Not surprisingly, the unfiltered arrays showed poorer overall correlation even between replicate arrays than did the analysis filtered for high intensity spots. Even the unamplified arrays had correlation coefficients somewhat lower than expected when array data was unfiltered, indicating that the filtration process removes the effect of background and nonspecific hybridization on the cDNA arrays. The filtration process did not change the individual expression ratios but rather focused the analysis on genes in which the hybridization gave a particularly strong signal, a strategy that seems logical when attempting to work with very rare clinical specimens to ensure a lower false call rate. Other groups have reported analysis of total RNA arrays with input as high as 20–40 ug per microarray, which may have improved the results for our total RNA arrays That being said, in real world situations researchers will not have abundant amounts of clinically relevant total RNA to use for comparisons to low-input RNA samples. Confirmation of array results will require reliance on functional assays such as QPCR to validate results, which is why we emphasized the QPCR component of our analysis. All three methods tested showed comparable accuracy rates that were within the range of previous reports based on amplification from picogram amounts of total RNA. However, our study also determined the minimal input requirements of each of the 3 amplification methods. Table 1 shows the threshold for each technique - for example, above 250 pg, the Arcturus method worked reproducibly, while the modified T7 and balanced PCR techniques required approximately 1 ng of total RNA.
RNA amplification technologies serve translational clinical research well. Linear amplification already has enabled examination of gene expression in clinical core needle biopsies, surgical biopsies, fine needle aspirates and even single human cells. Our analysis of the high intensity spots demonstrates that amplification is reproducible and highly correlated with QPCR measurements using sub-nanogram input RNA samples. These results clearly demonstrate that data-processing has a marked impact on expression array results, particularly when working with very low-input samples.
While each method was able to provide data in the sub-nanogram range, certain methods are advantageous over others in terms of lower limit of RNA that can reliably be amplified, cost per reaction and the number of days required for processing of samples. The Arcturus RiboAmp HS method was more reliable at generating expression arrays at the threshold of below a nanogram of total RNA. For modified T7 and balanced PCR, a nanogram of total RNA will ensure reliability. The Arcturus amplification procedure was able to produce very substantial amounts of nucleic acids from just 250 pg of total RNA and therefore should be considered more reliable than the other two methods at lower input thresholds. Caretti et al. performed a comparison of amplifications of a colon biopsy subjected to laser capture microdissection with purification of an estimated 1 nanogram of RNA per specimen; they compared the two cycle Arcturus OA and the one cycle Nugen amplification and found that the Arcturus method showed the lowest variance and highest correlation. However, our data is distinct in that we examined the Arcturus Ribo Amp HS kit, which permits RNA amplification from lower input samples than the Arcturus OA or Nugen kits. In addition, our study provides detailed comparisons of amplified RNA to both total RNA and QPCR, which is not available in the study by Caretti et al. Wilhelm et al compared Arcturus Ribo Amp (a less sensitive kit than Ribo Amp HS) to SMART PCR (Clontech) and concluded that SMART is preferable when working with less than 200 ng of total RNA. Our study is distinct in that we show that IVT based amplification (Ribo Amp HS) is more reliable than PCR based amplification at the sub-nanogram input level. Clontech's SMART PCR is marketed for use at the nanogram low-input range, not the sub-nanogram range, which is why we did not include SMART PCR in our comparison.
Shearstone et al performed a comparison of their laboratory's novel IVT amplification, termed BIIB, to Arturus Ribo Amp HS, Nugen Ovation, Affymetrix One and Affymetrix Two Cycle. Their method, BIIB, proved to be capable of successful amplification with good reproducibility from 50 pg of total RNA, and could amplify from lower input amounts of RNA than commercially available kits. They emphasized that T7 based linear amplification has advantages over PCR based amplification because T7 accurately retains the transcript stoichiometry of the original sample. Our study is similar to that of Shearstone et al in the design by which they use a reference RNA for the baseline from which to determine which of several amplification techniques performs the best at low inputs of RNA. Our study is distinct in that we found that Arcturus Ribo Amp HS had acceptable correlation coefficients in the 250 pg range when attention was focused to the high intensity spots for data analysis. Our study provides additional insight on the impact of data processing on microarray results at low input of total RNA, which was not the focus of Shearstone et al. Although BIIB performed well in their hands, Arcturus Ribo Amp HS is currently commercially available, kit based, and takes less time to perform than BIIB. Further studies are needed to compare these two methods and perhaps improve upon them with the ultimate goal being consistent successful amplification from even a single cell.
Our paper describes a platform independent measure, the FER, useful in comparing amplification methods based on QPCR versus microarray. FER was calculated based on QPCR of total RNA rather than amplified RNA, since there was not sufficient amplified product available for performing QPCR on all 37 primer probe pairs. In addition, QPCR of amplified RNA is biased towards recapitulating the results of the microarray experiment due to truncation of the RNA products.
Each of the three amplification techniques yielded fairly consistent expression results within the constraints of each technique's input threshold of total RNA, both on microarray analysis and when compared to QPCR. Based on these excellent correlations, it is feasible to reproducibly perform high fidelity amplifications by a variety of techniques when starting with sub-nanogram input quantities of total RNA. However, when attempting expression arrays analysis from less than 500 pg input, linear amplification with Arcturus RiboAmp HS was more successful than the other methods studied.
Below 1 ng, the modified T7 method could not reproducibly amplify such that insufficient RNA was typically generated for even a single microarray hybridization. Only by performing multiple attempts at amplification were we able to achieve successful amplified product with this technique. While we were successful in hybridizing 3 arrays with this method at 500 pg we do not recommend this method below 1 ng of input total RNA as several operators quite experienced with this method could not repeat these results. One drawback of this technique is the greater length of time involved (3 days) compared to other amplification reactions (2 days) and the relative complexity of the protocol.
Arcturus RiboAmp HS was able to provide expression array data at a lower input concentration than any of the other tested methods and we were able to use smaller amounts than the manufacturer's recommended minimum sample input of 500 pg total RNA. Below 250 pg, this method typically failed to amplify in our hands (although the manufacturer validated the Arcturus method to a threshold of 100 picograms). This likely represents a theoretical limit of 25 cells total RNA content (for laser capture microdissection more would be required because of fractionation of cellular material), unless specialized tissues that bear more RNA such as oocytes are examined. Since each somatic mammalian cell is thought to have approximately 10 pg of total RNA, the Arcturus method is especially promising for limiting clinical samples, suggesting that clinical samples comprised of approximately 25 cells could be routinely analyzed with Arcturus, or 100 cells with modified T7 or balanced PCR, with profound implications for correlative science studies using gene expression profiling that have previously not considered using such small quantities. It is somewhat concerning, however, that the % FER was observed to increase from 10.8% at 500 pg to 19% at 250 pg in our study. This is a potential area of future investigation as scientists seek to push the lower limit of input RNA required
Balanced PCR is a promising technique for the amplification of low-input quantities of RNA. It maintains a high degree of accuracy with an input as low as 667 pg of RNA (FER 10.8–13.5%). While theoretical concern exists regarding the accuracy of logarithmic amplification methods, this method overcomes the potential problem by stopping the PCR reaction before the logarithmic phase of the PCR curve. This method had the lowest cost per reaction and also required the least amount of technician time compared to the other methods. In addition, it has been recently demonstrated that the same balanced-PCR protocol used for cDNA amplification may also be used for the unbiased amplification of whole genomic DNA followed by array-CGH analysis. However, several iterations were required for experienced personnel to learn to successfully perform balanced PCR.
These results have certain important limitations to consider. First, hybridizations were carried out sequentially over a period of several months rather than all at the same time. It is recognized that arrays that are hybridized together under identical conditions are more similar to each other than arrays hybridized on separate occasions. Additionally, no dye swap experiments were performed. The rationale for this is that since a standardized universal reference (StratRef) was employed, it has been demonstrated that this mitigates the effect of potential experimental bias introduced by separate hybridization reactions and even permits the comparisons of array data between members of the research community. Use of a universal reference may in some cases have advantages over dye swap experiments. 
Each laboratory will have to weigh their decision on which amplification technique is most suitable based on factors including amount of starting input total RNA, cost per reaction, technician time, and experience/comfort level with the techniques. Laboratories that routinely work with samples in excess of 1 ng starting material should focus on cost-savings as each of the methods tested proved to be reliable above this threshold. Balanced PCR could be further optimized to include amino-allyl-dUTP incorporation in the PCR reaction. This would facilitate indirect Cy dye labeling, which would reduce the labeling cost for this method.
It is important to ascertain the linearity of a chosen method at the low input range before going on to work with precious clinical specimens. Each of the 3 tested methods performed surprisingly accurately when amplifying from low inputs of total RNA based on microarray analysis validated with QPCR of 37 genes.