Fit REO-pairratio within-sample log-ratio discriminator

Learns a frozen penalized-logistic discriminator over selected pairwise log-ratios. Features are pre-screened using only the training data by the absolute two-sample t-statistic on log(X_train + pseudocount). All pairwise log-ratios among the screened features are then generated with ws_logratio, and glmnet::cv.glmnet() fits a binomial elastic-net model with deterministic folds. No test data or test-time batch statistics are used.

A retained ratio \(\log(x_a + c) - \log(x_b + c)\) is computed from one specimen's two features, so a frozen linear predictor over these ratios is single-sample deployable. Larger returned linear predictors are more case-like.

Usage

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

Arguments

X_train

Numeric matrix (samples \(\times\) features) of non-negative abundance values. Columns must be uniquely named feature ids.

y_train

Integer/numeric 0/1 labels aligned to X_train; 1 is case/disease and 0 is control.

meta_train

Optional per-sample metadata. Accepted for interface uniformity and ignored by this method.

hp

List of frozen hyperparameters: pseudocount (default 0.5, positive), m_features (default min(40L, ncol(X_train)), at least 2), alpha (default 1, in [0, 1]), nfolds (default 10L, at least 3), lambda_rule (default "1se", or "min"), and seed (default 1L, used only to make fold assignment reproducible while restoring the global RNG state).

Value

A plain list of class reo_pairratio_model containing selected_pairs, nonzero coefficients, intercept, pseudocount, screened_features, feature_universe, lambda, lambda_rule, fit_status, and resolved hp. If every training log-ratio is constant, or if the chosen lambda yields an intercept-only glmnet model, selected_pairs has zero rows and scoring returns the intercept for every specimen. The former state is recorded as "intercept_only_zero_variance" and has no fitted lambda.

References

Aitchison J. (1986) The Statistical Analysis of Compositional Data. Chapman and Hall.

Friedman J, Hastie T, Tibshirani R. (2010) Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software 33: 1-22.

Examples

if (FALSE) { # \dontrun{
set.seed(1)
X <- matrix(stats::rgamma(60 * 20, shape = 2), nrow = 60,
            dimnames = list(NULL, paste0("miR-", seq_len(20))))
y <- rep(c(0, 1), each = 30)
X[y == 1, "miR-1"] <- X[y == 1, "miR-1"] + 10
X[y == 0, "miR-2"] <- X[y == 0, "miR-2"] + 10
model <- fit_reo_pairratio(X, y, hp = list(m_features = 8L, nfolds = 5L))
score <- score_reo_pairratio(model, X)
} # }