RNA Preparation. Total RNA was extracted from histone deacetylase inhibitor (HDACi)-treated and control (DMSO)-treated CCRF-CEM and CEM-p16 cells (3 × 10⁷ cells per treatment time point) with the use of TRIzol reagent (Invitrogen) and Qiagen (Valencia, CA) RNeasy kit according to the manufacturer’s instructions. RNA integrity was verified by electrophoresis through denaturing agarose gels. RNA for quantitative real-time PCR was further treated with RQ1 RNase-free DNase (Promega) to remove any residual genomic DNA.

Microarray Normalization. Normalization and statistical analysis of the expression data were carried out by using the LIMMA software package for the R programming environment (1, 2). This package contains a number of analysis methods not found in other software. Local background subtraction usually produces log-ratios that are very variable at low intensities (3). Because it was desired to detect differential expression for genes that might not necessarily be highly expressed, filtering out low-intensity spots was avoided. Instead, a strategy of background correction was used that avoids exaggerated variability of log-ratios for low-intensity spots. Background correction was performed by using the "normexp" method in LIMMA to adjust the local median background estimates. This strategy is similar to the background correction method used by the popular RMA software for Affymetrix data (4). It avoids problems with background estimates that are greater than foreground values and ensures that there were no missing or negative corrected intensities. An offset of 100 was used for both channels to further damp down the variability of log-ratios for low-intensity spots. The resulting log-ratios were normalized by using the print-tip group loess method with span 0.4, as recommended by Smyth and Speed (5). Here and elsewhere, the small number of spots that were manually marked as "bad" on visual inspection of the scanned arrays were filtered out of the analysis, while spots that were flagged as "not found" by GENEPIX were kept in the analysis but downweighted. The arrays used in these experiments were from four different print runs that were all printed with the same elements but with slightly different print layouts. The arrays from the different print runs were therefore normalized separately and the normalized expression data were combined and aligned by probe for subsequent analysis.

Assessment of Differential Expression. The combined CEM-CCRF expression data were analyzed together by using linear model methods (6, 7), and the combined CEM-p16^INK4A expression data were then similarly analyzed. Each probe was tested for changes in expression over the six time points by using a moderated F test (7). This test gave an F statistic on 5° of freedom to test for any differences between the six time points. This test is similar to an ANOVA method for each probe except that the residual standard deviations are moderated across genes to ensure more stable inference for each gene. The moderated standard deviations are a compromise between the individual genewise standard deviations and an overall pooled standard deviation. Among other advantages, this method prevents a gene from being judged as differentially expressed with a very small fold change merely because of an accidentally tiny residual standard deviation. The F statistics were computed for each probe to test for a response over time to suberoylanilide hydroxamic acid (SAHA), for a response to depsipeptide, and for any differences in the pattern of response to SAHA as compared with depsipeptide. Each probe therefore gave three statistics and three P values, one for SAHA response, one for depsipeptide response, and one contrasting the SAHA and depsipeptide responses. Similarly for the CEM-p16^INK4A arrays. In that case, F tests were computed for each probe to test for any response over time with and without doxycycline and for any differences between the two response patterns over time.

The linear models allowed for general changes in gene expression between successive time points. The use of dye-swaps in the experimental design allowed a dye-effect to be estimated for each probe for each compound (SAHA, depsipeptide, CEM-p16^INK4A (-dox) and CEM-p16^INK4A (+dox)). Removing this technological artifact increased the precision with which differential expression could be detected. Each time course of six arrays hybridized at the same time was treated as a randomized block to allow for correlations within the biological replicates and between arrays hybridized together (8). This refinement allowed the models to reflect more realistically the experimental sources of variability and made the computed P values more reliable.

P values were adjusted for multiple testing by using the method of Benjamini and Hochberg (9) to control the false discovery rate. Tests were considered to be significant if the adjusted P values were <0.05, nominally controlling the expected false discovery rate to no more than 5%. As a check on the methodology, moderated t tests and P values were computed to test for differences between the treated and untreated cells at time 0 h, a comparison that should be null. No probes were found to have adjusted P values < 0.05 for this comparison for either the SAHA or depsipeptide array series. This result suggests that the estimate of 5% for the false discovery rate estimate is actually conservative for this data and that the assessment of differential expression is not overstated.

Quantitative Real-Time PCR. DNase-treated RNA (2 m g ) was reverse transcribed with M-MLV Reverse Transcriptase, RNase H Minus, Point Mutant (Promega). PCR primers for each gene were designed by using PRIMER EXPRESS software (Applied Biosystems, Foster City, CA) with a melting temperature at 58–60°C and a resulting product of between 75–150 bp. See Table 2 for primer sequences. Each PCR was carried out in triplicate in a 20 m l volume by using SYBR Green Master Mix (Applied Biosystems) for 15 min at 95°C for initial denaturing, followed by 35 cycles of 95°C for 30 s and 60°C for 30 s in the ABI Prism 7700 sequence detection system. The ribosomal gene L32 was used as the control gene because it showed little variation in expression across the microarray time course experiments. Values for each gene were normalized to the expression levels obtained for L32, and fold induction was calculated versus time 0 h. Each reaction was done in triplicate from at least two independent experiments.

1. Wettenhall, J. M. & Smyth, G. K. (2004) Bioinformatics 20, 3705–3706.

2. Gentleman, R., Bates, D., Bolstad, B., Carey, V., Dettling, M., Dudoit, S., Ellis, B., Gautier, L., Gentry, J., Hornik, K., et al. (2004) Genome Biol. 5, R80.

3. Yang, D. H., Buckley, M. J., Dudoit, S. & Speed, T. (2002) J. Comp. Graph. Stats. 11, 108–136.

4. Irizarry, R. A., Hobbs, B., Collin, F., Beazer-Barclay, Y. D., Antonellis, K. J., Scherf, U. & Speed, T. P. (2003) Biostatistics 4, 249–264.

5. Smyth, G. K. & Speed, T. (2003) Methods 31, 265–273.

6. Yang, D. H. & Speed, T. P. (2003) Design and Analysis of Comparative Microarray Experiments (Chapman & Hall, New York).

7. Smyth, G. K. (2004) Stat. Appl. Genet. Mol. Biol. 3, Article 3.

8. Smyth, G. K., Michaud, J. & Scott, H. (January 18, 2005) Bioinformatics, 10.1093/bioinformatics/bti270.

9. Benjamini, Y. & Hochberg, Y. (1995) J. R. Stat. Soc. B 57, 289–300.