Per-Feature Batch-vs-Case Signal Partitioning

For each feature, computes three quantities that jointly distinguish biological case signal from dataset/batch provenance:

• F_dataset_healthy: one-way ANOVA of feature expression against dataset identity, using only healthy samples (so case biology is removed);

• F_case: one-way ANOVA against case/control status, using all samples;

• dataset_case_corr: Pearson correlation between the sample-wise dataset-mean signature and the sample-wise case-mean signature.

Features with high F_dataset_healthy AND high dataset_case_corr are the most suspect: their case signal is colinear with dataset identity.

Usage

os_per_feature_batch_signal(X, y, cohort, feature_names = NULL)

Arguments

X

Numeric matrix or data frame (n samples x p features).

y

Binary case vector (0/1). Length n.

cohort

Factor or vector of cohort/dataset identity. Length n.

feature_names

Optional character vector of feature labels.

Value

A data frame with one row per feature, including a suspect_score that combines healthy-only dataset ANOVA magnitude with dataset_case_corr, ranked from most to least provenance-like.