MetaRNN

MetaRNN predicts the pathogenicity of human non-synonymous single nucleotide variants (nsSNVs) and non-frameshift insertion/deletions (nfINDELs) using deep learning to distinguish rare pathogenic variants from benign ones.

Key Features:

• Variant scope: MetaRNN addresses nsSNVs while MetaRNN-indel addresses nfINDELs (non-frameshift insertion/deletions).
• Annotation integration: Integrates 28 high-level annotation scores, including 16 functional prediction scores and allele frequency features.
• Model architecture: Employs a deep recurrent neural network (RNN) model.
• Context window: Incorporates a +/- 1 codon window around the affected codon for context-aware predictions.
• Ensemble approach: Uses an ensemble strategy to produce robust and interpretable pathogenicity scores.
• Pre-computed scores: Provides pre-computed pathogenicity scores for all possible human nsSNVs.
• Score comparability: Produces comparable prediction scores between nsSNV-based and nfINDEL-based models to support integrated genotype-phenotype analyses.
• Benchmarking: Evaluated on independent test datasets and reported superior performance and a more interpretable score distribution compared to existing state-of-the-art models.

Scientific Applications:

• Disease screening: Distinguishes rare pathogenic variants from benign variants for disease screening workflows.
• Clinical genetics: Supports prioritization of variants for diagnosis and clinical investigation.
• Genotype-phenotype association: Enables integrated genotype-phenotype association analyses using comparable scores for nsSNVs and nfINDELs.
• Genomics research and personalized medicine: Facilitates variant interpretation in genomics research and personalized medicine studies.

Methodology:

Integrates 28 high-level annotation scores (including 16 functional prediction scores) and allele frequency features into a deep recurrent neural network (RNN) model that applies a +/- 1 codon window and an ensemble approach for context-aware pathogenicity prediction.