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

Transposable elements (TEs) are parasitic genomic elements that are ubiquitous across the tree of life and play a crucial role in genome evolution. Advances in long-read sequencing have allowed highly accurate TE detection, though at a higher cost than short-read sequencing. Recent studies using long reads have shown that existing short-read TE detection methods perform inadequately when applied to real data. In this study, we use a machine learning approach (called TEforest) to discover and genotype TE insertions and deletions with short-read data by using TEs detected from long-read genome assemblies as training data. Our method first uses a highly sensitive algorithm to discover potential TE insertion or deletion sites in the genome, extracting relevant features from short-read alignments. To discriminate between true and false TE insertions, we train a random forest model with a labeled ground-truth dataset for which we have calculated the same set of short-read features. We conduct a comprehensive benchmark of TEforest and traditional TE detection methods using real data, finding that TEforest identifies more true positives and fewer false positives across datasets with different read lengths and coverages, while also accurately inferring genotypes and the precise breakpoints of insertions. By learning short-read signatures of TEs previously only discoverable using long reads, our approach bridges the gap between large-scale population genetic studies and the accuracy of long-read assemblies. This work provides a user-friendly tool to study the prevalence and phenotypic effects of TE insertions across the genome.