Transcriptional profiling of formalin fixed paraffin embedded tissue: pitfalls and recommendations for identifying biologically relevant changes

Abstract

Expression profiling techniques have been used to study the biology of many types of cancer but have been limited to some extent by the requirement for collection of fresh tissue. In contrast, formalin fixed paraffin embedded (FFPE) samples are widely available and represent a vast resource of potential material. The techniques used to handle the degraded and modified RNA from these samples are relatively new and all the pitfalls and limitations of this material for whole genome expression profiling are not yet clarified. Here, we analyzed 70 FFPE tongue carcinoma samples and 17 controls using the whole genome DASL array covering nearly 21000 genes. We identified that sample age is related to quality of extracted RNA and that sample quality influences apparent expression levels in a non-random manner related to gene probe sequence, leading to spurious results. However, by removing sub-standard samples and analysing only those 28 cancers and 15 controls that had similar quality we were able to generate a list of 934 genes significantly altered in tongue cancer compared to control samples of tongue. This list contained previously identified changes and was enriched for genes involved in many cancer-related processes such as tissue remodelling, inflammation, differentiation and apoptosis. Four novel genes of potential importance in tongue cancer development and maintenance, SH3BGL2, SLC2A6, SLC16A3 and CXCL10, were independently confirmed, validating our data. Hence, gene expression profiling can be performed usefully on archival material if appropriate quality assurance steps are taken to ensure sample consistency and we present some recommendations for the use of FFPE material based on our findings.

Figure 1. Flowchart illustrating experimental procedure.

Description of the procedures used to assess samples from RNA extraction to acquiring of gene lists, including samples removed following each step of analysis. Tumour samples are abbreviated T and controls C.

Figure 2. Linear regression analysis.

Regression model describing how much of the variation in sample quality (Ct_diff) can be explained by sample storage time prior to extraction.

Figure 3. Linear regression analysis.

Regression model describing how much of the variation in number of detected genes on the array can be explained by sample quality after extraction (Ct_diff).

Figure 4. Example of three genes and how their expression was affected by difference in sample quality (Ct_diff).

(A), (B) and (C) Linear regression analysis of the genes YPEL5, TRPM4 and the short non-coding gene SNORA10 describing the relationship between expression of the gene and sample quality (CT_diff). All 78 samples are included and analysis was performed using non-normalized data. (D), (E) and (F) Linear regression analysis of the same genes using normalized data. Samples and regression line for tumours are denoted in blue and sample and regression line for controls in red.

Figure 5. Dendogram from unsupervised hierarchical clustering.

Including the 43 samples selected for differential gene expression analysis and all 12579 genes detected in these samples. Control samples are denoted by C and tumours by T. Pearson correlation was used as a measurement of similarity.