Proteins are long chains of amino-acid residues that fold into specific shapes. Properly folded proteins function normally whereas misfolded ones can lead to debilitating diseases. Since these chains are quite long, a given protein can actually fold into one of a very large number of shapes — yet it makes a beeline for a specific shape while avoiding all the others. How and why this happens constitute an important mystery in structural biology called the protein-folding problem. In 2018, five decades after it was mooted, a Google subsidiary named DeepMind developed a purpose-built AI tool to predict the shapes into which different proteins could fold, called AlphaFold. The upgraded AlphaFold 2 followed two years later. Many scientists and technologists acknowledge that these two deep-learning systems have transformed human awareness of protein structures, a feat the machines demonstrated in the biennial Critical Assessment of Protein Structure Prediction contest. Recently, DeepMind launched AlphaFold 3, which can reportedly predict the shapes with nearly 80% accuracy as well as model DNA, RNA, ligands, and modifications to them. As with the first two AlphaFolds, no. 3 is great for being able to elucidate the folded proteins’ structures in seconds rather than the years humans have required with advanced microscopic techniques.

Not surprisingly, the excitement that followed the release of AlphaFold 3 has been unable to escape the hype and overblown expectations that dogged the launches of its predecessors. These machines can predict protein structures with relatively high accuracy but they cannot say why they are folded that way; this is still the task of human scientists. How the AlphaFolds will catalyse drug discovery is also unclear. Many drugs fail to make it to the market from the laboratory because medical researchers are unable to anticipate all the interactions between the drugs’ various components and various parts of the body. The protein-folding problem is important to crack but it will not magically improve drugs’ chances in human clinical trials. It is a step in that direction. Finally, the free use of AlphaFold 3 is limited while its inner mechanisms are unavailable for public exploration or scrutiny, so far. While the motivation to innovate of DeepMind is laudable, the cutting-edge value AlphaFold 3 presents to health care means the company should explore alternative revenue models in which the system is not trapped behind paywalls or exorbitant prices — a fate that has already befallen scientific papers and medicines born of publicly funded research. Recall that the AlphaFolds’ training data itself includes protein structures first elucidated by such research.