Leveraging long-read assemblies and machine learning to enhance short-read transposable element detection and...

Area of Science:

• Genomics
• Bioinformatics
• Evolutionary Biology

Background:

• Transposable elements (TEs) are mobile genetic sequences crucial for genome evolution.
• Long-read sequencing improves TE detection but is costly; short-read methods are less accurate.
• Existing short-read TE detection tools show inadequate performance on real genomic data.

Purpose of the Study:

• To develop a machine learning approach (TEforest) for accurate discovery and genotyping of TE insertions and deletions using short-read sequencing data.
• To leverage TEs identified from long-read assemblies for training a robust short-read detection model.
• To provide a user-friendly tool for studying TE prevalence and phenotypic effects.

Main Methods:

• Utilized a machine learning approach, TEforest, employing a gradient-boosted decision tree model.
• Trained the model using a labeled ground-truth dataset of TEs identified from Drosophila melanogaster long-read assemblies.
• Extracted features from short-read alignments to identify potential TE insertion/deletion sites and discriminate true from false positives.

Main Results:

• TEforest demonstrated superior performance compared to traditional methods in comprehensive benchmarks using D. melanogaster and human data.
• The tool identified more true positives and fewer false positives across various read lengths and coverages.
• TEforest accurately inferred genotypes and precise insertion breakpoints, bridging the gap between short-read scalability and long-read accuracy.

Conclusions:

• TEforest offers a cost-effective and accurate solution for TE detection using short-read sequencing.
• The developed method enhances the study of TE insertions in large-scale population genetics and evolutionary studies.
• This tool facilitates research into the prevalence and phenotypic consequences of transposable elements across genomes.