Fig. 2.

Batch effect-correction does not work optimally in practice. Top row: Feature selection without multiple-test correction. Bottom row: Feature selection with FDR correction. First column shows base performance with only class effects (No batch). Second column shows performance with batch effects incorporated. Third, fourth and fifth columns show recovery using various batch effect-correction methods, (COMBAT, quantile normalization and linear-scaling). Feature-selection test used here is the two-sample t-test (Abbreviations: P, Precision; R, Recall; F, F-score)