell type. So what is the output? Well, how active its genes are. But wait, scientists have already measured this in a stupendously large number of cells. So we have tons of training data and it turns out we can use that data for things other than serving ads to sell more Nissans. Great. Now if you do that, there are two big questions. One, is AI powerful enough to learn these interactions? And if yes, question number two, can it even work with cell types it hasn't seen yet? So can it generalize? Can it go beyond the training data? Is that even possible? Dear fellow scholars, this is two minute papers with Dr. Károly Zsolnai-Fehér. The paper is 78 pages. It is a lovely read, perfect bedtime reading if you don't intend to sleep. And I am not an expert. I am just a student. I could be wrong, but I am trying my best. So let's have a look. In one test, Alpha Genome Two could identify double the number of mutations known to affect gene activity with 90% precision. For every 10 mutations the AI marked, about nine were likely to affect gene activity. That is incredible. Now question two, does it generalize? Now hold on to your papers, fellow scholars, because the answer is yes. In a Parkinson's disease study, a previous specialist AI could correctly predict whether gene activity went up or down for 17% of the known genetic effects, and AlphaFold two had never trained on that study. So, how many could it predict? What? Are you seeing what I am seeing? 41% more than double, both at the same high level of accuracy. So this AI performed better on completely new information than a specialist AI that was trained on this information. Folks, this is science fiction level of goodness. Absolutely insane work, okay? But we do the real talk here, limitations. It is far from perfect. Even its best predictions still miss many disease-related mutations, and it's still not great at rare cell types. So, please do not walk away from this video thinking that they solved genetic diseases. Not even close. But when AlphaFold first came out, I remember thinking that it had all the makings of a Nobel Prize. And looking at this new work, I wouldn't be surprised if DeepMind wins another. I would like to congratulate Sir Demis Hassabis and Pushmeet Kohli and the whole team. Sometimes the difference between healthy and sick is just a single letter, and understanding that letter can change millions and millions of lives. And I [clears throat] hope that this work will save us from terrible diseases in the future. And that is the best application of AI ever. Oh, and the best part, AlphaFold two is already available to researchers for free through an API, and DeepMind plans to release the code and model weights too. What a time to be alive. And I have a recent interview with the creator Pushmeet talking about AlphaFold one in the video description if you are interested. And subscribe for more. I use Lambda to reproduce AI research papers often in minutes. It's great to train your own models, fine-tune an existing one, run inference, text-to-image or video, easy-peasy. Running a deep seek chatbot or agent, super fast, super reliable. Lambda gives you powerful NVIDIA GPUs to run your own experiments. I test ideas from papers I cover and moments later, results. Love it. Seriously, try it out now at lambda.com/papers.
Generated algorithmically for Search Engine
Indexing.