The utility of MAS5 expression summary and detection call algorithms

Abstract

Background: Used alone, the MAS5.0 algorithm for generating expression summaries has been criticized for high False Positive rates resulting from exaggerated variance at low intensities.

Results: Here we show, with replicated cell line data, that, when used alongside detection calls, MAS5 can be both selective and sensitive. A set of differentially expressed transcripts were identified that were found to be changing by MAS5, but unchanging by RMA and GCRMA. Subsequent analysis by real time PCR confirmed these changes. In addition, with the Latin square datasets often used to assess expression summary algorithms, filtered MAS5.0 was found to have performance approaching that of its peers.

Conclusion: When used alongside detection calls, MAS5 is a sensitive and selective algorithm for identifying differentially expressed genes.

Figure 1

A comparison of fold changes found by RMA and MAS5 for MCF7 and MCF10a cell line data. Absent-flagged probesets show significant disagreement between the two algorithms. First panel: all data. Remaining panels are separated by detection calls. AA: probesets flagged Absent by MAS5 detection call. PA: probesets flagged Absent in one cell line, Present in the other. PP: probesets flagged Present in both cell lines. Lines represent 2-fold thresholds (data are on a log2 scale). Details of Present/Absent flagging can be found in Methods.

Figure 2

MA plots of the MAS5 and RMA processed data showing probesets selected for real time PCR. Light grey points all data. Circles: PA probesets. Diagonal Crosses: AA-RMA probesets. Triangles: AA probesets. Vertical Crosses: PP probesets. RMA reports the fold change for the PA, AA and AA-RMA probesets as low in comparison to MAS5. AA and PA probesets are of low intensity. Additional MA plots, stratified by detection call can be found in the supplementary data.

Figure 3

Comparison between fold changes from MAS5 and RMA and those found using real time PCR (rt1, rt2, rt3). PA: Present-Absent probesets. PP: Present-Present probesets. AA: Absent-Absent probesets. AA-RMA: Absent-Absent probesets with fold changes > 2 according to RMA. Data are on a log2 scale.

Figure 4

MA plots for the Latin square data, generated by affycomp [4]. Points represent probesets targeting transcripts not expected to change, numbers represent probesets targeting transcripts spiked into the dataset at different concentration. The number represents the expected fold change. Fold change is represented by M on the y-axis. A: RMA processed data. B: Raw MAS5 processed data. C: filtered-MAS5 data. Raw MAS5 data suffers from a significant number of False Positives (i.e. probesets recording a differential expression greater than 2-fold), while RMA shows much better performance on these data. Filtering MAS5 data by detection call as described in Methods, significantly reduces the number of False Positives and brings the results much closer to those of RMA.