Predicting gene regulatory regions with a convolutional neural network for processing double-strand genome sequence...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Advances in sequencing technology have generated vast amounts of genomic data.
• Annotating biological functions, especially in non-protein-coding regions, remains a significant challenge.
• Deep learning methods show promise for predicting gene regulatory regions from genomic sequences.

Purpose of the Study:

• To improve the prediction accuracy of gene regulatory regions using deep learning.
• To develop and introduce novel convolution layers for enhanced genomic sequence processing.
• To provide a software tool (DeepGMAP) for training and comparing deep learning models.

Main Methods:

• Designed novel convolution layers, termed forward- and reverse-sequence scan (FRSS) layers, to process genomic sequence information.
• Integrated both forward and reverse strand information within the convolution layers.
• Developed the DeepGMAP software for model training and comparison.
• Assessed previous studies to identify and address issues causing overfitting in data structures.

Main Results:

• The FRSS layers demonstrated enhanced capability in predicting gene regulatory regions.
• Identified and mitigated overfitting issues related to data structures in previous deep learning approaches.
• Developed visualization methods to interpret the learning process of the deep learning models.

Conclusions:

• The novel FRSS layers significantly improve the prediction accuracy for gene regulatory regions.
• DeepGMAP provides a valuable tool for researchers in genomics and bioinformatics.
• The findings contribute to better interpretation of genomic sequence data, particularly non-coding regions.