This column discusses types of
malperforming probes, illustrates their impact on identifying
differentially expressed genes and describes a simple method for handling
them when performing data analysis of Affymetrix microarrays.
| Jun
1, 2005 |
| By:
Wendell
Jones |
| Pharmaceutical
Discovery |
|
Wendell Jones
|
Several years ago, when commercial microarrays first were becoming widely
accepted as a discovery platform for understanding critical aspects of
biology at the molecular level, the actual content of the physical probes
on the arrays was not publicly known. Affymetrix (Santa Clara, California,
USA) was one of the first major microarray providers to fully disclose
probe content. Most recently, in April 2005 Agilent Technologies (Palo
Alto, California, USA) announced that they would be disclosing the
detailed probe sequence information of their microarrays. In parallel,
several articles have been published recently on cross-platform
comparisons (1–4) with positive results, and it is hoped that the
disclosure of the detailed microarray content by the major microarray
providers will improve efforts related to cross-platform reproducibility.
Knowledge of microarray probe content also enables more refined
analysis of a specific platform. Unlike many other microarray platforms,
Affymetrix GeneChips contain multiple distinct 25-mer oligonucleotide
probes per transcript. They are arranged in pairs of perfect match (PM)
probes that are complementary to the target sequence and mismatch probes
(MM) probes, which contain a single base-pair mismatch and act as negative
controls. One expression value is calculated for each transcript
associated with a particular probe set using the probe intensity values of
the PM and MM probes. The more recent Affymetrix designs, such as the
HG-U133 family, typically contain 11 PM and 11 MM probes per probe set. In
general, the multi-probe design provides a degree of robustness to
probe-specific effects and resistance to minor defects in the
hybridization image as these probes are scattered throughout the surface
of the chip.
Steve Casey
|
Because each probe in a probeset has a unique sequence and distinct
hybridization characteristics, not all probes in the same probeset
hybridize equally well to their intended target — a characteristic
termed probe affinity. In the differential expression context, the probe
affinity effect can be reduced to a great degree with appropriate
algorithms, especially ones that attempt to model probe affinity or remove
its effect directly. However, studies at Expression Analysis (Durham,
North Carolina, USA) and other institutions have discovered several
situations where certain probes in a given probeset perform so poorly that
they have a negative impact on the analysis of differential expression. We
term these probes as malperforming probes.
Malperforming Probes There are
several possible reasons for malperforming oligonucleotide probes,
including:
- the probe is based on poor quality sequence data from the respective
Unigene build (or some other source is used as a reference for the
synthesized sequence)
- the probe has undesirable hybridization properties, such as a strong
affinity to non-specific targets or a tendency to not hybridize to any
target, including its intended target
- the probe is not constructed to full length
- the probe or its target have a tendency to form secondary structures
that hinder hybridization
- the probe is 5' biased while Affymetrix chip architecture and
protocols are heavily 3' biased.
Gautier et al. (5) and Zhang et al. (6) have compared Affymetrix's
human U95 and U133 probe sequence with reference sequences from RefSeq, a
curated sequence database available from the national Center for
Biotechnology Information (NCBI, Bethesda, Maryland, USA) as well as other
sources. For example, Gautier et al., found that, in one experiment, only
61 out of 163 Affy probe sets that were significantly different between
conditions and that were associated with reference sequences still would
be intact (that is, composed of all 11 PM probes), given our current
knowledge of the transcripts. This implies that many probe sets have one
or more probes that are based on earlier, poorer-quality sequences. Since
the content of these chips was designed more than four years ago, this
finding is not completely surprising. When Expression Analysis examined
many of these types of sequences in detail, we frequently found that the
reference sequence used as a basis for the probeset had been updated
several times in recent years.
In addition to comparing sequence
information, Expression Analysis also has analyzed thousands of HG-U133A
hybridizations from a large variety of mRNA source materials (multiple
human tissues and cell lines) and found that approximately 70,000 of the
240,000+ PM probes on the HG-U133A array may be classified as
malperforming, primarily due to reasons a) and b) described earlier. The
signal intensity of malperforming probes typically did not vary or varied
only slightly under very different conditions when compared to their
within-probeset probe cohorts. Probe invariance generally obscures
differential expression when different transcript abundance levels are
present.
Figure 1. Sequence alignment for
reference sequence NM_000918 and the sequence basis for Affy
probeset 200654_at, where the actual probes are shown in red.
|
Examples
To illustrate malperforming probes with a concrete example, consider the
probeset denoted by Affymetrix as 200654_at, which is present on both the
HG-U133A and HG-U133 Plus 2.0 array. This probeset is annotated as P4HB,
procollagen-proline, 2-oxoglutarate 4-dioxygenase (proline 4-hydroxylase),
beta polypeptide (protein disulfide isomerase; thyroid hormone binding
protein). Figure 1 shows a sequence alignment of the Reference Sequence
for the P4HB (NM_000918) with the corresponding Unigene 133 sequence
(gb:J02783.1) on which the 200654_at Affymetrix probeset was originally
based. Probe locations are indicated by number and the NM_000918 sequence
is on the top. As Figure 1 indicates, probes 1–7 from this probeset show
excellent alignment with the reference sequence while probes 8–11 show
little or no alignment at all. A BLAST on the sequences from probes 8–11
did not match any known human mRNA sequence (although two did match a
mouse sequence).
Figure 2. A scatterplot matrix of
the 11 probes of the 200654_at the probe set over thousands of
hybridizations. Each graph inside the matrix is a scatterplot of
one probe's intensity values versus another probe's. Probes that
tend to agree would form a 45o line. Probes that disagree would
show a random pattern or a pattern other than a 45o line.
|
To validate further that probes 8–11
do not detect the same transcript as probes 1–7, consider the
scatterplot matrix of Figure 2. This scatterplot matrix illustrates the
correlation in log intensity for every probe pair within the 200654_at
probeset. The matrix shows individual scatterplots of each pairing of
probes within the 200654_at probeset for thousands of HG-U133A
hybridizations performed at Expression Analysis. Probes that generally
agree with each other would create a cloud of points that resembles a 45o
line. Probes that do not agree would show either a random pattern or a
pattern other than a 45o line. We would expect probes in the same probeset
to agree since they are targeting the same transcript, just different
sections. One can easily see from the scatterplot matrix that the first
seven probes show agreement, but the agreement changes dramatically when
we examine the bottom or right sides of the scatterplot matrix. The
correlation of the first seven probes with the last four often is zero, or
can even be negative. Cleary, using the last four probes with the other
seven will tend to obscure any differential expression that may be
indicated by an experiment for this probe set. The first seven probes will
have to prevail over the disinformation provided by the last four.
We have provided a detailed
illustration of only one probeset. Many probe sets have much better
overall behavior, but many also have much worse. Our analysis shows that
almost half of the HG-U133A probe sets have at least four malperforming
probes and that these malperforming probes often mask true differential
expression no matter what signal algorithm is used to summarize the
various probe intensities within a probe set.
Although malperforming probes may be
difficult to find or characterize, the solution is surprisingly simple:
ignore or mask the malperforming probes when computing probeset signal.
There are many ways of doing this within the multiple software packages
available for computing signal for Affymetrix chips, some being simpler
than others. For some, rebuilding a library file may be required. For
others, it is as simple as removing entries in a table. Surprisingly,
malperforming probes may provide utility when it performs chip
normalization, as these probes tend to be relatively invariant — a
desired attribute during normalization.
To understand the impact of these
malperforming probes on detecting differentially expressed genes, a
particular study conducted by Expression Analysis on human subjects
uncovered 110 transcripts that were considered statistically significant
using the robust multi-array average (RMA) for signal computation and
using SAM significance analysis of microarrays (SAM) to assess
significance and estimate false positives (which were estimated to have at
least a two-fold difference between experimental groups). This type of
analysis is fairly standard when comparing two classes of samples with
Affymetrix arrays. When malperforming probes were removed and the same
analysis steps were repeated with all other factors the same, our list
grew to 251 transcripts. Moreover, all 110 transcripts in our original
list were contained in the new list of 251. This implies a large number of
transcripts had their estimated fold change compressed below two-fold by
malperforming probes. When we repeat the analysis with other signal
measures, such as MAS5 or dChip, we see a similar impact.
When examining a wider array of
experiments, we (7) have seen that malperforming probes could be obscuring
20%–75% of the differential expression that is detectable in an
experiment using techniques that address multiple testing and that use
thresholds for fold change and/or significance (i.e., techniques that
customarily are used to reduce noise and false positives).
Expression Analysis recently has
completed additional studies on malperforming probes for the rat and mouse
in addition to our studies of the most current human arrays. Not
surprisingly, the results for rat and mouse are very similar to human,
both in their nature and magnitude.
Conclusion
The good news is that microarrays, given sufficient and accurate
information, perform as expected; what, in essence, are random sequences
show relatively invariant hybridization intensities across a range of
tissues. On the other hand, accurate probe sequences within the same
probeset generally have intensities that positively correlate, as
expected. For this reason, we feel that, both for understanding your
experimental results, and when comparing your results with others,
microarray probe content should be made available to users. In addition,
when we analyze probe content and compare the content data with what is
currently known, we are much more sensitive and efficient in detecting
differential expression.
References
1. J.E. Larkin, B.C. Frank, J. Quackenbush et al., Nature Methods 2,
337–344 (2005).
2. B.H. Mechanm, G.T. Klus, Z.
Szallasi et al., Nucleic Acids Res. 32, 9e74 (2004).
3. R.A. Irizarry, D. Warren, W. Yu et
al., Nature Methods 2, 345–350 (2005).
4. J. Parker, Pharmaceutical
Discovery 5(2), 22–26 (2005).
5. L. Gautier, M. Moller, L. Friis-Hansen
and S. Knudsen, BMC Bioinformatics 5, 111 (2004).
6. J. Zhang, R.P. Finney, H. Buetow
et al., Genomics 85, 297–308 (2005).
7. W.J. Jones, Identification of
malperforming short oligonucleotide probes and their quantitative impact
on Affymetrix expression algorithms and differential expression
experiments. In preparation.
Wendell Jones is senior
research statistician and Steve Casey* is the founder and COO for
Expression Analysis Inc., a provider of regulatory compliant genomic
processing services. Steve Casey can be reached at scasey@expressionanalysis.com
.
*To whom all correspondence should be
addressed.
|
