The past decade – and especially the last two years – have witnessed the recognition of the role of artificial intelligence and machine learning in biomedical science, and drug discovery in particular.

Yet, based on some headlines, a casual onlooker would be forgiven for thinking that drug discovery scientists are now simply sitting around, awaiting a panacea to be spewed out by their AI-models. This is, of course, nonsense.

Data science and machine learning require two main elements: large, sufficiently organized and annotated data sets to ‘learn from’, and a training framework (designed by a human) to allow the computer to learn from data and make predictions. Drug discovery – including molecular design – is one of the fields with the best foundation of data and machine learning which is why it’s poised to benefit from AI. Yet we are approaching a major plateau in the benefits we’ll see for drug discovery – one that cannot be overcome by smarter AI alone.

AI and the protein folding problem

A protein’s 3D structure provides immeasurable insight into normal function, pathogenicity, and – for drug discovery scientists – rational drug design. However, structural biology is complex, costly and technically challenging, which means that the ability to predict protein structure is essential.

In 1968, American molecular biologist Cyrus Levinthal stipulated that the 3D protein structure and folding pathway must be encoded within its amino acid sequence. Randomly sampling all conformations to reach a physiologically stable 3D fold would – remarkably – take longer than the age of the universe. Yet, proteins fold in milliseconds. With this realisation, the protein folding problem and the Levinthal Paradox emerged. Identifying the rules governing protein folding will have profound implications on biology and drug discovery.

And its clear data science and AI are having a huge impact here. The 2024 Nobel Prize for chemistry was awarded to the Alpha Fold team (along with David Baker and his team for developing Rosetta, a protein structure prediction computational method) for their computational prediction of countless hitherto unsolved protein structures. The excitement in the field is still felt two years later.

So, have these teams solved the protein folding problem? Sadly for all of us, no they have not.

To my knowledge, nor have these innovators made such a claim. Excited media coverage maybe responsible for the gross misunderstanding.

The great achievement is not actually that of being able to accurately predict every part of every protein. Far from it. Take c-Myc for example – perhaps one of the most well-known oncogenes, even referred to as the grand orchestrator of cancer – the AlphaFold 3 server is unable to predict its structure. The achievement of AlphaFold – to date at least – is that we can uncover structures that are similar to those we already know. Our technologies prior to AlphaFold were incapable of detecting such similarities.

While AlphaFold and similar algorithms are not, as yet, a magic bullet they nonetheless they can be profoundly useful. For example, my own  lab’s analysis shows that using the publicly available AlphaFold 2 models has doubled the number of druggable proteins available to the drug discovery community.