Comparison of signal measures computed by different array normalization procedures in control and degraded samples

Signal calculation for each array image was performed in MAS5.0 in three different ways. First, signals were calculated by omitting any normalization procedure, yielding unadjusted signal intensities (raw signals). Second, images were re-analyzed using the all probe sets scaling option (globally scaled signals). Third, signals were computed using the selected probe sets scaling option with the help of a mask file, which combines all the 27 available probe sets for the spiked-in control RNAs (spike mask scaled signal). As the external spike controls were present in equal amounts in all RNA samples, the latter scaling procedure ensures that the normalization step corrects for any unknown variables but not for systematic parameters linked to the 3'-bias in cRNA targets prepared from partially degraded RNA. Thus, signal intensities obtained by this method should reflect the genuine hybridization signals of the labeled cRNAs more precisely than the raw signals or the signals computed by global scaling.

Statistical analysis of the raw signal data revealed that mean (83%), trimmed mean (82%), and cumulative raw signals (83%) were significantly decreased in degraded RNAs (Table 10.1). Scale factors generated by scaling to the spike mask were highly similar in degraded and control samples, demonstrating that the spiked-in controls were indeed equally abundant in all samples. Consistent with the observations already made at the level of raw signals, resulting mean (81%), trimmed mean (80%), and cumulative signals (81%) after the spiked-in control probe-set scaling proved to be highly significantly decreased in degraded samples (p <0.001). A detailed analysis of the size distribution of the mean raw and spike-mask-scaled probe-set signals showed a general downshift to lower intensities in degraded samples (Figure 10.2A, B). In contrast, size distributions of globally scaled signals of control and degraded samples did not correctly reflect

(C) 180016001400-|> 1200-■§ 1000-•2" 800600400 200

(C) 180016001400-|> 1200-■§ 1000-•2" 800600400 200

Figure 10.2.

The effects of various array normalization methods on the distribution of resulting signal measures in control and partially cleaved RNA samples. (A) Unadjusted mean raw signal measures of control (n =6) and degraded samples (n =6) were logarithmically transformed (base 2) and sorted into bins of equal size (0.5). Increasing bin numbers represent higher signal bins. (B) Signal measures resulting from spike-mask scaling are represented as in (A). Signals from degraded samples show a general downshift due to the presence of partial, non-productive cRNA copies, which is also evident in the distribution of the raw signals. (C) Signal measures resulting from global scaling are shown as in (A). Note, that this procedure arithmetically compensates some of the signal drop observed in degraded samples. Thus, signal measures of degraded samples do not accurately reflect hybridization signals any more. (D) Spike-mask-scaled signals of 12 118 reliably measured probe sets (>50% present detection calls in the 12 arrays hybridized) were binned as shown in (C). The resulting distributions show that the peaks observed at lower signal intensities in (A)-(C) represent probe sets which received absent and/or marginal detection calls in the majority of analyses.

this signal drop any more (compare Figure 10.2B and C), as the difference is largely compensated by correspondingly increased scale factors (p =0.02). Because all arrays were taken from the same batch and very little intra-group variation was observed, raw as well as spike-mask-scaled data indicate a significant drop in signal intensities on arrays hybridized with targets from partially cleaved RNA samples. Since all arrays were hybridized with equal amounts of labeled cRNA, we conclude that due to the partial cleavage mimicking RNA degradation and the 3'-biased amplification procedure, there is a considerably increased fraction of cRNA molecules present in targets prepared from the degraded starting material that does not generate hybridization signals. As indicated in Figure 10.1B, such non-productive or less productive molecules are generated during the cDNA and cRNA synthesis steps, when RNA cleavage occurred at a position located within or 3' of the sequences represented by the corresponding probe set.

0 0

Post a comment