TERL

TERL classifies transposable elements by converting one-dimensional TE sequences into two-dimensional image-like representations and applying deep convolutional neural networks to assign hierarchical labels such as superfamilies and orders.

Key Features:

• Two-dimensional transformation: Preprocesses one-dimensional TE sequences into two-dimensional image-like data for downstream modeling.
• Deep convolutional neural networks (CNNs): Leverages CNNs to learn optimal sequence representations and to handle non-homologous sequences that traditional methods struggle with.
• RepBase performance: Achieves macro mean accuracies and F1-scores of 96.4% and 85.8% at the superfamily level, and 95.7% and 91.5% at the order level on RepBase sequences.
• Cross-database performance: On sequences from seven databases, reports macro mean accuracies of 95.0% and F1-scores of 70.6% for superfamilies, and 89.3% and 73.9% for orders, while surpassing other methods in accuracy, recall, specificity, and computational efficiency.
• Runtime efficiency: Runs approximately 20 times faster than TEclass and about three orders of magnitude more efficient across the reported experiments compared to PASTEC.

Scientific Applications:

• Hierarchical TE annotation: Provides superfamily- and order-level classification labels for transposable element annotation in genomes.
• Evolutionary studies: Supports analyses of the distribution and evolution of TEs across eukaryotic genomes.
• Genomic diversity and stability analyses: Facilitates studies on genetic diversity and genome stability related to TE content.
• Gene regulation investigations: Enables identification of potential functional roles of TEs and their impacts on gene regulation.

Methodology:

Converts one-dimensional TE sequences into two-dimensional image-like representations and trains deep convolutional neural networks to learn sequence representations for hierarchical classification, with validation performed on RepBase and six additional databases.