To define a measure of expression representing the amount of the corresponding mRNA species it is necessary to summarize probe intensities for each probe set. Several model-based approaches to this problem have been proposed. We have developed an effective expression measure motivated by a log scale linear additive model. This summary statistic is referred to as the log scale robust multi-array analysis (RMA).

A recent paper (7) reported that variation of a specific probe across multiple arrays could be considerably smaller than the variance across probes within a probe set. In the log₂ scale, the between-array standard deviation (SD) is in general five times smaller than the within-probe set SD (4,7). To account for this strong probe affinity effect, a multiplicative model, PM_ij – MM_ij = θ_ij + ε_ij, i = 1,…, I, j = 1,…, J, was proposed (7). The probe affinity effect is represented by _j. For analyses where multiple arrays are available a model-based expression index is defined as the maximum likelihood estimate (under the assumption that the errors follow a normal distribution) of the expression parameters θ_i. This estimate will depend on the probe affinity effects _j, which we can estimate if we have enough arrays. The software package dChip (http://www.biostat.harvard.edu/complab/dchip/) can be used to fit this model and obtain what we refer to as the dChip expression measure. Outlier probe intensities are removed as part of the estimation procedure (7).

Using data from a spike-in experiment (described in more detail below) we found that appropriately removing background and normalizing probe level data across arrays results in an improved expression measure motivated by a log scale linear additive model. The model can be written as T(PM_ij) = e_i + a_j + ε_ij, i = 1,…, I, j = 1,…, J, where T represents the transformation that background corrects, normalizes, and logs the PM intensities, e_i represents the log₂ scale expression value found on arrays i = 1,…, I, a_j represents the log scale affinity effects for probes j = 1,…, J, and ε_ij represents error as above. Notice that this is an additive model for the log transform of (background corrected, normalized) PM intensities. It is quite different from the additive model in PM – MM that was found unsatisfactory in Li and Wong (7), most likely because of the very strong mean variance dependence that would be present in such an additive model. A robust linear fitting procedure, such as median polish (8), was used to estimate the log scale expression values e_i. The resulting summary statistic is referred to as RMA. The normalization and background correction procedures used are reported elsewhere (4,9). Recent results (4,10) suggest that subtracting MM as a way of correcting for non-specific binding is not always appropriate. It is possible that information about non-specific binding is contained in the MM values, but empirical results demonstrate that mathematical subtraction does not translate to biological subtraction. We have found that, until a better solution is proposed, simply ignoring these values is preferable.

There is no gold standard to compare and test summaries of probe level data. For this reason, data from spike-in experiments have been used to assess the technology and to motivate normalization procedures (1,11,12). In a recent paper (13) a dilution/mixture experiment was used to compare existing expression measures. In a similar way, we used data from spike-in and dilution experiments to conduct various assessments on the MAS 5.0, dChip and RMA expression measures. Specifically, we compare the measures of expression according to three criteria: (i) the precision of the measures of expression, as estimated by standard deviations across replicate chips; (ii) the consistency of fold change estimates based on widely differing concentrations of target mRNA hybridized to the chip; (iii) the specificity and sensitivity of the measures’ ability to detect differential expression, presented in terms of receiver operating characteristic (ROC) curves.

For the dilution study (http://qolotus02.genelogic.com/datasets.nsf/), two sources of cRNA, human liver tissue and a central nervous system cell line (CNS), were hybridized to human arrays (HG-U95A) in a range of dilutions and proportions (4). We studied data from six groups of arrays that had hybridized liver and CNS cRNA at concentrations of 1.25, 2.5, 5.0, 7.5, 10.0 and 20.0 µg total cRNA. Five replicate arrays were available for each generated cRNA (n = 60 total).

For the spike-in studies, different cRNA fragments were added to the hybridization mixture of the arrays at different pM concentrations. The cRNAs were spiked-in at a different concentration on each array (apart from replicates) arranged in a cyclic Latin square design with each concentration appearing once in each row and column. All arrays had a common background cRNA. We used data from two different studies, one from Affymetrix (http://www.affymetrix.com/analysis/download_center2.affx) where 14 human genes were spiked-in at concentrations ranging from 0 to 1024 pM and one from GeneLogic (http://qolotus02.genelogic.com/datasets.nsf/) where 11 control cRNA fragments were spiked-in at concentrations ranging from 0 to 100 pM.

The GeneLogic spike-in experiment consists of a number of arrays each hybridized to samples with suitable concentrations of 11 different cRNA fragments added to a hybridization mixture consisting of cRNA from the same AML tissue. The 11 control cRNAs were BioB-5, BioB-M, BioB-3, BioC-5, BioC-3 and BioDn-5 (all Escherichia coli), CreX-5 and CreX-3 (phage P1), and DapX-5, DapX-M and DapX-3 (a Bacillus subtilis gene) (11,14,15). The cRNA were chosen to match the target sequence for each of the Affymetrix control probe sets. For example, for DapX (a B.subtilis gene), the 5′, middle and 3′ target sequences (identified by DapX-5, DapX-M and DapX-3) were each synthesized separately and spiked-in at a specific concentration. Thus, for example, DapX-3 target sequence may be added to the total hybridization solution of 200 µl to give a final concentration of 0.5 pM. The 11 control cRNAs were spiked-in at a different concentration on each array (apart from replicates). The 12 concentrations used were 0.5, 1, 1.5, 2, 3, 5, 12.5, 25, 37.5, 50, 75 and 100 pM, and these were arranged in a 12 × 12 cyclic Latin square, with each concentration appearing once in each row and column. The 12 combinations of concentrations used on the arrays were taken from the first 11 entries of the 12 rows of this Latin square. Three replicated hybridizations were carried out for each combination of concentrations of the spiked-in material.

The Affymetrix spike-in experiment was done in a similar fashion. It consists of a series of human genes spiked-in at known concentrations. They represent a subset of the data used to develop and validate the MAS 5.0 algorithm. The Latin square consists of 14 spiked-in gene groups in 14 array groups. The concentration of the 14 groups in the first array group are 0, 0.25, 0.5, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 and 1024 pM. Each subsequent array group rotates the spike-in concentrations by one group, i.e. array group 2 begins with 0.25 pM and ends at 0 pM, on up to array group 14, which begins with 1024 pM and ends with 512 pM. There were three replicates for each concentration combination, except for two combinations for which 12 replicates were formed.

A basic application of the GeneChip technology is to study differences in gene expression between different RNA samples. Observed fold change in expression measures is used to assess differential expression (3). While the Affymetrix protocol calls for 15 µg of RNA, in practice the amount of target mRNA available for the hybridization reactions can differ greatly depending on the cells or tissue type under study. In some cases the available RNA will be amplified, and in others the hybridization will be carried out with <15 µg. It is desirable to have estimated fold changes in expression largely independent of the amount of target mRNA used. For an extreme example, suppose that one series of experiments is done with 20 µg of RNA in each hybridization, and another series is identical, but uses just 1.25 µg of RNA. Ideally, the answers should be very similar. For each gene we computed fold change estimates between the liver and CNS samples using the 10 arrays in the 1.25 µg concentration group for each of the three expression measures. We then computed estimates using the arrays in the 20 µg concentration group. Because fold change is a relative measure, estimates should be independent of the amount of RNA that is hybridized to the arrays. Log (base 2) fold change estimates of gene expression between liver and CNS samples computed from arrays hybridized to 1.25 µg of cRNA were plotted against the same estimates obtained from arrays hybridized to 20 µg for all three measures (Fig. (Fig.2).2). The correlation of fold change estimates from the different concentrations (Table 1) demonstrated that RMA and dChip provided more consistent estimates than MAS 5.0. RMA was slightly better than dChip. Using MAS 5.0 (Fig. (Fig.2A),2A), 1223 genes had at least a 2-fold discrepancy (shown with larger dots) between the two fold change estimates. For dChip there were 302 (Fig. (Fig.2B)2B) and for RMA (Fig. (Fig.2C)2C) there were only 22. This assessment demonstrated that RMA provides more consistent estimates of fold change.

A typical application of GeneChip technology is finding genes that are differentially expressed in different tissues. Successful fold change analysis will detect all and only genes that are differently expressed due to biological variation. Because in the spike-in experiments arrays were hybridized to the same background, successful differential expression analyses should identify only the spiked-in genes as being differentially expressed. The absence of a batch mode in MAS 5.0 and dChip made running comparisons for all pairs prohibitive due to time. We therefore chose 10 pairs of arrays at random from both Affymetrix and GeneLogic spike-in studies. For each of these pairs we computed estimates of fold change using the three expression measures. Then, for a large range of cut-off values we computed the number of false positives (non-spiked-in genes with fold change estimates larger than the cut-off) and the number of true positives (spiked-in genes with fold change estimates larger than the same cut-off). ROC curves were created by plotting the true positive rates (sensitivity) versus false positive rates (1 – specificity). The true positive rates were estimated for the filtering operation, Observed Fold Change > cut-off, for a large range of cut-off values, by calculating the proportion of genes spiked-in at different concentrations that satisfy the filtering criterion. False positive rates were calculated in a similar way by computing the proportion of non-spiked-in genes that satisfy the filtering criteria. Areas under ROC curves can be used to compare specificity and sensitivity of competing tests. The fact that the RMA curves dominated the dChip and MAS 5.0 curves demonstrated that the differential expression calls obtained with RMA have higher sensitivity and specificity than those obtained with the other two measures (Fig. (Fig.3A3A and B). The true fold changes resulting from our random choice of pairs ranged from 3/2 to 1024. The task of detecting fold changes much larger than 2 might be considered less important than that of reliably detecting changes 2-fold or less, so we chose 10 pairs where the true fold changes were exactly 2 and repeated the analysis. The superiority of RMA appears even greater in this assessment (Fig. (Fig.3C).3C). For comparisons of two arrays, Affymetrix software provides an alternative to fold change analysis based on the P value of a non-parametric test statistic (5). Test statistics can be created for RMA and dChip based on estimates of standard error obtained from probe level data (4,7). We repeated the above analysis for the test statistic and found the Affymetrix’s P value approach to work as well as the test statistic based on RMA and better than dChip’s version (Fig. (Fig.3D3D and E). However, Affymetrix’s P value analysis can only be used when comparing two arrays. We performed fold change analyses on two sets of 12 arrays with the same spiked-in concentrations and found RMA to have almost perfect sensitivity and specificity (Fig. (Fig.3F).3F). In this comparison, dChip performed almost as well as RMA and significantly better than MAS 5.0. This assessment demonstrated that using RMA provides higher specificity and sensitivity when using fold change analysis to detect differential expression.

To understand why fold change analysis using RMA has better sensitivity and specificity we looked at M_g = log₂(Y_g/X_g) versus A_g = log₂√X_gY_g = (logX_g + logY_g)/2, (MvA) plots for expressions X_g and Y_g from two arrays being compared for all genes, g = 1,…, G. Log scale scatter plots of Y_g versus X_g are commonly seen in the literature. MvA plots are 45° rotations of these scatter plots (18). We found MvA plots useful because log fold change (the quantity of most interest) is represented on the y-axis and average absolute log expression (another quantity of interest) on the x-axis. We selected one array from one of the Affymetrix spike-in experiments to use as a reference and then computed M_g and A_g for the comparisons of that array with all other arrays in the experiment using MAS 5.0 (Fig. (Fig.4A),4A), dChip (Fig. (Fig.4B)4B) and RMA (Fig. (Fig.4C).4C). In these plots, the colored numbers represent the log (base 2) fold change in concentrations of all 14 spiked-in genes. Each distinct fold change is represented with a different color as a visual aid. The -∞ and ∞ represent fold changes with a zero in the numerator or denominator, respectively. The red points represent non-spiked-in genes with a fold change larger than 2. Except for the colored numbers, including ∞, genes should have log fold changes of 0. The fact that using RMA resulted in plots with fewer red points demonstrated that its smaller variance, especially for genes with lower absolute expression (Fig. (Fig.4A–C)4A–C) resulted in better detection capability of genes spiked-in at different concentrations in the different arrays. Most of the genes having log fold changes of 2 when 0 was expected (red points in Fig. Fig.4A)4A) for MAS 5.0 were due to this large variance at the low end. Color box plots (Fig. (Fig.5)5) of fold change estimates demonstrated that RMA produces fold changes closer to 1 for genes that are not changing than those for MAS 5.0, with those for dChip being in between. In particular, the interquartile ranges of log₂ fold change for equivalently expressed genes were 0.92, 0.22 and 0.19 for MAS 5.0, dChip and RMA, respectively.