Google’s DeepMind AI Breakthrough: Modeling DNA and Protein Structures

AlphaFold 3: Expanding the Horizons of⁣ Protein⁢ Structure Prediction

While⁢ Google⁣ has been actively promoting‍ its AI assistant,⁣ Bard, the company has also been quietly working⁣ on enhancing a more specialized ⁣AI tool. AlphaFold, a software developed by Google’s DeepMind AI​ unit⁣ to predict the 3D structure of proteins, has ⁣received a significant upgrade. ‍The latest version, AlphaFold ⁢3, can now model other biologically important molecules, including DNA, and the interactions between antibodies produced by the immune system and the molecules of disease organisms.

Borrowing Techniques from AI Image Generators

DeepMind incorporated new capabilities into AlphaFold 3 by borrowing techniques from AI image generators. Demis⁢ Hassabis, the head of DeepMind, explained that the company discovered that deep learning, a technique at the heart​ of many recent AI advances, could predict the shape ‍of proteins simply from their constituent amino acids by learning from ⁢experimentally verified structures.

The Evolution of AlphaFold

AlphaFold’s journey began in 2018 ⁣when Google DeepMind revealed it was working on AI software to accurately⁤ predict the shape of proteins. In 2020, AlphaFold 2 made a breakthrough⁢ by predicting the structures of proteins with unprecedented accuracy. The following year, DeepMind ⁢released an open-source ‍version of AlphaFold ‍for anyone to use, along with 350,000 predicted protein structures, including almost every⁣ known protein in the human ‌body. In 2022, the company further expanded its database​ by releasing more than‍ 2 million protein structures.

The Role of Diffusion Models

AlphaFold 3’s ability to model⁢ different proteins was improved in part through an algorithm called a diffusion model, which ⁢helps AI image‍ generators like Midjourney create unique and sometimes photo-realistic imagery. The diffusion model inside AlphaFold 3 sharpens the molecular structures the software generates⁣ by‍ learning from patterns in a collection of verified protein structures, much like how an​ image generator learns from real ​photographs to render realistic-looking images.

Confidence Scale and Limitations

While AlphaFold 3 represents a significant advancement, it is not perfect. The software offers a color-coded confidence scale for its predictions, with blue areas indicating high confidence and red areas showing less certainty.

The ​structure prediction performance of AlphaFold 3 is very impressive.

David Baker, a professor at the University of Washington⁤ who leads a group working on techniques⁤ for protein design, acknowledges the capabilities of AlphaFold 3. His own lab recently released a diffusion model to help​ model a wider⁢ range of molecular structures, but he ⁢concedes that AlphaFold 3 is more capable. However, Baker adds that it‌ is a shame that‍ the source code for AlphaFold 3 has not been released⁤ to‍ the scientific community.

The Potential for AI in​ Scientific Research

Hassabis, who leads ⁢all of Alphabet’s AI⁢ initiatives, has long been interested in ⁢the potential for AI to accelerate scientific research. He believes that the latest⁢ techniques being developed for AlphaFold, a ⁤highly⁣ specialized AI system, could prove ‍useful for building more general systems that aim to exceed human capabilities in many dimensions. As AI programs like Google’s Gemini become more capable over the next decade, Hassabis envisions them using‌ tools like AlphaFold to achieve other goals.