TERL: classification of transposable elements by convolutional neural networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transposable elements (TEs) are abundant in eukaryotic genomes, but their classification into deeper levels like superfamilies remains challenging.
• Existing methods often rely on handcrafted features or homology searches, limiting their efficiency with non-homologous sequences.

Purpose of the Study:

• To introduce a novel method, Transposable Elements Representation Learner (TERL), for accurate and efficient classification of TEs.
• To evaluate TERL's performance against established methods across different classification levels and datasets.

Main Methods:

• TERL preprocesses one-dimensional TE sequences into two-dimensional, image-like data.
• Deep convolutional neural networks are applied to learn representations for classification.
• Performance was assessed through six experiments comparing TERL with other classification tools.

Main Results:

• TERL achieved high macro mean accuracies and F1-scores, reaching 96.4% for superfamilies and 95.7% for orders on RepBase data.
• On broader datasets, TERL obtained 95.0% accuracy for superfamilies and 89.3% for orders.
• TERL significantly outperformed other methods in accuracy, recall, and specificity, and was orders of magnitude faster.

Conclusions:

• TERL demonstrates a powerful capability to learn and predict TEs at any hierarchical classification level.
• The method offers a substantial speed advantage over existing tools like TEclass and PASTEC.
• TERL provides an efficient and accurate solution for detailed transposable element classification in genomics.