May 8, 2024

expert reaction to the latest version of DeepMind’s AlphaFold

A study published in Nature looks at the new version of AlphaFold for protein structure predictions.

Dr Nicole Wheeler, Birmingham Fellow, Institute of Microbiology and Infection, University of Birmingham, said:

“Since the release of AlphaFold in 2020, protein structure prediction has underpinned a wide range of advances in the biological sciences and biotechnology. Being able to predict the structure of complex biological parts helps us understand what function they perform and how, and helps us engineer better versions of existing biological machines or invent new biological parts that perform new functions.

“It’s impressive to see the more generalised modelling approach taken by AlphaFold 3 outperform a diverse range of specialised applications built on AlphaFold 2. This finding indicates we still have a lot of progress to make in “foundation” models in biology. AlphaFold 2 itself required a number of specialised modules that leverage expert knowledge of the physics and chemistry of proteins to produce good results, but AlphaFold 3 has managed to abstract much of this away by using a generative AI component, an area of AI that is also producing impressive results in producing images, video, and other types of data.  

“The new AlphaFold 3 has extended its capabilities to a wide range of biomolecules, like DNA and small molecules, indicating that the generalisability of AI in bio is expanding, similar to what we’re seeing in other multi-modal models like those that can model text, images and sound in one framework. This offers a lot of promise in expanding what we can do with these AI tools for understanding and engineering biology, like designing biological parts to control the expression of genes or designing small molecules to treat disease.

“Unlike general trends toward adding more parameters at each new iteration of a model to achieve higher performance, AlphaFold 3 actually reduces the number of steps and computational complexity of their model. This is an important advance as the US government start placing reporting requirements on models that use a lot of compute, and people become more conscious of the carbon footprint of training models with increasing levels of complexity.

“The fact that the training data has been enriched by AlphaFold-Multimer predictions demonstrates the potential for new generations of AI to be built on training data generated by the current generation of AI. The potential to reduce the problem created by hallucinations of non-functioning proteins that pose challenges for the downstream viability of protein designs is also exciting, creating the possibility that people using these tools to design new biological parts won’t need to spend as long testing designs that won’t work. Physically producing and testing biological designs is a big bottleneck in biotechnology at the moment, so this is very encouraging for the prospect of rapidly prototyping biological parts for new applications, ranging from medicines, to food, to environmental applications.”

‘Accurate structure predictions of biomolecular interactions with AlphaFold 3’ by John Jumper et al. was published in Nature at 16:00 BST Wednesday 08 May.

DOI: 10.1038/s41586-024-07487-w

Declared interests

For all experts, no reply to our request for DOIs was received.