Format

Send to

Choose Destination
Hum Hered. 2018;83(3):163-172. doi: 10.1159/000493215. Epub 2019 Jan 25.

tRNA-DL: A Deep Learning Approach to Improve tRNAscan-SE Prediction Results.

Author information

1
Department of Computer Science, New Jersey Institute of Technology, Newark, New Jersey, USA.
2
Department of Computer Science, New Jersey Institute of Technology, Newark, New Jersey, USAzhiwei@njit.edu.
3
The Center for Applied Genomics, The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, USA.
4
Department of Pediatrics, University of Pennsylvania School of Medicine, Philadelphia, Pennsylvania, USA.

Abstract

BACKGROUND:

tRNAscan-SE is the leading tool for transfer RNA (tRNA) annotation, which has been widely used in the field. However, tRNAscan-SE can return a significant number of false positives when applied to large sequences. Recently, conventional machine learning methods have been proposed to address this issue, but their efficiency can be still limited due to their dependency on handcrafted features. With the growing availability of large-scale genomic data-sets, deep learning methods, especially convolutional neural networks, have demonstrated excellent power in characterizing sequence patterns in genomic sequences. Thus, we hypothesize that deep learning may bring further improvement for tRNA prediction.

METHODS:

We proposed a new computational approach based on deep neural networks to predict tRNA gene sequences. We designed and investigated various deep neural network architectures. We used the tRNA sequences as positive samples, and the false-positive tRNA sequences predicted by tRNAscan-SE in coding sequences as negative samples, to train and evaluate the proposed models by comparison with the conventional machine learning methods and popular tRNA prediction tools.

RESULTS:

Using the one-hot encoding method, our proposed models can extract features without involving extensive manual feature engineering. Our proposed best model outperformed the existing methods under different performance metrics.

CONCLUSION:

The proposed deep learning methods can substantially reduce the false positive output by the state-of-the-art tool tRNAscan-SE. Coupled with tRNAscan-SE, it can serve as a useful complementary tool for tRNA annotation. The application to tRNA prediction demonstrates the superiority of deep learning in automatic feature generation for characterizing sequence patterns.

KEYWORDS:

Deep learning; Genomics; Machine learning; Multilayer neural network; tRNA prediction; tRNAscan-SE improvement

PMID:
30685762
DOI:
10.1159/000493215
Free full text

Supplemental Content

Full text links

Icon for S. Karger AG, Basel, Switzerland
Loading ...
Support Center