Neural ADMIXTURE

Neural ADMIXTURE performs unsupervised global ancestry inference from genotype, variant genotyping, or sequencing data using a neural-network autoencoder to estimate fractional cluster assignments that represent vectors of DNA marker frequencies.

Key Features:

• Unsupervised clustering: Employs ADMIXTURE- and STRUCTURE-like unsupervised clustering to decompose individual genomes into fractional cluster assignments representing population structure.
• Neural-network autoencoder: Uses a neural network autoencoder that adheres to the same modeling assumptions as traditional ADMIXTURE for ancestry estimation.
• Computational efficiency: Accelerates inference on large datasets, reducing compute time compared with traditional methods for large biobanks and extensive variant genotyping or sequencing data.
• Multiple outputs and hyperparameter flexibility: Can generate multiple outputs corresponding to different hyperparameter configurations (e.g., different K) within a single run.
• Scalability and storage: Produces storable models that enable linear-time computation for subsequent cluster assignments on new samples.

Scientific Applications:

• Population ancestry inference: Characterizes global and subcontinental ancestry structure in population biobanks and cohort studies.
• Genetic association and prediction studies: Provides ancestry estimates used to account for population stratification in association analyses and polygenic prediction.
• Sample-level substructure interpretation: Supplies interpretable fractional assignments for individual-level analyses of genetic substructure.
• Large-scale cohort analysis: Enables analysis of complex ancestry patterns across massive cohorts generated by genotyping or sequencing projects.

Methodology:

Trains a neural-network autoencoder under ADMIXTURE-like modeling assumptions to perform unsupervised clustering that decomposes genotypes into fractional cluster assignments (vectors of DNA marker frequencies), supports producing multiple hyperparameter outputs in one run, and allows storing trained models for linear-time later assignments.