Human tissue RNA was obtained from Ambion (Applied Biosystems, Carlsbad, CA) or Stratagene (Agilent Technologies Inc., Santa Clara, CA). The catalog and lot numbers of the products tested are indicated in Table 1. RNA was also prepared from human skeletal muscle tissue using a Qiagen TissueLyser and Qiagen RNeasy miniprep kits (Qiagen, Valencia, CA). The donor tissue was obtained from amputation procedures through the Cooperative Human Tissue Network which is funded by the National Cancer Institute. The protocol for this study was reviewed by the FDA Research Involving Human Subjects Committee.
RNA quantification and assessment of purity was performed on a NanoDrop spectrophotometer (ThermoScientific, Wilmington, DE). RNA quality was assessed using an Agilent RNA 6000 Nano kit, an Agilent Bioanalyzer (Agilent Technologies, Inc.), and the manufacturer's software to assign RINs . The RINs that were measured for different lots of commercial human RNA are in Table 1.
UHRR, HBRR, liver RNA, and skeletal muscle RNA were individually labeled using the 3' IVT Express kit (Method 1A) and hybridized to Human Genome U133A 2.0 arrays (Affymetrix, Santa Clara, CA). Three replicate datasets were created by labeling and hybridizing the four tissue RNA samples on separate dates to introduce technical variation. The commercial lots of RNA used for analyte selection are indicated in Table 1. For each replicate dataset, the log2 signal intensities were derived and quantile normalized using the Robust Multichip Analysis (RMA) algorithm in the Affymetrix Expression Console software. A mean TSI for each probe set was calculated from the three replicate datasets and the threshold for tissue selectivity was a mean TSI greater than 3.22 log2 units. For the MTRC-4, 429 HBRR-, 255 liver RNA-, 197 UHRR-, and 154 skeletal muscle RNA-selective analytes were identified (see Additional file 1 - Lists of MTRC-4 analytes for Affymetrix HG-U133A 2.0 arrays). With omission of the UHRR from the comparison, 586 HBRR-, 384 liver RNA-, and 181 skeletal muscle RNA-selective analytes were identified as analytes for the MTRC-3 (see Additional file 2 - Lists of MTRC-3 analytes for Affymetrix HG-U133A 2.0 arrays).
Four batches of MTRC-4 were prepared using total RNA from four human tissues from different lots (Table 1). Batches 1-3 were tested in singlicate measurements (Trials 1-3) and Batch 4 was used in all of the other MTRC-4 datasets. A 100 μg batch of MTRC-4 Mix1 contains 20 μg UHRR, 30 μg liver RNA, 40 μg HBRR RNA, and 10 μg skeletal muscle RNA. A 100 μg batch of MTRC-4 Mix2 contains 20 μg UHRR, 20 μg liver RNA, 20 μg HBRR RNA, and 40 μg skeletal muscle RNA. The ratio of total RNA contributed by UHRR, liver, HBRR and skeletal muscle between Mix1 and Mix2 in the MTRC-4 is 1:1, 1.5:1, 2:1, and 1:4, respectively. UHRR was used for the 1-to-1 component in the MTRC-4 to minimize the impact on the observed mixed ratio due to the higher fraction of mRNA in the cell line derived total RNA than in sources of tissue RNA like the HBRR . Human liver RNA, which was found to have a higher RIN value across commercial lots than the other components, was used for the ratio more sensitive to performance differences (1.5-fold). Of the remaining two components, the HBRR was selected for measuring 2-fold changes because it contained the larger number of tissue-selective analytes.
MTRC-3 were prepared from total RNA from three human tissues using the lots indicated for Batch 5 in Table 1 and mixed to provide a 1:1 ratio for HBRR and reciprocal 1.5:1 ratios for liver and muscle RNAs. A 100 μg batch of MTRC-3 Mix1 contained 25 μg HBRR RNA, 45 μg liver RNA, and 30 μg skeletal muscle RNA. A 100 μg batch of MTRC-3 Mix2 contained 25 μg HBRR RNA, 30 μg liver RNA, and 45 μg skeletal muscle RNA. With the HBRR as the 1-to-1 component in the MTRC-3, a similar number of true negatives (586 brain RNA-selective analytes) and true positives (565 combined liver RNA- and skeletal muscle RNA-selective analytes) could be used in the ROC-plot analyses.
Labeled target was prepared from total RNA samples using one of the following three reagent kits: (1) 3' IVT Express kit (Affymetrix Part No. 901229), (2) IVT kit (Affymetrix Part No. 900449), or (3) Ovation RNA Amplification System V2 and the FL-Ovation cDNA Biotin module kits (Nugen Catalog Nos. 3100 and 4200). Labeled target was hybridized to Affymetrix Human Genome U133A 2.0 arrays in a GeneChip Hybridization Oven 640. The arrays were washed on an Affymetrix GeneChip Fluidics Station 450 using fluidics protocol FS450-002 and scanned using an Affymetrix GeneChip Scanner 3000 7G. The variations made in protocols for each target labeling method are listed in Table 3. For each method, a pair of MTRC was run in triplicate but processed on different days to create sets of technical replicates. The microarray data from this study are available in the ArrayExpress Archive at the European Bioinformatics Institute through accession number E-TABM-1091 http://www.ebi.ac.uk/arrayexpress.
Each dataset that was created using a different target labeling protocol or different batch of MTRC was processed separately with RMA. The selective analytes for the 1-to-1 component (UHRR for the MTRC-4 and HBRR for the MTRC-3) were used to normalize Mix2 with respect to Mix1. For each technical replicate, the difference in the 10% trimmed mean intensity between the Mix1 and Mix2 data for analytes in the 1-to-1 component was calculated and used to correct the Mix2 signal data.
For ROC-plot calculations using the MTRC, the true negatives are the tissue selective analytes for tissue RNA that is mixed in a 1-to-1 ratio between Mix1 and Mix2. Separate ROC-plots are generated for each true positive subset in the MTRC. For MTRC-4, the true positives are the 1.5-, 2-, and 4-fold changes and for the MTRC-3, the true positives are the 1.5-fold changes in both directions. For singlicate assays, analytes are ranked by log2 ratio . For replicate assays, analytes are ranked by p-values calculated using a paired t-test comparison of the three Mix1 and Mix2 signals. AUCs were calculated using the trapezoidal method. Statistical analyses of differences in AUCs were performed as previously described  using the method of Hanley and McNeil .
For each dataset, the normalized log2
signals were averaged across technical replicates of either Mix1 or Mix2. The average log2
) for Mix1 and Mix2 were then used to calculate the ratio
(R) and intensity
(I) for each analyte as follows:
For MTRC-4 datasets, the analytes (excluding the 1-to-1 component) were distributed into 19 bins of 42, with a single bin of 40 at the lowest intensity. For MTRC-3 datasets, the analytes (excluding the 1-to-1 component) were distributed into 20 bins of 28 analytes, omitting the 5 lowest intensity analytes. The standard deviation (s) ranged from 0.30 - 0.34 for MTRC-4 datasets and from 0.16 - 0.21 for MTRC-3 datasets.