Inbreeding coefficient estimation with dense SNP data: comparison of strategies and application to HapMap III

If the parents of an individual are related, it is possible for the individual to have received at 1 locus 2 identical-by-descent alleles that are copies of a single allele carried by the parents' common ancestor. The inbreeding coefficient measures the probability of this event and increases with increasing relatedness between the parents. It is traditionally computed from the observed inbreeding loops in the genealogies and its accuracy thus depends on the depth and reliability of the genealogies. With the availability of genome-wide genetic data, it has become possible to compute a genome-based inbreeding coefficient f, and different methods have been developed to estimate f and identify inbred individuals in a sample from the observed patterns of homozygosity at markers.

For this paper, we performed simulations with known genealogies using different SNP panels with different levels of linkage disequilibrium (LD) to compare several estimators of f, including single-point estimates, methods based on the length of runs of homozygosity (ROHs) and different methods that use hidden Markov models (HMMs). We also compared the performances of some of these estimators to identify inbred individuals in a sample using either HMM likelihood ratio tests or an adapted version of the ERSA software.

Single-point methods were found to have higher standard deviations than other methods. ROHs gave the best estimates provided the correct length threshold is known. HMMs on sparse data gave equivalent or better results than HMMs modeling LD. Provided LD is correctly accounted for, the inbreeding estimates were very similar using the different SNP panels. The HMM likelihood ratio tests were found to perform better at detecting inbred individuals in a sample than the adapted ERSA. All methods accurately detected inbreeding up to second-cousin offspring. We applied the best method on release 3 of the HapMap phase III project, found up to 4% of inbred individuals, and created HAP1067, an unrelated and outbred dataset of this release.

We recommend using HMMs on multiple sparse maps to estimate and detect inbreeding in large samples. If the sample of individuals is too small to estimate allele frequencies, we advise to estimate them on reference panels or to use 1,500-kb ROHs. Finally, we suggest to investigators using HapMap to be careful with inbred individuals, especially in the GIH (Gujarati Indians from Houston in Texas) population.