DeepHel: improving the accuracy and completeness of Helitron annotation with a multi-stage deep learning framework.

DeepHel, a multi-stage deep learning framework, improves accurate and complete Helitron annotation across five eukaryotic genomes by better recovering intact elements and refining HLE1/HLE2 boundaries.

• Why it matters: Helitron annotation is difficult because structural signals are weak, termini are ambiguous, fragments and nested insertions complicate detection, and existing tools trade off sensitivity and precision.
• What they did: DeepHel uses four coordinated modules—candidate identification, intact element recognition, boundary refinement, and Helitron classification—plus nested insertion removal, integrating sequence features, structural signals, and multiple sequence alignment information.
• The result: Benchmarking across five eukaryotic genomes under three evaluation settings shows DeepHel achieves a superior sensitivity–precision balance versus HiTE, HELIANO, HelitronScanner, RepeatModeler2, and EDTA, including in low-copy and complex nested cases.

AI summary of the abstract · may contain errors · not medical advice · read the paper.