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

Abstract

Background

Methods

Results

Combat batch adjustment

Moment-based diagnostics for batch effects

Sample-level moments

Gene-level moments

Robust F-test

Mean-only adjustment for batch effects

Reference batch adjustments

Software implementation

Dataset simulation

Pathway simulation

Bladder cancer

Nitric oxide

Oncogenic signature

Lung cancer

Moments-based tests of significance for batch effects

Mean-only batch adjustments

Selecting the appropriate ComBat version for each dataset

Higher order moment-based batch adjustments

Batch adjustments based on a reference batch

Simulation study

EGFR signature and drug prediction

Discussion

Conclusions

Proposed diagnostic tools

Improved models based on ComBat

For certain situations

Significance test for batch differences

Proposed approach

Mean only ComBat for less severe adjustment

Reference-batch can leave one unchanged - Useful for reproducibility, for instance in context of biomarkers

Outperforms ComBat in particular cases

Contributing multiple methods for improved combination and analysis of data