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

Area of Science:

• Biochemistry and Biophysics
• Computational Biology
• Bioenergetics

Background:

• Iron-Sulfur (Fe-S) proteins are crucial for numerous biological processes, including energy conversion and DNA repair.
• Their function relies on finely tuned reduction potentials (RP) determined by metal cofactors, but predicting RP from structure is challenging.
• This difficulty impedes systematic modulation of RP for protein design.

Purpose of the Study:

• To introduce FeS-RedPred, a Machine Learning (ML) framework for accurate and scalable prediction of RP in Fe-S proteins.
• To provide a tool that aids in understanding the determinants of RP and guides protein engineering efforts.
• To enable high-throughput prediction of redox potentials for diverse metalloprotein families.

Main Methods:

• Developed a Machine Learning (ML) framework, FeS-RedPred, utilizing Extreme Gradient Boosting (XGB) models.
• Employed structure-derived molecular descriptors computed at multiple spatial scales (local to global).
• Focused on mono- and binuclear Fe-S clusters (e.g., rubredoxins, [2Fe-2S] ferredoxins) with available data.

Main Results:

• Achieved a mean absolute error of approximately 40 mV in RP prediction, competitive with state-of-the-art methods.
• Demonstrated a highly efficient balance between predictive accuracy and computational cost.
• The model provides insights into the key determinants of RP, facilitating interpretation.

Conclusions:

• FeS-RedPred offers a valuable foundation for understanding metalloprotein redox behavior.
• Enables high-throughput prediction of redox potentials, informing data-driven protein design.
• Advances the field of bioenergetics and human health by improving our ability to engineer Fe-S proteins.