Firstly, the quality of the data must be assessed in order to detect problematic raw probe-level data, such as spatial artifacts on the chip or poor quality hybridisation. Indeed, gene expression experiments suffer from many sources of technical and experimental variation. Removing noise and systematic biases is performed in order to both improve the biological signal and make all the arrays comparable. This is the so-called normalisation step. Secondly, we propose to discard the probesets with very low signal across the samples (i.e. genes unexpressed or below detection threshold). This filtering step leads to both a noise reduction in the data and an increase in the statistical power of the subsequent analysis.

Then, exploratory approaches are classically used to find clusters of genes (or samples) with similar profiles. Note that here, biological interpretation depends on the choice of the similarity metrics. These approaches potentially highlight outliers and/or non relevant effects (batch effect for example), which can be subsequently estimated and/or removed from the data thanks to appropriate methods.

Finally, supervised approaches aim at the identification of differentially expressed genes (DEG), or deregulated pathways by taking into account the multiple testing issues. The biological interpretation of the differential analysis results can be performed thanks to functional and gene set enrichment analyses. Sample class prediction (eg good vs poor clinical outcome) based on supervised classification methods can also be performed to highlight genes signatures.

For the data quality assessment, we recommend to use the arrayQualityMetrics package 3, which performs a powerful, easy-to-use and comprehensive data quality estimation as well as an automatic html report. The EMA package proposes the most famous techniques for Affymetrix GeneChip normalisation: MAS5.0 4, RMA 5 and GCRMA 6. We recommend to use GCRMA because it outperforms the other approaches (by ignoring the mismatch intensities and taking into account the probe sequence information) and allows an efficient filtering of irrelevant probesets thanks to its bimodal distribution of probesets expression values (Figure 1a). Other packages such as limma 7, vsn 8 or lumi 9 can be used to normalise non Affymetrix data. After this first step, the main EMA functions can be used for any type of expression data, using a simple data expression matrix as input.

The EMA package provides functions to perform exploratory analyses such as Principal Component Analysis (PCA, Figure 1b), hierarchical clustering (Figure 1c) or Multiple Factor Analysis. They are based on R packages such as FactoMineR 10, cluster 11, or mostclust 12. The use of linear model is proposed to estimate and to remove the non relevant effects potentially detected.

Various methods are proposed to perform differential analysis and their choice depends on the sample size. The multtest package provides standard approaches like Student or Mann-Whitney test associated with multiple testing correction methods. The Significance Analysis of Microarrays (SAM) approach 13 (siggenes package) is also very interesting because it both estimates the null distribution and takes into account the correlation between probesets (Figure 1d). The rank product method 14 (RankProd package) dedicated to small sample size dataset is also offered, as well as some linear model (ANOVA) functions. Alternatively, the user can apply the limma package which is a very powerful tool to assess differential expression by linear models.

The functional enrichment of the DEG list is assessed based on the GeneOntology 15, and KEGG 16 pathways annotation terms. The hyper-geometric test of the GOstats package is used to test the over-representation of the functional terms in the gene list.

For sample class prediction, we suggest to use the CMA package 17 including the most popular machine learning and gene selection algorithms. In the context of censored data, the EMA package supports Kaplan Meier and log-rank analyses using the survival package.

The proposed analysis strategy was applied to the breast cancer gene expression dataset 18 comparing 12 Basal-like carcinomas (BLCs) and 11 HER2 positive carcinomas (HER2+). Some graphical outputs for data preprocessing, exploratory analysis and differential analysis steps are displayed in Figure 1. The RNA profiles were analysed using U133 plus 2.0 Affymetrix GeneChip. Three genes (P-cadherin, v-kit, FOXC1) were reported by the authors to be associated to a genes cluster over-expressed in the basal-like carcinomas and three genes (PTEN, Her2 and GRB7) to a genes cluster over-expressed in the Her2+ carcinomas. All these genes but one (v-kit) were found to be differentially expressed using the EMA package. This discrepancy is easily explained because in spite of v-kit belongs to a basal-like expression cluster, no change in v-kit expression can be observed between the two groups in this clustering analysis. This is because the hierarchical clustering was performed on genes (such as v-kit) not necessary differentially expressed between the two populations.

The R scripts used to analyse this gene expression dataset can be found in [Additional file 1]. Transcriptomic data used in this application are publicly available at Gene Expression Omnibus (Accession number: [GSE13787]) and are part of the package.