Methods and protocols for RNA isolation from formalin-fixed tissues have been published since almost 20 years [16–32].
RNA was quantified by dot blot hybridization , semi-quantitative PCR  and more recently, by QPCR [24, 18, 26, 13, 17, 33, 32] and other methods [28–30]. RNA derived of FFPE material is not only partially hydrolyzed but also chemically modified: formalin reacts with nucleotides leading to the formation of methylol groups in nucleobases. These groups tend to further react and form intra- and inter-molecular methylene bridges in RNA, DNA [34, 35, 31] and protein . As a result, reverse transcription is impaired and threshold cycle values (Ct values) increase during subsequent QPCR.
The protocol for RNA isolation described here was complemented by adding a separate demodification step which involves incubation at elevated temperature in a buffer containing ammonium chloride which favors the reversion of methylol groups to amino groups in nucleobases. It does not only improve the efficiency of downstream applications (mainly reverse transcription), it also improves the recovery of RNA from FFPE sections. RNA yield and quality can be further improved by extensive digestion of FFPE material with protease in a buffer containing guanidinium thiocyanate. Reverse transcription in the presence of gene-specific primers prevents the initiation of cDNA synthesis inside amplicons and therefore, cDNA made in the presence of gene-specific primers is a better template for QPCR than cDNA made from random primers (Fig. 2). Several papers have demonstrated that QPCR with primers coding for short amplicons are more efficient than primers coding for long amplicons [17, 20, 24, 13, 32].
Finally, normalization of raw data is used to eliminate or at least reduce the effect of poorer quality of starting RNA. Various approaches of normalization were proposed in the literature [37, 14, 38, 32]. They are based on calculating relative expression values: expression levels of genes of interest are expressed relative to the expression of one or a panel of several suitable reference genes. An ideal reference gene has a stable expression level in all the samples under investigation. As such "ideal" reference gene normally does not exist, the mean or median expression level of several suitably chosen reference genes is used as a relatively stable reference Ct value. We used a formalized approach to characterize all candidate reference genes. Candidate reference genes were ranked according to their standard deviations of raw Ct values in RNA from FF and FFPE material. The final rank of each candidate reference gene was taken as the mean of the two ranks obtained with RNA from intact and FFPE material. Genes with higher ranks were excluded as reference genes.
We also applied GeNorm  to characterize candidate reference genes: ACTB and RPS11 had poorest stability measure M  for FFPE-derived RNA and RPL7A had a poor stability measure when RNA from FF material was tested (data not shown). For these reasons GAPDH, GUSB, RPLP0, TFRC and UBB were used as reference genes in this study.
Our own RNA isolation protocol was compared to RNA that was isolated from the same material but according to commercial protocols and products (Qiagen RNeasy FFPE and ncLysis system of Applied Biosystems). Additional products for FFPE material from commercial providers (e.g. Stratagene, Ambion) were tested and the results obtained with our own protocol were superior to all tested commercial products (data not shown).
We determined module scores for each of the 14 tumors in this study. The limited number of samples does not allow statements about the clinical significance of module scores but they can be used to compare scores computed from intact RNA from FF material and RNA isolated from FFPE according to our own protocol. Pearson correlations between these RNAs in the 14 tumors were 0.966, 0.856 and 0.833 for ER, HER2 and Proliferation scores, respectively. As kryo-preserved RNA and RNA from FFPE material always originated from different portions of the same tumor, a certain variation of gene expression cannot be excluded and, as a consequence, part of the observed variability between kryo and FFPE material may be attributed to biological heterogeneity in the tumors. The three module scores were combined to a Total score. The Total score is similar to the recurrence score described by Paik , with high expression of genes related to proliferation and HER2 and low expression of ER-related genes indicating higher risk.
The data generated from FF and FFPE material were also compared to ER and HER2 levels assessed by IHC results from the same tumors. Three tumors (#15, #18, #20) were ER-negative and one was strongly HER2-positive (#6) (Tab. 2). The same tumors had low ER scores when assessed by QPCR (Fig. 8). Tumor #6 had a high HER2 score and an intermediate ER score. These results are in good agreement with the expected distribution of the three scores [15, 39]. By comparing QPCR based data with well known tumor subtypes allowed to validate the protocols developed here, even if no new biological findings are provided. The primary issue of this work was to document that stable and robust expression values can be determined from FFPE-derived RNA which are close to the values computed from intact RNA of the same tumors. The optimization and validation of the scoring procedure remains an important issue but obviously, the available number of samples is not sufficient to deal with this aspect and it will be addressed separately and on a larger collection of samples.
While IHC results are at most semi-quantitative, QPCR-based results reflect more accurately the expression level of genes in question. The module scores proposed here integrate quantitative gene expression data from several genes, this makes the resulting scores more robust than measurements based on single genes. QPCR is not only quantitative, it is also very sensitive over a large dynamic range. The number of genes which can be measured by QPCR is not limited and additional genes and module scores can be included in the analysis if this will be required.
Importantly, certain predictive parameters still cannot be determined with current technologies. For example, breast cancers are classified into histological grade 1, 2 or 3. This grading most likely reflects the proliferative state of tumor cells . Grading may be especially important as high grade tumors seem to respond more favorably to chemotherapy than low grade tumors. Unfortunately, many tumors are histological grade 2 and for those tumors the benefit is not clear. Paik and co-workers documented that their recurrence score (RS) was also predictive for a response to chemotherapy . The RS defined by Paik and coworkers is composed of 16 test genes mainly representing ER response genes, proliferation-associated genes, HER2-related genes and invasion genes and 5 genes for normalization [11, 41].
The genes selected for this study (Tab. 1) were selected from published DNA chip studies with breast cancer samples . They mostly coincide with the genes used by Paik ad co-workers.