Fit REO-MetaKTSP meta-analytic pair-order discriminator

Learns a meta-analytic k-Top-Scoring-Pairs discriminator across training cohorts named by meta_train[[cohort_col]]. Each unordered feature pair receives an avgTSP score: the simple equal-cohort-weight mean of per-cohort case-control differences in feature_i < feature_j. This is a transfer-estimand method at training because pair selection uses cross-cohort information, but it remains single-sample at inference because score_reo_metaktsp uses only one specimen's own pair order. When no usable cohort metadata is available, or only one non-missing cohort is present, fitting gracefully degenerates to ordinary single-cohort fit_reo_ktsp behavior over the training rows. Rows with missing cohort labels are dropped from meta-analytic selection, cohorts with only cases or only controls are skipped, and a zero-usable-cohort meta split falls back to the single-cohort path. At score time, no usable retained pairs returns the neutral vote fraction 0.5.

Usage

fit_reo_metaktsp(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. When it contains cohort_col, non-missing values define the training cohorts used for avgTSP pair selection.

hp

List of frozen hyperparameters: k, the number of oriented pairs requested (default 11L), and cohort_col, a single non-empty character string naming the cohort/study column in meta_train (default "accession"). These are resolved once during fitting and are not tuned inside fit_reo_metaktsp().

Value

A plain list of class reo_metaktsp_model containing retained oriented pairs, feature_universe, retained pair count k, cohort_col, number of cohorts used in selection n_cohorts, and the resolved hp. n_cohorts is the auditable meta-engagement signal: n_cohorts > 1 means avgTSP selection ran across multiple cohorts, while n_cohorts == 1 means fitting degenerated to single-cohort k-TSP (NULL/missing/single-level cohort_col, or a zero-usable-cohort fallback). A caller that intends cross-cohort meta-analysis should check n_cohorts > 1 so a mis-wired meta_train cannot silently demote this transfer method to pooled within-cohort k-TSP.

References

Kim S, Lin C-W, Tseng GC. (2016) MetaKTSP: a meta-analytically derived k-top-scoring-pair classifier for robust prediction of human cancer subtypes. Bioinformatics 32: 1966-1973. PMID: 27153719.

Geman D, d'Avignon C, Naiman DQ, Winslow RL. (2004) Classifying gene expression profiles from pairwise mRNA comparisons. Statistical Applications in Genetics and Molecular Biology 3: Article19.

Tan AC, Naiman DQ, Xu L, Winslow RL, Geman D. (2005) Simple decision rules for classifying human cancers from gene expression profiles. Bioinformatics 21: 3896-3904.

Examples

if (FALSE) { # \dontrun{
X <- matrix(stats::rgamma(60 * 12, shape = 2), nrow = 60,
            dimnames = list(NULL, paste0("miR-", seq_len(12))))
y <- rep(c(0, 1), times = 30)
meta <- data.frame(accession = rep(paste0("GSE", 1:3), each = 20))
X[y == 0, "miR-1"] <- X[y == 0, "miR-1"] + 5
X[y == 1, "miR-2"] <- X[y == 1, "miR-2"] + 5
model <- fit_reo_metaktsp(X, y, meta_train = meta, hp = list(k = 3L))
score <- score_reo_metaktsp(model, X)
} # }