FPE RNA is largely fragmented, making it a poor substrate for poly-dT primed reverse transcription. Random priming in combination with poly-dT priming has proven to be somewhat more successful for generating cDNA from FPE RNA [6,11]. However, with highly fragmented RNA, even random priming does not yield satisfactory RT-PCR data [see Additional file 1], as compared to that obtained with in tact RNA [see Additional file 2]. We therefore chose to maximize assay sensitivity by using gene specific priming for each target sequence included in the RT-PCR screening panel. To evaluate the quality and consistency of cDNA from this complex reverse transcription reaction, we compared the expression values resulting from 8 smaller reverse primer pools (each containing 94 to 96 primers) to those from the single pool with all of the primers combined. In every case, cDNA from multiplexed RT reactions was combined with PCR master mix and distributed into 384 well plates with each well containing the primers and probe for 1 gene assay. High quality commercial RNA (Universal RNA, Stratagene, La Jolla, CA.) was used as the template for this evaluation. The resulting data are shown in Figure 1. Each priming condition was repeated twice and the mean C_T values were plotted. The mean C_Tvalues were very similar: C_T = 29.40 versus C_T = 28.98 for the small and large gene-specific primer (GSP) pools, respectively. Additionally, the Pearson's correlation for these two priming methods was 0.98. The median difference between priming methods was 0.4 C_T units. The large majority of assays (97.5%) had a C_T difference less than 2. For 2 of the assays which had a C_T difference greater than 2, this could be accounted for by 1 failed reaction (C_T = 40). The priming efficiency for all other assays where alternative priming lead to a C_T difference greater than 2 appears to be reproducible so the result may be accounted for by cross-priming events in the more complex GSP pool.

While measuring the realistic limit of multiplexed RT-priming in FPE RNA we did not want the data complicated by biological variability contributed by individual samples. We therefore created a sample consisting of pooled FPE RNA that represented a range of high and low expression values across the gene panel and reflected the quality of samples that would be used in the gene identification study. This was done by pooling 36 breast FPE RNA samples from tumor blocks over 10 years old. Two of the 8 primer sub-pools were selected at random to verify that the priming reaction remained consistent as the RNA template type changed from a high quality sample to a highly fragmented sample. The results, shown in Figure 2, indicate that the reaction using FPE RNA as a template was consistent with that seen for the high quality commercial RNA. The Pearson's correlation was 0.95 and the mean C_T values obtained for the entire gene panel were almost identical: C_T = 31.84 versus C_T = 31.79 using the small and large gene-specific primer (GSP) pools, respectively. The average raw C_T obtained with the FPE RNA is higher than that obtained with the high quality RNA. This is expected since the FPE RNA is highly fragmented resulting in fewer available targets for priming [5]. Although this shift does result in FPE samples being somewhat skewed toward the upper end of the assay dynamic range, typically at least a 2000 fold range in expression can still be reliably measured for most genes before reaching the assay limit of quantitation [12].

Once we confirmed the high complexity priming reaction was working consistently in FPET RNA, we prepared a gene specific primer pool containing 761 unique reverse primers for the biomarker discovery study using NSABP C01 and C02 clinical trial specimens.

To be included in the final study analysis, samples had to pass pathology, clinical and laboratory data QC requirements. To meet pathology acceptance criteria, a minimum of 5% of the tissue present in each sample was required to be invasive cancer cells. All samples were dissected to enrich tumor tissue and minimize non-tumor elements.

RNA was extracted from 354 FPE tissues cut from the NSABP C01/C02 clinical study samples (archived from 1977–1983) [13,14]. All samples yielded RNA using the semi-automated extraction process described below, however a total of 38 samples yielded less than the 1069 ng RNA required to ensure a final concentration of 1 ng/assay well. We performed RT-PCR on 23 of these samples, but none of these yielded acceptable RT-PCR data. Of the 306 samples taken through RT-PCR with the standard RNA load, 21 (< 7%) were excluded from analysis due to unacceptable RT-PCR data. Although the precise age of each excluded block was not available, it is reasonable to assume that the samples with unacceptable RT-PCR data may have been from the oldest blocks [5]. After other pre-specified exclusion criteria were applied, (pathological and clinical ineligibility), the final number of evaluable patients was 270. Table 1 provides a detailed breakdown of all excluded samples. It should be noted that the overall exclusion rate in this study is higher than we have experienced in other studies and is very likely related to the age of these samples, which is the oldest set we have analyzed.

Using a semi-automated extraction process, two lab technicians extracted RNA from 48 samples per day. Paraffin was removed from the tissue manually with xylene, followed by ethanol washes. After a proteinase K digestion and phenol-chloroform purification, the remaining RNA extraction steps were performed on an automated liquid handler using a plate-based protocol. The purified RNA was then quantified using an automated RiboGreen fluorescence assay (Invitrogen, Carlsbad, CA.). The resulting files containing the RNA quantification data were automatically collected by the laboratory information management system (LIMS). The LIMS used those data to generate an RNA concentration normalization work-list, which was then executed by an automated liquid handler during the reverse transcription reaction assembly procedure. The reverse transcription reaction was completed using an MJ Research thermocycler (Bio-Rad, Hercules, CA). Finally, quantitative PCR reactions were assembled at a rate of 64 plates (32 patient samples) per day. All automated liquid handlers were obtained from Tecan (TECAN Schweiz AG, Männedorf, Switzerland). This system yielded a total of 24,576 real time quantitative PCR reactions to be performed per day using four ABI PRISM 7900 HT instruments (Applied Biosystems, Foster City, CA.).

Process reproducibility was monitored throughout the study. Two Tecan Genesis liquid handling robots were used to assemble plates for 335 patient samples over a period of five weeks, for a total of 670 assay plates (this includes samples run with an RNA concentration of less than 1 ng/assay well). Given the large scale of the project and the number of potential sources of variability, we monitored the process performance and reagent stability throughout the duration of the study. This was done by using a single RNA reference sample made by pooling RNA extracted from 80 FPE colon specimens (block age ranging from 3 years – 6 years). As described previously, a pooled FPE sample was selected as a reference RNA since it more representative of a typical study sample than commercially available RNA, yet sufficiently abundant to provide a stable reference baseline throughout the study period. This sample was analyzed 3 times on the same ABI 7900 HT instrument to set a baseline reference profile for the study genes and then was assayed at intervals throughout the duration of the study. Figure 3 shows boxplots of the averaged expression values of all 761 genes over time for the reference sample. The baseline data (3 runs) and all 8 repeat analyses done during the 5 weeks of the study are shown. The outer fence of 3*IQR (Inter Quartile Range) was used to flag outliers. These data demonstrate that assay reagents and the assay processes remained stable throughout the study period. This measurement has been repeated for this sample during subsequent studies and continues to give consistently reproducible results (data not shown).

Figure 4 shows the same reference sample data with the associated standard deviations (SD) obtained for all 11 data points for each gene assay, plotted against the mean C_T expression value. The majority of assays (737 out of 761) showed a SD less than 1.0. The genes indicated by symbol "Δ"are those where mispipetting by the robot resulted in a single empty well. These wells were given a C_T value of zero. The gene indicated by symbol "□" was caused by a single spurious failed reaction well which was assigned a C_T value of 40.

The four ABI PRISM 7900 HT machines used for the study underwent qualification to meet internal performance specifications prior to use in clinical studies. Figure 5 represents typical data from a comparison (raw C_T values) of the reference sample on two different machines. The Pearson's correlation is 0.97. The variance in RT-PCR data increases as the C_T value increases and the assay approaches its "limit of quantitification" (LOQ). This has been described in detail elsewhere [12] and is consistent with the variance observed in this study.

Two 384 well plates were prepared per patient RNA specimen to screen all 761 candidate marker genes. Assays were randomly assigned between the two plates except for reference genes which were present on both assay plates. The average C_T values for each plate pair were compared and the data are illustrated in Figure 6. This graph demonstrates that plates from the same patient give consistent raw mean C_T values prior to reference normalization. The one outlier specimen was repeated, and on repeat fell into alignment with the other plate results. It should be noted that Figure 6 represents all patient samples taken through RT-PCR, including the 23 samples run with less than 1 ng/well input RNA. These specimens, which had high average plate C_T values were excluded from the final study analysis due to unsatisfactory RT-PCR data.

Although the two assay plates for each patient RNA were run on the same ABI Prism 7900 instrument to minimize process variability, data from each sample had to be internally normalized to allow all study specimens to be compared without being confounded by relative variability in RNA quality, quantity or process variability. This was accomplished by subtracting the averaged expression values from 6 reference genes (CLTC, NEDD8, RPLPO, RPS13, UBB, UBC) from the expression values for each gene in each sample. This method has previously been shown to effectively compensate for variability associated with RNA degradation in FPE material of different ages and qualities [5]. Normalization genes were chosen on the basis of low expression variation among patient samples, robust C_T signals (C_T < 35) and lack of association with clinical outcome. As part of our quality control and acceptance criteria, the average expression C_T value for the 6 reference genes for each patient sample had to be less than 35. Within the final evaluable dataset of 270 patient samples, the average reference signal was 29.5 C_T (SD = 1.39 C_T) with the lowest observed reference signal at 33.7 C_T, indicating good overall signal response and relative invariance in reference genes among patients.

A high degree of reproducibility was seen for these reference genes, as shown by the RPLPO example, in Figure 7. Because this gene was assayed on both QPCR plates for each patient sample, a direct comparison of C_T values between plates could be made. The intra-specimen Pearson's correlation for RPLPO was 0.98.

As an indication of the robustness of this high complexity assay we sought to identify co-expressed groups of genes, that are plausible based on known biological pathways. Unsupervised cluster analysis was performed on the final sample set using all 761 genes, resulting in several distinct clusters representing known biological pathways. Figures 8, 9, 10, 11, 12, 13 display sub-clusters observed among the 761 genes. We identified the following gene groups and pathways: cell cycle and proliferation, epithelial markers (or products secreted by epithelial cells), focal adhesion, stromal response, early response and immune/interferon inducible genes. The most highly correlated gene set from among these groups was the stromal response gene group, where the 1-Pearson's distance for 26 genes was less than 0.5. Within this group are many genes that encode extracellular matrix (ECM) proteins and regulators thereof – a signature similar to that seen during wound healing [15]. There is also a very high concordance in gene expression between p16ink4 and p14arf. This is to be expected since these are alternative transcripts of the same gene and the primer probe set used for p16ink4 also amplifies the p14arf variant. Additionally, correlations among genes within families were observed; for example, CDX1 and CDX2 (Pearsons correlation = 0.67), FUT3 and FUT6 (Pearsons correlation = 0.59), AREG and EREG (Pearsons correlation = 0.73), HSP1A1 and HSP1AB (Pearsons correlation = 0.44).

After data quality control and reference normalization, approximately 19% of the cancer related genes tested in this study were found to have a significant (p < 0.05) correlation with Recurrence Free Interval (RFI) by univariate regression analysis [10]. Approximately one quarter of these genes would be expected to be false positives [10]. While a large number of candidate genes showed significant correlations with clinical outcome (on both the raw and normalized C_T measurements), the average reference signal was shown not to be correlated with disease recurrence.

Archival colon tumor FPE tissue blocks were provided by the NSABP from both the C01 "A Clinical Trial To Evaluate Postoperative Immunotherapy And Postoperative Systemic Chemotherapy In The Management Of Resectable Colon Cancer" and C-02 "A Protocol To Evaluate The Postoperative Portal Vein Infusion Of 5-Fluorouracil And Heparin In Adenocarcinoma Of The Colon" clinical trials. Patients were enrolled in these trials between 1977–1983. Samples used in this study are representative of the general study populations for both trials.

Histotechnologists wore gloves at all times when handling tissue blocks. Before and after each block was sectioned, any debris was removed from the microtome with a disposable cotton swab or brush. All appliances (brush, forceps, knife, knife holder base) were wiped with an RNase Zap wipe followed by a soft cloth wetted with de-ionized water. A tissue floatation bath was used to help eliminate wrinkles and distortions in sections being mounted to glass microscope slides. When dissection was required to remove significant non-tumor elements, a representative H&E stained slide was used as a guide to mark tumor and non-tumor portions of three 10 micron unstained slides. Tumor tissue was then scraped away from the non tumor material and placed into an extraction tube. The tumor tissue from all 3 sections was placed into the same tube.

RNA was extracted from the tumor-enriched portion of three 10 micron sections per patient block. In order to scale sample throughput to 48 samples/batch, the RNA extraction procedure used a semi-automated method performed on a TECAN robotic liquid handler (TECAN Schweiz AG, Männedorf, Switzerland). The samples originated in individual 1.5 ml Eppendorf tubes and paraffin was removed by incubating with lab grade xylene for 5 minutes. Tissue was then pelleted by centrifugation at room temperature (+18°C to +25°C) for 5 minutes at approximately 14,000 RPM. The xylene was removed and the procedure repeated. The tissue pellet was washed by inverting several times with 200 proof ethyl alcohol and again pelleting by centrifugation at room temperature. The ethyl alcohol wash step was repeated. Immediately prior to adding Proteinase K, samples were inspected for residual alcohol; if any alcohol was visible, it was aspirated without disturbing the tissue pellet. Proteinase K digestion was performed using reagents from the MasterPure^® Purification kit (Epicentre, Madison, WI). Samples were incubated at +65°C for 2 hr with Proteinase K. Protein and genomic DNA were removed by manual addition of an equal volume of acid-phenol: chloroform and the removal of the upper aqueous phase after centrifugation for 5 minutes at approximately 10,000 RPM. Samples tubes were transferred to a TECAN liquid handler where purification was completed using the mirVana™ RNA purification kit (Ambion, Austin, Texas) on a 96 well glass fiber filter plate. Purification was followed by DNase I treatment on the same 96 well filter plate.

RNA was quantified using the RiboGreen fluorescence method as described by the kit manufacturer (Invitrogen, Carlsbad, CA.).

Candidate genes were obtained from published gene expression profiling data relating to colon cancer prognosis and response to therapy, biological pathways known to be important in cancer [23-28] and suggestions by our collaborators at the NSABP. The full gene list, accession numbers and associated primer and probe oligo sequences are provided [see Additional file 3].

The reference sequence for each gene included in the study was obtained from the NCBI Entrez website. TaqMan^® RT-PCR primers and probes were designed using an automated in-house primer design module. The complete list of assay primers and probes is shown in Additional file 3. Oligonucleotides were purchased from Biosearch Technologies Inc. (Novato, CA), Integrated DNA Technologies (Coralville, IA) and Eurogentec (San Diego, CA). Dual labeled TaqMan^® probes had 5'FAM as a reporter and either 3'BHQ-1 or 3'BHQ-2 as a quencher. Amplicon size was limited to a maximum of 90 bp.

Reverse transcription was performed using the Omniscript kit (Valencia, CA) for RT-PCR. For initial investigations, reverse primers (each at 100 μM) were collected into sub-pools of 94–96 primers. An aliquot from each sub-pool was added together to create the final GSP pool (each primer at 100 nmol/L). For the clinical study, reverse primers (each at 1 mM) were first collected in sub-pools of 89–96 primers and tested for priming performance before being combined into a master 761 gene-specific primer pool (each primer at 1 μmol/L). For both the initial investigations and the clinical study, each primer in the RT reaction was at a concentration of 50 nmol. Our standard procedure was to add RNA to the RT reaction at 12.5 ng/ul which equates to 1 ng/well (cDNA) for quantitative PCR. In each case, the RT reaction was performed in a single tube with the GSP pool (ie up to 761 reverse primers). The resulting cDNA was distributed equally among the wells of a 384 well plate, and the appropriate forward and reverse primer and probe were added to each assay well.

For each patient sample, two 384-well plates were used. Assays for 7 potential reference genes were included on both plates, with all other gene assays randomly distributed in single assay wells. RT-PCR assays for three K-ras gene mutations and one BRAF gene mutation were included in the assay panel along with each corresponding wild type allele assay. Therefore, 757 normal gene alleles were assayed and 4 mutant genes were assayed, bringing the total number of unique assays to 761. TaqMan^® RT-PCR was performed according to instructions of the manufacturer, using Applied Biosystems Prism (ABI) 7900 HT instruments. Reactions were performed in a 5 μl volume with cDNA equivalent to 1 ng total RNA. Final primer and probe concentrations were 0.9 μmol/L (primers) and 0.2 μmol/L (probe). For the K-ras mutation assays a blocker oligomer was added to the primer and probe pool at a final concentration of 3.6 μmol/L (mutant 1) 3.6 μmol/L (mutant 2) and 12.95 μmol/L (mutant 3). These blockers are added to inhibit amplification of the non-mutant allele, permitting specific amplification of the mutant allele. PCR cycling conditions were 95°C for 10 minutes for one cycle, 95°C for 20 seconds, and 60°C for 45 seconds for 40 cycles. A reference sample (pooled colon FPET RNA) was assayed throughout the study to ensure reagent and process stability. As a negative control, wells without any template were also assayed every two weeks to ensure that no exogenous nucleic acid contaminations occurred.

The unsupervised hierarchical clustering of genes was performed using 1-Pearson R as the distance measure for gene expression and the un-weighted pair-group average as the amalgamation method [29].