mirDNN

mirDNN employs a convolutional deep residual neural network to predict pre-miRNA sequences from raw genomic data by integrating sequence and secondary-structure information to identify novel microRNA (miRNA) precursors.

Key Features:

• Deep learning architecture: Employs a convolutional deep residual neural network to model sequence context and structural patterns for pre-miRNA prediction.
• Secondary structure integration: Incorporates pre-miRNA secondary-structure properties into the predictive model.
• Automatic feature extraction: Learns intrinsic structural characteristics and contextual features from raw genomic data without manual feature engineering.
• Performance: Demonstrated high precision and recall across animal and plant genomes, achieving precision up to five times higher than other algorithms at equivalent recall levels.
• Validation on novel species: Includes a real-world validation methodology confirming performance on species not included in training datasets.
• Input and interpretability outputs: Operates on sequences in FASTA format and produces nucleotide-level importance scores/plots.

Scientific Applications:

• Functional Genomics: Enables identification of novel pre-miRNA candidates to study miRNA-mediated gene regulation.
• Comparative Genomics: Supports exploration of evolutionary conservation and divergence of miRNA genes across species.
• Biomedical Research: Aids discovery of miRNA biomarkers and potential therapeutic targets.

Methodology:

Implements a convolutional deep residual neural network that learns sequence context and secondary-structure features from raw genomic sequences without manual feature engineering.