Fit a training-frozen trimmed-rCLR signed-panel score

Fits the canonical trimmed-rCLR reference used by the OmicSelector single-sample benchmark. Each training specimen is transformed independently with [ws_rclr_trimmed()]. Features are ranked by the absolute standardized case-control mean difference in the training data, the first `panel_size` features are retained, and each feature receives the sign of its training case-minus-control median difference. No held-out value or outcome is used.

Usage

fit_ws_rclr_panel(X_train, y_train, meta_train = NULL, hp = list())

Arguments

X_train

Numeric training matrix with specimens in rows and named features in columns.

y_train

Binary training labels coded 0/1.

meta_train

Unused; accepted for the canonical roster fit contract.

hp

Optional named list. Supported fields are `panel_size` (default 20), `pseudocount`, `trim_upper` (0.10), `trim_lower` (0.05), `exclude_features`, `zero_policy` (`"pseudocount"`), and `min_centering_size` (8).

Value

A `ws_rclr_panel_model` containing the frozen feature panel, signs, training effects, and transform settings.

See also

[score_ws_rclr_panel()], [ws_rclr_trimmed()]