Google DeepMind, Google LLC’s artificial intelligence research unit, today detailed a new version of its AlphaFold neural network for biologists.

One of the factors that influence the behavior of a biomolecule such a protein is its physical shape. Historically, mapping out molecules’ shape required years of highly complicated research. Google’s AlphaFold family of artificial intelligence models can significantly speed up the task by automating manual work for scientists.

The first iteration of the model that the company debuted in 2021 focused on one specific type of biomolecule: proteins. AlphaFold 3, the latest version of the AI that Google DeepMind detailed today, can also predict the structure of DNA, RNA, ligands and other biological building blocks. Moreover, it does so with higher accuracy than existing methods.

In one internal test, Google DeepMind researchers used AlphaFold 3 to predict the structure of an enzyme comprising a protein, simple sugar molecules and an ion, an atomic with an electric charge. This enzyme is found in a type of soil-borne fungus that can damage plants. According to Google, the findings uncovered by AlphaFold 3 could potentially help researchers develop healthier crops.

Besides estimating the shape of biomolecules, the model can also predict how they will behave under different conditions. “AlphaFold 3 can model chemical modifications to these molecules which control the healthy functioning of cells, that when disrupted can lead to disease,” Google DeepMind researchers explained in a blog post.

The search giant carried out a series of evaluations to compare AlphaFold 3 against existing molecule analysis methods. According to Google, the AI proved to be at least 50% better than earlier technologies at predicting proteins’ interactions with other molecules. When analyzing certain types of chemical interactions, AlphaFold 3 can provide a more than 100% increase in accuracy.

“AlphaFold 3 is 50% more accurate than the best traditional methods on the PoseBusters benchmark without needing the input of any structural information, making AlphaFold 3 the first AI system to surpass physics-based tools for biomolecular structure prediction,” Google DeepMind detailed.

The search giant achieved that accuracy improvement by enhancing a component of AlphaFold 3 known as the Evoformer. This module, which was carried over from the previous iteration of the model, performs several of the initial steps involved in predicting a biomolecule’s structure.

The Evoformer implements a so-called attention mechanism, a feature associated with large language models. An attention mechanism allows a neural network to take a large number of contextual details into account when interpreting a piece of data. Additionally, Google DeepMind equipped the Evoformer with features that allow it to carry out “direct reasoning about the spatial and evolutionary relationships” of molecules.

AlphaFold 3 combines the module with a diffusion network. This is a type of AI most commonly used for image generation tasks. According to Google, the diffusion model starts the molecule analysis process with a digital object that represents a cloud of atoms. It then turns this cloud into the structure of the biomolecule being studied through a gradual, multistep process.

The search giant is making AlphaFold 3 available to scientists through a new cloud service called AlphaFold Server. According to Google, it provides access to most of the model’s features at no charge. 

In conjunction, Alphabet Inc.’s Isomorphic Labs drug design unit is using the AI to support its research efforts. The subsidiary is collaborating with multiple pharmaceutical companies to identify ways that AlphaFold 3 could speed up their work. Isomorphic Labs will also apply the AI to several internal research projects. 

“AlphaFold 3 brings the biological world into high definition,” Google’s researchers wrote. “It allows scientists to see cellular systems in all their complexity, across structures, interactions and modifications.”

Image: Google DeepMind