A deep learning model, specifically a gated recurrent neural network developed at Oregon State University, has been used to decipher which ribonucleic acids have the potential to encode proteins. Details from the endeavor were published in Nucleic Acids Research earlier this month.
"Deep learning may seem scary to some people, but at the end of the day, it's just crunching numbers," said David Hendrix, the study's lead author. "It's a tool just like calculus or linear algebra, one that we can use to learn biological patterns. The amount of sequencing data we have now is huge, and deep learning is well suited to face the challenges associated with the vast amount of data and to learn new biological rules that characterize the function of these molecules."
Hendrix and colleagues gave the gated neural network training on both noncoding and messenger RNA sequences, then turned it loose on the data to "learn the defining characteristics of protein-coding transcripts on its own." "It's really exciting," Hendrix added. "With the competing programs, developers would tell the program what an open reading frame is, what a start codon is, what a stop codon is. We thought it would be better to have a more de novo approach where the neural network can learn independently."
The team reported that compared to other approaches, the model they developed, known as mRNN, was better by a statistically significant margin in nearly every available metric.
"It not only found stop codons, it distinguished real stop codons from other trinucleotides that match stop codons and recognized long-range dependencies in the sequences," Hendrix said. "It doesn't wait to see a stop codon—we found it makes its decision long before the stop codon, 200 nucleotides from the start codon. And it learned a subset of codons that were highly predictive of protein-coding potential when observed in a potential open reading frame."