How to choose sets of ancestry informative markers: A supervised feature selection approach

Abstract

Highlights:• We provide AIMsetfinder, a tool to systematically select ancestry informative markers (AIMs).• Simulations of human population structure can be used to assess the performance of AIM selection procedures.• 17 SNPs identified by AIMsetfinder suffice to classify all african, european, east asian, and south asian individuals in the 1000 Genomes project correctly.

Inference of recent admixture using genotype data

“…The former comes with 128 SNPs, and we ignore seven tri-allelic SNPs (rs17287498, rs2069945, rs2184030, rs433342, rs4540055, rs5030240, rs12402499), since our methods currently rely on bi-allelic SNPs. It was designed to distinguish Africa, Europe, East Asia, Native America, and Oceania, but was shown to perform well on the 1000 genomes dataset, also for distinguishing South Asia, even when ignoring the tri-allelic SNPs [6]. The latter comes with 55 bi-allelic SNPs and was introduced as a global AIMset differentiating between 73 populations.…”

“…More precisely, we simulate a sample of 400 individuals per island, each with 20 recombining chromosomes, each with about 2.5•10 4 SNPs. From these ∼ 5 • 10 5 SNPs, we use the step-wise approach from [6] to look for 10 Ancestry Informative Markers (AIMs). When using a naive Bayes approach as in SNIPPER [5], this AIMset gives a vanishing misclassification error for the task of classifying the 3×400 simulated, non-admixed individuals.…”

“…To get a picture of all non-admixed and recently admixed samples, we computed errors for estimating IA as given in (6) for the admixture and recent-admixture model. As described in the MM section, we average estimatesq M andq P from the recent-admixture model, in order to compare to the true IAs.…”

“…Africa, Europe, East Asia, Native America, South-East Asia and Oceania; see e.g. [5,6]), or it is assumed that it consists of a mixture of ancestral genetic material originating in several groups. For the inference of such admixture proportions of individual ancestry (IA), STRUCTURE [4] and the faster ADMIXTURE have become the de-facto standards [3]; see also [5] for the same model.…”

“…the continental BGA of both parents differs. Since methods used in BGA classification [5,6] can only result in single population/class label, the possibility of recent admixture is usually not even implemented by such methods, making results on classification of recently admixed individuals hard to interpret. In mixed membership models as implemented in STRUCTURE or ADMIXTURE for estimating continental (or other scales of geography) IA of a trace, the genome is thought to be a mosaic of stones of different (continental) origin.…”