The mitochondrion, often referred to as the powerhouse of the cell, plays critical roles in cellular function, making it a prime organelle to target for fundamental studies, metabolic engineering, and disease therapies. With only a limited number of existing mitochondrial targeting sequences, a new study from the Carl R. Woese Institute for Genomic Biology demonstrates the utility of generative artificial intelligence for designing new ones.
Much like each organ plays an important role in the human body—the heart for pumping blood or the lungs for breathing—cells contain different compartments called organelles that contribute to overall cellular function. These organelles have distinctive characteristics and environments for performing specific tasks for the cell.
The mitochondrion is a specialized organelle for generating energy for cells, and its unique environment is also the ideal location for various cellular processes including metabolic pathways. Dysfunctional mitochondria have also been associated with aging and disease states.
"Researchers want to study the biology of the mitochondria which can't be done efficiently without using targeting sequences," said Huimin Zhao (BSD leader/CABBI/CGD/MMG), the Steven L. Miller Chair of Chemical and Biomolecular Engineering at the University of Illinois Urbana-Champaign. "But we are currently limited by the availability of these mitochondrial targeting sequences, or MTSs."
In order to maintain cellular organization and processes, there are complex mechanisms in place to ensure that protein cargo is delivered to the correct location. But rather than using an address and a stamp to send these packages throughout the cell, proteins are tagged for delivery to a particular organelle by unique amino acid targeting sequences.
MTSs found in nature range from 10 to 120 amino acids in length, averaging around 35 amino acids. Currently only a handful of MTSs have been identified and used, and there is a lack of predictable patterns in their sequences, making it difficult to design new artificial ones. "There are only a few MTSs that have been characterized, and people use the same sequence again and again," said Aashutosh Boob, first author of the publication and former doctoral student in Zhao's group.
"One of the issues is that for different passenger proteins, there's a different optimal targeting sequence. Secondly, if the same sequences are used often, particularly in metabolic engineering, it can actually lead to homologous recombination and then genetic instability. So ideally there would be a library of diverse MTSs at hand to test and use."
The challenge is that mitochondrial targeting abilities of an MTS arise from its chemical and structural characteristics in 3D space rather than its 2D amino acid sequence. Generative AI can solve this problem by finding intricate patterns in the training data—in this case MTSs found in nature—that are difficult for humans to recognize and connect.
Using an unsupervised deep learning framework called Variational Autoencoder, the research team identified key features of MTSs including being positively charged and amphiphilic and tending to form an α-helix. They then designed a million AI-generated MTSs and experimentally tested the mitochondrial targeting abilities of 41 of them. Using confocal microscopy for the validation studies, they achieved a 50 to 100 % success rate in yeast, plant cells, and mammalian cells.
To further demonstrate the utility of the AI-generated MTSs, the researchers applied the targeting sequences for both metabolic engineering and protein delivery—the latter which could be beneficial for therapeutics. They also illustrated how AI can help to and understand the evolution of dual-targeting sequences for both the mitochondria and chloroplasts, highlighting the breadth of scientific questions that could be studied using this technology.
Overall, this research marks an important milestone for the Zhao research group as the first generative AI publication from the lab. The study was especially unique in the depth of experimental work that was done to validate the AI findings.
"Characterizing the targeting sequences in the lab took us a lot of time, but we wanted to highlight their application, both in terms of metabolic engineering and therapeutics," Boob said. "This project spanned a significant portion of my PhD, challenging me to broaden my expertise beyond the lab. It strengthened my ability to think critically and design a rigorous scientific study, while also giving me the chance to work with great people in a fun, fast-paced environment that made the experience both enjoyable and rewarding."
Zhao said, "AI is so hot right now, and people are really interested in knowing potential applications of AI, particularly in the scientific domain. This project clearly demonstrates that generative AI is a useful tool for synthetic biology and biotechnology."
The publication, "Design of diverse, functional mitochondrial targeting sequences across eukaryotic organisms using variational autoencoder" can be found at https://doi.org/10.1038/s41467-025-59499-3 and was funded by the U.S. Department of Energy Center for Advanced Bioenergy and Bioproducts Innovation.
