It is with drum rolls that DeepMind, Google’s artificial intelligence (AI) subsidiary, unveiled in the magazine Nature on May 8 AlphaFold 3, the new version of its software applied to biology. “Faster, more precise, capable of much more complex tasks”, rejoiced, during a press conference on May 7, Demis Hassabis, the director of DeepMind, about this new opus which makes it possible to model almost all proteins and their interactions, opening promising perspectives for the development of new drugs. But the reception given to him by the scientific community is more reserved. And for good reason: unlike the previous version, Google has kept the code of the famous algorithm secret.
Modeling the structure of proteins may seem trivial. However, the stakes are immense. Proteins are large molecules essential to the functioning of living organisms. These include antibodies, enzymes and hemoglobin, which transports oxygen in our blood. Proteins come in the form of long chains of amino acids, similar to strings of pearls. They quickly fold up in a complex way, resembling a bundle of cables that you have left tangled in your pocket. This folding is not random: each protein adopts a specific shape, which allows it to accomplish its function. A poorly formed protein will therefore not be able to fulfill its role. This is the case, for example, in sickle cell disease (or sickle cell anemia): a defect in the shape of hemoglobin reduces its ability to transport oxygen in the blood.
However, although we have access to the amino acid composition of any protein, determining its folding remains a long and complicated adventure. This requires months, even years, of laboratory work for a single protein… knowing that there are nearly 20,000 different ones in humans!
Disconcerting precision
The release of AlphaFold 2 in 2021 had the effect of a bomb. AI made it possible to predict the folding of any protein in less than twenty-four hours. Simply providing the amino acid sequence was enough for the program to predict the three-dimensional structure, often with astonishing accuracy and confidence. The secret ? Deep learning, or deep learning. The AI first trained on all proteins whose structure was known, and now predicts the folding of other proteins by analogy.
You have 64.39% of this article left to read. The rest is reserved for subscribers.
