Biotechnology (Commonwealth Union) – South Korean scientists have made a significant scientific finding with the formation of a circMAC model that has the ability to forecast the joining sites nucleotide by nucleotide, as it engages with the molecule as the loop.
A majority of RNA molecules are linear strings that has a start point and an end point. When it comes to circRNAs (circular RNAs) they do not equate as the same. Life scientists have noted that the rings can be a sponges for microRNAs, short regulatory molecules that dial back gene activity. A circRNA that acts as a sponge for a particular microRNA can free genes that would otherwise be suppressed by that microRNA.
Researchers at Pusan National University in South Korea say they have moved a step closer to solving the difficult puzzle of where on a circular RNA a microRNA binds with an artificial intelligence tool called circMAC.
Loop as a line: the problem of solving
Many computational approaches treat a circRNA as a regular linear transcript, with a start and end. But that shortcut is missing something. These molecules are defined by the point at which the ring closes, the back-splice junction, and sequences crossing the back-splice junction can be important for binding. Cut the ring at an arbitrary point and the pattern that goes round the seam is cut in half.
To make things even more complicated, a single circRNA gene can produce many different full-length versions (termed isoforms) that vary in their composition.
How circMAC works
Circular RNAs are tiny ring-shaped molecules found in cells. They can be locked up by tiny molecules called microRNAs, a bit like a key going into a lock. “The key to understanding gene regulation is knowing where this latching occurs, but finding those spots is difficult.
You give the tool two things: the complete genetic spelling of a circular RNA and the spelling of one microRNA. It scores each letter of the ring, giving a measure of the likelihood of this position being the binding site for the microRNA.
The tool uses three strategies The first one allows it to sense connections between distant parts of the ring. The second is a newer technique called Mamba that helps it remember the order of the letters. The third is looking for short repeated letter sequences, often attachment sites.
The tool takes into account the fact that the molecule is a loop, so the end of the sequence is considered joined to the start, avoiding the interpretation of a pattern crossing the point where the ring closes as two broken pieces. It considers the specific microRNA, so the same ring can give different predictions depending on the microRNA tested. Before its main training, it was taught how letters pair with each other in RNA, giving it a sense of the shape of the molecule.
In the lab, the researchers trained and tested the tool using circular RNAs that have experimentally validated binding sites (a technique that shows the location of protein complexes carrying microRNAs on RNA). That was a tricky task, because only about six out of every 100 letters were true attachment sites, so the tool had to find a few needles in a very big haystack.
For the 1000-nucleotide configuration, the F1-macro score, AUROC and AUPRC of the circMAC model pretrained with base-pairing prediction were 0.762, 0.902 and 0.545 respectively. That beat the pretrained RNA language models it was tested against, the team said. The model also performed best overall among the full-length models tested under more stringent splits, e.g. separating isoforms and separating back-splice junctions between training and testing. The rationale for these splits is to prevent a model from just memorising what it has already seen.
