Drug development in the AI era: AlphaFold 3 is coming!

For decades, researchers have sought a convenient, reliable, and fast track to drug development. Usually, it takes about 12–15 years from the initial drug discovery stage to the point at which the drug gets approval by national drug administration agencies. Moreover, when we consider the costs of both successful and failed programs, it costs about 2.5 billion US dollars to bring a drug to market (Figure 1). How to reduce the time and costs required for drug development is a subject that we are always thinking deeply about.

Generally, drug development will undergo several specific steps. It often starts with the discovery stage, when we identify the biological target responsible for a disease, possibly a protein receptor or an enzyme, and then perform screening for molecules that might interact with the target. Once we get the resulting candidates, we further work to improve their activity and reduce any associated side effects. If this is successful, we enter the next stage, preclinical testing, which helps to understand how a drug candidate is transported and metabolized in an animal’s body and answers questions of safety and the dosage required for approval for clinical trials.

In the past years, we have used structural biology methods, including crystallography and cryoelectron microscopy (cryo-EM), and biophysical methods to investigate drug-target interaction details at atomic levels, which take more time and cost more. When the deep neural networks AlphaFold 2 (AF2) and RoseTTAFold (RF) came out in 2021, they raised a revolution in the modeling of protein structures and their interactions, and in a faster and cheaper way, they enable wide applications in the field of protein design and medicine. However, AF2 and RF have their limitations in the modeling of complexes, with low accuracy in structure prediction. Recently, in early 2024, the introduction of the AF3 model with substantially updated diffusion-based architecture1 and RF All-Atom (RFAA)2 has greatly changed this situation, and the model enables the joint structure prediction of complexes including proteins, nucleic acids, small molecules, ions, and modified residues. The AF3 and RFAA models have taken a large step toward understanding the complex atomic interactions of biological systems.