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

## TL;DR

This paper introduces a machine learning model to predict redox potentials in iron-sulfur proteins, aiding in understanding and designing these crucial biological systems.

## Contribution

The novel contribution is FeS-RedPred, a machine learning framework for accurate and efficient prediction of redox potentials in iron-sulfur proteins.

## Key findings

- FeS-RedPred achieves a mean absolute error of ∼40 mV in predicting redox potentials.
- The model uses structure-derived molecular descriptors across multiple spatial scales.
- It offers insights into the determinants of redox potentials for guiding protein engineering.

## Abstract

Iron–Sulfur (Fe–S) proteins play essential
roles
in a wide range of biological processes, from energy conversion and
respiration to DNA repair and redox signaling, making them highly
relevant to both bioenergetics and human health. These proteins mediate
electron transfer through finely tuned reduction potentials (RP) defined
by their metal cofactors. However, predicting RP from protein structures
remains a significant challenge due to the complex electronic nature
of Fe–S clusters and their intricate coupling with the surrounding
protein environment. This complexity limits our ability to systematically
modulate RP, hindering efforts in high-throughput and rational protein
design. In this study, we introduce a Machine Learning (ML) framework,
FeS-RedPred, for accurate and scalable prediction of RP in Fe–S
proteins. We focus on mono- and binuclear clusters, such as rubredoxins
and [2Fe–2S] clusters of ferredoxins, Rieske, and mitoNEET-type,
which serve as ideal model systems thanks to the availability of abundant
structural and electrochemical data. Our approach relies on structure-derived
molecular descriptors computed across multiple spatial scales, from
local atomic environments to global protein-level features. Using
Extreme Gradient Boosting (XGB) models, we achieve a mean absolute
error of ∼40 mV, which is competitive with state-of-the-art
computational approaches, while also providing a highly efficient
compromise between accuracy and computational cost. Beyond predictive
accuracy, our model also offers indications about the determinants
of RP, enabling a basis for interpretation and potentially guiding
protein engineering. This work provides a valuable foundation for
understanding the redox behavior of metalloproteins, enabling the
high-throughput prediction of redox potentials and informing data-driven
design across diverse protein families.