Analysis of oxidative stress through deep atomic embeddings

Vanderbilt University, Nashville TN

Investigators

Abstract

PROJECT SUMMARY This project aims to computationally assess the impact aging-related oxidative stress has on the human proteome via a novel deep learning methodology. As a person ages, oxidative damage accumulates within their cells through exposure to both endogenous and exogenous sources. While the correlation between oxidative damage and aging is well-known, the specific mechanisms through which oxidation leads to aging-related ailments is complex and not well understood. Therefore, there exists a need for datasets which comprehensively enumerate how the human proteome is modified under oxidative stress. Although there exist mass spectroscopic methods for detecting these modifications, they are highly cost and labor intensive to do on a proteome-wide scale; computational prediction methods can help generate these datasets efficiently and complement existing datasets. However, oxidation modification prediction faces its own challenges due to the relative lack of known modification sites for training models when compared to other types of post-translational modifications. Therefore, I propose to build predictors for this problem by first constructing a deep learning- based atomic protein structure embedding method using a graph transformer autoencoder architecture. This autoencoder will be trained to embed and recover the atomic features of a set of high-resolution protein structures and high-confidence AlphaFold2 models. The resulting embedding model will then be benchmarked against state-of-the-art tools in a few protein function prediction tasks to ensure the information captured within the embedding is useful for biochemical analysis. This embedding approach not only allows the downstream classifiers to use the rich information contained within the training set of the embedding model, but also demonstrates the potential to extend to other biomolecules outside of the protein training set. Once the embedding procedure has been developed, a set of classifiers will be trained based on the deep learning-based embeddings to predict the likelihood of oxidation-related modifications occurring at a given protein residue. These classifiers will then be applied to the entire human proteome, resulting in a comprehensive proteome- wide prediction dataset of oxidation modification sites. This dataset will then be analyzed for functional pathway enrichment and individual proteins of interest. The most salient of these predicted modifications will be confirmed via mass spectroscopy. Finally, as an application to aging-related disease, proteins implicated in the progression of Alzheimer’s disease will be investigated in depth. Through successful completion of this research, a versatile deep learning-based method for encoding protein geometry at atomic resolution will be created, and novel functional insights and therapeutic targets will be derived from a comprehensive dataset of the predicted oxidation modification of the human proteome.