Single-sample batch-correction methods (Module C)

Frozen batch-correction primitives introduced in the single-sample scoring bank (Module C). All methods follow a fit-then-freeze paradigm: factor loadings or basis vectors are estimated once on training data, then applied to test samples without re-estimation, making them suitable for clinical deployment where a population-matched training cohort may not be available at prediction time.

Methods provided:

• fit_frozen_ruv / apply_frozen_ruv: frozen RUV (Remove Unwanted Variation) using empirical negative-control features. Reference: Risso et al. (Nat Biotechnol 2014).

• fit_robust_pca_residual / apply_robust_pca_residual: MAD-scaled SVD residual filter. NOT the Candès RPCA (ALM sparse-low-rank) decomposition.

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.

Hubert M, Rousseeuw PJ, Vanden Branden K. (2005) ROBPCA: A New Approach to Robust Principal Component Analysis. Technometrics 47(1): 64–79.