Regression-based Z-score for trisomy prediction

Implements NIPTeR's stepwise linear regression approach. Builds multiple regression models that predict the focus chromosome's fraction from other chromosomal fractions, then computes Z-scores based on the prediction residuals.

Usage

nipter_regression(
  sample,
  control_group,
  chromo_focus,
  n_models = 4L,
  n_predictors = 4L,
  exclude_chromosomes = c(13L, 18L, 21L),
  include_chromosomes = NULL,
  train_fraction = 0.6,
  overdispersion_rate = 1.15,
  force_practical_cv = FALSE,
  seed = NULL
)

Arguments

sample

A NIPTeRSample object.

control_group

A NIPTeRControlGroup object.

chromo_focus

Integer; the target chromosome (1-22).

n_models

Number of regression models to build (default 4).

n_predictors

Maximum number of predictor chromosomes per model (default 4).

exclude_chromosomes

Integer vector of chromosomes to exclude from the predictor pool (default c(13, 18, 21)).

include_chromosomes

Integer vector of chromosomes to force-include.

train_fraction

Fraction of control samples used for training (default 0.6).

overdispersion_rate

Overdispersion multiplier for theoretical CV (default 1.15).

force_practical_cv

Logical; always use practical CV regardless of theoretical CV? Default FALSE.

seed

Random seed for the train/test split. Default NULL (no seed set).

Value

A list of class "NIPTeRRegression" with elements:

models

A list of n_models sublists, each containing: z_score (numeric), cv (selected CV), cv_type ("practical" or "theoretical"), predictors (character vector of predictor chromosomes), shapiro_p_value, control_z_scores (named numeric vector).

focus_chromosome

Character.

correction_status

Character vector.

sample_name

Character.

Details

For SeparatedStrands samples the predictor pool is doubled (each autosomal chromosome contributes both a forward and reverse fraction), and complementary-strand exclusion within each model prevents selecting both strands of the same chromosome as predictors in the same model. Across models, only the exact predictor string is excluded; the complementary strand remains available.

The algorithm:

1. Split the control group 60/40 into train and test sets.

2. Compute chromosomal fractions for all samples.

3. For each of n_models models, greedily select n_predictors chromosomes by maximising adjusted R-squared (forward stepwise selection). Predictors already used in previous models are excluded.

4. For each model, fit lm(focus ~ predictors) on the training set, predict on the test set, and compute:

   • Practical CV: sd(obs/pred) / mean(obs/pred) on the test set.

   • Theoretical CV: overdispersion_rate / sqrt(total_reads_focus) for the test sample.

   • The larger CV is used (unless force_practical_cv = TRUE).

   • Sample Z-score: (sample_ratio - 1) / cv.

See also

nipter_z_score(), nipter_ncv_score()

Examples

if (FALSE) { # \dontrun{
reg21 <- nipter_regression(sample, cg, chromo_focus = 21)
reg21$models[[1]]$z_score
} # }