Affymetrix expression console

Total RNA was used to synthesize double-stranded complementary DNA (cDNA) utilizing a random priming method, and then double-stranded cDNA was fragmented, labelled and hybridized to the Affymetrix GG-H Array. After hybridization and washing, processed slides were scanned with the Affymetrix GeneChip Scanner 3000 7G (Affymetrix, Santa Clara, CA). The Affymetrix Expression Console (version 1.2.1) implementation of RMA was used for quantile normalization and background correction. All gene level files were imported into Affymetrix Expression Console (version 1.2.1) and normalized by the quantile method. Combat Software was used to adjust the normalized intensity to remove batch effects. A random variance model (RVM) [23] (link)
t-test was applied to discriminate differentially expressed genes between controls and MDD patients because the RVM t-test can raise degrees of freedom effectively in the case of small samples. After false-discovery rate (FDR) analysis, we selected differentially expressed genes only if p value <0.05 and the change in expression was ≥1.5-fold. Hierarchical clustering was performed using EPCLUST. The microarray analysis was performed by Gminix, Shanghai, PR China.

Previously obtained microarray data using total RNA prepared from human temporal (10 AD, 19 non-AD cases) and frontal cortices (15 AD and 18- non-AD cases)^{18 (link)} are available from the GEO database (accession number GSE36980). CEL files were imported into the Affymetrix Expression Console (Affymetrix Japan K.K., Tokyo, Japan) and CHP files were obtained using a Gene Level-RMA-Sketch method. CHP files were input into the Affymetrix Transcriptome Analysis Console (TAC) software and a gene level differential expression analysis was performed according to the user’s guide. One-way between subject ANOVA was performed between AD and non-AD subjects and a list of transcripts was created. Principal Component Analysis (PCA) and hierarchical clustering was performed in the Affymetrix Expression Console and TAC software, respectively, and several samples were found to be outliers, likely owing to biological heterogeneity or technical issues (Supplementary Figs S1 and S2). These outliers were excluded to avoid undesirable artefacts during the profiling analyses.

All datasets were obtained from the EMBL-EBI ArrayExpress [70 (link)], or NCBI Gene Expression Omnibus (GEO) [71 (link)] websites. Accession numbers for the datasets and experimental details can be found in Table 5. Microarray raw intensity files were MAS5 normalized (Affymetrix Expression Console, Affymetrix, CA, USA) [72 (link)] and log2 transformed to obtain the raw data datasets. We used only the HG-U133A data for the main HD dataset [27 (link)]. For the neurodegenerative diseases dataset [35 (link)], the array files were normalized using the Rosetta error model (Rosetta Biosoftware, WA, USA) and log2 transformed to obtain the raw data dataset. All raw datasets were collapsed to a one probe per gene level using the R function collapseRows [73 (link)]. Microarray probes were matched to gene names and Entrez gene IDs (NCBI) of homo sapiens genome build hg19 (Consensus CDS, NCBI), if the annotation was not provided by the affymetrix array annotation file (Affymetrix Expression Console, Affymetrix, CA, USA). Probes with ambiguous gene annotations were removed. Outlier samples were removed by a completely unbiased method, which ignores phenotypic traits. To this end the Euclidian distance between samples in a network and their connectivity was calculated. Subsequently, samples with a standardized connectivity of less than −2.5 were removed.