PheWAS in both directions

Choose the scan direction from the scientific question:

Question	Direction	Repeated model
Is one PRS, exposure, or biomarker associated with many phenotypes?	phenotypes_as_outcomes	Each phenotype is the outcome; the anchor is the predictor
Is each phenotype associated with one clinical endpoint?	phenotypes_as_predictors	The clinical endpoint is the outcome; each phenotype is a predictor

The examples below use synthetic participants so they can run without access to clinical data.

library(phewasFlow)

example <- phewas_example_data(n = 300)
data <- assemble_phewas_data(
  example$sample, example$phenotypes, example$anchors,
  example$covariates, id = "id"
)

eligibility <- list(
  min_n = 100L,
  min_cases = 10L,
  min_controls = 10L,
  min_outcome_levels = 3L
)

PRS or PGS as the predictor

The phenotype metadata explicitly declares binary, continuous, count, and ordinal outcomes. The anchor is standardized once on the common base cohort, before phenotype-specific missingness is considered.

forward_spec <- phewas_spec(
  analysis_id = "pgs_to_phenome",
  anchor = "pgs",
  direction = "phenotypes_as_outcomes",
  phenotypes = example$forward_metadata,
  covariates = c("age", "sex"),
  anchor_type = "numeric",
  anchor_transform = "zscore",
  eligibility = eligibility,
  fdr_threshold = 0.05
)

forward <- run_phewas(data, forward_spec)
forward[, .(phenotype, description, engine, estimate, native_effect, status)]
#>           phenotype          description  engine  estimate native_effect status
#>              <char>               <char>  <char>     <num>         <num> <char>
#> 1:     pheno_binary     Binary phenotype logistf 0.5653552     1.7600729     ok
#> 2: pheno_continuous Continuous phenotype      lm 0.6765848     0.6765848     ok
#> 3:      pheno_count      Count phenotype  glm.nb 0.3124958     1.3668323     ok
#> 4:    pheno_ordinal    Ordinal phenotype     clm 0.7604123     2.1391579     ok

Binary and ordinal effects are odds ratios, count effects are incidence-rate ratios, and continuous effects remain regression coefficients. The estimate column always contains the coefficient on the model’s link scale.

For a count outcome with unequal exposure time, name a column containing the already-log-transformed exposure in that phenotype row’s offset field. Leave offset empty when no offset belongs in the model.

Phenotypes as predictors

This example asks whether each phenotype is associated with a continuous clinical endpoint. The endpoint stays fixed while phewasFlow changes the phenotype predictor.

reverse_spec <- phewas_spec(
  analysis_id = "phenome_to_continuous_endpoint",
  anchor = "endpoint_continuous",
  direction = "phenotypes_as_predictors",
  phenotypes = example$reverse_metadata,
  covariates = c("age", "sex"),
  outcome_type = "continuous",
  anchor_type = "numeric",
  eligibility = eligibility,
  fdr_threshold = 0.05
)

reverse <- run_phewas(data, reverse_spec)
reverse[, .(phenotype, description, estimate, native_effect, status)]
#>           phenotype          description  estimate native_effect status
#>              <char>               <char>     <num>         <num> <char>
#> 1:     pheno_binary     Binary phenotype 1.0625758     1.0625758     ok
#> 2: pheno_continuous Continuous phenotype 0.3610119     0.3610119     ok
#> 3:    pheno_ordinal    Ordinal phenotype 0.2145129     0.2145129     ok

Missing phenotype values are excluded association by association. They are never converted to a binary reference group.

For a binary clinical endpoint, set outcome_type = "binary", anchor_type = "binary", and give its reference value with anchor_reference. Count and ordinal endpoints use outcome_type = "count" or "ordinal"; ordinal endpoints also require anchor_levels in their true order. Use a separate analysis_id and result directory for each clinical endpoint because each one has its own multiple-testing correction.