Griffith University researchers use AI to predict RNA structures
Griffith University researchers developed an artificial intelligence(AI) method to predict the secondary structures of RNA in a better way. The researchers hope that the new tool will help in a better understanding of RNA’s role in various diseases and cancer.
Being one of the four major macromolecules, RNA (ribonucleic acid) plays a vital role in decoding, coding, expression, and regulation of genes in all living organisms.
The group of researchers implemented deep learning, a subset of machine learning in artificial intelligence, to derive complicated mathematical functions for carrying out specific tasks automatically without direct human instructions. They created a precise model for the functional relationship between the sequence and structure of RNA.
This is the first time in the world scientists are using artificial intelligence to predict RNA structure. And this advancement comes after decades of unaltered performance by previous techniques to predict RNA structure.
Professor Zhou believes that this new technique will help to design new RNA molecules of therapeutic potentials.
Since the number of proteins outnumbers almost 10 times the number of RNAs in the body, scientists had been focusing on proteins for decades. Consequently, we were unaware of why these RNAs are present in our body.
This is why Griffith University researchers developed this tool to get some clues about RNA structures. According to Professor Zhou, clues are essential since the involvement of more and more RNAs is being identified in various diseases, including cancer.
The sequence of RNA is encoded in the genome, but how RNAs function through their structures was unexplored. The new tool will enable scientists to link the structure and sequence of RNA better.
Using the deep learning technique, researchers will now be able to understand the relationship between sequence, structure, and function of RNAs. Once the link is understood, we will be able to design RNAs for a particular purpose, like therapeutics or molecular sensors.
To develop the model, the team analyzed the existing data sets for known RNA structures and expanded the data set by refining the training method.
To predict the structure of 30 million RNAs by analyzing that of just 250 known RNA structures could be done only through deep learning. This was a complex task as the nucleotides can theoretically pair with any other base within the RNA, and deep learning was used to find out which nucleotides are paired together. The algorithm had to work through millions of RNA sequences.
To date, people were relying on comparative studies based on RNA families and statistics-based algorithms. Although these methods could somewhat model the specific functions linking the structure and sequence, they could only reach up to 80% accuracy in basepair predictions.
Deep learning brought an opportunity to overcome all the shortcomings, including the accuracy to boost to 93%.
The research ‘RNA secondary structure prediction using an ensemble of two-dimensional deep neural networks and transfer learning’ has been published in Nature: Communications.