Fit a frozen RUV factor model from training data

Estimates latent factors of unwanted variation from empirical negative-control features on training data, then freezes the factor loadings for deployment. This is the frozen-mode extension of RUVSeq (Risso et al., Nat Biotechnol 2014): for a new sample \(x\), the unwanted variation component is estimated via the frozen pseudo-inverse of training latent factors, then subtracted.

If fewer than 5 named control features are found in colnames(X_train), falls back to the top-30% lowest-variance features as proxy controls, with a warning.

Usage

fit_frozen_ruv(
  X_train,
  control_features = NULL,
  k = 2:6,
  select = c("BIC", "fixed", "explicit")
)

Arguments

X_train

Numeric matrix (samples \(\times\) features). Training cohort.

control_features

Character vector of feature names expected to be non-cancer-relevant (empirical negative controls). Default = a curated set of stably-expressed serum/plasma miRNAs.

k

Integer or integer vector (range). Number of unwanted factors to estimate. Default 2:6 — BIC selects within this range.

select

Method for selecting k. "BIC" (default), "fixed" (uses the smallest value in k), or "explicit" (if select is a single numeric).

Value

Object of class frozen_ruv_fit with components: feat_mean, full_loadings, latent_train_inv, k_selected, bics, n_train, p_full, n_controls.

References

Risso D, Ngai J, Speed TP, Dudoit S. (2014) Normalization of RNA-seq data using factor analysis of control genes or samples. Nature Biotechnology 32(9): 896–902.

Examples

if (FALSE) { # \dontrun{
set.seed(42)
X <- matrix(rnorm(200 * 50), nrow = 200,
            dimnames = list(NULL, paste0("mir-", seq_len(50))))
fit <- fit_frozen_ruv(X, k = 2:4)
fit$k_selected
} # }