Sequencing millions of animals for genomic selection 2.0

Genomic selection (GS) has made animal breeding an extremely exciting field to have been part of in recent years. Breeding programmes have been redesigned, new genotyping technology, imputation algorithms and statistical methods developed, operational and logistical constraints addressed, and greater understanding obtained about the underlying information that drives the accuracy, the requirements for a successful GS training population design, and the impact of selecting on genomic breeding values on the genetic resources available. In short, a relatively simple recipe and set of easy-to-use utensils are now available that, within the biological, logistical and economic constraints of a breeding programme, are such that GS has a high probability of being very useful. For the purposes of debate, one could claim that there are now three types of GS. The original GS model could be labelled GS0.0 and assumed that linkage disequilibrium between markers and causative mutations would drive prediction. Current implementations of GS could be labelled GS1.0, and they primarily utilize linkage information via realized relationships with close relatives because training populations, although large by historical standards, are far from sufficiently large for linkage disequilibrium information to be very useful for making predictions about quantitative traits. GS2.0 is a label that could be given to the type of GS that will emerge in the next 5 years. Exactly, how GS2.0 will be is of course unknown, but we can be confident that it will involve huge volumes of data; potentially, millions of animals will have data obtained by sequencing. If this is the case, GS2.0 will accumulate the information required for utilizing both linkage disequilibrium and causative nucleotides when making predictions about breeding value. One could be pessimistic about the prospects for sequence data given that it is expensive and has initially shown little increase in accuracy [Clark et al. (2011) Genet. Sel. Evol., 43, 18]. Furthermore, in many breeding programmes, the overwhelming success of GS1.0 has left little room for GS2.0 to improve accuracy. However, with progress in research, this lack of additional benefit from sequence data can be overcome within the next 5 years in the major breeding programmes that will have the ability to generate sequence information on millions of animals. Generating sequence data for millions of individuals will require that the costs per individual be low. In GS1.0, genotyping costs were reduced through the use of genotyping strategies and imputation by capitalizing on some key features of animal breeding populations (large family sizes, large numbers of close relatives, abundant recording of pedigree) and on the ability to genotype very few markers (e.g. 384 SNP) at very low cost. GS2.0 will use the same features to drive sequence imputation and low-cost sequencing. It has been proposed that one way to do this would be to sequence the key ancestors at high-coverage and to then impute this sequence using SNP chips. An alternative strategy may be to sequence all individuals at low-coverage [Li et al. (2011) Genome Res., 21, 940–950], use an imputation like approach to build consensus haplotypes and then impute full sequence information on the basis of these consensus haplotypes. This strategy means that the actual haplotypes of individuals would be sequenced at low-coverage using the individuals own DNA, but high-coverage sequencing of these haplotypes would be obtained by accumulating the low-coverage reads from all the individuals that carry the same haplotypes. This approach has three potential advantages over the approach that uses key ancestors sequenced at high-coverage. Firstly, it seems inefficient to sequence the same haplotype at high-coverage multiple times in different individuals. Secondly, accurate imputation involves the resolving of the combination of haplotypes that an individual carries, and in the era of sequence data, the refinement of recombination locations will be the most important component of this. For an individual to be imputed, low-coverage sequence data can generate many more markers than SNP chips for the same cost. This vastly greater number of markers will allow the recombination locations to be much more precisely located. Thirdly, in GS1.0, accurate genotype imputation was not possible unless very close ancestors (e.g. sires and grandsires) of individuals were genotyped at high-density. Only sequencing key ancestors implies that not all sires and grandsires will be sequenced at high-coverage, and this would severely weaken the power to do imputation. To make a low-coverage sequencing and imputation strategy work, much research must be performed. Most importantly, new imputation algorithms are needed that can account for the uncertainty and randomness of the missing data that arise with low-coverage sequencing. Strategies at the level of the population need to be developed to determine which individuals should be sequenced and at what depth, and these need to account for the different costs and benefits of accumulating sequence data with different qualities in the short, medium and long term. New forms of low-coverage sequencing are needed. Costs of low-coverage sequencing are uncertain and widely debated. I have seen figures of between $15 and $70 for library preparation and approximately $130 for a 1.0x sequence of the genome. On this basis, 0.1x coverage of the genome of an individual would cost as little as $28. There is a possibility to drive these costs down further if large volumes were to be undertaken or if greater research effort was brought to bear upon the topic. Therefore, low-coverage sequencing can be of comparable cost to current low-density genotyping platforms. At such costs, major breeding programmes could genotype perhaps 200 000 or more individuals per year, many of which will be phenotyped for at least some traits. Therefore, across a 10-year period, perhaps several millions of individuals will be genotyped and phenotyped. If imputation algorithms and strategies both at the population and molecular level can be optimized such that the quality of the data and the costs in the short term are competitive with what is currently performed, then huge data sets with sequence information can be generated. Huge data sets offer several possibilities. If these are phenotyped, then larger proportions of the genetic variation for quantitative traits can be fine mapped to causative nucleotides. As data sets get bigger, GWAS and GS will coalesce. This will increase the accuracy and persistency of predictions, could rescue the promise of across breed prediction and make the explicit use of the millions of de-novo mutations that arise naturally in our breeding populations possible. Having large proportions of the genetic variation mapped to causative nucleotides will allow the identification of lots of new targets for drugs or for genome editing (Marx, 2012, Nature Methods, 9:1044-1059), a technique that can be used to create new genetic variation in targeted ways or 'repair' old genetic variation. Finally, with huge volumes of data the accuracy of GS2.0 will asymptote; generation interval cannot be shortened much more in-vivo; maximum selection intensity is restricted by the cost of genotyping candidates; so therefore, the only way to obtain greater rates of genetic gain using the parameters of the breeders equation is through the creation of more genetic variation. Because of the infrequency (and likely cost of artificial induction) of de-novo mutations and because of the resources required to estimate their effects, it is perhaps better to tackle the problem of faster genetic gain through faster release of the standing genetic variation. One way to do this would be to increase the rate of recombination. Recombination is under both environmental and genetic control. It is heritable, and GWAS hits have been discovered [Fledel-Alon et al. (2011) Plos ONE, 6, e20132]. Selection pressure could be placed on recombination, which in turn would allow more permutations of the current standing variation be released in each generation. GS2.0 could be the enabling technology and thus increase the speed of our journey towards the continually moving target that is the ideotype! Genetics is ultimately about recombination, mutation and segregation. In a breeding programme with 200 000 candidates each year, there could be as many as 12 million new genomic segments and four million de-novo mutations. With sequence data for whole populations, these events can be utilized. This will bring benefits if we devote more research on genetics and breeding than to statistical issues!! I would like to acknowledge the many useful discussions and support of Dave McLaren, Kellie Watson, Matthew Cleveland, Andreas Kranis, Birgit Gredler, John Woolliams, former colleagues at Armidale and CIMMYT, and Gregor Gorjanc.