[Paper Review] Fast Inference of Admixture Coefficients Using Sparse Non-negative Matrix Factorization Algorithms

Éric Frichot, François Mathieu+3 more|arXiv (Cornell University)|Sep 24, 2013

Genetic Mapping and Diversity in Plants and Animals29 references3 citations

TL;DR

This paper introduces sNMF, a fast and accurate method for estimating individual admixture coefficients using sparse non-negative matrix factorization (NMF) with least-squares optimization. It achieves 10–30× speedup over ADMIXTURE while maintaining high accuracy, as validated on human and plant genomic data sets with R² > 0.96 and RMSE < 5.5%.

ABSTRACT

Inference of individual admixture coefficients, which is important for population genetic and association studies, is commonly performed using compute-intensive likelihood algorithms. With the availability of large population genomic data sets, fast versions of likelihood algorithms have attracted considerable attention. Reducing the computational burden of estimation algorithms remains, however, a major challenge. Here, we present a fast and efficient method for estimating individual admixture coefficients based on sparse non-negative matrix factorization algorithms. We implemented our method in the computer program sNMF, and applied it to human and plant genomic data sets. The performances of sNMF were then compared to the likelihood algorithm implemented in the computer program ADMIXTURE. Without loss of accuracy, sNMF computed estimates of admixture coefficients within run-times approximately 10 to 30 times faster than those of ADMIXTURE.

Motivation & Objective

• To address the computational bottleneck of likelihood-based methods like ADMIXTURE in estimating admixture coefficients from large-scale genomic data.
• To develop a fast, scalable alternative to likelihood-based population structure inference that maintains high accuracy.
• To evaluate the performance of sparse NMF as a non-parametric, robust method for ancestry coefficient estimation in diverse population genetic data sets.
• To compare sNMF’s accuracy and runtime against ADMIXTURE using real and simulated human and plant genomic data.
• To validate the method’s ability to recover true population structure and ancestry proportions under varying levels of admixture, inbreeding, and missing data.

Proposed method

• Employs sparse non-negative matrix factorization (NMF) to decompose the genotypic matrix into a sparse admixture proportion matrix Q and a gene pool frequency matrix G.
• Solves a penalized least-squares minimization problem using an active set method to optimize Q and G, with sparsity enforced via regularization.
• Uses a cross-entropy criterion based on masked genotypes to select optimal K (number of ancestral populations) and regularization parameter α.
• Implements a 5% random masking of SNPs to create test sets for evaluating prediction accuracy of allelic probabilities.
• Predicts missing genotypes using the formula $ p^{ ext{pred}}_{i heta}(j) = \sum_{k=1}^{K} q_{ik} g_{k heta}(j) $, where j is the genotype (0,1,2).
• Evaluates model performance using the cross-entropy between predicted and true allelic distributions, minimizing $ H(p^{ ext{sample}}, p^{ ext{pred}}) = -\sum_{j=0}^{2} p^{ ext{sample}}(j) \log p^{ ext{pred}}_{i heta}(j) $.

Experimental results

Research questions

• RQ1 Can sparse NMF provide a computationally efficient alternative to likelihood-based methods like ADMIXTURE for estimating admixture coefficients?
• RQ2 How accurately does sNMF reproduce admixture coefficients estimated by ADMIXTURE across diverse human and plant genomic data sets?
• RQ3 To what extent does sNMF maintain accuracy under varying levels of admixture, inbreeding, and missing data in simulated data?
• RQ4 Can the cross-entropy criterion reliably select the correct number of ancestral populations (K) and optimal regularization parameter (α) in real and simulated data?
• RQ5 How does the runtime of sNMF compare to ADMIXTURE across different data set sizes and SNP counts?

Key findings

• sNMF achieved a median correlation of R² > 0.96 with ADMIXTURE estimates across 480 runs on 6 HGDP panels, indicating high accuracy in reproducing admixture coefficients.
• The root mean squared error (RMSE) of sNMF estimates remained below 5.5% across all 480 runs, demonstrating consistent precision.
• On the 1000 Genomes data set, sNMF computed admixture coefficients approximately 10 to 30 times faster than ADMIXTURE, with no loss in accuracy.
• The cross-entropy criterion successfully identified the correct K and optimal α in simulated data, even under moderate to high inbreeding (F_IS = 25% and 100%).
• sNMF maintained high predictive accuracy on masked genotypes, with cross-entropy values indicating reliable estimation of allelic probabilities.
• The method performed robustly on both human (HGDP, 1000 Genomes) and plant (A. thaliana) data, correctly identifying population structure with low admixture levels.