Contents

compute$genome$convergence

Convergent outcome signal across the model library

Shows where along the genome the model library's outcome signal concentrates. For each variant it sums |weight x outcome effect| over the models that carry it -- the same quantity compute$genome$attribution lays out per trait, collapsed across traits.

Usage

compute$genome$convergence(
  models,
  gwas,
  associations = NULL,
  outcome = NULL,
  fdr = 0.05
)

Arguments

Argument	Description
`models`	A PGSLibrary or a single PGS, whose weights are attributed.
`gwas`	A single-variant [PolyGeniusAssociation](/reference/polygeniusassociation/) (associate$singleVariant() or its meta) supplying the per-variant outcome effect. Required; aborts when none of its variants overlap models.
`associations`	A per-model PGS regression [PolyGeniusAssociation](/reference/polygeniusassociation/) carrying predictor and pval (e.g. from associate$regression() or associate$meta()), or NULL (default) to let every model contribute. Its predictor names must equal the (uniquified) model names.
`outcome`	Character scalar, or NULL (default). Selects one outcome when an association input holds several, and is required in that case.
`fdr`	Numeric scalar, default 0.05. False-discovery-rate threshold for the associations gate. The association's own adj.pval is used only when it is populated for every model; otherwise Benjamini-Hochberg is applied across the per-model strongest effects, so fdr always controls the FDR. Aborts when no model clears it, and when no variant of the surviving models overlaps gwas.

Value

A PolyGeniusGenomeSignal with statistic = "convergence", resolution = "variant", bin.reduce = "sum" and frame = "absolute". $results has one row per variant that at least one contributing model carries and gwas covers, with chr, position, nea, ea, value and n.models. $diagnostics holds n.ambiguous (strand-ambiguous model rows dropped) and gated (whether associations was supplied).

Details

A locus is tall only where models place real, outcome-relevant weight, so -- unlike a sum of raw -log10 p -- a cloud of near-null LD/clumping variants does not inflate it. Supplying associations gates the sum to the FDR-significant models, so the track shows where the outcome-predictive PRSs concentrate; omitting it sums over the whole library. n.models travels with each row, so a reader can tell one strongly-attributing model from many moderate ones.

Examples

gwas <- associate$singleVariant(data, phenotypes = AD)
prs  <- associate$regression(data, outcomes = AD)

# Every model contributes
sig <- compute$genome$convergence(data$library, gwas = gwas)

# Only the models whose PRS is outcome-predictive at 5% FDR
sig.gated <- compute$genome$convergence(data$library, gwas = gwas, associations = prs)
visualize$genome$convergence(sig.gated)