Predicting Metalloprotein Redox Potentials with Machine Learning: A Focus on Iron–Sulfur Systems

Iron-sulfur (Fe-S) proteins play crucial roles in key biological processes by mediating electron transfer through finely tuned redox potentials (RP). However, predicting RP from protein structure remains challenging due to the complex electronic properties of Fe-S clusters and their interactions with protein environment. In our study, we developed a machine learning (ML) framework to efficiently and accurately predict RP values of Fe-S proteins. By designing 3D structure-based molecular descriptors across multiple spatial scales (from local atomic environments to global protein properties) we trained gradient boosting models (XGBoost), achieving a mean absolute error below 40 mV, outperforming conventional quantum mechanical approaches. Then, we identified key structural features driving RP variation, including steric hindrance, backbone flexibility, and ligand composition. Our interpretable and scalable model lays the foundation for high-throughput RP prediction and rational protein design across diverse metalloprotein classes. Current efforts focus on extending this framework to other Fe-S cluster types and evaluating alternative ML architectures to further improve predictive performance and generalizability.