A paper presents AF3-ReD, a method for finding multiple protein conformations with AlphaFold3

A paper presents AF3-ReD, a method for finding multiple protein conformations with AlphaFold3

On August 26, JACS Au published a paper on AF3-ReD. The method modifies the AlphaFold3 inference procedure: each new run is steered away from similar structures found in previous runs, while the neural network parameters remain unchanged.

Some proteins function by switching between several conformations. ATP binding moves the β-subunit of F1-ATPase, part of a molecular motor, from an open conformation to a closed one. In transporter proteins, different conformations alternately expose a pathway for a molecule to one side of the membrane and then the other. A single most probable structure cannot represent the full transition.

AlphaFold3 generates possible atomic arrangements for a complex between a protein and a ligand, a molecule that binds to the protein. Its diffusion model starts from a random arrangement of atoms and removes noise step by step. Repeated runs often converge on the same conformation. AF3-ReD retains similar structures obtained in earlier runs and adds a mathematical repulsion from them during the next calculation. This encourages the model to search for other possible conformations without changing the neural network itself.

The authors tested whether the additional predictions resembled states already identified experimentally. In the β-subunit experiment, standard AlphaFold3 predicted the open conformation even when ATP was bound to the protein. AF3-ReD produced open, closed, and intermediate conformations that were close to experimental structures. For comparison, the authors reduced the number of sequences in the multiple sequence alignment, which is a set of related proteins from which the model derives information about their evolution. This approach also found closed conformations, but some predictions differed from both experimentally known conformations.

The transporter proteins OxlT and NarK move molecules across membranes. When the substrate is bound, the protein blocks access to it from both sides. A reduced multiple sequence alignment broadened the set of predictions, while AF3-ReD identified conformations close to known structures of this state and to structural ensembles obtained through molecular dynamics, a computational method for simulating atomic motion.

The authors propose using the intermediate conformations as starting points for molecular dynamics calculations. Structure prediction can then serve as the first step in studying the transition between states of a protein bound to a ligand.