Evaluation of AlphaFold3 for Predicting Human Heme-Binding Protein Structures

Heme is a crucial cofactor involved in various biological processes, such as electron transport, catalysis, and oxygen binding. Understanding the binding of heme to heme-binding proteins (HBPs) is essential for clarifying their functions and molecular mechanisms and for applications in enzyme engineering and therapeutic development. AlphaFold3 (AF3), an artificial intelligence (AI)-based macromolecular prediction tool, has been applied to the structural modeling of various proteins, including HBPs. Nevertheless, whether AF3 provides reliable structural information for HBPs has not yet been investigated. To determine the AF3 predictions for HBPs, the apo and holo states of four human HBPs, including the cytochrome b5 domain of sulfite oxidase (SO-b5), the cytochrome b5 domain of NADH cytochrome b5 oxidoreductase (Ncb5or-b5), cytochrome b5 type B (CYB5B), and neuroglobin (NGB), generated by AF3, were examined and compared with experimental HBP structures. The overall positions of the heme molecules were well docked into the heme-binding pockets of HBPs; however, there were differences in the heme-binding configurations, including pocket geometry and coordination environment, which are crucial for functional interpretation. The experimental NGB structure contains a disulfide bond near the heme-binding region, whereas the AF3-predicted model lacks this bond, causing differences in local folding that affect the heme-binding environment. Molecular dynamics simulations demonstrated that the AF3-predicted NGB structure exhibited distinct molecular conformations and flexibility compared with the experimental structure. These data indicate both the potential and the limitations of using AF3-predicted structures to model the heme-binding states of HBPs.