Logo of geneticsGeneticsCurrent IssueInformation for AuthorsEditorial BoardSubscribeSubmit a Manuscript
Genetics. 2008 Apr; 178(4): 2289–2303.
PMCID: PMC2323816

Reproducing Kernel Hilbert Spaces Regression Methods for Genomic Assisted Prediction of Quantitative Traits


Reproducing kernel Hilbert spaces regression procedures for prediction of total genetic value for quantitative traits, which make use of phenotypic and genomic data simultaneously, are discussed from a theoretical perspective. It is argued that a nonparametric treatment may be needed for capturing the multiple and complex interactions potentially arising in whole-genome models, i.e., those based on thousands of single-nucleotide polymorphism (SNP) markers. After a review of reproducing kernel Hilbert spaces regression, it is shown that the statistical specification admits a standard mixed-effects linear model representation, with smoothing parameters treated as variance components. Models for capturing different forms of interaction, e.g., chromosome-specific, are presented. Implementations can be carried out using software for likelihood-based or Bayesian inference.

A massive quantity of genomic information is increasingly available for several species. For example, Wong et al. (2004) reported 2.8 million single-nucleotide polymorphisms (SNPs) in the chicken genome, and Hayes et al. (2004) found 2507 putative SNPs in salmons. Hundreds of thousands of SNPs have been identified in humans (e.g., Hartl and Jones 2005). It is natural to consider use of this information as an aid in genetic improvement of livestock or plants or in molecular classification (or prediction) of diseases. In medicine and agriculture, for example, genomic information could also be used for designing diet or plant fertilization regimes that are genotype specific.

Early discussions on the use of molecular markers in genetic selection programs are given by Soller and Beckmann (1982) and Fernando and Grossman (1989). Subsequently, much work has addressed determining location and use of a single or a few quantitative trait loci (QTL). However, Dekkers and Hospital (2002), in a review of many studies, observed that there are an abundant number of loci associated with variation in quantitative traits. These authors noted that most statistical methods for marker-assisted selection proposed so far do not deal adequately with the complexity (in the sense of number of loci) posed by many traits. A relevant issue to be addressed is how a massive number of SNPs, viewed as covariates with potential explanatory power, can be incorporated reasonably into a statistical model specification. Some hurdles in the process of model building include multiple testing, strong dependence of inferences on assumptions, ambiguous interpretation of effects in a multiple-marker analysis due to collinearity, the famous “curse of dimensionality,” as the number of markers, e.g., SNPs, exceeds by far the number of cases in a sample, and the handling of nonadditive gene action. Balding (2006) discusses many of these problems.

A main challenge is how the many interactions between genotypes at different loci ought to be dealt with. A stylized treatment of epistatic variability from an evolutionary perspective is presented by Cheverud and Routman (1995). Translating this into whole-genome data analysis is another matter: if thousands of marker genotypes are fitted in a model for genomic-assisted prediction, the number of potential interactions and their interpretation can be mind boggling.

First, consider an analysis with random-effects models, so that the variance component parameterization or, more generally, the dispersion structure becomes the focus of the problem. Due to smoothing or “regularization” induced by, e.g., a multivariate normal assumption, all random effects can be predicted uniquely. This is illustrated in Meuwissen et al. (2001), Gianola et al. (2003), and Xu (2003). For instance, animal breeders typically infer a number of breeding values that amply exceed the number of observations available (Quaas and Pollak 1980). However, coping with nonadditive genetic variability introduces additional difficulty. Theoretically, epistatic variance can be partitioned into orthogonal additive × additive, additive × dominance, dominance × dominance, etc., variance components, only under idealized conditions. These include linkage equilibrium, absence of mutation and of selection, and no inbreeding and assortative mating (Cockerham 1954; Kempthorne 1954). These assumptions are violated in nature and in breeding programs. Also, estimation of nonadditive components of variance is very difficult, even under standard assumptions (Chang 1988), leading to imprecise inference. Therefore, whether or not standard random-effects models for quantitative genetic analysis account for nonadditive relationships between genotypes and phenotypes accurately remains an open question.

Second, interactions between markers could be studied using fixed-effects models; this is what Cheverud and Routman (1995) refer to as “physiological epistasis,” to disassociate inference from the gene and genotype frequencies that generate a probability distribution. Such an analysis “runs out” of degrees of freedom quickly in a whole-genome treatment, because there are 2 d.f. per biallelic SNP locus. Even if the number of parameters is reduced in some manner, estimates of effects are expected to be unstable and imprecise, due to severe lack of orthogonality induced, partly, by extant linkage disequilibrium. Also, interactions involving more than three loci are very difficult to interpret. A standard parametric treatment may require a formidable model selection exercise, with any model in particular probably having little plausibility or predictive power. Bayesian model averaging (e.g., Hoeting et al. 1999) is an option, but how can this be made free from some strong and possibly untestable parametric assumptions?

A third and distinct avenue is to explore model-free approaches, which may be useful for phenotypic prediction under subtle or cryptic forms of epistasis. There is little evidence that such methods have been considered in quantitative genetics. Gianola et al. (2006) discussed semiparametric procedures for analysis of complex phenotypic data involving massive genomic information. These authors argued that application of the parametric additive genetic model in selective breeding of livestock produced tangible dividends, as shown in Dekkers and Hospital (2002), and proposed combining a nonparametric treatment of effects of molecular SNPs with features of the additive polygenic mode of inheritance.

The objective of this article is to develop further a reproducing kernel Hilbert spaces (RKHS) mixed model proposed by Gianola et al. (2006), with a focus on its theoretical aspects. The accompanying article by González-Recio et al. (2008, this issue) presents an application of the methodology to data on chicken mortality.

This article is organized as follows. The semiparametric mixed model section sets the stage and introduces notation. The nonparametric treatment (RKHS) adopted here is sketched in the reproducing kernel hilbert spaces regression section, where the main theoretical results are presented; additional details are in the appendix. dual formulation shows how the problem can be embedded into a mixed-effects model structure and discusses how statistical learning proceeds in a penalized-likelihood framework. The rkhs chromosome mixed model section presents a linear model aimed at capturing interactions between many loci at different chromosomes and presents a Bayesian implementation. The article concludes with a discussion of some standing issues.



The notation follows that of Gianola et al. (2006). Each of n individuals possesses a measurement for some quantitative trait denoted as y and information on a possibly massive number of SNP genotypes represented by a vector x. An SNP locus is considered biallelic, so at most three genotypes are observed. Genotype instances can be coded uniquely via two linearly independent variables per locus as in an analysis-of-variance setting, i.e., with 2 d.f. per locus. In standard quantitative genetics settings, the two dummy variates are coded such that the corresponding effects are interpretable as “additive” and “dominance.” This is irrelevant from the predictive point of view taken here, in the sense that parameters (most of which lack a mechanistic interpretation) serve as transition tools, to pass from observed to predicted data.

Suppose, temporarily, that there are no nuisance variables and that the focus is on discovering a function relating xi to yi. Three alternative modeling possibilities are considered, for illustrative purposes.

  1. Let the relationship between y and x be represented as
    equation M1
    where yi is a measurement on the quantitative trait for individual i, xi is a p × 1 vector of dummy SNP instance variates observed on i, and equation M2 is some unknown function relating genotypes to phenotypes. Define equation M3 as the conditional expectation function, that is, the mean phenotypic value of an infinite number of individuals, all possessing the p-dimensional genotypic attribute vector xi. equation M4 is a random residual, distributed independently of xi and with variance equation M5. Typically, the residual distribution is assumed normal.
    The vector x may have a probability distribution reflecting frequencies of the SNP attributes in the population. However, the prediction problem normally centers on what can be expected about the phenotypic distribution, given some specific configuration x = x*, say. In nonparametric regression, equation M6 is left unspecified and estimated as a smooth equation M7; this function represents pertinent signals on the phenotype from elements of xi, acting either additively or as members of some genetic network. Several techniques for inferring equation M8 are described in Takezawa (2005).
  2. A second specification is the additive regression model
    equation M9
    (Hastie and Tibshirani 1990; Fox 2005), where xij is the value of attribute j in individual i. Each of the “partial-regression” functions equation M10 allows exploration of effects of individual attributes on phenotypes. This model is expected to pick up additive and dominance effects at each of the marker loci, but not epistatic interactions. It does not possess any clear advantage over a standard regression model with additive and dominance effects, the main difference residing in the nonparametric treatment that (2) would receive.
  3. One could also think in terms of an additive “chromosome” model, as follows. Let C be the number of pairs of chromosomes, and partition vector xi as equation M11, so that equation M12 contains the values of the SNP instance variates at chromosome pair j equation M13, and so on. If the number of SNPs in chromosome pair j = equation M14, then the order of xij is pj, and the dimension of x is equation M15. The additive chromosome model is
    equation M16
    with xij being the attributes observed on chromosome pair j for individual i. This model would account for chromosome-specific signals (reflecting additive, dominance, and any relevant epistatic effects involving genes in chromosome pair j) and combine all these additively over pairs of chromosomes. Examples of tightly linked genes having epistatic effects are the major histocompatibility complex and the lac operon in Escherichia coli. Evidence of epistatic interactions among linked loci in plants is in Fenster and Galloway (2000), who studied fitness traits in the annual legume Chamaecrista fasciculata. The interplay between epistasis, linkage, and linkage disequilibrium is an old topic in population genetics (Kimura 1965; Franklin and Lewontin 1970).
    Another modeling option consists of dividing all chromosomes somehow into R genomic regions of equal or different sizes and then combining the R-region-specific signals additively.

Models (1)–(3) are nonparametric descriptors of situations in which epistasis plays different roles, i.e., a major one in (1), none in (2), or involving only linked genes in (3). In what follows, model (1) is retained for presentation of theoretical developments, which are extended to model (3) later on.

Additional structure:

Animal breeders have exploited to advantage the additive model of quantitative genetics, embedding it into a mixed-effects linear model specification. Basing selection of parents on predictions of additive genetic values, notable genetic progress has been attained in many species, such as dairy cattle, pigs, and poultry. While it is possible to accommodate some types of nonadditive gene action in a parametric manner, the assumptions are very strong. Further, construction and inversion of “epistatic relationship matrices” are daunting and a realistic parametric treatment is simply not available. Hence, as argued by Gianola et al. (2006), it seems reasonable to expand (1) as

equation M17

where equation M18 is an f × 1 vector of nuisance location parameters and u is a q × 1 vector containing additive genetic effects of q individuals (these effects are assumed here to be independent of those of the markers), some of which may lack a phenotypic record, so typically n < < q; equation M19 and equation M20 are known nonstochastic incidence vectors. As before, g(xi) is an unknown function of the SNP data, to be inferred. It is assumed that equation M21, where equation M22 is the additive genetic variance due to unmarked polygenes and A is the additive relationship matrix, whose entries are twice the coefficients of coancestry between individuals. Let equation M23 be the n × 1 vector of residuals, and take equation M24. In matrix notation

equation M25

where equation M26 and equation M27 are incidence matrices of appropriate order. Further, equation M28 is a vector of order n × 1, an unknown function of marker matrix X, with n rows and p columns; a row of X contains the p SNP instance variates (two per marker locus) observed in individual i.

Gianola et al. (2006) suggested backfitting-type algorithms in which, first, g(xi) is estimated for i = 1, 2, …, n, via some nonparametric estimate equation M29, and then a standard (frequentist or Bayesian) mixed-model analysis is carried out using the “corrected” data vector and pseudomodel

equation M30

where equation M31 is a residual vector. The pseudomodel ignores uncertainty about equation M32, because equation M33 is treated as if it were the true regression (on SNPs) surface and equation M34 is regarded as having the same distribution as e, which is of course not true in finite samples. Subsequently, some estimates of equation M35 and u are obtained, and the offset equation M36 is evaluated at these estimates, to produce a new fit of g(xi). The backfitting algorithm iterates back and forth between the nonparametric and parametric phases. At convergence, the “total” genetic value of individual i is assessed as equation M37, where equation M38 is the converged value of the empirical best linear unbiased predictor (or of a posterior mean in a Bayesian analysis) of ui and equation M39 is the converged nonparametric smooth of g(xi). Instead, a self-contained approach for inferring u and equation M40 is discussed in what follows.



A precise account of the theory is beyond the scope of this article, so only essentials are given here. Foundations and some applications are in Aronszajn (1950), Kimeldorf and Wahba (1971), and Wahba (1990, 1999, 2002). Some essential theoretical details and term definitions are presented in the appendix.

Consider inferring a function g from data y, without any assumptions. The problem is ill-posed, because any function passing through the data would be acceptable (Rasmussen and Williams 2006). Bayesians introduce assumptions via a prior over functions, but this problem has also been tackled using “regularization,” i.e., by imposing some smoothness assumptions on g. This second approach starts by considering the functional (a function containing functions as part of an argument)

equation M41

where equation M42; equation M43 is some function of the data and of equation M44 a is a positive smoothing parameter (typically unknown); and equation M45 is some norm or “stabilizer” under a Hilbert space H, a space of functions on a set having an inner product equation M46 and a norm equation M47 for g1, g2epsilonH (Wahba 2002; Mallick et al. 2005).

Optimizing function:

Consider functional (4), and let

equation M48

which is a deviance measure, assuming temporarily that u is a fixed parameter in the frequentist sense; subsequently, a random-effects treatment of u is made. Making explicit the dependency of the functional on the positive smoothing parameter a, write

equation M49

where the factor equation M50 is introduced for convenience. The second term in (5) acts as a penalty because it adds up to the deviance. It is also known as a regularizer, representing smoothness assumptions encoded in the RKHS. The issue here is finding the function g(x) that minimizes (5), which is a calculus-of-variations problem over a space of smooth curves. The solution is given by the representer theorem of Kimeldorf and Wahba (1971); see Wahba (1999) for a more recent account and O'Sullivan et al. (1986) for extensions to generalized linear model deviances. The representer theorem states that the minimizer has the form

equation M51

where the α's are unknown coefficients and the basis function equation M52 is a reproducing kernel, possibly dependent on some parameter h. While x is p × 1, there are n + 1 coefficients in the function. The intercept α0 can be included as part of equation M53, so that the focus is on α1, α2, …, αn. A possible kernel to be used as a basis function (Mallick et al. 2005) is the single-smoothing-parameter squared exponential (Gaussian) function

equation M54

The values of equation M55 range between 0 and 1, so the kernel is positive definite and acts as a correlation, in the sense that the closer xj is to x, the stronger the correlation is. Parameter h controls the rate of decay of the correlation: smaller h values produce a sharper correlogram. Define now the 1 × n row vector

equation M56

the n × n symmetric matrix equation M57 of kernels, which can be interpreted as a correlation matrix; and the n × 1 column vector equation M58, j = 1, 2, …, n. Then, the minimizing function (6) can be expressed in vectorial manner as the linear function of equation M59:

equation M60

These results can now be employed in (5), leading to a function having equation M61, and equation M62 as arguments, given a and h. One obtains

equation M63

Using (A1) in the appendix,

equation M64

Now, a vectorial generalization of the result in (A4) of the appendix is

equation M65

This can be used in (9), because the integral is a vector valued inner product of the kernel and of each minimizer (7); that is,

equation M66

Finally, this can be employed in (8), producing

equation M67

Note that (10) does not include a penalty for the random vector u. This is added later on.

Minimizer of the penalized sum of squares:

The gradients of equation M68 with respect to the parametric equation M69 and nonparametric equation M70 coefficients are

equation M71
equation M72


equation M73

Noting that Kh is symmetric (so that equation M74; the notation equation M75 is retained to facilitate analogies with mixed-model methodology), the first-order condition is satisfied by the system

equation M76

with the notation emphasizing the dependence of the solutions on a and h. There is no unique solution to this system, because the number of equations (p + q + n) exceeds the rank of the coefficient matrix. This problem is solved by a random-effects treatment of u via the assumption equation M77, stated earlier. Under this assumption and the penalized-likelihood framework, the objective function to minimize becomes

equation M78

Taking derivatives as before, setting to zero and rearranging produces now

equation M79

which has full rank if the elements of equation M80 are defined uniquely, i.e., as a set of linearly independent estimable functions. There is a clear parallel between the forms of the u-equations and of the α-equations. In particular, in the nonparametric part, equation M81 plays the role of equation M82. Hence, one can arrive at representation (12) by making the assumption equation M83, where equation M84 is the “variance” of the equation M85-effects, and equation M86 is their correlation matrix. Fortunately, this inverse is not needed for solving (13) as Kh is a dense n × n matrix. The α-equations can be rearranged such that the solution for the nonparametric coefficients is

equation M87

Large linear systems such as (13) have been solved routinely in animal breeding since the 1980s (e.g., Quaas and Pollak 1980). System (13) differs from Equation 26 in Gianola et al. (2006), which has equation M88 instead of equation M89. The latter is the correct RKHS representation.


The linear model:

The preceding results imply that the RKHS approach is equivalent (this is referred to as a “dual” formulation) to the linear model

equation M90

under the assumptions that equation M91 is a “fixed” vector and that the random effects equation M92, and e are distributed independently as equation M93, and equation M94, respectively. Hence, given h, implementation of the RKHS regression is as in a standard mixed-effects linear model, especially if the kernel matrix does not involve any parameter(s) h. For instance, variance components can be estimated via restricted maximum likelihood; subsequently, point estimates of equation M95, and equation M96 are obtained by solving (13) evaluated at the variance-components estimates. If equation M97 is large (implying that a is large), the estimated equation M98-coefficients are expected to be near 0. A remarkable aspect of the RKHS procedure is the mutual exchange of information between α-coefficients due to the nontrivial correlation structure induced by Kh. This is similar to the exchange of information between relatives induced by A in the classical additive genetic model.

Effective number of parameters:

In a standard regression model (equation M99-coefficients only) the degrees of freedom of the model are given by equation M100, provided the incidence matrix has full-column rank (this can be assumed without loss of generality). The fitted value is equation M101, and the n × n matrix equation M102 is called the smoother operator (Hastie and Tibshirani 1990). Note that the degrees of freedom can also be arrived at by taking equation M103.

Let now equation M104 so that in the context of (14), the vector of fitted values is equation M105, where C is the coefficient matrix in (13). Hence, a measure of the effective number of parameters fitted in RKHS regression

equation M106


equation M107


equation M108

where Cuu and Cαα are the u- and α-blocks of C−1. Then, in some sense equation M109 is the effective number of additive genetic effects fitted and equation M110 is the effective number of α-coefficients. If in (12) equation M111 and a = 0 (equivalently, equation M112 = ∞), (15) is equal to p + q + n and, in the limit, the degrees of freedom of the model are given by the rank of equation M113, so that the model interpolates the data. On the other hand, as equation M114 and equation M115 tend to 0 (equation M116), the effective number of parameters fitted decreases, and the model becomes less capable of reflecting potentially existing patterns in the data.

Uncertainty about predictions:

Given h and ignoring the error of estimation of variance components, an estimator of the variance–covariance matrix of the estimates of equation M117, and of the prediction errors of u and equation M118, is given by

equation M119

Let Sh = QhC−1equation M120 be the smoothing matrix. The variance–covariance matrix of the vector of fitted values is

equation M121

and the variance–covariance matrix of the fitted residuals is

equation M122

In a genetic context, a relevant prediction problem is that of inferring a vector of future observations y* in individuals possessing marker genotypes X*, so that the unknown molecularly marked genetic value, or contribution to phenotype, is equation M123. The model for the future observations is

equation M124

with its coefficients inferred from currently available data y. The point predictor is equation M125, and the prediction error variance–covariance matrix is

equation M126

where I* is an identity matrix with as many rows and columns as there are elements in y*. A confidence band for the elements of y* based on the pointwise standard errors of prediction (Hastie and Tibshirani 1990) is given by equation M127, where equation M128 denotes a vector whose elements are equal to twice the square root of the diagonal elements of equation M129. The confidence band does not consider the uncertainty in the estimates of variance components as well as that associated with h.

Tuning parameter:

If the kernel matrix involves one or more h's, some value(s) needs to be arrived at. Typically, cross-validation (CV) is used (e.g., Craven and Wahba 1979; Wahba 1990; Golub et al. 1999). The simplest method (albeit computationally intensive) is the leave-one-out cross-validation measure

equation M130

where equation M131 is a vector of fitted values resulting from n fits obtained from deleting y1, y2, … , yn, respectively. For instance, equation M132 is the fitted value of observation 1 using all data other than y1, and so on. The value of h chosen results from minimizing equation M133 over a grid. Clearly, this is not computationally feasible in most quantitative genetic data sets, where n can range from hundreds to millions of observations. In such a situation, one may wish to carry out a cross-validation involving a less intensive level of folding, e.g., a leave 20%-out assessment. A more appealing (and, on some grounds, theoretically firmer) criterion is the generalized cross-validation

equation M134

Using (15), the statistic becomes

equation M135

The main difficulty here is the calculation of the inverse matrices under the trace operator. These traces may be approximated (animal breeders have proposed approximations required for REML computations) or estimated via Monte Carlo sampling; e.g., in some Bayesian contexts the diagonal elements of Cuu and Cαα are proportional to posterior variances.


The linear model:

Consider again the additive chromosome specification (3), but now in the context of linear model (14). For ease of presentation, let the number of pairs of chromosomes be C = 2; generalization is straightforward. The unknown function of SNP genotypes to be inferred is

equation M136

Using the dual formulation, the model can be written as

equation M137

where equation M138 is an n × n matrix with typical element equation M139, equation M140 is also n × n with typical element equation M141 and h2 are the decay parameters corresponding to SNPs in chromosome pairs 1 and 2, respectively, and equation M142 and equation M143 are each n × 1 vectors of coefficients. As before, xi1 and xi2 are p1 × 1 and p2 × 1 genotype incidence vectors pertaining to the appropriate chromosomes; recall that the number of markers in the two chromosomes is given by equation M144 and equation M145, because two dummy variates are needed for coding the three genotypes uniquely. The counterpart of objective function (12) is

equation M146

Using the dual formulation, the form of the penalty is equivalent to making the assumption

equation M147

that is, equation M148 and equation M149 are independently and normally distributed with covariance matrices equation M150, i = 1, 2. Letting equation M151, equation M152, the counterpart of (13) is

equation M153

where equation M154, etc., denote that the solution vector in question depends on equation M155.


The procedure can be carried out in a non-Bayesian manner as follows:

  • Define a grid of h1, h2 values.
  • For each point in the grid, estimate equation M156, and equation M157 and (19).
  • For each point in the grid, calculate the fitted values
    equation M158
    the smoothing matrix equation M159, and the generalized cross-validation criterion.
  • Choose the combination of h1, h2 values optimizing equation M160, and predict future observations as outlined previously.

Bayesian approach:

The procedures described for the “global” and chromosome models (14) and (17), respectively, do not take into account uncertainty about unknown parameters. This can be addressed by adopting a Bayesian perspective; see Mallick et al. (2005) and Gianola et al. (2006). Here, a Bayesian analysis of the chromosome model (16) using the dual formulation (17) is outlined.

Let the collection of unknowns be equation M161, where, as before, y* is a vector of future phenotypic values to be predicted and

equation M162

Assume the joint prior density has the form

equation M163

where H denotes all hyperparameters (whose values are fixed a priori) and equation M164 indicates a multivariate normal distribution with appropriate mean vector and covariance matrix; the prior equation M165 is discussed below. The four variance components equation M166 are assigned independent scaled inverse chi-square prior distributions with degrees of freedom ν and scale parameters S2, with appropriate subscripts. Assign an improper prior distribution to each of the elements of equation M167 and, as in Mallick et al. (2005), adopt independent uniform priors for h1 and h2 with lower and upper boundaries hmin and hmax, respectively.

Given equation M168, observations are assumed to be conditionally independent, so the distribution of the observed (y) and future (y*) data is

equation M169

where n* is the order of the future data vector. Given equation M170, and equation M171, future observations are independent of past ones. Let equation M172 and equation M173.

Given h1, h2 the setting is that of a Bayesian analysis of a mixed linear model, and Markov chain Monte Carlo procedures for this situation are well known (e.g., Wang et al. 1993, 1994; Sorensen and Gianola 2002). All conditional posterior distributions are known, except those of h1, h2. A Gibbs–Metropolis sampling scheme can be used in which conditional distributions are used for drawing equation M174, and y*, and a Metropolis–Hastings update is employed for h1 and h2. The distributions to be sampled from are considered successively.

  • Draw location effects equation M175 from a multivariate normal distribution with mean vector
    equation M176
    and covariance matrix equation M177.
  • Sample additive genetic effects from a normal distribution centered at
    equation M178
    and with covariance matrix equation M179.
  • The conditional posterior distributions of each of the coefficients equation M180 are multivariate normal as well, with mean vectors
    equation M181
    equation M182
    and variance–covariance matrices
    equation M183
    equation M184
  • The variance components have scaled inverse chi-square conditional posterior distributions and are conditionally independent. The conditional posterior distributions to sample from are as follows, where ELSE denotes all parameters other than those being sampled,
    equation M185
    equation M186
    equation M187
    equation M188
    where equation M189 is the vector of residuals evaluated at the current sample values of the location effects.
  • The most difficult parameters to sample are h1 and h2. However, if the kernels do not involve h's this step is omitted from the sampling process. If uniform (bounded) priors are adopted, their conditional posterior density is
    equation M190
    where equation M191 is an indicator function taking the value 1 if both h parameters are between the bounds and 0 otherwise. Further, the residual vector equation M192 has as its ith element
    equation M193
    recalling that
    equation M194
    equation M195
    Density (32) is not in a recognizable form. However, a Metropolis algorithm (Metropolis et al. 1953), as suggested by Mallick et al. (2005) and Gianola et al. (2006), can be tailored for obtaining samples from the distribution equation M196. Let the Markov chain be at state equation M197 and draw proposal values equation M198 and equation M199 from some symmetric candidate-generating distribution. The proposed values are accepted with probability
    equation M200
    If the proposal is accepted, then set equation M201 otherwise, the chain stays at equation M202.
  • Finally, the vector of yet to be observed phenotypes is inferred from samples drawn from the conditional distribution
    equation M203
    with the values of equation M204, and h2 evaluated at the corresponding realizations from the current round of Markov chain Monte Carlo sampling. When the algorithm converges to the equilibrium distribution, the samples of y* drawn are from the predictive distribution equation M205, which fully takes into account the uncertainty about all unknown model parameters.


This article presents a procedure for quantitative genetic analysis using information on whole-genome markers, such as SNPs, and phenotypic measurements for some complex candidate trait. Focus is on theory and methods based on RKHS regression, arguably the state of the art in functional data analysis (Wahba 1990; Gu 2002; Wood 2006). The model contains a parametric component, represented by the classical additive genetic model of quantitative genetics, and an unknown function or set of functions of SNP genotypes that is dealt with nonparametrically, as in the generalized additive models of Hastie and Tibshirani (1990). The number of nonparametric functions employed is a model choice issue, and many alternative specifications can be formulated. Here, a global function was considered, expected to reflect all relevant “genetic signals,” e.g., dominance and various forms of epistasis, as well as a sum of chromosome-specific functions, each of which is expected to capture dominance as well as epistasis involving linked loci at the corresponding chromosome.

The parametric component includes additive genetic effects only. Dominance and (some) epistatic effects can be handled parametrically. However, a standard treatment requires constructing and inverting large and possibly dense matrices such as, e.g., A # A # D # D if an additive × additive × dominance × dominance variance component were to enter into the parameterization; # denotes Hadamard product. Further, the Cockerham–Kempthorne decomposition machinery available for dealing with epistatic variance collapses under inbreeding and selection; e.g., all sorts of covariances between genetic effects crop up, rendering the parametric approach invalid. A related point is related to limitations of the orthodox view of nonadditive genetic effects in quantitative genetics. The classical definitions of epistatis pertain to a model in which effects enter linearly when forming the genotype. However, one could argue that biology is far from linear. For instance, it may not be enlightening to think in terms of variance components in situations in which a phenotype results from a sum of sine and cosine waves or when nonadditivity enters via terms such as equation M206, where the a's are effects of alleles at some different loci.

Gianola et al. (2006) discuss how a “nonparametric analog of breeding value” can be derived via a Taylor series expansion; this may have merit in genomic selection contexts in which the objective is to increase (or decrease) additive genetic value for some quantitative trait. Caveats and generalizations of the procedures are discussed next.

Filtering SNPs:

Availability of a massive number of SNPs does not necessarily imply that all markers should be included in a prediction model. Apart from standard preprocessing based on minimum allele frequency or information content, it may well be that predictive ability is enhanced by reducing the dimension of the features used as input in a model. For example, Long et al. (2007) described a machine-learning technique based on filtering (using entropy reduction) and wrapping (Bayesian classification performance), to process >5000 SNPs genotyped in broiler families. Predictive ability of bird mortality was increased when the top (based on information gain) 50 SNPs were downsized to 24. More research is needed in regard to the strategic use of markers, e.g., using a few ones vs. all, or on the assignment of different window-width parameters to genomic regions in a whole-genome treatment.

Model choice:

A natural question is that of the model to be used for predicting phenotypes. For instance, should a chromosome model be adopted, instead of a global specification? Comparison of models is a complex issue, as some specifications are better for describing observed data, while others may have a superior predictive ability. See Sorensen and Waagepetersen (2003) for a case study using several criteria for Bayesian model comparison. Some non-Bayesian techniques are discussed by Hastie and Tibshirani (1990) and Wood (2006). The latter include likelihood ratios evaluated at the penalized likelihood estimates and differences in deviances using approximations to the model degrees of freedom; none takes into account the uncertainty associated with the estimates of the smoothing parameters. A different approach, called BRUTO (Hastie and Tibshirani 1990) is based on iterative minimization of a modification of the generalized cross-validation statistic GCV(.) discussed earlier. For example, in a model with additive functions for each of C chromosomes, there would be 2C tuning parameters equation M207 involved in the BRUTO iteration. Bayesian methods of model comparison have a stronger formal justification.

Incomplete genotyping:

In animal breeding, it is not feasible to genotype all individuals for the SNPs. For instance, poultry breeding and cattle artificial insemination companies typically genotype sires only. The number of animals with phenotypic information can be in the order of hundreds of thousands, and genotyping is selective, so animals with SNPs are not a random sample from the population. Methods for dealing with incomplete molecular information are discussed by Gianola et al. (2006), and some include sampling of genotypes. Here, the problem is revisited in the light of RKHS regression, primarily to illustrate difficulties.

Assume a global model with a single h parameter. The vector of phenotypic values is partitioned as equation M208, where y1 (n1 × 1) consists of records of individuals lacking SNP data, and y2 (n2 × 1) includes phenotypic data of genotyped individuals. In animal breeding n1 > p > > n2. Write the model

equation M209

where equation M210 is an unobserved n1 × n1 matrix of kernels, equation M211 is also an unobserved n1 × n2 matrix, equation M212 is the n2 × n2 matrix of observed kernels, and equation M213 and equation M214 are n1 × 1 and n2 × 1 vectors of coefficients. Specifically

equation M215
equation M216
equation M217

where equation M218 denotes the vector of unobserved SNP genotypes in individuals with phenotypes y1. Assign the multivariate normal distribution (suppressing dependence on h in the notation)

equation M219

where equation M220. Given equation M221, h, equation M222, equation M223, equation M224, and the phenotypes y1 and y2, the best linear unbiased predictor (conditional posterior mean) of u and equation M225 is the solution to

equation M226


equation M227

As shown in (34) both the coefficient matrix and the vector of right-hand sides depend on equation M228. From a Bayesian perspective under Gaussian assumptions,

equation M229

where xmiss is the collection of unobserved SNPs equation M230 and x denotes the SNPs of all genotyped individuals. Assuming equation M231, equation M232, and h are known, for simplicity, one is interested in arriving at the unconditional expectation

equation M233

so the Bayesian solution requires averaging over the conditional distribution equation M234. This is a formidable probabilistic imputation, although some simplification is possible. It would seem reasonable to approximate this distribution by equation M235, arguing that, conditionally on x, phenotypic values do not provide much additional information about SNP genotypes. Subsequently, form the Bayesian classifier

equation M236

where equation M237 is the prior probability of observing genotype configuration equation M238 in the population. The prior probability can be estimated using models with various degrees of refinement; e.g., one can assume linkage equilibrium and estimate the joint prior probability from the product of the marginal distributions at individual SNP loci. Likewise, equation M239 can be approximated by some naïve probability calculation, e.g., assuming independence between individuals and loci. The denominator can also be approximated as

equation M240

where S is a set of missing values having relatively high plausibility. Naïve Bayesian classifiers have been enormously successful in the machine-learning literature (e.g., Elkan 1997) and may prove to be competitive against some involved genotype sampling procedures that have been suggested (Van Arendonk et al. 1989; Fernando et al. 1993; Sheehan and Thomas 1993; Jensen et al. 1995; Kerr and Kinghorn 1996; Jensen and Kong 1999; Fernández et al. 2002; Stricker et al. 2002). Once an approximation to equation M241 is arrived at, this can be used as mixing distribution in (35). Alternatives are discussed in Gianola et al. (2006), including fitting a bivariate model.

Choice of kernel and discreteness of genotypes:

Following Hastie and Tibshirani (1990) and Wood (2006), the developments carry when the distribution of y is a member of the exponential family. As in density estimation, many candidate kernels are available, and attaining a good predictive behavior is critically dependent on the choice of kernel. As noted, some kernels do not involve tuning parameters; for instance, see González-Recio et al. (2008).

The theory of RKHS regression holds for a continuously valued x. It is unknown if the procedures are robust with respect to using a Gaussian kernel function when, in fact, SNP genotypes or haplotypes are discrete. Silverman (1986) discussed univariate density estimation and concluded that various kernels differed little in mean squared error. It is unknown if this robustness argument holds for RKHS regression (clearly the inner product arguments are valid for the continuous case), but a discrete approximation may work.

A kernel suitable for discrete covariates is proposed here. For a biallelic SNP, there are three possible genotypes at each “locus.” Suppose the elements of x are coded as 0, 1, 2, to denote the appropriate genotypes. For an x vector with p coordinates, its statistical distribution is given by the probabilities of each of the 3p outcomes. With SNPs, p can be very large (possibly much larger than n), so it is hopeless to estimate the probability distribution of genotypes accurately from observed relative frequencies, and smoothing is required (Silverman 1986). The number of disagreements between a focal x and the observed xi in subject i is given by

equation M242

where equation M243 takes values between 0 and 4p. As an illustration, if “genotype” AABbccDd is the focal point and individual i in the sample is aaBbccDD, then, equation M244. Following Silverman (1986), one could use the “binomial” kernel

equation M245

with equation M246; alternative forms of the kernel function are discussed by Aitchison and Aitken (1976) and Racine and Li (2004). The (pseudo-)RKHS dual formulation would take the form

equation M247

Alternatively, incidence of the three genotypes at a locus can be described with two “free” dummy variates, as in standard ANOVA. There are p predictor variates, where p/2 is the number of markers. Let x1k and x2k be the two dummy variates at locus k (equation M248). Define

equation M249

which give the number of disagreements between a focal equation M250 and the observed equation M251 in subject i. Each of the d's varies between 0 and 1. Then, let

equation M252

each varying between 0 and p/2. Subsequently, consider the “trinomial” kernel

equation M253

where xi is the “focal” p × 1 vector of covariates and xj is the observed value in individual j. If each of the h1, h2 parameters takes values in 0–1, such that 0 < h1 + h2 < 1, then k takes values in 0–1 and is a suitable candidate as kernel, because the matrix equation M254 would be positive definite. The dual (pseudo-)RKHS representation would be

equation M255

Then, procedures for inferring the variance components and h parameters outlined in this article would be followed. Research is needed for assessing the adequacy of these approximations. González-Recio et al. (2008) present an application of the trinomial kernel.


The methods presented here take the whole-genome view of those of Meuwissen et al. (2001), Gianola et al. (2003), Xu (2003), Yi et al. (2003), Ter Braak et al. (2005), Wang et al. (2005), and Zhang and Xu (2005). The main difference is the attempt to capture unknown forms of interaction between many loci that, arguably, parametric models are not able to explore properly, due to either violation of assumptions (for decomposition of epistatic variance) or inadequate statistical machinery for understanding high-level “physiological epistasis.”

An application of the theory presented here is in González-Recio et al. (2008). Using mortality rates observed in 200 families of paternal half-sib broilers, these authors compared the predictive ability of a standard parametric mixed model against that from linear regression (fitting additive effects of 24 SNPs), kernel regression, RKHS, and the Bayesian regression model of Xu (2003). For RHKS, the kernel consisted of a similarity score between any two SNP sequences. The five models were contrasted in terms of a fivefold predictive cross-validation. Results indicated an advantage of RKHS, which had a global “accuracy” that was twice as large as the one from the mixed model, was 2.5 times larger than the one attained with the linear regression specification, and exceeded that attained with the procedure of Xu (2003) by 25%. However, predictive cross-validation accuracy was not large, probably due to the very low heritability of the trait used for the case study, chick mortality.

It should be noted that a kernel function explores commonalities in some sense; e.g., in a chromosome model markers in a contiguous position borrow information. In spirit, this is similar to the use of relationship or identity-by-descent matrices between individuals or cultivars in animal and plant breeding, respectively.

Recent developments in nonparametric statistics and machine learning offer exciting avenues for whole-genome analysis of quantitative traits and perhaps suggest a change in analytical paradigms. This theoretical article intends to make a contribution in this direction.


Gustavo de los Campos and Oscar González-Recio are thanked for useful discussions. Support by the Wisconsin Agriculture Experiment Station and by National Science Foundation (NSF) grant DMS-NSF DMS-044371 to D. Gianola is acknowledged. Research was completed while D. Gianola had a Senior Researcher position within a Marie Curie European Transfer of Knowledge–Development project with contract no. MTKD/I-CT-2004-14412. J. B. C. H. M. van Kaam acknowledges support by grant D.M. 305/7303/06 of the Ministero delle Politiche Agricole e Forestali.


In an Euclidean space of dimension n, the dot product between vectors v and w is equation M256, and the norm is equation M257. The inner product generalizes the dot product to vectors of infinite dimension. For instance, in a vector space of real functions with domain equation M258, the inner product is

equation M259

and the norm is equation M260. If x is a continuous random variable with probability density function equation M261, the inner product (Hastie and Tibshirani 1990) is

equation M262

Consider now the choice of basis functions for x, that is, a transformation of the input (SNP) space to be used as regressors in the nonparametric regression. A kernel equation M263 is a function that maps inputs α, x into some space, and the kernel is said to be symmetric if equation M264 (Rasmussen and Williams 2006). A kernel equation M265 involving random vectors x, t is positive definite if

equation M266

for all functions g, where equation M267 is a joint density. An eigenfunction of positive-definite kernel k with eigenvalue λ satisfies the equation

equation M268

There are an infinite number of eigenfunctions equation M269 with an ordering corresponding to λ1 ≥ λ2 …. The eigenfunctions are orthogonal and can be normalized to satisfy

equation M270

Mercer's theorem (Rasmussen and Williams 2006) enables expressing a kernel in terms of its eigendecomposition such that

equation M271

If all eigenvalues are positive, the sum is infinite. If, on the other hand, the sum terminates at some value p, say, this yields a degenerate kernel of rank p. For example, in a linear random-regression model with coefficients equation M272 in which the x variables are transformed into orthonormal basis functions equation M273, the regression function equation M274 evaluated at points x1, x2, … , xn generates the covariance matrix equation M275 of rank n; recall that, in our situation, n < < p. Here, the eigendecomposition becomes that of a covariance matrix of finite dimension. In general, equation M276 is called a covariance function, and K is an infinite-dimensional generalization of a covariance matrix. The kernel is clearly symmetric, as equation M277.

The notation equation M278 and equation M279 will denote that, at some fixed point x*, these functions take values equation M280 and equation M281, respectively. A space of functions H is a RKHS with kernel k if the two following conditions are met (Wahba 2002; Mallick et al. 2005; Rasmussen and Williams 2006):

  • For every x, equation M282 is in the Hilbert space H.
  • For all x and for every g in H the inner product equation M283 holds; this is a reproducing property, in some sense.

Consider now a Hilbert space constituted by linear combinations of the orthonormal eigenfunctions; e.g., equation M284 and equation M285, where fi and vi are loadings or regression coefficients on the eigenfunctions. Using definition (A1) and orthonormality condition (A2), the inner product is

equation M286

Because equation M287, the Hilbert space has a norm. In its infinite-dimensional form equation M288, this implies that the sequence of fi coefficients must decay quickly, which imposes smoothness conditions (Rasmussen and Williams 2006).

Next, examine if equation M289 is in the Hilbert space spanned by functions such as f and v above. First, recall the eigendecomposition of kernel equation M290. Using again the orthogonality property, the inner product

equation M291

This shows that the inner product between the function and the kernel reproduces the function. Also

equation M292

Because the eigenfunctions are orthogonal, terms where ij vanish, and recall from (A2) that equation M293, so

equation M294

Hence, the inner product of kernels equation M295 produces kernel equation M296. This demonstrates that the Hilbert space constituted by linear combinations of the eigenfunctions of k has the reproducing kernel properties.


  • Aitchison, J., and C. G. G. Aitken, 1976. Multivariate binary discrimination by the kernel method. Biometrika 63 413–420.
  • Aronszajn, N., 1950. Theory of reproducing kernels. Trans. Am. Math. Soc. 68 337–404.
  • Balding, D. J., 2006. A tutorial on statistical methods for population association studies. Nat. Rev. Genet. 7 781–791. [PubMed]
  • Chang, H. L. A., 1988. Studies on estimation of genetic variances under nonadditive gene action. Ph.D. Thesis, University of Illinois, Urbana-Champaign, IL.
  • Cheverud, J. M., and E. J. Routman, 1995. Epistasis and its contribution to genetic variance components. Genetics 139 1455–1461. [PMC free article] [PubMed]
  • Cockerham, C. C., 1954. An extension of the concept of partitioning hereditary variance for analysis of covariances among relatives when epistasis is present. Genetics 39 859–882. [PMC free article] [PubMed]
  • Craven, P., and G. Wahba, 1979. Smoothing noisy data with spline functions. Num. Math. 31 377–403.
  • Dekkers, J. C. M., and F. Hospital, 2002. The use of molecular genetics in the improvement of agricultural populations. Nat. Rev. Genet. 3 22–32. [PubMed]
  • Elkan, C., 1997. Boosting and naïve Bayesian learning. Technical Report. University of California, San Diego.
  • Fenster, C. B., and L. F. Galloway, 2000. Population differentiation in an annual legume: genetic architecture. Evolution 54 1157–1172. [PubMed]
  • Fernández, S. A., R. L. Fernando, B. Gulbrandtsen, C. Stricker, M. Schelling et al., 2002. Irreducibility and efficiency of ESIP to sample marker genotypes in large pedigrees with loops. Genet. Sel. Evol. 34 537–555. [PMC free article] [PubMed]
  • Fernando, R. L., and M. Grossman, 1989. Marker assisted selection using best linear unbiased prediction. Genet. Sel. Evol. 21 467–477.
  • Fernando, R. L., C. Stricker and R. C. Elston, 1993. An efficient algorithm to compute the posterior genotypic distribution for every member of a pedigree without loops. Theor. Appl. Genet. 87 89–93. [PubMed]
  • Fox, J., 2005. Introduction to nonparametric regression. Lecture Notes. http://socserv.mcmaster.ca/jfox/Courses/Oxford.
  • Franklin, I., and R. C. Lewontin, 1970. Is the gene the unit of selection? Genetics 65 707–734. [PMC free article] [PubMed]
  • Gianola, D., M. Perez-Enciso and M. A. Toro, 2003. On marker-assisted prediction of genetic value: beyond the ridge. Genetics 163 347–365. [PMC free article] [PubMed]
  • Gianola, D., R. L. Fernando and A. Stella, 2006. Genomic assisted prediction of genetic value with semi-parametric procedures. Genetics 173 1761–1776. [PMC free article] [PubMed]
  • Golub, T. R., D. Slonim, P. Tamayo, C. Huard, M. Gasenbeek et al., 1999. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 286 531–537. [PubMed]
  • González-Recio, O., D. Gianola, N. Long, K. A. Weigel, G. J. M. Rosa et al., 2008. Non-parametric methods for incorporating genomic information into genetic evaluations: an application to mortality in broilers. Genetics 178 2305–2313. [PMC free article] [PubMed]
  • Gu, C., 2002. Smoothing Spline ANOVA Models. Springer, New York.
  • Hartl, D. L., and E. W. Jones, 2005. Genetics: Analysis of Genes and Genomes, Ed. 6. Jones & Bartlett, Boston.
  • Hastie, T. J., and R. J. Tibshirani, 1990. Generalized Additive Models. Chapman & Hall, London.
  • Hayes, B., J. Laerdahl, D. Lien, A. Adzhubei and B. Høyheim, 2004. Large scale discovery of single nucleotide polymorphism (SNP) markers in Atlantic Salmon (Salmo salar). AKVAFORSK, Institute of Aquaculture Research, Aas, Norway. www.mabit.no/pdf/hayes.pdf.
  • Hoeting, J. A., D. Madigan, A. E. Raftery and C. T. Volinsky, 1999. Bayesian model averaging: a tutorial. Stat. Sci. 14 382–417.
  • Jensen, C. S., and A. Kong, 1999. Blocking Gibbs sampling for linkage analysis in large pedigrees with many loops. Am. J. Hum. Genet. 65 885–901. [PMC free article] [PubMed]
  • Jensen, C. S., A. Kong and U. Kjaerulff, 1995. Blocking Gibbs sampling in very large probabilistic expert systems. Int. J. Hum. Comp. Stud. 42 647–666.
  • Kempthorne, O., 1954. The correlation between relatives in a random mating population. Proc. R. Soc. Lond. Ser. B 143 103–113. [PubMed]
  • Kerr, R. J., and B. P. Kinghorn, 1996. An efficient algorithm for segregation analysis in large populations. J. Anim. Breed. Genet. 113 457–469.
  • Kimeldorf, G., and G. Wahba, 1971. Some results on Tchebycheffian spline functions. J. Math. Anal. Appl. 33 82–95.
  • Kimura, M., 1965. Attainment of quasi-linkage equilibrium when gene frequencies are changing. Genetics 52 875–890. [PMC free article] [PubMed]
  • Long, N., D. Gianola, G. J. M. Rosa, K. Weigel and S. Avendaño, 2007. Machine learning classification procedure for selecting SNPs in genomic selection: application to early mortality in broilers. J. Anim. Breed. Genet. 124 377–389. [PubMed]
  • Mallick, B. K., D. Ghosh and M. Ghosh, 2005. Bayesian classification of tumours by using gene expression data. J. R. Stat. Soc. B 67 219–234.
  • Metropolis, N., A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller and E. Teller, 1953. Equations of state calculations by fast computing machines. J. Chem. Phys. 21 1087–1091.
  • Meuwissen, T. H. E., B. J. Hayes and M. E. Goddard, 2001. Is it possible to predict the total genetic merit under a very dense marker map? Genetics 157 1819–1829. [PMC free article] [PubMed]
  • O'Sullivan, F., B. S. Yandell and W. J. Raynor, 1986. Automatic smoothing of regression functions in generalized linear models. J. Am. Stat. Assoc. 81 96–103.
  • Quaas, R. L., and E. J. Pollak, 1980. Mixed model methodology for farm and ranch beef cattle testing programs. J. Anim. Sci. 51 1277–1287.
  • Racine, J., and Q. Li, 2004. Nonparametric estimation of regression functions with both categorical and continuous data. J. Econometrics 119 99–130.
  • Rasmussen, C. E., and C. K. I. Williams, 2006. Gaussian Processes for Machine Learning. MIT Press, Cambridge, MA.
  • Sheehan, N., and A. Thomas, 1993. On the irreducibility of a Markov chain defined on a space of genotype configurations by a sample scheme. Biometrics 49 163–175. [PubMed]
  • Silverman, B. W., 1986. Density Estimation for Statistics and Data Analysis. Chapman & Hall, London.
  • Soller, M., and J. Beckmann, 1982. Restriction fragment length polymorphisms and genetic improvement, pp. 396–404 in Proceedings of the 2nd World Congress on Genetics Applied to Livestock Production, Vol. 6. Editorial Garsi, Madrid.
  • Sorensen, D., and D. Gianola, 2002. Likelihood, Bayesian, and MCMC Methods in Quantitative Genetics. Springer-Verlag, New York.
  • Sorensen, D., and R. Waagepetersen, 2003. Normal linear models with genetically structured residual variance heterogeneity: a case study. Genet. Res. 82 207–222. [PubMed]
  • Stricker, C., M. Schelling, F. Du, I. Hoeschele, S. A. Fernandez et al., 2002. A comparison of efficient genotype samplers for complex pedigrees and multiple linked loci. Proceedings of 7th World Congress on Genetics Applied to Livestock Production, INRA, Castanet-Tolosan, France, CD-ROM Communication no. 21–12.
  • Takezawa, K., 2005. Introduction to Non-Parametric Regression. Wiley-Interscience, Hoboken, NJ.
  • Ter Braak, C. J. F., M. Boer and M. Bink, 2005 Extending Xu's (2003) Bayesian model for estimating polygenic effects using markers of the entire genome. Genetics 170 1435–1438. [PMC free article] [PubMed]
  • Van Arendonk, J. A. M., C. Smith and B. W. Kennedy, 1989. Method to estimate genotype probabilities at individual loci in farm livestock. Theor. Appl. Genet. 78 735–740. [PubMed]
  • Wahba, G., 1990. Spline Models for Observational Data. Society for Industrial and Applied Mathematics, Philadelphia.
  • Wahba, G., 1999. Support vector machines, reproducing kernel Hilbert spaces and the randomized GAVC, pp. 68–88 in Advances in Kernel Methods, edited by B. Schölkopf, C. Burges and A. Smola. MIT Press, Cambridge, MA.
  • Wahba, G., 2002. Soft and hard classification by reproducing kernel Hilbert spaces methods. Proc. Natl. Acad. Sci. USA 99 16524–16530. [PMC free article] [PubMed]
  • Wang, C. S., J. J. Rutledge and D. Gianola, 1993. Marginal inferences about variance components in a mixed linear model using Gibbs sampling. Genet. Sel. Evol. 25 41–62.
  • Wang, C. S., J. J. Rutledge and D. Gianola, 1994. Bayesian analysis of mixed linear models via Gibbs sampling with an application to litter size in Iberian pigs. Genet. Sel. Evol. 26 91–115.
  • Wang, H., Y. M. Zhang, X. Li, G. Masinde, S. Mohan et al., 2005. Bayesian shrinkage estimation of quantitative trait loci parameters. Genetics 170 465–480. [PMC free article] [PubMed]
  • Wong, G. K., B. Liu, J. Wang, Y. Zhang, X. Yang et al., 2004. A genetic variation map for chicken with 2.8 million single-nucleotide polymorphisms. Nature 432 717–722. [PMC free article] [PubMed]
  • Wood, S. N., 2006. Generalized Additive Models. Chapman & Hall/CRC Press, Boca Raton, FL.
  • Xu, S., 2003. Estimating polygenic effects using markers of the entire genome. Genetics 163 789–801. [PMC free article] [PubMed]
  • Yi, N., G. Varghese and D. A. Allison, 2003. Stochastic search variable selection for identifying multiple quantitative trait loci. Genetics 164 1129–1138. [PMC free article] [PubMed]
  • Zhang, Y., and S. Xu, 2005. A penalized maximum-likelihood method for estimating epistatic effects of QTL. Heredity 95 96–104. [PubMed]

Articles from Genetics are provided here courtesy of Genetics Society of America
PubReader format: click here to try


Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...


Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...