EukaUTR: a foundation model unifying functional modelling and design of eukaryotic 3′ UTRs
EukaUTR is a foundation model trained on diverse eukaryotic 3′ UTR sequences that unifies functional prediction and de novo design, outperforming existing baselines across multiple regulatory tasks and enabling targeted sequence optimization to reduce mRNA degradation rates.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Inside every living cell, the instructions for building proteins are written in a long, complex code called messenger RNA. For decades, scientists focused almost entirely on the beginning of this code, the part that tells the cell which protein to build. However, the end of the message, a section known as the three-prime untranslated region, holds a different kind of secret. This tail does not code for a protein itself; instead, it acts like a sophisticated control panel. It contains signals that determine how long the message survives, where it travels within the cell, and how efficiently it is read. If this tail is too short or carries the wrong signals, the message might be destroyed too quickly, or the cell might produce too much or too little of the final protein. Understanding how this region works is crucial because it governs the timing and quantity of life's essential processes, yet for a long time, the rules that dictate its behavior across the vast diversity of life remained a mystery.
A team of researchers has now built a powerful new tool to decode these rules, creating a system they call EukaUTR. This tool is a type of artificial intelligence trained specifically to read and understand the language of these genetic tails across a wide range of living things, from humans to fish to plants. Unlike previous models that could only speak the language of humans or mammals, this new system learned from a massive library containing nearly seven million genetic sequences from over a thousand different species. By studying this vast diversity, the model learned to recognize the subtle patterns that control how genetic messages are managed. The researchers found that this system could predict with high accuracy how a specific genetic tail would behave, such as how quickly it would break down or how well it would help produce a protein. In tests across thirteen different scenarios, the new model consistently outperformed existing tools, often improving prediction accuracy by more than twenty-five percent.
The power of this system goes beyond just reading what already exists; it can also write new genetic instructions from scratch. The researchers asked the model to generate entirely new genetic tails that had never been seen before. The results were striking. The new sequences looked and behaved like natural genetic material, possessing the same chemical balance and structural properties as those found in nature, yet they were completely unique and did not simply copy any existing sequence from the training library. This suggests the model truly understood the underlying grammar of genetic life rather than just memorizing a dictionary. Furthermore, the team developed a method to use this understanding to fix existing genetic messages. By analyzing a small set of sequences that were known to be stable, they taught the system to identify which parts of a shaky genetic tail were causing the instability. When they applied this guidance to edit unstable sequences, the model successfully reduced the predicted rate of degradation by thirty-seven percent, effectively making the genetic messages much more durable.
Despite these impressive capabilities, the researchers are careful to note what their work does and does not prove. The model excels at predicting how a sequence will behave based on its structure and composition, and it can identify regions where specific regulatory signals are likely to be found. However, it does not yet explain the exact mechanical steps of how these signals interact with cellular machinery, nor has the biological activity of the newly generated sequences been tested in a living cell. The improvements in stability are currently predictions based on the model's understanding of sequence patterns, and the true test will come when these engineered sequences are put to work in a laboratory. Nevertheless, this work marks a significant step forward, providing a unified framework that moves genetic design from simple prediction to active creation. It offers a way to not only understand the complex regulatory language of life but to speak it fluently enough to design new instructions that could one day be used to treat diseases or improve biological research.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.