Determining whether the non-protein-coding DNA sequences are in a complex interactive relationship by using an artificial intelligence method

TL;DR

This study uses deep neural networks to analyze non-protein-coding DNA, revealing potential long-range interactions and distinguishing real from scrambled sequences with high accuracy.

Contribution

It demonstrates that convolutional neural networks can learn complex features of non-coding DNA without prior motif specification, suggesting long-range interactions.

Findings

CNN can distinguish real vs scrambled DNA sequences effectively.

Non-coding DNA may have meaningful interactions beyond 100 bp.

Deep learning outperforms linear SVMs in this task.

Abstract

Non protein coding regions of the human genome contain many complex patterns which regulate the cellular activity. Studying the human genome is limited by the lack of understanding of its features and their complex interactions. However, recent advances in AI research have enabled automatically learning representations of high dimensional complex data without feature engineering, using deep neural networks. Therefore, in this paper, we demonstrate that a convolutional neural network can learn a representation of DNA sequence without specifying any motifs or patterns, such that it becomes capable of predicting whether a DNA sequence is natural or artificial. The trained model could distinguish scrambled vs real DNA sequences for scrambling lengths of 2 bp, 10 bp, 50 bp and even 100 bp, with a significantly higher accuracy than linear SVMs. With this study, we have discovered that regions of non protein coding DNA might have meaningful interactions at even longer than 100 bp distances even though they do not code proteins.