Josch K. Pauling

Systematic biases introduced throughout every step of clinical studies affect high-throughput omics data. Normalization methods (NMs) aim to adjust for these biases to make the biological signal more prominent. Due to the NM’s potential impact on data analysis, the NM needs to be carefully chosen. For this, we systematically evaluated 16 NMs on four tandem mass tag (TMT) and six label-free protein quantification datasets.

Opposed to state-of-the-art normalization practice, we found that successful reduction of intragroup variation of samples is not directly related to NM effectiveness. Instead, differential expression (DE) analysis revealed that the number of DE proteins depends significantly on the NM. Thus, and because of unexpectedly high numbers of false positives in spike-in datasets, we recommend F1 scores as a reasonable performance evaluation metric.

Moreover, NM effectiveness is dataset-specific for the majority of methods. In particular, we demonstrate the benefit of combining NMs with batch effect correction methods for TMT datasets. Notably, the novel approach RobNorm performed consistently well on numerous datasets.

Finally, experimental data analysis demonstrated that the choice of NMs affects downstream steps, such as DE, statistical enrichment analysis, and related insights into biological mechanisms. Therefore, we propose to comparatively evaluate NMs beyond F1 scores including their impact on downstream analysis.