Probe set algorithms: is there a rational best bet?
© Seo and Hoffman; licensee BioMed Central Ltd. 2006
Received: 13 April 2006
Accepted: 30 August 2006
Published: 30 August 2006
Affymetrix microarrays have become a standard experimental platform for studies of mRNA expression profiling. Their success is due, in part, to the multiple oligonucleotide features (probes) against each transcript (probe set). This multiple testing allows for more robust background assessments and gene expression measures, and has permitted the development of many computational methods to translate image data into a single normalized "signal" for mRNA transcript abundance. There are now many probe set algorithms that have been developed, with a gradual movement away from chip-by-chip methods (MAS5), to project-based model-fitting methods (dCHIP, RMA, others). Data interpretation is often profoundly changed by choice of algorithm, with disoriented biologists questioning what the "accurate" interpretation of their experiment is. Here, we summarize the debate concerning probe set algorithms. We provide examples of how changes in mismatch weight, normalizations, and construction of expression ratios each dramatically change data interpretation. All interpretations can be considered as computationally appropriate, but with varying biological credibility. We also illustrate the performance of two new hybrid algorithms (PLIER, GC-RMA) relative to more traditional algorithms (dCHIP, MAS5, Probe Profiler PCA, RMA) using an interactive power analysis tool. PLIER appears superior to other algorithms in avoiding false positives with poorly performing probe sets. Based on our interpretation of the literature, and examples presented here, we suggest that the variability in performance of probe set algorithms is more dependent upon assumptions regarding "background", than on calculations of "signal". We argue that "background" is an enormously complex variable that can only be vaguely quantified, and thus the "best" probe set algorithm will vary from project to project.
Mini-intro to the Affymetrix array
Current Affymetrix microarrays  use chemical synthesis of 25 mer oligonucleotides in 8 μm2 features on glass slides, resulting in a density of approximately 1 million oligonucleotides per 1.3 cm2 microarray. The impressive density of oligonucleotides provides adequate space on the chip ("real estate") for use of multiple probes per mRNA transcript. Thus, each transcript is queried by a probe set (typically 22 oligonucleotides on current arrays). The assessment of the amount of a specific mRNA in a solution from a tissue or cell is determined by the amount of hybridization to the different oligo features for that mRNA. While "hybridization" is a highly sensitive and specific assay, it is not perfectly sensitive or perfectly specific. Thus, it is critical to provide measures of non-specific hybridization and probe binding (background), and one method is to print mismatch probes, where a single base mutation in the center of each 25 mer probe should disrupt most "specific" hybridization. Current expression profiling microarrays use 11 perfect match and mismatch probe pairs for each probe set, although there may be many probe sets per transcript unit.
When the same data set is then normalized to time 0, the Y axis becomes a measure of "fold-change" between baseline (time 0), and peak expression (day 3) (Figure 2). As the graph now becomes a ratio of day 3 (numerator) and day 0 (denominator), the very different values for day 0 expression lead to very different ratios for the different probe set algorithms (Figure 2). In the example shown, the PCA probe set algorithm calculates a 90-fold increase in expression of myogenin (PCA), while the perfect match only probe set algorithms show only a 2-fold increase (RMA, dCHIP PMonly) (Figure 2). The volatility of the fold-change is a consequence of different calculations of the denominator for a ratio, and this is due to different assumptions concerning what is "background", and what is true "signal". If one takes this single example and extrapolates to all 40,000 probe sets on a microarray, one quickly understands why different probe set algorithms give very different interpretations of the same data.
Comparison studies of probe set algorithms
There has been growing interest on the effect of probe set algorithms on data interpretation, with development of new probe set algorithms, and also comparisons between probe set algorithms (see  and  for reviews). A number of studies have found the lower variance of the perfect-match only algorithms, and other methods of reducing variance, to lead to more precision, with the conclusion that these provide the best performance [2, 8–12]. However, one recent paper studying the issues of effects of normalizations methods on variance and data interpretation noted, "Any normalization reduces variance and increases sensitivity with no clear winner" . Other authors have similarly cautioned against equating increased precision with improved performance, as precision may come at the cost of accuracy. The dramatic differences in fold-change interpretation of myogenin by different probe set algorithms shown in Figure 2 illustrate this point. The signal of the mismatch probes for myogenin is relatively high at time 0, but this mismatch signal does not rise proportionately with the increased myogenin perfect match signal seen at day 3. Thus, the greater precision of RMA and dCHIP PMonly comes at a cost of underestimating the fold-change at day 3. In other words, the balance of precision and accuracy would fall in favor of the greater accuracy of PCA, dCHIP diff, and MAS5.0 in this specific instance. Thus, appropriate assessments of "performance" may depend on a balance of precision and accuracy, and that this is likely project and probe set dependent [3, 14, 15].
An excellent R package to compare the performance of different probe set algorithms (Affycomp II)  has been used to compare over 50 probe set algorithms in a paper describing the effects on accuracy, precision, and sensitivity into low signals . Consistent with expectations, the authors found that background correction (e.g. using mismatch signal) improved accuracy but worsened precision. They proposed that GC-RMA, PLIER, ZL, and RSVD algorithms provided the best balance of accuracy and precision . This interpretation was based upon a spike-in data set, where a limited number of synthesized mRNAs were studied at varying concentrations.
We recently took a different approach, where three different large data sets with different biological variance intrinsic to each organism and experimental system (grass, rat, and human biopsy samples) were studied. We built an interactive analysis tool to query the statistical power of each probe set on the arrays in the project, with dynamic user-definition of sample size, effect size (fold-change), and statistical significance . We found that the same algorithms that were found to provide excellent accuracy using spike in data , tended to reduce variance to inappropriate levels, leading to high false positives . A recent paper found similar results where the majority of RMA-specific probe sets were unable to be validated by RT-PCR . A generalization from all of these studies is that probe set algorithms that utilize only the perfect match probes give much better sensitivity with less variance (enhanced accuracy), particularly at low signal levels. However, these same low signals are particularly susceptible to strong effects from cross-hybridization from related mRNA sequences (non-specific hybridization), and other sources of undesired background signal. It should be kept in mind that in "real experiment" (e.g. not spike in data), the levels of both the desired (specific) transcript are changing, as well as the levels of the cross-hybridizing (non-specific) transcripts. This serves to make the system highly (if not impossibly) complex, and the ideal balance of precision and accuracy could be project and organism specific. Thus, all probe set algorithms should be viewed as a necessary but imperfect compromise in data interpretation.
Is there a best bet?
One could envision that the ideal probe set algorithm would optimize background correction and normalization for each project and each probe set in that project. While such approaches are doomed to be "data starved", recently, new "hybrid" probe set algorithms have emerged that attempt to better combine the precision of model based approaches, with more judicious use of background correction (mismatch penalty). These include the Affymetrix PLIER algorithm  and GC-RMA . The new hybrid algorithms have the potential to achieve a more ideal balance of accuracy and precision. To test this, we took a novel approach; namely statistical power analysis of "absent call only" probe sets in a microarray project. To explain this approach, the original MAS5.0 probe set algorithm looks at the signals from the 11 probe pairs in a probe set, and determines if the signals from perfect match probes exceed the signals from the paired mismatch probes. The statistical confidence that the ensembled perfect match signals from a probe set are significantly above background is derived as a "detection p value". If, on average, the perfect match probes show greater signal intensity than the corresponding mismatch, then the detection p value improves, to a threshold to where a "present call" is assigned (e.g. the target transcript is likely "present" in the sample tested). Whereas poor signal from the perfect match, and increasing signal from the mismatch, causes the detection p value to worsen, leading to an "absent call" determination. The "present call" is thus a statistically derived threshold reflecting a relative confidence that the desired mRNA is indeed present in the RNA sample being tested at a level significantly above background hybridization. An "absent call" suggests that the target mRNA is not present in the sample, or that there is non-target mRNAs binding to the probe set (non-desired cross-hybridization), or both. Thus, limiting data analyses to "present calls" can restrict analyses to the better performing probe sets and target mRNAs . dCHIP algorithms also have an option to utilize the MAS5.0 "present calls" as a form of noise filter, and we have recently shown that use of the detection p value as a weighting function improves the performance of all probe set algorithms .
We reasoned that use of only those probe sets that showed 100% absent calls by MAS5.0 algorithms would enrich for poor quality signals with considerable noise components, due to poorly performing probe sets, low level or absent target mRNAs, or cross-hybridization. By limiting our statistical power analysis to these 100% absent call probe sets, we hypothesized that this would lead to a sensitive test for performance of probe set algorithms with regards to balance of precision and accuracy, and further, an assessment of the potential for false positives.
Another method of illustrating this point is to analyze sample mixing data sets. We used a publicly available dataset of two different human tissues (PBMC and placenta) mixed in defined ratios (100:0, 95:5, 75:25, 50:50, 25:75, 0:100) . To enrich for signals near detection threshold, we selected all probe sets that were "present calls" in all 100% PBMC profiles, but absent calls in 75%, 50%, 25%, and 0% PBMC profiles. We then took all probe sets appropriately powered at n = 2 for PLIER and GC-RMA, and studies the ability of the two probe set algorithms to track dilutions of these low level PBMC-specific probe sets. Correlation analysis between PBMC RNA concentration and expression levels of the selected probe sets showed a statistically significant difference (F-Test p value = 1.43449E-05). As expected, GC-RMA (average R2 = 0.244803) was considerably less accurate than PLIER (R2 = 0.324908).
Efforts to use a single "best" probe set algorithm for all probe sets and projects may be an unattainable goal. Instead, selection of analysis methods for any particular project should be viewed as striving towards an appropriate balance of precision and accuracy. Emerging hybrid probe set algorithms, such as PLIER, may prove particularly powerful in reducing false positives. New tools such as Affycomp and HCE-power can be used to assess this balance.
- Irizarry RA, Wu Z, Jaffee HA: Comparison of Affymetrix GeneChip expression measures. Bioinformatics 22(7):789–94. 2006 Apr 1 2006 Apr 1 10.1093/bioinformatics/btk046View ArticlePubMedGoogle Scholar
- Seo J, Gordish-Dressman H, Hoffman EP: An interactive power analysis tool for microarray hypothesis testing and generation. Bioinformatics 22(7):808–14. 2006 Apr 1 2006 Apr 1 10.1093/bioinformatics/btk052View ArticlePubMedGoogle Scholar
- Li C, Wong WH: Model-based analysis of oligonucleotide arrays: expression index computation and outlier detection. Proc Natl Acad Sci USA 98(1):31–6. 2001 Jan 2 2001 Jan 2 10.1073/pnas.011404098PubMed CentralView ArticlePubMedGoogle Scholar
- Probe Profiler Software[http://www.corimbia.com/Pages/ProbeProfiler.htm]
- Irizarry RA, Bolstad BM, Collin F, Cope LM, Hobbs B, Speed TP: Summaries of Affymetrix GeneChip probe level data. Nucleic Acids Res 31(4):e15. 2003 Feb 15 2003 Feb 15 10.1093/nar/gng015PubMed CentralView ArticlePubMedGoogle Scholar
- Li C, Hung Wong W: Model-based analysis of oligonucleotide arrays: model validation, design issues and standard error application. Genome Biol 2001, 2(8):RESEARCH0032.PubMed CentralPubMedGoogle Scholar
- Millenaar FF, Okyere J, May ST, van Zanten M, Voesenek LA, Peeters AJ: How to decide? Different methods of calculating gene expression from short oligonucleotide array data will give different results. BMC Bioinformatics 7(1):137. 2006 Mar 15 2006 Mar 15 10.1186/1471-2105-7-137PubMed CentralView ArticlePubMedGoogle Scholar
- Wu Z, Irizarry RA: Stochastic models inspired by hybridization theory for short oligonucleotide arrays. J Comput Biol 2005, 12(6):882–93. 10.1089/cmb.2005.12.882View ArticlePubMedGoogle Scholar
- Bolstad BM, Collin F, Simpson KM, Irizarry RA, Speed TP: Experimental design and low-level analysis of microarray data. Int Rev Neurobiol 2004, 60: 25–58.View ArticlePubMedGoogle Scholar
- Irizarry RA, Hobbs B, Collin F, Beazer-Barclay YD, Antonellis KJ, Scherf U, Speed TP: Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics 2003, 4(2):249–64. 10.1093/biostatistics/4.2.249View ArticlePubMedGoogle Scholar
- Bolstad BM, Irizarry RA, Astrand M, Speed TP: A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 19(2):185–93. 2003 Jan 22 2003 Jan 22 10.1093/bioinformatics/19.2.185View ArticlePubMedGoogle Scholar
- Freudenberg J, Boriss H, Hasenclever D: Comparison of preprocessing procedures for oligo-nucleotide micro-arrays by parametric bootstrap simulation of spike-in experiments. Methods Inf Med 2004, 43(5):434–8.PubMedGoogle Scholar
- Harr B, Schlotterer C: Comparison of algorithms for the analysis of Affymetrix microarray data as evaluated by co-expression of genes in known operons. Nucleic Acids Res 34(2):e8. 2006 Jan 23 2006 Jan 23 10.1093/nar/gnj010PubMed CentralView ArticlePubMedGoogle Scholar
- Huang S, Yeo AA, Gelbert L, Lin X, Nisenbaum L, Bemis KG: At what scale should microarray data be analyzed? Am J Pharmacogenomics 2004, 4(2):129–39. 10.2165/00129785-200404020-00007View ArticlePubMedGoogle Scholar
- Affycomp II[http://affycomp.biostat.jhsph.edu/]
- Guide to Probe Logarithmic Intensity Error (PLIER) Estimation[http://www.affymetrix.com/support/technical/technotes/plier_technote.pdf]
- McClintick JN, Edenberg HJ: Effects of filtering by Present call on analysis of microarray experiments. BMC Bioinformatics 7: 49. 2006 Jan 31 2006 Jan 31 10.1186/1471-2105-7-49PubMed CentralView ArticlePubMedGoogle Scholar
- Seo J, Bakay M, Chen YW, Hilmer S, Shneiderman B, Hoffman EP: Interactively optimizing signal-to-noise ratios in expression profiling: project-specific algorithm selection and detection p-value weighting in Affymetrix microarrays. Bioinformatics 20(16):2534–44. 2004 Nov 1 2004 Nov 1 10.1093/bioinformatics/bth280View ArticlePubMedGoogle Scholar
- HCE-power: An Interactive Power Analysis Tool for Microarray Hypothesis Testing and Generation[http://www.cnmcresearch.org/bioinformatics/power/power.html]
- GEO Accession Viewer[http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE3077]
- Inoue M, Nishimura S, Hori G, Nakahara H, Saito M, Yoshihara Y, Amari S: Improved parameter estimation for variance-stabilizing transformation of gene-expression microarray data. J Bioinform Comput Biol 2004, 2(4):669–79. 10.1142/S0219720004000806View ArticlePubMedGoogle Scholar
- Durbin BP, Rocke DM: Variance-stabilizing transformations for two-color microarrays. Bioinformatics 20(5):660–7. 2004 Mar 22 2004 Mar 22 10.1093/bioinformatics/btg464View ArticlePubMedGoogle Scholar
- Rocke DM, Durbin B: Approximate variance-stabilizing transformations for gene-expression microarray data. Bioinformatics 19(8):966–72. 2003 May 22 2003 May 22 10.1093/bioinformatics/btg107View ArticlePubMedGoogle Scholar
- Zhao P, Hoffman EP: Embryonic myogenesis pathways in muscle regeneration. Dev Dyn 2004, 229(2):380–92. 10.1002/dvdy.10457View ArticlePubMedGoogle Scholar
- PEPR:Public Expression Profiling Resource[http://pepr.cnmcresearch.org/]
- Bakay M, Wang Z, Melcon G, Schiltz L, Xuan J, Zhao P, Sartorelli V, Seo J, Pegoraro E, Angelini C, Shneiderman B, Escolar D, Chen YW, Winokur ST, Pachman LM, Fan C, Mandler R, Nevo Y, Gordon E, Zhu Y, Dong Y, Wang Y, Hoffman EP: Nuclear envelope dystrophies show a transcriptional fingerprint suggesting disruption of Rb-MyoD pathways in muscle regeneration. Brain 2006, 129(Pt 4):996–1013. 10.1093/brain/awl023View ArticlePubMedGoogle Scholar
- The Erin Godla Research Fund for Neuromuscular Diseases[http://www.juvenileals.org]
- National Center for Medical Rehabilitation Research[http://www.ncmrr.org]
- Wellstone Muscular Dystrophy Center – Washington, DC[http://www.wellstone-dc.org]
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.