An RNA Language Model trained on sequence alone reveals the structural logic of Internal Ribosome Entry Sites
The authors introduce Albatross, an RNA language model trained solely on sequence data that significantly outperforms existing methods in predicting Internal Ribosome Entry Site structures, enabling the discovery of novel functional subclasses and accelerating antiviral target identification.
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 This is an AI-generated explanation of a preprint that has not been peer-reviewed.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
Viruses are masters of disguise, but their most critical secrets are often hidden in plain sight within their genetic code. For many viruses, including the picornaviruses that cause everything from the common cold to devastating neurological diseases, the instructions for making new viral particles are not just a simple list of letters. They are folded into complex three-dimensional shapes. These shapes act as switches and handles, allowing the virus to hijack the cell's protein-making machinery. One of the most important of these shapes is called an Internal Ribosome Entry Site, or IRES. Unlike human cells, which usually need a specific cap at the start of their genetic instructions to begin reading them, these viral IRESes act as a direct invitation, allowing the virus to start building proteins immediately and independently. Understanding exactly how these IRESes fold is vital for figuring out why some viruses infect specific tissues, why some cause severe disease while others do not, and how we might design drugs to stop them. However, these structures are notoriously difficult to map. They are long, highly variable, and change shape depending on the environment, making traditional methods of prediction slow and often inaccurate.
A team of researchers has now developed a new way to see these hidden shapes, bypassing the need for expensive, time-consuming experiments. They created an artificial intelligence system, which they named Albatross, trained exclusively on the raw genetic sequences of thousands of these viral elements. The system did not receive any information about how RNA folds, the rules of chemistry, or the physics of stability. It was simply fed millions of sequences and asked to learn the patterns within them. Remarkably, the model learned to predict the three-dimensional structure of these viral elements with high precision, simply by recognizing the statistical relationships between the letters in the sequence. When the researchers tested Albatross against a library of 96 different full-length viral IRESes, the model's predictions were far more accurate than those of the best existing methods. While other tools correctly identified less than half of the actual structural connections, Albatross was right about 80 percent of the time.
The power of this approach lies in what the model actually learned. It did not just memorize the shapes; it learned the logic of the virus. By analyzing how a change in one part of the sequence affected the prediction of another part, the model identified which pieces of the RNA must be close together to function. This allowed the researchers to distinguish between shapes that are merely stable and those that are essential for the virus to work. In fact, the model was so good at spotting functional structures that it could predict which parts of the RNA were most likely to be involved in starting the infection process. This capability allowed the team to build a massive atlas of over 75,000 predicted structures, covering a vast diversity of viruses that had never been studied in this detail before.
Using this new map, the researchers discovered something entirely new. They found a previously unknown sub-group of viral structures that included an extended stem, a specific folding pattern that had not been seen in this class of viruses before. They tested this discovery in the lab, confirming that the extended stem was real and that it played a crucial role in the virus's ability to function. This finding suggests that the model can uncover biological truths that were previously invisible to scientists. Furthermore, the team showed that this method is not limited to just one type of virus. When they applied the same training strategy to the genetic material of hantaviruses and bacterial sensors, the model successfully identified complex, long-range interactions and even the ability of a single RNA molecule to switch between two different shapes depending on its environment.
The implications of this work extend far beyond just mapping viruses. For decades, scientists have struggled to predict the structure of RNA because the molecules are so flexible and the rules governing them are so complex. Traditional methods often rely on calculating the energy of every possible shape, a process that becomes impossible for long sequences. Other methods require comparing many similar viruses to find patterns, which fails when the viruses are too different. Albatross sidesteps these problems by learning directly from the sequence data itself. The researchers found that the model's accuracy improved as they increased the size of the training data and the size of the model, suggesting that this approach can continue to get better with more data.
This study represents a significant shift in how we approach RNA biology. Instead of treating these molecules as static puzzles to be solved with physics equations, the researchers treated them as a language to be learned. The results show that the evolutionary history of a virus is encoded in its sequence in a way that an artificial intelligence can decode. By revealing the structural logic of these viral elements, the work provides a powerful new tool for understanding how viruses operate and how they might be stopped. The ability to predict these structures at scale means that scientists can now explore the functional potential of thousands of viral sequences that were previously too difficult to study, opening the door to new discoveries in virology and medicine.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.