Batch Effects in Cancer Genomics

RNA-seq data is produced across different hospitals, machines, dates, and protocols. Those technical differences leave fingerprints unrelated to biology. A model can quietly learn the batch, score beautifully on familiar data, and collapse on samples from a new institution. I set out to measure how much these effects distort breast-cancer subtype prediction - and whether standard corrections fix the problem or merely hide it.

I worked with TCGA-BRCA, retaining 842 tumors and the 5,000 most variable genes after filtering and variance stabilization. The task was to predict five PAM50 molecular subtypes, judged with macro-F1 and balanced accuracy because the classes are uneven.

The core design was leave-one-batch-out cross-validation: hold out an entire institution, train on the other fifteen, and always evaluate on an unseen source. Across 16 institutions, I compared an uncorrected baseline, ComBat-seq, and per-batch mean-centering with logistic regression, random forest, and SVM - 144 experiments in total. PCA, ANOVA, Wilcoxon signed-rank tests, Kruskal-Wallis, and Spearman tests completed the analysis.

The headline result contradicted my expectation: biology dominated. Subtype explained roughly 22 times more variation in the leading principal component than institution did - an R² near 0.75 versus 0.03. Batch effects were real, but they were not the principal structure in the data.

Correction’s value appeared in consistency rather than average performance. Uncorrected macro-F1 varied by around 0.50 across held-out hospitals; ComBat-seq roughly halved that spread for logistic regression, down to about 0.23, while sacrificing only a few average F1 points. It traded a little headline accuracy for much steadier cross-hospital behavior.