Alternative empirical Bayes models for adjusting for batch effects in genomic studies

BMC Bioinformatics

Table 2 Comparison between five batch correction methods in predicting pathway activity and drug efficacy

		Correlation: EGFR protein expression		Drug response in cell lines
ComBat version	Common genes (cell line vs. TCGA)	Cell line	TCGA	Erlotinib	GSK1120212
Original ComBat	20 (40%)	0.316	0.132	0.360	0.401
Mean-only ComBat	44 (88%)	0.331	-0.042	0.294	0.407
Reference-batch ComBat	50 (100%)	0.442	0.299	0.415	0.520
Frozen SVA	50 (100%)	0.115	0.092	-0.09	-0.131
RUV	40 (80%)	0.287	0.182	0.332	0.145

We combined the oncogenic signature dataset with the cell line and TCGA patient data separately to adjust for batch effect and enable the profiling of EGFR signatures from the oncogenic data to the test sets. We observed the set bias using original ComBat (40% same signature genes), mean-only ComBat (88% same genes), and RUV (80% same genes) to combine the datasets. Reference-batch ComBat and frozen SVA kept the same signature genes. Also, using reference-batch ComBat gave the highest correlations of prediction scores with both protein expression and drug response, among all five batch correction methods. These results support the benefit of using reference-batch ComBat in this context

ISSN: 1471-2105