Evaluating Reduced and Oxidized Abeta Peptides Using Machine Learning

Abstract

Abeta peptides are implicated in cytotoxicity related to Alzheimer’s disease. Because Abeta is readily oxidized, it is critical to understand how the conformations of oxidized Abeta differ from those of the reduced form. To explore this, we generated structures of oxidized and reduced Abeta10-40 peptides using replica-exchange molecular dynamic simulations. Analysis of the secondary structure revealed that Abeta oxidation, which occurs at methionine-35, reduced helical content in the C-terminal and enhanced turn and coil conformations. Tertiary structure, computed using the intrapeptide distance matrix between C-alpha atoms, shows a mixture of stabilizing and destabilizing effects by Abeta oxidation. In particular, oxidation of methionine-35 largely increases its distances to other amino acids compared to the reduced peptide indicating a reduction in internal structure. Additionally, we utilized the random forest machine learning algorithm to classify Abeta peptides using the distance matrix. With our machine learning model, we were able to discriminate between reduced and oxidized states and identify important amino acid pairs in the distance matrix that contributed to this classification. The success of the model was evaluated using the receiving operator characteristic and confusion matrix. Given that oxidized Abeta aggregates more slowly than the reduced form, we propose our machine learning model can predict aggregation-prone peptide structures and thus has applications in understanding Alzheimer’s disease severity.