Patients and Tissue Specimens
Patient tumor samples were collected retrospectively from Roswell Park Cancer Institute (RPCI), Buffalo, NY and from Aarhus and Odense University Hospitals in Denmark. The use of all included samples in this study was approved by the institutional review board at Roswell Park Cancer Institute, and, in Denmark, by Den Videnskabsetiske Komité. The study was conducted in accordance with the Helsinki declaration. In total 75 NSCLC specimens were included in the study. These comprised 68 stage I NSCLC from a US cohort collected and treated at RPCI [6] and seven NSCLC specimens collected in Denmark. Clinical data were obtained only for the US cohort, from the tumor registry at RPCI and through chart reviews [6]. Approximately half of the patients from the US cohort were known to have had a recurrence. The recurrence-free cases were followed for at least 32 months, with approximately half of them followed for at least 5 years [6]. For the US specimens and for two Danish specimens, tissue cores were sampled from FFPE tissue blocks from areas with > 70% tumor cell content (as verified by HE-stain), and subsequently cores were re-embedded in paraffin. For the remaining four FFPE specimens, tissues sections (20 μm thick) were obtained for the comparison of the two RNA extraction kits. One NSCLC specimen was collected within two hours of surgery and was preserved in RNAlater (Ambion, Inc 2130 Woodward St. Austin, TX) with approval by Den Videnskabsetiske Komité in Denmark and with informed consent obtained from the patient.
RNA extraction
In a PubMed http://www.ncbi.nlm.nih.gov/pubmed search using the search terms "miRNA" AND "Cancer" AND "FFPE", 33 publications were retrieved covering the period from 2009 to 2011. Of these, 16 publications described the use of global miRNA profiling, and in over half of these (i.e. in 9 studies), RNA was extracted using the RecoverAll kit (Ambion). In the present study we compared two different RNA extraction methods. In addition to the widely used RecoverAll kit ("RA Kit", Ambion), we included the High Pure miRNA Isolation Kit ("HP Kit", Roche Applied Science, 68298 Mannheim, Germany). For the HP Kit, RNA was extracted from deparaffinized and proteinase K-treated FFPE core tissues (20-40 mg) or sections according to the manufacturer's instructions. In approximately one-third of the cases, RNA preparations were of poor quality. Consequently, RNA was extracted again from FFPE tissue. For the RA Kit, RNA was extracted from deparaffinized and protease-treated FFPE core tissues (20-40 mg) or sections with on column DNAse digestion according to the manufacturer's instructions. RNA concentration and quality was assessed by absorbance spectrometry and electrophoresis using the NanoDrop (Thermo Fisher Scientific, Wilmington, DE) and Bioanalyzer 2100 (Agilent Technologies) instruments.
Labeling of RNA
Two different RNA labeling methods were compared in this study; the FlashTag™ Biotin RNA Labeling Kit (Genisphere, PA), and the new HSR version of the same kit were used according to the manufacturer's instructions unless otherwise stated. Poly (A)-tailing: 100-800 ng RNA including the small RNA species was used as starting material for the polyA-tailing reaction with different ATP-mix dilutions (from 1:10-1:500). Poly (A)-tailing was performed at 37°C for 15 minutes (GeneAmp PCR System 9700, Applied Biosystems, Foster City, CA). Ligation: Using the entire reaction product from the poly (A) tailing reaction, a biotin-labeled 3DNA® dendrimer was ligated to the poly (A) tails using a T4 DNA Ligase and incubated at 25°C for 30 minutes (GeneAmp PCR System 9700, Applied Biosystems®) following the manufacturer's instructions.
Hybridization
Of the 16 studies from the PubMed http://www.ncbi.nlm.nih.gov/pubmed search describing global miRNA profiling, 12 studies used six different commercially available platforms and the remaining four studies used different custom platforms. Between 365 and 847 human miRNAs were profiled using from 20 ng to 5 ug totalRNA as input. In the present study we used the platform with the broadest coverage, i.e. the Afymetrix GeneChip® miRNA Array that interrogates 847 human miRNAs and about an equal number of small non-coding RNAs. Hybridization washing and staining were performed using the Affymetrix GeneChip Hybridization, Wash and Stain Kit (Affymetrix, CA). Briefly, the hybridization cocktail containing the biotin labeled RNA was heated to 99°C for 5 minutes and then to 45°C for 5 minutes (GeneAmp PCR System 9700, Applied Biosystems®) before loading onto the Affymetrix probe array cartridge (GeneChip® miRNA Array). The volume of the hybridization cocktail loaded on the chip was changed from 100 ml to 80 ml in order to improve movement/flow of the cocktail in the hybridization chamber, ensuring a better and more even hybridization process. Thus, in effect only 80% of the labeled RNA was placed on the chip. The probe array was incubated for 17 hours at 48°C with constant rotation (60 r.p.m.). The probe array was incubated for 17 h at 48°C at constant rotation (60 r.p.m.). The biotin labeled RNA was stained with a streptavidin-phycoerythrin conjugate and the signals amplified using a biotinylated goat antibody against streptavidin. Finally, the samples were stained with a streptavidin-phycoerythrin conjugate.
Scanning
The probe arrays were scanned using a confocal laser-scanning microscope (Affymetrix GCS3000Dx2). The readings from the quantitative scanning were analyzed using the Affymetrix Molecular Diagnostics Software (AMDS). The microarray data was deposited in the Array Express public database http://www.ebi.ac.uk/arrayexpress/ and has been assigned accession number E-MTAB-618 (under experiment name: Laboratory assays for prediction of relapse in stage I non-small cell lung cancer (NSCLC)).
RNA input amount
For comparisons, RNA was extracted from a sample collected under conditions where RNA degradation is expected to be minimal. Thus, a sample of a T2 NSCLC post-resection surgical specimen was collected and preserved in RNAlater (Ambion) within 2 hours of after surgery. RNA was subsequently extracted using the HP Kit. The effect of varying RNA input amounts (80, 100, 160, 400 and 640 ng) for hybridization was examined. In addition, RNA was prepared from ten stage I NSCLC surgical FFPE specimens using the HP Kit (100 and 600 ng) and these samples were labeled and hybridized to miRNA arrays. Signal intensities were examined on the chip using laser scanning microscopy. Background intensity ranged from 31 to 36 and was unrelated to the amount of input RNA (results not shown). Subtraction of background intensities did not affect the results (data not shown).
ATP-mix dilution
The effect of different ATP-mix dilutions (1:10; 1:50 and 1:100) was examined using 100 ng RNA extracted from a single NSCLC specimen using the HP Kit and labeled using the FlashTag™ Biotin RNA Labeling Kit (Genisphere). In addition, different ATP-mix dilutions (1:50; 1:150 and 1:500) were examined using two preparations of 600 ng RNA extracted from two different NSCLC FFPE specimens using the RA Kit and labeled using the FlashTag™ Biotin HSR RNA Labeling Kit (Genisphere).
Different chip lot numbers
Correlations in signal intensities were examined across two different lots of arrays that were hybridized (in triplicates) to 100 ng of labeled RNA (FlashTag™ Biotin RNA Labeling Kit, ATP-mix dilution 1:50) from of a single RNA preparation of a T2 NSCLC tumor using the HP Kit.
Comparisons of two different RNA extraction kits
The High Pure miRNA Isolation Kit ("HP Kit ", Roche) and RecoverAll ("RA Kit", Ambion) extraction methods were compared in an RNA extraction experiment with four FFPE NSCLC specimens. From each specimen, 20 sections (20 mm thick) were cut, every other slide being used for one extraction method and the remaining slides for the other. For each extraction method, 1000 ng RNA of each specimen was polyadenylated (ATP-mix dilution 1:50) and labeled with biotin using the FlashTag™ Biotin HSR RNA Labeling Kit (Genisphere, PA).
Comparisons of two different labeling kits
The FlashTag™ Biotin RNA Labeling Kit (Genisphere), and the new HSR version of this kit were compared in an RNA labeling experiment using RNA from four NSCLC specimens from the RPCI cohort extracted using the HP Kit. For each labeling kit, 600 ng RNA from each specimen was polyadenylated (ATP-mix dilution 1:50) and labeled with biotin.
Assay I and II for prognostication in stage I NSCLC
For assay I, total RNA including small RNA was extracted from 68 stage I NSCLC specimens using the High Pure miRNA Isolation Kit (Roche) and 600 ng was labeled using the FlashTag™ Biotin RNA Labeling Kit (Genisphere). For assay II, total RNA including small RNA was extracted from 63 of the 68 stage I NSCLC specimens using RecoverAll (Ambion) and 600 ng was labeled using the FlashTag™ Biotin HSR RNA Labeling Kit (Genisphere).
Prognostic profile based on non-coding small RNA species
We tested the performance of a SVM classifier on both assays using a Monte Carlo approach. First, two-thirds of a data set was randomly chosen and used as a training set. Second, training was done by setting the number of features (i.e. non-coding RNAs) to 30, chosen according to a highest t statistic, MCRestimate package [21], in a LOOCV-loop. The 30 most frequent features were used for training of the SVM classifier. Third, performance was measured on one-third of the data set left out. Finally, the above procedure was repeated 1000 times. A nested LOOCV approach was further used to identify the optimal number of non-coding RNAs in the following way: First, a test sample was held out in the outer loop (leaving N-1 samples). In the inner loop LOOCV on N-1 samples was used to determine the accuracy for a range of selected features. Here, features were first selected based on the highest t statistic mentioned above. Next, selected features were used as input to the SVM classifier to classify each left out sample in turn. Subsequently, the number of features yielding the highest accuracy was used to classify the test sample that was held out in the outer loop. Ultimately, the most frequent number of features yielded the chosen prognostic profile.
Impact of analytical conditions on the robustness of the prognosticator
To qualify the impact of any variances in the analytical conditions of the assay, the calls (i.e. "recurrence" or "no recurrence") of the prognosticator were examined after varying RNA input amount, ATP-mix dilution, chip lot numbers, RNA extraction kit and RNA labeling kit for eleven selected NSCLC samples. The prognosticator was trained using either 68 (assay I) or 63 (assay II) stage I NSCLC samples using a SVM classifier and the identified prognostic profiles of each assay. Second, the trained SVM classifier was used to predict the outcome and thus examine the robustness of the prognosticator.
Statistical analysis and Bioinformatics analysis
For testing the effects of RNA amount; ATPmix dilution; extraction kit and labeling kit on signal intensity, background intensity, the numbers of detected probes and probe sets, arrays were pre-processed using Affymetrix miRNA QC Tool 1.0.33 (with workflow set to "default"). T-tests and ANOVA, assuming equal variances were performed using the R software package [22]. For correlation analysis, data normalization was performed using the justRMA procedure in Bioconductor [21] generating expression indexes (log with base 2) for all human features on the Affymetrix GeneChip miRNA Arrays. Principal component analysis (PCA) was performed using all human non-coding RNAs and the extracted signature for prognostication. For the prognostic profiles we used the raw miRNA data without background correction. We used perfect match probes only and summarized with average difference.