Microarray Data Analysis

The microarrays were scanned with GenePix 400B (Molecular Devices, Union City, CA, USA) and normalization of fluorescence intensities was accomplished by using the GenePix Pro 6.0 scanning software (Axon Instruments, Union City, CA, USA). Each scanned image was examined thoroughly and dust particles and spots with high background were flagged and removed from analysis. The background and standard deviation were calculated for each raw data file after scanning, and only those spots with intensities three standard deviations above background were considered “expressed” and loaded into Genespring 12.1 (Agilent Technologies Palo Alto, CA, USA). Loess normalization was applied to all microarrays before statistical comparisons. In the analysis, each probe was considered individually.

In the post-normalization evaluation of the probes on the microarrays, 12,274 probes present in either the standard reference or the sample on 90% of the microarrays underwent further analysis. We wished to quantify the effect of temperature, recovery time, and HHP on the gene intensities. As these factors are all categorical, an ANOVA model was the natural choice. Considering a combination of the nature of the experiment and the biological focus, we decided to omit one variable from consideration to simplify the analysis. A separate ANOVA model was fit each of the possible covariates: HHP, recovery time and temperature. We looked at two metrics, firstly the number of genes for which we found that factor significant using a significance level of 0.01 and secondly the sum of the P-values for all probes. By taking into consideration both significant factors and that the re-expansion rates were not significantly affected by temperature, we chose to combine the data from the two temperatures to increase the statistical power and allow for a comprehensive analysis. We fit an ANOVA model with the covariates of HHP and recovery time. The ANOVA model returned a single P-value per probe. In order to account for multiple comparisons, we used the Benjamini Hochberg procedure to control for a false discovery rate (FDR) of 0.05. Hierarchical clusters were generated using Genespring GX 12.1 with the K-means clustering algorithm. Heatmaps and Venn diagrams of differentially expressed genes were developed with R.