LncRNAs imported into mitochondria possess distinct features stratified by machine learning that promote interaction with the mitochondrial import protein PNPase

• Published in: The FASEB journal 2022-05, Vol.36 (S1), p.n/a
• Main Authors: Taylor, Andrew, Hathaway, Quincy, Robart, Aaron, Cook, Chris, Kunovac, Amina, Durr, Andrya, Rizwan, Saira, Cramer, Evan, Starcovic, Sarah, Hollander, John
• Format: Article
• Language: English
• ISSN: 0892-6638; 1530-6860
• DOI: 10.1096/fasebj.2022.36.S1.R1986

Understanding the localization and regulatory activity of extra‐nuclear long non‐coding RNAs (lncRNAs) is especially critical in the context of the mitochondrion, which possesses a genome separate from the nucleus that dictates bioenergetic function and is influenced by a broad range of pathologies. Prior research including efforts from our own laboratory have identified Polynucleotide Phosphorylase (PNPase) to be critical for selective RNA passage through the mitochondrial membrane, including microRNAs and other ncRNAs. However, the dynamics of lncRNA binding with PNPase have not yet been determined. Understanding how lncRNAs interact with PNPase is a crucial initial step for predicting their ability to be imported into mitochondria and how the process can be manipulated to regulate mitochondrial genome expression. The objective of this study was to evaluate lncRNAs present in mitochondria and classify sequence‐based features that might permit interaction with PNPase. Sequencing performed on mitochondrial and cytoplasmic isolates from human and mouse cardiac tissue identified over 500 mitochondrially‐localized nuclear genome‐encoded lncRNAs. Crosslinked immunoprecipitation (CLIP) of mitochondrial isolates using antibodies for PNPase pulled down predominantly lncRNAs in both human and mouse. The most highly bound lncRNA sequences were run though supervised machine learning using 10‐fold cross validation through Support Vector Machines (SVM) and Classification and Regression Trees (CART) machine learning algorithms, which identified stratification of primary and secondary sequence features as compared to random RNA and low binding lncRNA sequences, with an accuracy of 82% for SVM and an area under curve value of 0.89 for CART. Primary sequence features such as dissimilarities from coding sequences, various k‐mer frequencies, and overall GC content correlated with increased lncRNA interaction with PNPase (p