compute_gwas_weights: Compute Per-SNP GWAS Effect Size Weights

In RAS: Regional Association Score for Genome-Wide Association Studies

Compute Per-SNP GWAS Effect Size Weights

Description

Runs a per-SNP regression on the training split to obtain effect-size estimates used as polygenic score (PGS) weights in the forward scan.

Usage

compute_gwas_weights(
  geno,
  phenotype1,
  this.sample,
  this.df,
  is_continuous,
  covariate_formula = NULL
)

Arguments

geno	Matrix of genotype dosages (n samples \times N variants). Only the rows indexed by this.sample are used.
phenotype1	Numeric vector of length |\texttt{this.sample}|. Training-split phenotype values. For continuous traits these should be covariate-residualised values; for binary traits, the raw 0/1 indicator.
this.sample	Integer vector. Row indices of the training split in geno.
this.df	Data frame of covariates with rows aligned to geno. Required for binary traits; used to subset and build the regression data frame. Ignored for continuous traits.
is_continuous	Logical. TRUE for quantitative traits, FALSE for binary (case/control) traits.
covariate_formula	Character. The right-hand side of the regression formula including the SNP term this.x. For example, "this.x + age + sex + pc1". If NULL, a default formula with sex, age, age-squared, age-sex interaction, and the first ten ancestry PCs is used.

Details

For continuous traits (is_continuous = TRUE) the function fits lm(phenotype1 ~ this.x) for each SNP, excluding samples with missing dosage via na.omit. The covariates are assumed to have already been residualised from phenotype1 before this function is called (as done in ras_scan).

For binary traits (is_continuous = FALSE) the function builds a data frame combining the training-split rows of this.df with the dosage column (this.x) and phenotype (phenotype1), removes incomplete cases with na.omit, and fits lm(phenotype1 ~ covariate_formula). Note that lm rather than glm(family = binomial) is used at the GWAS weight stage to obtain a linear probability approximation to the effect size; this is consistent with standard GWAS practice for weight derivation.

Both branches are vectorised: instead of fitting one lm per SNP, all NA-free SNPs are solved in closed form with a single BLAS cross-product (continuous) or a batched Frisch-Waugh-Lovell residualisation (binary). SNP columns that contain missing dosages are handled individually so the result matches lm + na.omit exactly. A single “Starting GWAS ...” message is printed.

Value

A numeric matrix with N rows and 4 columns:

Estimate

Per-SNP effect-size estimate (slope of the dosage term). The first column is used as PGS weights by compute_pgs_matrix.

Std. Error

Standard error of the estimate.

t value

t-statistic.

Pr(>|t|)

Two-sided p-value.

See Also

compute_pgs_matrix for the next step, which multiplies these weights by the hold-out genotype matrix. ras_scan for the recommended high-level entry point that calls this function automatically.

Examples

set.seed(1)
n_samp <- 60; n_snp <- 10
geno <- matrix(sample(0:2, n_samp * n_snp, replace = TRUE,
                      prob = c(0.6, 0.3, 0.1)),
               nrow = n_samp, ncol = n_snp)
pheno_train <- rnorm(n_samp)

coef_mat <- compute_gwas_weights(
  geno              = geno,
  phenotype1        = pheno_train,
  this.sample       = seq_len(n_samp),
  this.df           = data.frame(row.names = seq_len(n_samp)),
  is_continuous     = TRUE
)
dim(coef_mat)        # n_snp x 4
head(coef_mat, 3)

RAS documentation built on July 24, 2026, 5:06 p.m.