RFMix

RFMix infers local ancestry in fully sequenced human genomes using discriminative machine-learning approaches to resolve fine-scale ancestral segments.

Key Features:

• Discriminative Modeling Approach: Employs a discriminative modeling technique that outperforms existing methods in speed and accuracy, achieving approximately 30 times faster processing without compromising precision.
• Conditional Random Field with Random Forests: Integrates a conditional random field (CRF) framework parameterized by random forests and utilizes reference panels for training.
• Adaptive Learning from Admixed Samples: Learns directly from admixed samples to adapt to dataset-specific genetic characteristics and autocorrect phasing errors.
• High Sensitivity and Specificity: Demonstrates high sensitivity and specificity across populations including simulated Hispanics/Latinos, African Americans, and admixed groups of Europeans, Africans, and Asians.

Scientific Applications:

• Population Genetics: Provides precise local ancestry assignments to study human genetic diversity and evolutionary history.
• Admixture Detection: Identifies subtle admixture events at finer scales than continental-level methods.
• Disease Susceptibility Studies: Supports analyses that associate ancestry-resolved genomic segments with disease risk.
• Pharmacogenomics and Personalized Medicine: Informs pharmacogenomic investigations and personalized-medicine research by delineating ancestry-specific genomic backgrounds.

Methodology:

Uses discriminative modeling implemented as a conditional random field parameterized by random forests, trained on reference panels and able to learn from admixed samples to autocorrect phasing errors.