Score the V-REx cross-cohort transfer single-sample discriminator

Scores each row of X independently with the frozen model from fit_dro_vrex, in PURE R (no python). Each query is mapped to the per-sample robust CLR over the frozen feature universe (absent universe features carry the neutral rCLR value 0), embedded by the exported encoder forward, and scored by the frozen linear logit $w^\top z + b$. Larger = more case-like. No RNG is used at scoring.

Queries with fewer than model$hp$min_features universe features present (a column-overlap floor), or empty positive support over the frozen universe on the (pre-rCLR) aligned abundances (!any(X_use[i, ] > 0)), return the neutral score 0. A FLAT all-equal-positive composition maps to the rCLR origin but is a VALID specimen and is scored normally. The score of a row depends only on that row and the frozen model and is exactly invariant to per-specimen positive scaling.

Usage

score_dro_vrex(model, X, meta = NULL)

Arguments

model: A dro_vrex_model object from fit_dro_vrex.
X: Numeric matrix (samples $\times$ features) of non-negative abundances with named feature columns.
meta: Optional per-sample metadata. Accepted for interface uniformity and ignored by this method (the transfer aspect is in the fit, not the score).

Value

Plain finite numeric vector of length nrow(X); larger values are more case-like.

References

Krueger D, et al. (2021) Out-of-Distribution Generalization via Risk Extrapolation (REx). ICML. arXiv:2003.00688.

Examples

if (FALSE) { # \dontrun{
model <- fit_dro_vrex(X, y, meta_train = meta)
score_dro_vrex(model, X[1, , drop = FALSE])
} # }