Proteins, DNA and Drugs in One Model: AlphaFold 3 Predicts Complex Structures

The problem of predicting the folded shape of a single protein was largely solved with AlphaFold 2, but real biological processes happen when proteins interlock with DNA, RNA, small molecules and metal ions. Researchers at Google DeepMind and Isomorphic Labs reported in Nature AlphaFold 3, which predicts these complexes together with a single neural network.

Switching to a Diffusion Model That 'Generates' Structures

AlphaFold 3 substantially reworked the earlier architecture and adopted a diffusion-based architecture of the kind used in image generation. As a result, it outputs the coordinates of complexes that mix proteins, nucleic acids, small-molecule ligands, ions and chemically modified residues all at once. The authors reported that its accuracy for protein-ligand binding is much higher than existing docking programs, that it outperformed nucleic acid-specific predictors for protein-nucleic acid binding, and that antibody-antigen prediction also improved markedly over AlphaFold-Multimer version 2.3.

The significance is that it makes it possible to handle 'interactions', not just single molecules, computationally. Before any experiment, researchers can narrow down which compound enters a receptor in which pose, and which groove of DNA a transcription factor grabs.

Predictions Are Hypotheses; Experiments Do the Verifying

However, all results are computational predictions and do not tell binding strength or drug efficacy. Errors can be larger for molecule types with little training data, for flexible regions, and for unusual chemical structures, and structural biology experiments are needed to confirm them.