B-RANKO: An ncRNA Foundation Model with Bidirectional Decoder-Only Architecture
B-RANKO is a novel 150M-parameter bidirectional decoder-only foundation model that unifies de novo RNA sequence generation and transferable representation learning, demonstrating strong performance in generating biologically plausible sequences and achieving competitive results on RNA engineering tasks.
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 This is an AI-generated explanation of a preprint that has not been peer-reviewed.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, a molecule called ribonucleic acid, or RNA, acts as a vital messenger and manager. While DNA holds the master blueprint for life, RNA is the active worker that reads those instructions, builds proteins, and even catalyzes chemical reactions. In recent years, scientists have realized that RNA is not just a passive carrier of information but a programmable tool with immense potential for medicine and biotechnology. Because of this, researchers have turned to artificial intelligence to understand and design new RNA molecules, hoping to create therapies that can target diseases with high precision. However, teaching a computer to both understand the complex language of RNA and invent new, functional sequences has proven difficult. Most existing AI models are good at one task or the other, but rarely both, leaving a gap in our ability to fully harness the power of these molecules.
To bridge this divide, a team of researchers from the University of Zagreb and the Genome Institute of Singapore has developed a new artificial intelligence system called B-RANKO. This model is designed to do two things at once: learn the deep patterns of natural RNA to make accurate predictions, and generate entirely new RNA sequences from scratch. The researchers trained B-RANKO on a massive collection of known RNA sequences, drawing from public databases that contain millions of examples. The core innovation lies in how the model reads and writes these sequences. Unlike standard AI models that read text strictly from left to right, B-RANKO uses a unique method where it generates a sequence from both ends simultaneously, moving inward until the two halves meet in the middle. This approach mimics the way RNA molecules naturally fold, where the two ends of the strand often come together to form stable structures. By training the model to predict the next piece of the puzzle from both directions at once, the researchers enabled a single architecture to master both the creation of new molecules and the analysis of existing ones.
When the team tested the model's ability to create new RNA, the results were striking. The computer generated sequences that looked and behaved like real biological molecules. They had the right chemical balance and folded into shapes that were physically stable, much like the RNA found in nature. Crucially, these new sequences were not just copies of what the model had seen during training. When the researchers compared the generated molecules against the vast library of known RNA, they found that a large portion of the new creations were entirely novel, representing unique variations that did not exist in the training data. The model did not simply memorize the past; it learned the underlying rules of RNA chemistry well enough to imagine new possibilities.
The researchers then put B-RANKO to the test in practical applications, specifically looking at its ability to design aptamers, which are RNA molecules engineered to bind to specific targets, much like a key fitting into a lock. After fine-tuning the model on a dataset of known aptamers, it successfully generated new candidates that matched the desired properties. In a broader evaluation across thirteen different tasks, ranging from predicting how RNA folds to determining how it interacts with cellular machinery, B-RANKO performed with notable strength. It excelled particularly in engineering tasks, such as predicting how stable a vaccine molecule would be or how well a CRISPR tool would target a specific gene without hitting the wrong ones. In these areas, it outperformed many other leading models, demonstrating that its ability to generate new sequences did not come at the cost of understanding.
While the model showed great promise, the researchers were careful to note its limitations. It was not equally perfect at every single task; for instance, it was slightly less accurate at predicting the fine details of RNA's internal structure compared to its success in broader engineering applications. This suggests that while the model has captured the general rules of RNA behavior, there is still room for improvement in understanding the most intricate local details. Furthermore, the sequences generated by the computer have so far been evaluated only through simulation. The true test will come when these molecules are built in a laboratory and tested in real biological systems to see if they function as predicted.
The development of B-RANKO marks a significant step forward in the field of computational biology. By proving that a single model can effectively learn from nature and then create new biological tools, the researchers have provided a unified framework for RNA design. This approach offers a powerful new way to explore the vast landscape of RNA molecules, potentially accelerating the discovery of new medicines and biotechnologies. The work suggests that the future of RNA engineering may not require separate tools for analysis and creation, but rather a single, versatile system capable of doing both, opening the door to a new era of programmable biology.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.