HarmonizR enables data harmonization across independent proteomic datasets with appropriate handling of missing values

Abstract

Dataset integration is common practice to overcome limitations in statistically underpowered omics datasets. Proteome datasets display high technical variability and frequent missing values. Sophisticated strategies for batch effect reduction are lacking or rely on error-prone data imputation. Here we introduce HarmonizR, a data harmonization tool with appropriate missing value handling. The method exploits the structure of available data and matrix dissection for minimal data loss, without data imputation. This strategy implements two common batch effect reduction methods—ComBat and limma (removeBatchEffect()). The HarmonizR strategy, evaluated on four exemplarily analyzed datasets with up to 23 batches, demonstrated successful data harmonization for different tissue preservation techniques, LC-MS/MS instrumentation setups, and quantification approaches. Compared to data imputation methods, HarmonizR was more efficient and performed superior regarding the detection of significant proteins. HarmonizR is an efficient tool for missing data tolerant experimental variance reduction and is easily adjustable for individual dataset properties and user preferences.

Multiomic profiling of medulloblastoma reveals subtype-specific targetable alterations at the proteome and N-glycan level

“…Missing value tolerant data harmonization 27 adjusted sample specific mean and CV values across studies (Supplementary Figure 3A) and enabled a clear separation of MB groups. (Figure 1E-F).…”

“…To further increase the cohort size, we integrated protein abundances with reanalyzed LC-MS/MS data of FF-MB from public repositories 18 ; 19 ; 17 , Figure 1D-F). Missing value tolerant data harmonization 27 adjusted sample specific mean and CV values across studies (Supplementary Figure 3A) and enabled a clear separation of MB groups. (Figure 1E-F).…”

“…Using the HarmonizR strategy for the integration of independently generated proteome datasets 27 , we established an integrated dataset of 167 MBs. We identified six different proteome MB subtypes, in line with previously published results, obtained from smaller fractions of the analyzed cases 18 .…”

“…Using a missing-value tolerant pipeline for integrating proteome data 27 , we incorporated publicly available, smaller MB proteome datasets [17][18][19] with data of 62 FFPE MB cases and established a joint MB proteome dataset (n = 176). We comprehensively correlated proteome data with DNA methylation dataa current gold standard for molecular brain tumor classification 28 .…”

Persistence of targetable lesions, predicted therapy sensitivity and proteomes through disease evolution in pediatric acute lymphoblastic leukemia

“…Precursor, and protein identifications false discovery rate (FDR) threshold was set to 1%, while the threshold for peptide was 0.5%. The data was normalized in Spectronaut based only on proteins identified in all samples and then further processed for batch effect removal by HarmonizR (41).We utilized HarmonizR for batch effect correction is that it allows the retention of the proteins that otherwise would be dropped due to containing missing values (Supplementary Fig. S5).…”