Score sliced random-projection Gaussian LRT discriminator — score_ot

Scores each row of X independently with a model learned by fit_ot_slicedlrt. Each specimen is transformed to its own rCLR vector over the features shared with the frozen training universe, projected onto deterministic random unit directions, and assigned the average per-direction Gaussian log-likelihood ratio of the case class versus the control class. Larger values are more case-like.

Usage

score_ot_slicedlrt(model, X, meta = NULL)

Arguments

model: An ot_slicedlrt_model object returned by fit_ot_slicedlrt.
X: Numeric matrix (samples $\times$ features) of non-negative abundance values. Columns must be uniquely named feature ids.
meta: Optional per-sample metadata. Accepted for interface uniformity and ignored by this method.

Value

Plain finite numeric vector of length nrow(X). If fewer than model$hp$min_features model-universe features are present in X, returns a neutral vector of zeros.

Details

At scoring time, the present feature set is intersect(model$feature_universe, colnames(X)). The sliced Gaussian representation is then re-derived exactly once from frozen case/control anchors restricted to that present set plus the frozen seed. A scored row contributes only its own rCLR vector; no statistic is estimated from any other row of X. This row-local computation is compatible with singleton deployment and is invariant to row order, duplication, and batch composition.

The default hp$min_features is 3L. Note that a present overlap of exactly one feature is degenerate: the single-feature per-sample rCLR is identically zero ($\log v - \log v = 0$), so every projection is zero and the score is a constant neutral 0 with no discrimination. Keep min_features $\ge 2$ (the default 3L already does) for a discriminating configuration.

For one row with projection coordinates $s_m$, case moments $\mu_{1m}, \sigma^2_{1m}$, and control moments $\mu_{0m}, \sigma^2_{0m}$, the score is $$\frac{1}{M} \sum_m \left[-\frac{(s_m-\mu_{1m})^2}{2\sigma^2_{1m}} -\frac{1}{2}\log\sigma^2_{1m} +\frac{(s_m-\mu_{0m})^2}{2\sigma^2_{0m}} +\frac{1}{2}\log\sigma^2_{0m}\right].$$

References

Bonneel N, Rabin J, Peyre G, Pfister H. (2015) Sliced and Radon Wasserstein barycenters of measures. Journal of Mathematical Imaging and Vision 51: 22-45.

Neyman J, Pearson ES. (1933) On the problem of the most efficient tests of statistical hypotheses. Philosophical Transactions of the Royal Society of London. Series A 231: 289-337.

Examples

if (FALSE) { # \dontrun{
set.seed(1)
X <- matrix(stats::rgamma(80 * 30, shape = 2), nrow = 80,
            dimnames = list(NULL, paste0("miR-", seq_len(30))))
y <- rep(c(0, 1), each = 40)
X[y == 1, 1:6] <- X[y == 1, 1:6] + 25
X[y == 0, 7:12] <- X[y == 0, 7:12] + 25
model <- fit_ot_slicedlrt(X, y)
score_ot_slicedlrt(model, X[1, , drop = FALSE])
} # }