Design of diverse, functional mitochondrial targeting sequences across eukaryotic organisms using variational autoencoder

Abstract

Mitochondria play a key role in energy production and cellular metabolism, making them a promising target for metabolic engineering and disease treatment. However, despite the known influence of passenger proteins on localization efficiency, only a few protein-localization tags have been characterized for mitochondrial targeting. To address this limitation, we exploited Variational Autoencoder (VAE), an unsupervised deep learning framework, to design novel mitochondrial targeting sequences (MTSs). In silico analysis revealed that a high fraction of generated peptides are functional and possess features important for mitochondrial targeting. Additionally, we devised a sampling scheme to indirectly address biases arising from the differences in mitochondrial protein import machinery and characterized artificial MTSs in four eukaryotic organisms. These sequences displayed significant diversity, sharing less than 60% sequence identity with MTSs in the UniProt database. Moreover, we trained a separate VAE and employed latent space interpolation to design dual targeting sequences capable of targeting both mitochondria and chloroplasts, shedding light on their evolutionary origins. As a proof-of-concept, we demonstrate the application of these artificial MTSs in increasing titers of 3-hydroxypropionic acid through pathway compartmentalization and improving 5-aminolevulinate synthase delivery by 1.62-fold and 4.76-fold, respectively. Overall, our work not only demonstrates the potential of generative artificial intelligence in designing novel, functional mitochondrial targeting sequences but also highlights their utility in engineering mitochondria for both fundamental research and practical applications in biology.