miRstring: An RNA language model enables mature miRNA decoding and artificial small RNA design across species

Inside every living cell, a quiet but critical editing process takes place to ensure the right instructions are followed. Cells produce long strands of genetic material that act as blueprints, but these blueprints often come with extra, unnecessary sections that must be trimmed away before they can be used. Among the most important of these trimmed fragments are tiny molecules called microRNAs. These small pieces of genetic code do not build proteins themselves; instead, they act as precise switches that can turn off specific genes, effectively silencing the messages that would otherwise be carried out. To function, these molecules must be cut from their larger precursor forms at exact locations. If the cut happens in the wrong place, the resulting molecule may fail to work or could disrupt the cell's delicate balance. For scientists, predicting exactly where these cuts will occur across the vast diversity of life has remained a difficult challenge, limiting our ability to understand how these switches work in different animals or to design new ones for research and medicine.

A team of researchers has now developed a new computational tool called miRstring that addresses this long-standing puzzle by treating genetic sequences as a language to be read and understood. Rather than relying on rigid rules about how these molecules should look, this system learns the patterns of natural genetic code by studying a massive collection of data. The researchers trained the model on 77,708 precursor molecules found in 414 different species, allowing it to observe how nature consistently handles the cutting process across the tree of life. The result is a framework that can look at a raw genetic precursor and accurately predict the four specific boundaries that define the final, functional microRNA molecule. This capability allows the tool to identify the very first building block of the mature molecule with a high degree of precision, a detail that previous methods often missed.

What makes this approach particularly powerful is that the system does not just guess; it appears to understand the biological mechanics of the process. When the researchers examined how the model made its decisions, they found that its internal focus naturally highlighted the exact spots where cellular enzymes typically make their cuts. This suggests the tool has learned the genuine biological rules that govern how these molecules are processed, rather than simply memorizing patterns. The model proved robust even when tested on species it had never seen before, outperforming existing methods in identifying the correct boundaries. This reliability across different life forms means the tool can be used to annotate and understand microRNAs in organisms where they were previously difficult to map.

Beyond simply reading what exists in nature, the researchers demonstrated that this understanding could be used to create something new. They employed the model to design artificial precursors that would produce specific, custom-made microRNAs. When these designed molecules were tested, they successfully repressed their target messages, confirming that the computer-generated designs worked as intended in a biological setting. This work establishes a scalable path from analyzing natural genetic sequences to engineering new ones, offering a way to harness artificial intelligence for the design of small RNA tools. By bridging the gap between decoding natural biology and creating synthetic solutions, the study opens a route toward more precise and broadly applicable genetic engineering across many different species.