Fit POWERUP models — fit_powerup

Fit one POWERUP XGBoost model for each requested perturbation in prepared POWERUP data returned by prepare_powerup_data().

Usage

fit_powerup_models(
  reference_data,
  models_to_make = NULL,
  perturbation_tags = "ko_",
  response_cutoff = 0.5,
  decreasing = FALSE,
  nfolds = 3,
  nrepeats = 3,
  nrounds = 200,
  min_score = 0.01,
  max_depth = 3,
  f_subsample = 1,
  shuffle = FALSE,
  seed = 1L,
  n_threads = 4,
  xgb_params = NULL,
  cor_n_features = NULL,
  shap_top_n = 100L
)

Arguments

reference_data: Prepared POWERUP data returned by prepare_powerup_data(), or a data frame containing requested perturbation responses and predictors.
models_to_make: Optional character vector naming perturbations to model. When reference_data is prepared POWERUP data, defaults to all prepared perturbations.
perturbation_tags: Character vector of prefixes used to identify response columns when reference_data is a data frame.
response_cutoff: Default biological response cutoff used to define target events for classification-style summaries and downstream event probabilities.
decreasing: If TRUE, values at or below response_cutoff define the target event; otherwise values at or above the cutoff define the target event.
nfolds: Number of cross-validation folds.
nrepeats: Number of repeated cross-validation runs.
nrounds: Maximum number of XGBoost boosting rounds.
min_score: Minimum mean R-squared required to retain a fitted model.
max_depth: Maximum tree depth.
f_subsample: Fraction of training rows sampled for each boosting round.
shuffle: If TRUE, shuffle response values before model fitting.
seed: Base random seed. A deterministic perturbation-specific seed is derived from this value so model results do not depend on target order. Both values are stored in each returned model object as seed and target_seed.
n_threads: Number of threads passed to XGBoost.
xgb_params: Optional named list overriding supported XGBoost parameters.
cor_n_features: Optional number of target-correlated features to retain. When set, features are ranked by absolute correlation within each cross-validation training fold and reranked using all training samples for the final refit. NULL disables correlation filtering.
shap_top_n: Maximum number of non-intercept SHAP features retained per sample before remaining contributions are collapsed into __other__.

Value

A named list of POWERUP model objects, one per requested perturbation.

Examples

if (FALSE) { # \dontrun{
prepared <- prepare_powerup_data(gene_expression, response, user_matrix = user_matrix)
models <- fit_powerup_models(prepared)
} # }