PaRFR

PaRFR performs parallelized Random Forest regression to map multivariate quantitative trait loci and estimate pair-wise genetic similarity from large-scale genome-wide association study (GWAS) data.

Key Features:

• Parallel Processing: Implements the MapReduce programming model and can be deployed on Hadoop to distribute computations across multiple nodes for datasets with hundreds of thousands of SNPs and high-dimensional traits.
• Parallel Random Forest Regression: Adapts the Random Forests algorithm for regression with both univariate and multivariate responses by constructing an ensemble of decision trees.
• Distance-Based Node Splitting: Uses a distance-based criterion for node splitting during tree estimation to improve computational efficiency and decision-tree accuracy.
• Genetic Similarity Learning: Estimates pair-wise measures of genetic similarity that correlate with phenotypic traits and provides SNP ranking related to trait variation.

Scientific Applications:

• Neuroimaging genetics: Analysis of structural and functional variability derived from magnetic resonance imaging (MRI) as multivariate quantitative traits.
• Genome-wide association studies (GWAS): Scalable identification and ranking of SNPs associated with multivariate quantitative traits in large cohorts.
• Alzheimer's disease research: Application to longitudinal neuroimaging phenotypes that identified known AD-related variants such as APOE4 and TOMM40 and related top-ranked mutation patterns to disease severity.
• Genetic architecture inference: Derivation of genetic proximity measures that align with phenotypic similarities to inform the genetic basis of complex traits.

Methodology:

Constructs an ensemble of decision trees for regression and genetic-similarity learning, employs a distance-based node-splitting criterion, and parallelizes computation via MapReduce on Hadoop for GWAS-scale datasets.