Current software for genotype imputation

Software platforms and licence

Many potential users will depend on pre-compiled, ready-to-use binaries for their respective platforms to carry out imputation. Platform incompatibilities can therefore represent a serious obstacle for the application of the programs. All three studied tools are offered as pre-compiled versions on their corresponding websites (see Table 1). BEAGLE is written in Java and thus runs on all major computing platforms with the appropriate java interpreter. IMPUTE binaries are available for all major platforms, including Linux, MacOS X, Solaris and Windows. MACH only supports Linux and MacOS X, but offers support for other platforms on demand. Unfortunately, source code is not available for any of the programs. MACH developers have announced that they will share their sources at some point in the future.

BEAGLE software is free, without any restrictions. IMPUTE remains the property of the University of Oxford and is distributed solely for non-commercial use. Software licensing for the MACH program is unclear, but the developers have announced that this will be resolved soon. Optionally, MACH and IMPUTE ask users to register for receiving electronic notifications about future software updates.

Documentation

The extent of software documentation varies considerably between the programs. IMPUTE and MACH provide the user with README files and short web tutorials on how to carry out the main task with the programs. There is a lack of documentation for the MACH program, so the average user is incapable of using MACH with all its features appropriately. By contrast, BEAGLE comes with comprehensive and informative documentation that contains many real-world examples of how to prepare input files, how to deal with memory management, how to handle data and how to analyse the output. Each software package includes example input files.

Graphical user interface and web service

All tools are executed through the command line -- that is, graphical user interfaces (GUI) are not available. This is likely to represent a hurdle for some users interested in these programs, and we recommend that this should be addressed by future software development. Imputation with MACH can also be performed without a local installation, using an online version at the HapMap website (http://www.hapmap.org/); however, the online version imposes a limit of 5000 HapMap reference SNPs.

Authors' responsiveness

Input data format

Many users with an interest in imputation will have only limited experience with scripting and programming of tasks. This includes format conversions before and after imputation, the latter being only the first step in a sequence of analyses. The use of (quasi-) standard data formats that can be generated and processed with a variety of software programs is therefore a prerequisite for widespread software use in genetic epidemiological studies. This includes the Linkage (pre-makeped; http://linkage.rockefeller.edu/soft/linkage/) format, HapMap format (http://www.hapmap.org/) and the increasingly popular PLINK format (http://pngu.mgh.harvard.edu/~purcell/plink/data.shtml), which is a variant of the Linkage format.

Computational demands

Typically, genotype imputation runs are started on a per chromosome basis. Earlier versions of IMPUTE required a huge amount of working memory (RAM) for this task. Since version 0.4.0, it has been possible (and recommended) to carry out genome-wide imputation in chromosomal sub-regions, instead of imputing whole chromosomes. In order to do this, input files do not have to be split manually. Instead, the region of imputation can be specified by command line arguments, which is very convenient. There is also an additional option to avoid edge effects at the borders of the imputed sub-regions. Afterwards, the imputed sub-regions can be easily concatenated to generate imputed files for complete chromosomes, for example by using the 'cat' command under Linux or MacOS X, while redirecting the output into a text file. BEAGLE and MACH do not offer the imputation of particular chromosomal regions with a special treatment of region borders. Their memory requirements are much lower compared to IMPUTE, however, and they have implemented alternative algorithms which pass memory costs to runtime in order to reduce memory usage. The BEAGLE software also supplies the user with a tool and detailed instructions on how to divide the sample cohort (not the reference panel) into sub-samples and perform imputation on each sub-sample separately.

Posterior genotype probabilities

Imputed genotypes are predictions, not actual observations, as obtained from genotyping. Subsequent analyses, such as testing for phenotypic association, should incorporate the uncertainty of these predictions, to avoid spurious results. A general approach for this is the use of posterior probabilities for the imputed genotypes. A less general approach, albeit probably well suited for screening purposes, is to use allele dosage, which is defined as the estimated number of minor SNP alleles for a genotype. Imputation programs should always report values for at least one of these approaches. Fortunately, all three considered programs report the posterior probabilities of genotype calls. In the corresponding files, the posterior probability of observed genotypes (ie those without uncertainty due to imputation) will be 1.0. IMPUTE's genotype output file has exactly the same format as the input file. For each SNP and individual, the prior genotype probabilities have been replaced by the posterior probabilities after imputation. MACH and BEAGLE both have their own file format for the posterior probabilities, which are nevertheless very similar to that of IMPUTE. MACH also reports allele dosages.

Of note is the size of the posterior genotype probability files, which can be very large, depending on the number of imputed samples and SNPs. BEAGLE and MACH report these probabilities with a fixed accuracy of three decimal places. BEAGLE reports the posterior probabilities of all three possible SNP genotypes, resulting in a size of 1476 Mb (100 per cent) for chromosome 6. MACH reports only two of the three probabilities, reducing the size to 938 Mb (63.6 per cent). IMPUTE reports only two decimal places by default, although more can be specified through the respective command-line option, and no decimal places at all if the probability equals zero or one. The resulting output file had a size of 533 Mb (36.1 per cent). Due to the enormous redundancy in the output files, they can be compressed to a tenth of their size using gzip or similar utilities.

Prediction quality

Statistical predictions should always be accompanied by measures of their accuracy. IMPUTE generates an accuracy information file containing two measures. First, the confidence score, which is the average of the maximum posterior probabilities of the imputed genotypes for a SNP, and, second, the information measure, a measure of the observed statistical information associated with the estimate of the allele frequency, are given for each imputed and non-imputed SNP. MACH also produces two slightly different estimates of imputation quality. Its quality score is identical to the confidence score of IMPUTE, while the ratio of the variances of the observed and the estimated allele counts is denoted by R² (also termed r² or OEvar). MACH uses this measure to assess the imputation performance of a SNP. BEAGLE assesses the quality of imputed genotypes by estimating allelic R² for each SNP, which is the squared correlation of the allele dosage with the highest posterior probability and the true allele dosage. As with MACH, this metric can be used to exclude SNPs with poor imputation accuracy.

Association testing

Genotype imputation is typically the first step for subsequent testing of phenotypic association in the exploratory, hypothesis-generating stage of a genetic epidemiological study. Each of the three programs considered here generates output that can readily be used by other programs for such an analysis. Output from BEAGLE can be used with PRESTO, which implements permutation testing of order statistics [14]. For IMPUTE, the SNPTEST tool (available from the IMPUTE website) allows for numerous association tests of binary and quantitative features, as well as for covariate adjustment and the calculation of Bayes factors. MACH has recently been complemented by the programs MACH2DAT and MACH2QTL (available from the MACH website), which can be used to test for association with binary and quantitative traits via regression models. The analysis of imputed genotypes with the software environment R,[15] (eg for model selection purposes) requires the split of the posterior probability files into smaller files of a few thousand SNPs per file for convenient use inside the R environment.

X chromosome imputation

While all considered programs can impute autosomal genotypes, IMPUTE is the only program so far that supports X chromosome imputation. It also properly handles the pseudoautosomal, as well as the non-pseudoautosomal regions of chromosome X. The gender of individuals (male/female) must be specified in an additional file. IMPUTE reports posterior genotype probabilities for females and allele probabilities for males. X chromosome imputation has not yet been implemented in BEAGLE or MACH. There will be an announcement that MACH is to incorporate X chromosome imputation in the next release (Y Li, personal communication). BEAGLE can be expected to include this feature in the near future (B. Browning, personal communication).

Trio data

A unique feature of BEAGLE is its ability to process offspring-parent trio and offspring-parent pair data alone or in combination with unphased or phased data of unrelated individuals. At this point, MACH and IMPUTE can only handle data from unrelated individuals. MACH is expected to handle family data in the near future (Y. Li; personal communication).

Imputation of multi-allelic markers

So far, the principal targets for marker genotype information have been SNP markers. For most SNPs, only two common alleles exist, however, tri-allelic SNPs do exist and multi-allelic markers-- such as short tandem repeats (microsatellites)--are still in use. An example is marker rs2032582,[16] which is listed as tri-allelic in the SNP database (dbSNP), but only as biallelic in HapMap. While IMPUTE and MACH can only handle biallelic markers, BEAGLE is able to process markers with up to 128 different alleles. This makes it the only imputation program suitable for multi-allelic markers.

Estimation of imputation accuracy

Only IMPUTE reports on the estimated distribution of imputation accuracy measures when quitting. IMPUTE implements a leave-one-out approach to estimate the prediction error for SNPs with observed genotypes.

'Correcting' genotypes

In some cases, MACH 'corrects' observed (ie genotyped) SNP genotypes based on the HapMap reference when used with the -mle option. The fraction of corrected genotypes is relatively low (in our experience, more than 99.7 per cent of observed genotypes remain untouched), but still significant. Correcting a highly likely genotyping error could be a rationale for this action, however, neither the feature nor the reasoning is documented. In our view, this lack of information should be addressed by the developers.

Large reference panels

Currently, most genotype imputation is performed using small reference panels from HapMap Phase II. This implies a limitation for imputation accuracy, particularly for markers with a low minor allele frequency. A considerable strength of BEAGLE is its flexibility and capability to use reference panels much larger than the 90 or 120 phased HapMap haplotypes used so far. BEAGLE's computational runtime depends on the combined size of the sample and the reference panel,[7] whereas computational times for MACH and IMPUTE are approximately linear with the size of the sample and quadratic with the size of the reference [1, 8]. A comparison of BEAGLE and IMPUTE for increasingly large reference panels has been published recently [7]. Larger reference panels for many more populations are already available or will become so soon -- for example, from the HapMap Phase III data and from the 1000 Genomes Project (10-15 Mio. SNPs estimated). For such large reference panels, BEAGLE might become the program of choice in the near future.