LDpred2-grid Weights from Summary Statistics

LDpred2-grid (Prive et al. 2020) runs the point-normal prior's Gibbs sampler at every combination of the proportion of causal variants p, the heritability h2 and sparse, and returns one of the fits. Upstream picks it on a validation sample; here, as for the penalized summary-statistics methods, it is picked on held-out summary statistics by default (see Details of lassosumRssWeights). A grid point has diverged, and is not picked, once a kept draw's squared effects sum past twice the squared marginal effects; bigsnpr's own sampler tests every iteration, burn-in included.

Usage

ldpred2GridWeights(
  stat,
  ldReference,
  backend = NULL,
  p = exp(seq(log(1e-05), 0, length.out = 21)),
  h2 = NULL,
  sparse = c(FALSE, TRUE),
  nIter = 150,
  nBurnin = 50,
  shrinkCorr = 0.95,
  allowJumpSign = TRUE,
  selection = c("heldOut", "ldQuadratic"),
  fitRetention = c("none", "slim", "full")
)

Arguments

stat

A list of summary statistics: n (per-variant sample sizes) and the marginal effects as cor (standardized), z, b with seb, or b alone, taken as already standardized. The sampler reads z / sqrt(n) for each variant's own sample size.

ldReference

The LD reference the statistics' variants are matched to by stat$variantNames: a genotype panel (see readGenotypes), a GenotypeMixture or an LdMatrix.

backend

"bigsnpr" or "pecotmr". NULL (default) is pecotmr for a genotype reference and bigsnpr for an LdMatrix; pecotmr's sampler reads only a genotype reference.

p

The proportions of causal variants in the grid.

h2

The heritabilities in the grid. NULL (default) is the window's LAVA estimate (see h2Init of ldpred2AutoWeights) times 0.3, 0.7, 1 and 1.4, as the LDpred2 vignette scales LD score regression's.

sparse

Whether each fit sets an effect whose inclusion probability falls below p to zero; both by default.

nIter, nBurnin

Each chain's iterations, burn-in included, and its burn-in.

shrinkCorr

The LD shrinkage \(c\): both engines read \(cR + (1 - c)I\). Default 0.95, as for ldpred2AutoWeights; bigsnpr's grid itself has none.

allowJumpSign

As for ldpred2AutoWeights; bigsnpr's grid sampler takes only TRUE, the default.

selection

How the returned fit is chosen: "heldOut" (default) or "ldQuadratic", as for lassosumRssWeights. "heldOut" needs the reference's sample size and refits the grid four times.

fitRetention

How much of the fit is kept as the "fit" attribute of the returned weights. "none" (default) attaches nothing. "slim" keeps path, the grid (p, h2, sparse) with whether each point diverged and which was selected. "full" also keeps beta, every point's effects (variants x points).

Value

A numeric vector of weights, the selected fit's effects, with the selection recorded in attribute "ldpred2_selection". Unless fitRetention = "none", the grid is attached as attribute "fit".

Details

backend = "bigsnpr" runs bigsnpr's snp_ldpred2_grid() on the LD matrix, written to a temporary file for the call; "pecotmr" runs pecotmr's port of the same sampler on a genotype reference's factor of the LD. bigsnpr's grid sampler has no LD shrinkage of its own, so for it the matrix written is shrunk, \(cR + (1 - c)I\), as the Bayesian alphabet's matrix engines are given theirs; it always allows a drawn effect to change sign, so it takes no allowJumpSign. pecotmr's sampler takes both, as LDpred2-auto's does. See ldpred2AutoWeights for the temporary file, the foreach backend and the seeds.

Examples

if (requireNamespace("bigsnpr", quietly = TRUE)) {
  data(eqtlRegionExample)
  X <- eqtlRegionExample$X[, 1:30]
  y <- eqtlRegionExample$yRes
  stat <- list(
    cor = as.vector(cor(X, y)),
    n = rep(nrow(X), ncol(X)),
    variantNames = colnames(X)
  )
  ld <- LdMatrix(cor(X), nSamples = nrow(X))
  set.seed(1)
  w <- ldpred2GridWeights(stat, ld, p = c(0.01, 0.1, 1), h2 = 0.1)
  attr(w, "ldpred2_selection")
}
#>      mode         p        h2    sparse 
#> "heldOut"    "0.01"     "0.1"   "FALSE"