rnnimp

rnnimp performs genotype imputation using bidirectional recurrent neural networks to estimate genotypes of unobserved variants by encoding haplotype information as model parameters.

Key Features:

• Implementation: Implemented in Python.
• Model architecture: Uses bidirectional recurrent neural networks to estimate genotypes of unobserved variants.
• Input formats: Accepts phased genotype data in HAP/LEGEND format.
• Output format: Produces imputation results in Oxford GEN format.
• Haplotype encoding and privacy: Encodes haplotype information from a large cohort into neural network model parameters to enable parameter sharing without exposing individual-level genotypes.
• Performance: Evaluated on phased genotype data from the 1000 Genomes Project with accuracy comparable to existing methods for variants with MAF ≥ 0.05, slightly lower accuracy for MAF < 0.05, and improved performance relative to conventional methods when haplotype data are limited, particularly for variants with MAF ≥ 0.005.

Scientific Applications:

• Genomic studies with restricted haplotype access: Supports genomic studies requiring high-quality genotype data when direct access to extensive haplotype datasets is restricted due to donor consent.
• Genetics research: Facilitates genotype imputation to assess genetic variation in research cohorts.
• Personalized medicine: Supports personalized medicine studies that rely on imputed genotypes.

Methodology:

Training uses bidirectional recurrent neural networks implemented in Python to encode haplotype information as model parameters from phased HAP/LEGEND inputs and produce imputed genotypes in Oxford GEN format, with evaluation performed on phased genotype data from the 1000 Genomes Project stratified by minor allele frequency.