Functionally annotating cysteine disulfides and metal binding sites in the plant kingdom using AlphaFold2 predicted structures

Deep learning algorithms such as AlphaFold2 predict three-dimensional protein structure with high confidence. The recent release of more than 200 million structural models provides an unprecedented resource for functional protein annotation. Here, we used AlphaFold2 predicted structures of fifteen plant proteomes to functionally and evolutionary analyze cysteine residues in the plant kingdom. In addition to identification of metal ligands co-ordinated by cysteine residues, we systematically analyzed cysteine disulfides present in these structural pre-dictions. Our analysis demonstrates most of these predicted disulfides are trustworthy due their high agreement (-96%) with those present in X-ray and NMR protein structures, their characteristic disulfide stereochemistry, the biased subcellular distribution of their proteins and a higher degree of oxidation of their respective cysteines as measured by proteomics. Adopting an evolutionary perspective, zinc binding sites are increasingly present at the expense of iron-sulfur clusters in plants. Interestingly, disulfide formation is increased in secreted proteins of land plants, likely promoting sequence evolution to adapt to changing environments encountered by plants. In summary, Alphafold2 predicted structural models are a rich source of information for studying the role of cys-teines residues in proteins of interest and for protein redox biology in general.

Automated summary

Deep learning algorithms, such as AlphaFold2, accurately predict the three-dimensional structure of proteins, providing an unparalleled resource for annotating protein function. This study analyzed the function and evolution of cysteine residues in plants using AlphaFold2 predicted structures. The analysis identified metal ligands coordinated by cysteine residues and systematically analyzed cysteine disulfides. The results showed that most of the predicted disulfides were reliable, with high agreement with known protein structures and characteristic features.