Assessment of the relationship between pre-chip and post-chip quality measures for Affymetrix GeneChip expression data

Abstract

Background: Gene expression microarray experiments are expensive to conduct and guidelines for acceptable quality control at intermediate steps before and after the samples are hybridised to chips are vague. We conducted an experiment hybridising RNA from human brain to 117 U133A Affymetrix GeneChips and used these data to explore the relationship between 4 pre-chip variables and 22 post-chip outcomes and quality control measures.

Results: We found that the pre-chip variables were significantly correlated with each other but that this correlation was strongest between measures of RNA quality and cRNA yield. Post-mortem interval was negatively correlated with these variables. Four principal components, reflecting array outliers, array adjustment, hybridisation noise and RNA integrity, explain about 75% of the total post-chip measure variability. Two significant canonical correlations existed between the pre-chip and post-chip variables, derived from MAS 5.0, dChip and the Bioconductor packages affy and affyPLM. The strongest (CANCOR 0.838, p < 0.0001) correlated RNA integrity and yield with post chip quality control (QC) measures indexing 3'/5' RNA ratios, bias or scaling of the chip and scaling of the variability of the signal across the chip. Post-mortem interval was relatively unimportant. We also found that the RNA integrity number (RIN) could be moderately well predicted by post-chip measures B_ACTIN35, GAPDH35 and SF.

Conclusion: We have found that the post-chip variables having the strongest association with quantities measurable before hybridisation are those reflecting RNA integrity. Other aspects of quality, such as noise measures (reflecting the execution of the assay) or measures reflecting data quality (outlier status and array adjustment variables) are not well predicted by the variables we were able to determine ahead of time. There could be other variables measurable pre-hybridisation which may be better associated with expression data quality measures. Uncovering such connections could create savings on costly microarray experiments by eliminating poor samples before hybridisation.

Figure 1

Total RNA Gel images from the Bioanalyser (Agilent). Representative total RNA samples of varying quality, as assessed objectively by RIN and by subjective assessment. Selected corresponding pre- and post-chip variable assessments are also shown. Samples identified as outliers on post-chip quality control measures but not excluded are labelled (*); outliers excluded from expression analysis are labelled (**). Sample from case with prolonged agonal state (†). Sample not run on arrays due to poor quality total RNA (‡). n.d. is value not determined. CB, cerebellum; CN, caudate nucleus; MC, motor cortex.

Figure 2

Pairwise scatterplots for the first four principal components. Outlier chips (Table 3) are represented by blue triangles.

Figure 3

Pairwise scatterplots of PC1 vs IQR_LR1. Outlier chips (Table 3) are represented by blue triangles.

Figure 4

Pairwise scatterplots showing RIN plotted against (A) B_ACTIN, (B) GAPDH and (C) SF. Outlier chips (Table 3) are represented by blue triangles.

Figure 5

The effect of including poor quality chips in analyses on ability to detect differential gene expression. Fewer differentially expressed genes are detected when comparing male and female motor cortex if a chip that failed QC is included in the analysis, reflected in the at least 50% fewer probe sets detected as differentially expressed at the two different p-value thresholds (t-test nominal unadjusted p-values). Bad (B) indicates comparisons where one chip that failed QC (HC71: female) was included in the analysis; Good (G) indicates comparisons where all chips passed QC. Samples were matched for age. This effect is most marked with very small chip numbers and gradually becomes less as chip numbers increase.