Identifying genomic islands with deep neural networks.

Area of Science:

• Microbiology
• Bioinformatics
• Genomics

Background:

• Horizontal gene transfer (HGT) drives bacterial adaptability by spreading genetic information via genomic islands (GIs).
• Genomic islands are classified by gene content and mobility, with existing computational methods unable to detect all types.
• Detecting GIs is crucial for understanding bacterial evolution and adaptation.

Purpose of the Study:

• To develop a novel computational method for detecting genomic islands (GIs).
• To improve the accuracy and generalizability of GI detection compared to existing tools.
• To leverage deep learning for identifying novel GIs.

Main Methods:

• A deep learning model (Inception V3) was adapted for GI detection.
• Genomic fragments were represented as images for analysis.
• Transfer learning was employed, pre-training the model on a large dataset and fine-tuning it on generated genomic images.

Main Results:

• The Shutter Island method demonstrated superior generalization capabilities compared to existing GI detection tools.
• The image-based deep learning approach successfully identified known GIs and predicted novel ones.
• The model achieved high accuracy in predicting GIs.

Conclusions:

• Deep learning, particularly with an image-based approach, offers a powerful strategy for genomic island detection.
• The method's ability to generalize suggests its applicability to other biological problems with limited curated data.
• This approach enhances the discovery of GIs, contributing to a better understanding of bacterial evolution.