DeepGRP

DeepGRP predicts and classifies genomic repetitive elements at nucleotide resolution using deep learning with bidirectional gated recurrent units (GRUs) and an attention mechanism to improve repeat annotation.

Key Features:

• Deep Learning Architecture: Combines bidirectional GRUs with an attention mechanism inspired by neural machine translation for nucleotide-level annotation of repetitive elements.
• Lineage: Builds on dna-brnn (Li 2019), which used recurrent neural networks for repeat annotation.
• Predictive Performance: Evaluation on the human genome reports a 20% increase in the Matthews correlation coefficient compared with dna-brnn.
• Expanded Repeat Classification: Predicts two additional repeat classes and transfers RepeatMasker-derived annotations across species, e.g., from human to mouse.
• Database Integration: Recovers repetitive elements annotated in the Dfam database that are not detected by RepeatMasker.
• Scalability and Efficiency: Implemented in TensorFlow with GPU acceleration, achieving approximately 1.8× speed over dna-brnn, ~8.6× over RepeatMasker, and >100× over HMMER for Dfam model searches.
• Implementation: Provided as a Python-based bioinformatics package.

Scientific Applications:

• Genome Annotation: Nucleotide-level annotation and classification of repetitive elements in eukaryotic genomes.
• Comparative Genomics: Transfer of repeat annotations across species, exemplified by human-to-mouse annotation transfer.
• Repeat Discovery and Curation: Detection of Dfam-annotated repeats missed by RepeatMasker to expand repeat catalogs.

Methodology:

Combines bidirectional GRUs with an attention mechanism (informed by neural machine translation and dna-brnn), trained on RepeatMasker-derived data, implemented in TensorFlow with GPU acceleration, and evaluated on the human genome using the Matthews correlation coefficient with comparisons to dna-brnn, RepeatMasker, and HMMER.