AlphaFold 3, how DeepMind's AI predicts the structure of life's molecules works

Credits: Google DeepMind.

Google DeepMind And Isomorphic Labs in an article that appeared in the scientific journal Nature they presented AlphaFold 3, an artificial intelligence system – now in its third version – which according to an official note from the Google DeepMind team “is able to predict the structure and interactions of all the molecules of life with unprecedented precision”, with a 50% improvement over previous forecasting methods. Inside every human, animal and plant cell there are billions of elements – proteins, DNA and other molecules – that work together through millions of combinations. To fully understand vital processes and find drugs capable of treating various pathologies, it is essential to understand how these elements interact with each other: this is precisely the task that AlphaFold 3 is called upon to perform. This technology may prove useful in encouraging the development of new drugs.

What AlphaFold 3 does and how it works

Getting into the nitty gritty, once the AlphaFold 3 algorithm is given an input list of molecules, it is able to generate their joint 3D structure, thus showing how these fit together. This AI model is capable of combining large biomolecules (such as proteins, DNA and RNA) and smaller molecules, also known as ligands, which play a leading role in many drugs. According to what was stated by Google DeepMind «AlphaFold 3 is able to model the chemical changes of these molecules that control the healthy functioning of cells, which if interrupted can lead to diseases».

But how does the model manage to do all this? The credit goes to its architecture, capable of covering all the molecules of life. This bases its operation on Pairformera “revised and corrected” version of the Evoformer module (used on AlphaFold 2). This new system deep learningafter processing the inputs, goes on to assemble its predictions using a diffusion network somewhat similar to those of “classic” AI tools for image generation. The process begins with what Google DeepMind calls a “cloud of atoms” that, in a series of steps, “converges towards its final, most accurate molecular structure.”

Again according to Google DeepMind, AlphaFold 3's predictions on molecular interactions exceed the precision of all existing systems. It should be noted, however, that the model's predictions are not infallible.

GIF of a rotating structure of a DNA-binding protein. Credits: Google DeepMind.

What AlphaFold 3 is for: the development of new drugs

Given the potential of AlphaFold 3, the Google DeepMind team expects that it can totally transform ours understanding of the biological world and, consequently, to development of new drugs. The tests conducted so far on the model give rise to hope, at least according to what was declared by the Google DeepMind team:

AlphaFold 3 achieves unprecedented accuracy in predicting drug-like interactions, including the binding of proteins to ligands and antibodies to their target proteins. AlphaFold 3 is 50% more accurate than the best traditional methods based on the PoseBusters benchmark without the need to enter structural information, making AlphaFold 3 the first AI system to surpass physics-based tools for predicting biomolecular structure.

To put it bluntly, Isomorphic Labs is already working on drug ideation for internal projects with some pharmaceutical partners using AlphaFold 3 to accelerate and improve the process.

It has also been fine-tuned AlphaFold Serverthe free platform made available to scientists all over the world through which it is possible to carry out non-commercial research using the power of AlphaFold 3. Below you will find the demonstration video of AlphaFold Server prepared by the Google DeepMind team.