Mobile Menu

AlphaFold 3: Stepping into the future of structure prediction

We’ve all heard about the potential of artificial intelligence in the life sciences field. In 2020, the launch of AlphaFold 2, pioneered by Google DeepMind, took the world by storm and marked a new age in protein structure prediction. But now, AlphaFold 3 is transforming the landscape again. In this news highlight, we explore the new tech, compare it to its predecessor and take a look to the future.

What is AlphaFold?

Before the AI revolution, protein structure prediction heavily relied on experimental methods, such as X-ray crystallography, NMR spectroscopy and, later, some complex computational methods like homology modelling. These methods were time consuming and costly, and were a major limiting step in drug discovery and development processes in particular. For years, scientists have been attempting to integrate the latest and greatest AI models into the field, in order to speed up the process and improve accuracy.

Enter AlphaFold, an artificial intelligence tool developed by Google’s DeepMind. The first version of the technology was released in 2018, but it was 2020’s AlphaFold 2 that made headlines – winning the prestigious Critical Assessment of Structure Prediction (CASP) 14 competition. Having gone through multiple major iterations, the most recent release, AlphaFold 3, is set to further transform the protein space. But what does it do, and how may it outperform its predecessor?

A comparison

AlphaFold 2 achieved remarkable success in predicting the 3D structures of proteins from their amino acid sequences. It gained notoriety for its performance in the CASP 14 competition, where it significantly outperformed other methods, demonstrating the potential of powerful AI in solving a decades-long challenge.

AlphaFold 2 utilised a deep learning architecture called Evoformer. This model has been instrumental in advancing research into protein structures and folding mechanisms, leading to breakthroughs in subsequent areas such as drug discovery and vaccine development. The model uses a multiple sequence alignment, which compares sequences found throughout living organisms, and a pair representation based on proteins that may have a similar structure.

But what about AlphaFold 3? DeepMind state that the Evoformer model has been updated in this iteration of the technology. The model is now ‘simpler’, and focuses more on the pair representation. In addition, after processing inputs, AlphaFold 3 uses a ‘diffusion network’, described by DeepMind as being similar to the tools used in AI image generation. These improvements to the model have led the company to claim that AlphaFold 3 has the highest accuracy yet.

Expanding the scope

Perhaps the most exciting comparison that can be drawn between AlphaFold 2 and 3 is the latter’s ability to make predictions not just about protein structure, but about interactions between proteins and biological molecules such as DNA, RNA and ligands – an acknowledgment that no molecules operate on their own. These interactions are crucial in understanding the underpinnings of life, and in identifying potential drug candidates. This is reflected in AlphaFold 3’s large training set, which includes a wider range of molecules.

How will this technology be used? Through a collaboration with Isomorphic Labs, the tool is now finding its way into the commercial sphere, due to its improved ability to upscale the drug development process. Plus, DeepMind have also released the AlphaFold server, which they describe as ‘the most accurate tool in the world for predicting how proteins interact with other molecules throughout the cell.’ The server is free to use for non-commercial purposes, providing researchers with a fast-paced, accurate tool. The company are also aiming to expand access and education in the Global South, ensuring that the benefits of this technology can be felt globally.

Controversial publication

AlphaFold 3 was released two weeks ago, with a corresponding paper published in Nature on the same day. However, both DeepMind and Nature came under fire following the publication, due to the fact that the authors had chosen to publish only pseudocode. Nature have defended their editorial decision, but an open letter questioning verifiability and reproducibility has convinced DeepMind to release the model for academic use within the next few months. Nature has since elaborated on the initial decision to allow the publication without full code, but has acknowledged the need for better conversation surrounding these matters.

Where next?

Although the release of AlphaFold 3 may feel like a huge jump from where we were just a few short years ago, it’s likely that we’ll continue to see improvements to this technology in the not-too-distant future. Even with AI tools like AlphaFold, the latter stages of the drug discovery journey, in particular, still take a significant length of time. Whilst the emergence of new AI models will cut the time and cost required to an extent, there is a still a huge amount of work to be done before we truly see a transformation of the journey from bench to clinic.

Are you planning to use AlphaFold 3 in your work, or have something you’d like to share with us? We’d love to hear your thoughts on our new free, digital community, FOG Circle!

References and further reading

Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature (2024).

Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).

Google DeepMind AlphaFold Team and Isomorphic Labs. AlphaFold 3 predicts the structure and interactions of all of life’s molecules. (2024). Available online at: