YuGene: A simple approach to scale gene expression data derived from different platforms for integrated analyses

Abstract

Gene expression databases contain invaluable information about a range of cell states, but the question "Where is my gene of interest expressed?" remains one of the most difficult to systematically assess when relevant data is derived on different platforms. Barriers to integrating this data include disparities in data formats and scale, a lack of common identifiers, and the disproportionate contribution of a platform to the 'batch effect'. There are few purpose-built cross-platform normalization strategies, and most of these fit data to an idealized data structure, which in turn may compromise gene expression comparisons between different platforms. YuGene addresses this gap by providing a simple transform that assigns a modified cumulative proportion value to each measurement, without losing essential underlying information on data distributions or experimental correlates. The Yugene transform is applied to individual samples and is suitable to apply to data with different distributions. Yugene is robust to combining datasets of different sizes, does not require global renormalization as new data is added, and does not require a common identifier. YuGene was benchmarked against commonly used normalization approaches, performing favorably in comparison to quantile (RMA), Z-score or rank methods. Implementation in the www.stemformatics.org resource provides users with expression queries across stem cell related datasets. Probe performance statistics including poorly performing (never expressed) probes, and examples of probes/genes expressed in a sample-restricted manner are provided. The YuGene software is implemented as an R package available from CRAN.

The Transcriptome of the Zebrafish Embryo After Chemical Exposure: A Meta-Analysis

“…Additionally, those genes were determined which were identified as DEG in most treatments (across all treatments and subgroups). (2) In order to identify genes with a common trend of differential transcription, common significance testing was applied: Each array was normalized by cumulative proportion transformation using the R-package “YuGene”, a normalization method specifically designed for cross-platform normalization considering different dynamic ranges for different platforms (Lê Cao et al , 2014). Then, random effect models were used to determine a summary effect size for each gene based on the fold-change and significance was estimated using a permutation analysis (1000 permutations) based on Significance Analysis of Microarrays (Tusher et al , 2001) applying the R-packages “MAMA “(Ihnatova, 2013) in combination with “GeneMeta” (Choi et al , 2003; Lusa et al , 2015).…”

CellFinder's Molecular Database and its Application to Stem Cell Research

“…Following annotation of each sample for which microarray data was available, the raw microarray data was downloaded from public databases, and normalized for within-study datasets (using BioC packages "affy" [6] and "lumi" [7] for Affymetrix and Illumina arrays, respectively, in R version 3.1 or greater [8]) and normalized with YuGene [9] for inter-study comparisons. Differential gene expression was calculated in studies where there were at least two cell types and three samples for each cell type, using both the within-study normalization data values and the cross-normalization data values.…”

Universal transcriptomic signature of age reveals temporal scaling ofCaenorhabditis elegansaging trajectories

“…Analysis of such small datasets is very challenging [25,26] because it is prone to over-fitting even if all of the samples are collected in the same laboratory under the same conditions. A direct comparison of gene expression data obtained in different laboratories is notoriously hard due to the nature of experimental procedures, leading to batch effects requiring, but not necessarily cured by, extensive normalization [27][28][29].…”

“…For individual genes represented by multiple probesets, the probeset with the largest signal was used. Gene expressions in all datasets were normalized using the YuGene [27] algorithm. The final MetaWorm dataset represents a 3724 ⇥ 4861 matrix (samples-x-genes) and incorporates more than 400 transcriptomic experiments (see Electronic Supplementary Materials).…”