London based Google DeepMind and Isomorphic Labs have announced Alpha Fold 3, the latest iteration of their groundbreaking AI system designed to predict the 3D structures of proteins — and now, much more.
Predicting protein structure from amino acid sequences has challenged scientists for decades, as function depends on shape. Solving this crucial problem accelerates everything from disease understanding to drug design. We wrote about the generation of genetic sequences back in week 9 with the Evo model, and now AlphaFold provides scientists with the ability to see how these fundamental structures of life actually interact, albeit only in static form at this stage.
AlphaFold 2 changed the game in 2020 by predicting protein structures with unprecedented accuracy. Thee system provided hundreds of thousands of protein structures for researchers worldwide, boosting studies in fields as diverse as genetics and biochemistry. This new version incorporates a diffusion model — a concept borrowed from AI systems designed to generate images like the pictures at the top of our news articles, and others created by the likes of Dall-E or Midjourney. These models work by gradually refining their guesses, starting from a randomly noisy image (or in this case, a molecular structure) to produce detailed and accurate pictures. In AlphaFold 3, this approach means the software can now model not just proteins but also DNA, RNA, antibodies, and even tiny molecules like metal ions. This ability to handle a broader range of biological molecules means AlphaFold 3 can predict how all these different entities interact. This is crucial for understanding complex biological processes and diseases at a molecular level.
AlphaFold 3’s enhanced capabilities allow researchers to see how a potential drug fits into a protein, how strong that interaction is, and what unintended targets the drug might also affect. This insight is invaluable for developing more effective and safer medications. The implications of these advancements are huge. For instance, by understanding how proteins interact with DNA, researchers can explore fundamental questions about cellular functions, such as how cells repair damaged DNA and how various diseases arise from these processes going awry.
Aside from the practical value, Google’s AI chief Demis Hassabis says could be worth north of $100 billion to the firm. While AlphaFold 3 is a significant step forward, it’s not without its open questions. DeepMind has previously been criticized for overhyping their scientific discoveries, as seen with their AI system called, Gnome (Graph Networks for Materials Exploration) which was purported to have discovered 380,000 new materials. This was subsequently challenged by experts for the lack of evidence provided on the utility and credibility of its newly predicted compounds. One other decision is that is likely to draw much criticism is that AlphaFold 3 will not be open-sourced. DeepMind has made it accessible for free via the cloud to researchers worldwide. This approach aims to balance the need for openness with the complexities of developing such advanced tools, and no doubt helps to manage the potential negative dual uses in creating dangerous pathogens or toxins. Anyone can try AlphaFold Server here.
Takeaways: The use of diffusion models in AlphaFold 3 illustrates the incredible versatility of modern AI. Diffusion can create photorealistic images from pixel noise, and it seems it can also figure out complex molecular structures from limited data. This echoes the re-usability of other key AI concepts such as vector embeddings, high dimensional numeric representations that allow systems to understand anything from words to sounds and pictures to robotic telemetry.
Dr Jim Fan, embodied AI lead at Nvidia posted on X: “We live on a timeline where learnings from Llama and Sora [OpenAI’s video model] can inform and accelerate life sciences. I find this level of generality absolutely mind-boggling. The same transformer+diffusion backbone that generates fancy pixels can also imagine proteins, as long as you convert the data to sequences of floats accordingly. We are not there yet at a single AGI [human-level] model, but we have successfully built a menu of general-purpose AI recipes that transfer training, data, and neural architectures across domains. This should not work, but thank god it does!”
This adaptability shows how AI can be a powerful tool across many domains, not just biology, by choosing the right approach for the right task, many things are possible.