Guide to PLIER Estimation

A F F Y M E T R I X P R O D U C T F A M I LY >

I N F O R M AT I C S >

Technical Note

This guide is a reference tool for biologists utilizing the new probe logarithmic intensity error (PLIER) method for
calculating signal. Additional detail
regarding the algorithm appropriate
for statisticians and bioinformaticists
is provided in the appendix.

Guide to Probe Logarithmic Intensity Error

(PLIER) Estimation

Introduction
The probe logarithmic intensity error
(PLIER) method produces an improved
signal (a summary value for a probe set) by
accounting for experimentally observed
patterns for feature behavior and handling
error appropriately at low and high abundance. Resulting benefits include:
Higher reproducibility of signal (lower
coefficient of variation) without loss of
accuracy
Higher sensitivity to changes in
abundance for targets near background
Dynamic weighting of the most informative probes in an experiment to
determine signal
This method was developed by building
upon many of the concepts that have been
published recently within the field of
GeneChip microarray data analysis,

including model-based expression analysis

and robust multichip analysis. It also
builds upon the summarization algorithm
provided in Affymetrix Microarray Suite
5.0 (MAS 5) by taking into account the
experimentally validated value of weighting
feature intensities to determine an overall
probe set summary.
Similar to other model-based approaches,
PLIER accounts for the systematic differences in intensity between features by
including parameters describing these
differences. These parameters are termed
feature responses,1 and one such parameter is included in the model for each feature
(or pair of features, when subtracting
Mismatch (MM) intensities). Feature
responses represent the relative differences
in intensity between features hybridizing
to a common target (Figure 1).

Figure 1: A probe set containing probes with varying feature response is helpful in detecting a
range of abundance, as illustrated by spiked-in concentrations. For example, strongly responsive
features are informative at the low end, but saturate at the high end. Weakly responsive features
are uninformative at the low end, but are informative at the high end of the abundance range.
15
14

Three features with different feature responses and backgrounds

(HGU95 Latin Square, spiked-in clone to 36311_at probe set)

Least responsive feature

13
Typically responsive feature

log(PM) intensity

II
II

12
Most responsive feature

11
10
9
8
7
6
5
0
0.25 0.5
1
2
4
8
16
32
64 128 256 512 1024
Concentration of PM labeled spikes (Probe set 36311_at in Latin Square experiment)

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

Figure 2: Once intensities are scaled using feature response, it becomes easier to detect
non-systematic differences among features, and down-weight low performance features. Only the
rescaled Perfect Match probes intensity is plotted here, and the actual algorithm incorporates
background information. Note that background plays a large effect in the least responsive feature,
and that the most responsive feature saturates first.
Same three features adjusted for responsiveness

12
10
8
Least responsive feature

6
Typically responsive feature

4
Most responsive feature

2
0

0
0.25 0.5
1
2
4
8
16
32
64
128 256 512 1024
Concentration of PM labeled spikes (Probe set 36311_at in Latin Square experiment)

PLIER produces a more accurate probe set

signal by utilizing these feature responses
to interpret intensity data, dynamic
weighting by empirical feature performance, and handling error appropriately
across low and high target abundance.
Feature responses are calculated using
experimental data across multiple arrays.
PLIER also uses an error model that
assumes error is proportional to the
observed intensity, rather than to the background-subtracted intensity. This ensures
that the error model can adjust appropriately for relatively low and high abundance
of target nucleic acids.
1 In Irizarry et al, this term is called the "probe affinity",

by analogy to target binding. This is not physically

accurate, due to the many factors which interact to
produce measured intensity.

NOTE: Refer to Table 2 for a comparison of

PLIER to other analysis methods.

Calculation of Signal
PROBE BEHAVIOR AND WEIGHTING

The PLIER algorithm utilizes experimental

data generated across multiple arrays in
order to identify and account for observed
patterns in feature intensities. Feature
responses can be fit from experimental data,
which allows good calibration of performance in different ranges of abundance, as

II
II

well as making it easy to identify poorly performing features with erratic hybridization
behavior. Feature responses are a measure of
how much the relative intensity of a feature
is due to the feature itself, as opposed to the
common target of a probe set. This relative
response is influenced by how likely a probe
is to bind to a complementary sequence

Figure 3: The information from all three feature pairs is combined to generate the signal estimate. Each feature pair votes for a particular target response value (lines 1-3). Each possible
value for the target response (plotted along the x-axis) has a penalty (lack of fit) when compared
to the intensity of a given feature pair. The three individual penalties are added up (line 4), and
the PLIER estimate of signal is the target response value with the lowest penalty (best fit to
data). Note that the feature with the lowest feature response is only useful in this instance for
excluding high signal values (because the feature reponse is so low that at a 4 pM spiked-in
concentration the perfect match intensity is mostly background). The remaining two features
estimate slightly different values individually due to noise. Note that the x-axis (possible target
response values) is logarithmically scaled, so that 5 is 25.
"Lack of fit" of target response to observed intensities

log (PM) intensity adjusted for

feature response

14

across a range of abundance, as all probes

have different thermodynamic properties and
binding efficiencies, but also is influenced by
many other factors, such as non-equilibrium
washes, labeling, and density of synthesis.
Feature response is therefore an empirical
factor local to each probe set, not a global
measure of probe binding. Note that it is not
necessary to know the sequence of the probes
within a feature to fit feature responses.
By using a scaling factor (feature response)
to account for this difference in intensities,
the intensity of all of the features within a
probe set can be easily compared (Figure 2).
For example, if one feature (or feature pair) is
consistently twice as bright as others in the
set, accounting for this enables the intensity
data of that probe to be analyzed consistently with the others for their response to the
common target. In the case of a probe set,
this enables all set members to be compared
and combined accurately.
Once the systematic differences between
features have been accounted for in this
way, it becomes easier to detect non-systematic differences. Features can then be
classified as high or low performance fea-

-1

0.09
0.08

Lack of fit for the experiment with spiked-in concentration

4 PM for the clone to probe set 36311_at plotted for three
feature pairs (adjusted for feature response)
[least responsive feature only useful here to
exclude strong target response]

0.07
0.06
0.05
0.04

Least responsive feature

Typically responsive feature

Most responsive feature

Overall lack of fit

(total of three features)

0.03
0.02
0.01
0
1

5
7
9
Possible values for target response

11

13

15

INFORMATICS

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

Intensity (scaled by feature response)

Figure 4: The amount of uncertainty (or potential error in estimation) present is dependent
upon the total intensity, not just the background-adjusted intensity. The proportion of intensity
represented by background is dependent upon the abundance of the complementary target as
well as the feature response. Thus, the small PM-MM difference in the first feature pair has an
error bar comparable to that of the large PM-MM difference in the second feature pair, and so
is much less informative than might be expected. Small differences between relatively large
numbers contain very little information, and therefore are effectively down-weighted by PLIER.
2000
1800
1600
1400

Error is proportional to the full intensity, not the background adjusted intensity
PM-MM
MM (Background)

1200
1000
800
600
400
200
0

Feature pairs

tures. For example, a feature that is consistently twice as bright as the median feature
intensity is considered a high performance
probe (and strongly responsive). However,
a feature that is twice as bright only half
the time but is the same as the median feature intensity the other half of the time is
considered a low performance feature due
to inconsistent behavior. This type of
observation is often due to cross-hybridization effects creating erratic patterns.
Once the scaling factor is applied, the
intensity data from each feature can be
combined to generate a goodness of fit curve
for possible signal values (target responses,
see glossary). In this process, the most
informative features provide the strongest
contribution to the signal. At the low end
of target abundance, where background
dominates the intensity, these are most likely
the probes with the highest feature response.
Inconsistent features are down-weighted,
using the Geman-McClure function.
Through the scaling and down-weighting
process, the features that are most informative in each experiment contribute the
most to the final signal estimate (Figure 3).
ERROR MODEL

The challenge with error estimation is

developing a model that fits both low and
high abundance because the error is directly
proportional to the largest component of

INFORMATICS

intensity. At high abundance, error is

approximately proportional to backgroundadjusted intensity, since most of the intensity
is due to the response to the specific target.
At the low end, however, error is approximately proportional to background intensity (which varies from feature to feature), as it
is the largest component of the observed
intensity (Figure 4). Due to this latter effect,
it is inaccurate to treat error in target
response as a proportion of backgroundadjusted intensity in all circumstances.
Therefore, the error model must be able to
estimate error in the target response depending on the abundance. If not, the amount of
error at the low end will be underestimated.
The PLIER error model smoothly transitions between the low end, where error is
dependent upon background, and the high
end, where error is dependent on the
response to the target.

Input Data
FEATURE INTENSITY

The algorithm requires feature-level intensity

data as the input. From these feature-level
data across many experiments, the algorithm calculates feature response terms.
Once generated, these terms may be stored
in a file, termed a model file, for repeated
use on different data sets that are sufficiently
similar to the reference set.

MODEL FILES

As explained above, a model file can be

generated to store feature responses (scaling
factors) once the feature responses have been
generated using selected data. The most
informative model files are those that have
been generated using a large experimental
data set across multiple relevant samples.
Even though a model file can technically be
generated from as few as two arrays, the
minimum recommended number of arrays
is five, although this is dependent upon the
specific experiment. Additional recommendations include having:
A broad range of abundance across
experiments for each target of interest
Comprehensive coverage across genes
of interest
These criteria provide the algorithm with
a reasonable amount of data to estimate both
feature and target response. The higher the
number of reliable input data points, the
higher the accuracy of the resulting signal
calculations. For example, informative
model files can be generated using data from
a control tissue panel that contains a wide
variety of data points that represent genes of
interest. If genes of interest are not well represented within the model file data set, the
algorithm cannot model feature response
accurately for these genes. If an experiment
uses highly different tissues, these can be
added to the tissue panel and a new model
file can be generated.
If a model file is being generated from a
small number of arrays or set of arrays with
poor or uncertain abundance or gene coverage, the addition of a Bayesian probe penalty is recommended. This term informs the
algorithm how heavily to weight the input
experimental data. For example, a maximum value for this term will instruct the
algorithm to ignore empirical data and treat
all probes as though they were equally
responsive. Use of this penalty helps mitigate the risk of treating anomalies, such as
scratches or cross-hybridizations, as accurate
data. The challenge with a small data set is
not having enough data to recognize these
types of problems as outliers.

II
II

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

NOTE: In the PLIER SDK, model file

input/output, normalization, and background
calculation are left for implementation wihtin
the hosted application.
DATA PRE-PROCESSING

A normalization procedure should be

applied either to intensity data before running the PLIER algorithm or a normalization procedure can be applied to the signal
after the signal is generated. One common,
powerful normalization procedure that operates on the raw intensity data is quantile
normalization. Quantile normalization assumes
a common distribution of intensities across
arrays. Because of this assumption, it is
designed for use across reasonably homogenous samples. As this is most often the case,
quantile normalization is the preferred
option. If samples are derived from significantly different tissues, quantile normalization should not be applied at this stage and
normalization should be performed on the
summary values after running PLIER.
PROBES AND BACKGROUND

Input probe types and background are the

next primary consideration. By adjusting
the input probe options and type of background calculated, the algorithm can be
tuned to target sensitivity to low expressors
or identification of small differential change.

Figure 5: Probe and background comparison

Sensitivity
to Low
Expressors

Detection

Step-by-Step Selection of
Options

PM-MM

PM-B

Small
Differential
Change
between
Experiments

PM only

MM treated as PM
Low

High

Sample-to-Sample Variability

However, this selection should be guided by

the expected sample-to-sample variability.
MM probes as background, a global uniform
background, or a GC-content-based background can be provided as the user requires.
Recommended options are discussed in
detail below.

This option is designed to handle high

sample-to-sample variability and maximize
sensitivity to low expressors by minimizing
bias. For example, a neurogenesis study
utilizing samples from different locations
investigating low-level early responses
would utilize the PM-MM option.

PM-MM

PM-B

As the most conservative approach, the

default option is the perfect match probe
feature intensity minus a corresponding
mismatch probe (PM-MM) feature intensity.
Since perfect match features and mismatch
features are similar in both location and
sequence, the difference between the intensities of a matched pair isolates target
response from many background effects.

The perfect match minus background

option (PM-B) is useful for moderate sample-to-sample variability, and moderate sensitivity to low expressors. For example, an
experiment utilizing different tissues from
the same lab assessing low to medium
expressors would leverage this option. For
this option, the background subtraction can
be calculated using any standard host pack-

Table 1: Summary Table

Method

Assumptions

Benefits

Drawbacks

Background effects are large and

potentially variable between features
across experiments relative to
effects of interest

Background effects minimized

due to low bias
Sensitivity to low expressors

Slightly noisier when signal is higher

than background

PM-B

Features have approximately the same

background

Low noise

May not represent all probe sets

accurately, typically leading to
underestimated differential change

PM Only

Background variation is insignificant

Low noise
Approximately constant CV

All probe sets biased

Compression of differential
change at the low end

MM treated
as additional
PM

Background variation is insignificant

Abundances moderate to large

Added statistical power

Low noise
Constant CV

All probe sets biased

Compression of differential change
at the low end

PM-MM

II
II

INFORMATICS

A
AFFFFY
YM
MEETTR
RIIX
X PPR
RO
OD
DU
UC
CTT FFA
AM
MIILY
LY >> A
AR
RR
RAY
AYS
S

Table 2: Other analysis methods

Method

Assumptions

Benefits

Drawbacks

PLIER

Multiple array analysis

Mixed error model
PM-MM, PM only etc.
Multiple background options
Smoothly handles intensities
below background

Higher reproducibility of signal

(lower coefficient of variation)
without loss of accuracy relative to
single array analysis
Higher differential sensitivity for low
expressors

Computationally intensive
In cases where feature intensities
disagree, may have more than one
solution
Performance relative to amount of
model data provided

Lack of bias

Variance not stable on log scale

Multiple array analysis

Arithmetic error model
PM only (stanardly)
Multiple background options
(no background typical)

Higher reproducibility of signal over

single array analysis
Good differential change detection
Variance stable on log scale with no
background

In cases where feature intensities

disagree, may have more than one
solution
Performance relative to amount of
model data provided
Positive bias at low end
(compression of Fold Change)

Multiple array analysis

Multiplicative error
PM only
Attenuated global background
(single global background used to
adjust for each intensity)

Higher reproducibility of signal over

single array analysis
Good differential change detection
Variance stable on log scale

In cases where feature intensities

disagree, may have more than one
solution (mitigated by median polish)
Performance relative to amount of
model data provided
Positive bias at low end
(compression of Fold Change)

Single array analysis

Multiplicative error
PM-MM
Background imputed to handle
negative differences

Conservative
Smooth down-weighting of outliers
Positive output values
Minimal bias

Limited by single array analysis

Variance not stable on log scale
Some positive bias

dCHIP

RMA

MAS 5

age method, such as a uniform percentage or

spatially smoothed percentage method.
PM ONLY

The Perfect Match only option (PM Only) is

most often utilized for experiments where
the background is assumed to be minimally
variable across experiments. Identical tissues, organisms, or cell lines are all examples
of experimental designs that would benefit
from this analysis option.
MM TREATED AS ADDITIONAL PM

The Perfect Match added to the Mismatch

option (PM+MM) can be applied in cases
where background can be assumed to be
irrelevant to target response. In these cases,
the mismatch probe can be utilized to
increase statistical power by doubling the
number of data points. The use of feature
response terms in the algorithm enables the
use of the mismatch probe features as either
a measurement of background or a measure-

INFORMATICS

ment of signal, as applied here. When used

as a measurement of signal, the mismatch is
simply treated as a less responsive feature.
An example of this type of input sample set
would be tissues from the same organisms
from the same lab, such as inbred mice. This
option is suitable for detecting small differential changes in expression value.
Figure 5 summarizes the way in which
these options can be tuned. Table 1 summarizes the benefits and drawbacks of each
option. The description of assumptions
helps define the appropriate context for
application.
The last option is quick or full optimization. Full optimization provides better sensitivity for low expressors but requires a
slightly longer computational time. The
quick version provides an output similar to
RMA, and so is quite robust and requires
less computational time.

Post-Processing
Once the PLIER algorithm has been run,
the appropriate scaling and normalization
options can be applied.
INTERPRETING OUTPUT

The signal value provided by PLIER is an

estimate of the common response of the features in a probe set to a given target. Some
transcripts are absent, and by design, PLIER
provides near-zero values for targets corresponding to such transcripts. One of the
most common ways to analyze these values
is by using ratios. Ratios with denominators
near zero are inherently unstable. By design,
the raw PLIER signal values are not variance
stabilized. To calculate ratios, a variance stabilizing transformation, such as log of signal
plus a constant, should be applied to avoid
unstable values. For example, a small constant value, such as 16, can be added to all
values before taking the logarithm.
II
II

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

Comparison to Affymetrix
Microarray Suite 5.0
While Affymetrix Microarray Suite 5.0
(MAS 5) treats all features equally, PLIER
utilizes experimental evidence to 1) weight
features based on consistency, and then 2)
utilize the features that are most consistent
for that experiment to estimate signal.
Arrays analyzed using the MAS 5 algorithm
can be re-analyzed using the PLIER algorithm.

Table 3: Options

Options

RMA Like Methods

dChip Like Methods

Data Pre-Processing

Quantile normalization

Quantile normalization

Probe Type

Perfect match only

Perfect match only

Background

% Background (Uniform)

None

Optimization

Quick

Full

REFERENCES
Li, C. and W.H. Wong. Model based analysis of
oligonucleotide arrays: Expression index computation
and outlier detection. PNAS 98:31-36 (2001).
www.biostat.harvard.edu/complab/dchip

Rafael A. Irizarry, Benjamin M. Bolstad, Francois Collin,

Leslie M. Cope, Bridget Hobbs and Terence P. Speed.
Summaries of Affymetrix GeneChip probe level data.
Nucleic Acids Research 31(4):e15 (2003).

Other Popular Methods

PLIER is designed to maximize sensitivity
for low expressors. As explained above,
it builds upon concepts used in dChip
and RMA. Table 2 summarizes how these
methods compare to one another. For more
detailed information, please visit:
affycomp.biostat.jhsph.edu.
Versions of PLIER are also provided that
generate outputs similar to dChip and RMA
for experiments that would benefit from
either of those approaches. Table 3 summarizes the options for either output.

Conclusion
PLIER provides an advanced, flexible framework for primary analysis of GeneChip
microarray data incorporating recent
advances in expression analysis. These
advances include the use of feature response
terms to improve the interpretation of
intensities, a robust M-estimator to reject
outliers, and an improved model for intensity
level errors. The PLIER error model is the
recommended option, as it simultaneously
incorporates a smooth transition between
analysis of low and high expressors, rescaling
by feature responses, and resistance to outliers. However, options within the PLIER
framework allow simpler models to be used
that can replicate the features of other popular low-level analysis techniques, including
the analysis of perfect-match only arrays.

II
II

Glossary
M-Estimator An estimator obtained by minimizing a function. Similar to maximum-likelihood estimates.
Mean The familiar arithmetic mean of several numbers. Minimizes the sum of squares.
Median The middle rank value (or average of the two such values, if there are an even
number of data points). Minimizes the sum of absolute values.
Robustness Ability to produce reasonably accurate results in the presence of outliers.
Sensitivity The ability to detect an entity when it is present. Two types of sensitivity
are often discussed for microarrays. The first is the ability to detect a (specified) transcript
as being present in a sample. The second is the ability to detect a (specified) differential
change between two experiments when there is a change.
False positive When an entity is detected and it is not present. For microarrays, there
are two types often discussed. First, when an absent transcript is reflected as a significantly present output. Second, when a differential change is falsely indicated when there
is no change in an expression level.
False negative When an entity is not detected, and it is present. For microarrays, there
are two types of false negative rates. First, when a transcript is indicated as being absent
when it is expressed in the sample. Second, when a (specified) differential change is
falsely indicated as unchanged.
Specificity The ability to exclude an entity when it is not there. For microarrays, there
are two types of specificity. First, when an absent transcript is correctly indicated as absent.
Second, when an unchanged expression level is correctly indicated as unchanged.
Background Unwanted intensity observed on an array. Sources of background include
fluorescence of the glass, stray DNA, and many other sources. Varies with sequence of
DNA in the probe.
Residuals Real experiments have noise, and therefore any estimates of true values will
differ slightly from those observed in each data point. The difference between the
observed values and the predicted values for a particular estimate.
Error model A means of interpreting how well a model explains the observed data.
Signal A summary value for the observed intensities in a probe set reflecting a common transcript. In PLIER, estimated as the target response that best fits the data given
the feature responses.
Total response The intensity due solely to the target of interest interacting with a given
feature. Assumed in the PLIER model to consist of two components: feature response
and target response, multiplied together.
Feature response A relative measure within a probe set of the (multiplicative) difference in
intensity due to a given feature being different (in location, probe sequence, etc.) than another.
Assumed to be invariant across experiments for a given feature. Example: one PM-MM
difference is reliably twice as bright as another PM-MM difference across experiments.
The pairs differ in feature response. Feature responses cannot be compared across probe
sets due to their relative nature. Feature response is a dimensionless scaling factor.
Target response A measure within an experiment of the (multiplicative) difference in
intensity due to a given experiment having a different target abundance. Assumed to be
common to all features (background adjusted) within a probe set. Example: one PM-MM
difference is twice as bright in one experiment as another experiment. The experiments
differ in target response for this feature. Target responses cannot be compared across
probe sets due to the relative nature of feature responses.

INFORMATICS

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

Appendix: Additional Statistical Details for PLIER

At a high level, the output of the PLIER algorithm is an improved estimate of target variation across experiments (signal) for each probe
set. This signal improvement comes about in two ways:
1. By incorporating information about the individual feature(s) or feature-pairs, in particular, exploiting the fact that different features
have different feature responses.
2. By incorporating an improved error model that smoothly transitions between the arithmetic regime in which feature intensities
are near background, to the "multiplicative" regime in which feature intensities are far from background.
The algorithm can be run on a group of experiments. The feature response may be provided in a model file, in which case the program
will simply generate signal estimates for each experiment, or the feature response may be calculated for the current group of experiments
and saved for future use in a model file.
PLIER drafts an initial estimate of signal (target response) and feature response for each experiment and feature using the data provided,
and then attempts to find estimates that more closely fit the data given an approximate likelihood. The algorithm stops when it finds an
estimate that cannot be made to fit the data more closely.
Justification of the PLIER M-estimator
Estimators obtained by finding estimates that minimize (or maximize, depending on sign) some function are known as M-estimators.
Familiar examples of estimators falling into this class are means (minimize the sum of squares) and medians (minimize the sum of absolute
values). The function that an estimator minimizes can be chosen to have desirable properties, such as computational convenience and
resistance to outliers, while still usefully approximating the fit of a chosen error model.
PLIER is based on the following simple assumptions about the behavior of probes and targets. First, the abundance of a target is never
negative, but can be zero, therefore the target response is always non-negative (t 0). Second, there is a linear link between intensity
(total response) and target response (T~f*t), with the feature response term (f) as the slope of this relationship. That is, the true underlying
intensity is the product of the feature response (common across experiments for a given feature) and the target reponse (t) (common
across intensities in a probe set).
Third, the multiplicative intensity error is the most significant source of variation, that is, the error in repeated experiments for a feature
intensity is approximately log-normal (log(I) ~ normal). It is presumed that background adds to the total response to make the total intensity (I ~ T+B), but the background value can vary from feature to feature, experiment to experiment.
Further, the optimistic assumption that background is sequence and location dependent is made, so that mismatched probe sequences
have backgrounds closely related to their corresponding perfect match probe (when they are sufficiently close), and the pessimistic
assumption is made that the background varies across different samples and locations.
Thus, the reduced model (PM-MM = f*t) is considered, due to the variability of background between experiments and between features
containing probes of very different sequence. Improving the behavior of estimates of signal for low expressors is of greatest interest, and
therefore concentrating on background driven effects, such as the amount of positive bias in an estimate, and the uncertainty of values
near background, is important. The mismatch features do have some positive feature response for the target, and will increase the variability of estimates far from background, but are still the most useful way of removing bias from intensity data found so far.
Considering multiplicative error on intensities, the third assumption, leads to assigning error terms to the observed intensities of both the
Perfect Match and the Mismatch probe features, that is, errors e1 and e2 should be found that satisfy e1*PM - e2*MM = f*t. In general,
good estimates for feature response and signal (target response) are expected to lead to estimates for e1 and e2 near 1 (that is, the data
fit the observations with minimal errors).
If it were possible to have exactly log-normal errors for all the intensities on a microarray, optimizing the function log(e1)2 + log(e2)2 (by
analogy to typical least-squares fitting) could be attempted. However, it is known that there are outliers on the array, and that the distribution is not precisely log-normal. Also, it is computationally inconvenient to find e1 and e2 given this constraint. It is stressed that e1 and
e2 are not the actual errors on the array (since these are unknown) but simply values giving an estimate of goodness of fit. Therefore, a
simplified model of the error terms for computational convenience is sought, and the tails of the distribution are discounted to insulate
estimates from outliers.
One natural simplification is to assume log(e1)2 = log(e2)2. There are two ways this relationship can hold, either log(e1) = log(e2), or
log(e1) = -log(e2). In the first case (which corresponds to log-transformation), no solutions exist when MM>PM. Even in the case where
MM<PM, it can be shown that the total squared error log(e1)2+log(e2)2 is larger under the first constraint than when errors are fit under the
second constraint. Therefore, the second constraint log(e1) = -log(e2) is placed on the errors, which results in solutions for all positive PM
and MM values and all non-negative target responses (t) and feature responses(f). Recall that the approximate likelihood is being simplified for
computational convenience, and these terms should not be interpreted as the actual physical errors on each observed feature intensity.
[y+ y2+4*PM*MM ] ,
The reduced model equation with errors under this constraint is e*PM-MM/e = f*t, which is easily solved to produce e=
2*PM
where y = f*t. This is a very simple formula for an approximating the error in fitting estimates to the observed values. Thus, if there were
no outliers, the natural estimate for signal would minimize the sum of squared residuals r2, where r = log(e) is obtained from the formula.
[Aside: Note that r=glog(f*t, 4*PM*MM) - glog(PM-MM, 4*PM*MM), where glog denotes the generalized logrithm. This shows that the
effective variance of the target response (t) is arithmetic when measured near background.]

II
II

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

Appendix: Additional Statistical Details for PLIER (continued)

This approximation is likely to be best when errors are small, and there will be outliers that do not fit this model on arrays. Therefore,
outliers and areas where the model is problematic are discounted by using a function which looks like r 2 near r = 0, but downweights
the tails. The off-the-shelf function of Geman and McClure,
r2
h(r) =
, satisfies the requirements.
r2
1+ z

Thus, a simple formula for goodness of fit has been derived which can be used to estimate signal and feature response by finding those
non-negative values of signal and feature response that best fit the observed data. This is the M-estimator called PLIER. Some performance
metrics for PLIER (PLIER+16 - with simple variance stabilization applied) can be found at the AffyComp web site run by Rafael Irizarry.
Other presentations on PLIER can be found at the Affymetrix web site, including the presentation at the Low-Level Analysis workshop,
discussing bias and variance issues (www.affymetrix.com/corporate/events/seminar/microarray_workshop.affx).
Calculation of Signal
PLIER drafts an initial estimate of signal (target response) and feature response for each experiment and background-adjusted feature using the
data provided. Please note that a reference table to variables used in the subsequent equations (Table 4) is provided at the end of this section.
PLIER operates by finding target responses (t(i)) for each experiment i and feature responses f(j) for each feature (pair) j that minimize the
function LL(t,f) = sum H(PM,MM, BKG, f(j),t(i)) over all i,j.
Define the following component functions:

y(i,j) = f(j)*t(i)
q = (y*y) + 4*PM*MM
e = (y + q)
(2*PM)
or (if not using mismatch)

e = (y + BKG)
PM
r = log(e)
h=

r*r
1+ r*zr

[z a tuning constant for robustness]

H(PM,MM,BKG, f(j),t(i)) =h(r)

The algorithm will find a minimum of such a function by a high-dimensional search starting at some assignment of values for target
response and feature response, the algorithm proceeds to find better and better assignment of values.
A reference table of symbols utilized in the equations described in this and subsequent sections is provided below:
Table 4. Symbols used in the equations
Symbol
t(i)
f(j)
q
r
PM(i,j)
MM(i,j)
y
BKG
h
GT
GF
LL

II
II

Description
Target response at each experiment i (current signal estimate)
Feature response for each feature (or feature pair) j
Intermediate value in likelihood computation
Residual on log-scale
Intensity of the j-th perfect match feature in the i-th experiment
Intensity of the j-th mismatch feature, in the i-th experiment
Total response estimate with no background
Background value (if not using MM)
German-McClure function for discounting residuals as applied to one probe pair
Extra penalty term added to log likelihood for differences in target response
Extra penalty term added to log likelihood for differences in feature response
Log-Likelihood approximation used by PLIER

INFORMATICS

A F F Y M E T R I X P R O D U C T F A M I LY > A R R AY S

Appendix: Additional Statistical Details for PLIER (continued)

Data Augmentation
Because it is possible for a zero value to be reported as a probe intensity, and it is assumed that there are multiplicative errors on
probes, the first step is to add a small positive value to all the input intensity values to avoid problems.
Drafting an Initial Estimate
It is best to start the iterative procedure from a reasonably good initial estimate. For producing this estimate, the software uses Simplified
Expression Analysis (SEA). One way of coming up with reasonable estimates of target response and feature response is to look at transformed data (i,j) and fit a simple model (i,j)=log(t(i))+log(f(j))+e by use of median polish (this is similar to how RMA operates).
One particularly simple transformation is

(PM B) + (PM B)2 + 4*L*PM*B

(i,j) = log
2

where B is either BKG or MM, depending on which option has been selected by the user. L is an attenuation parameter ranging from 0.0
to 1.0 that controls how the data approach zero as PM approaches B (or becomes less than B).
Median polish is a procedure for fitting a simple additive model robustly by alternately taking the median of each column ((i,j)-log(f(j)) as
the estimate of log(t(i)), and then taking the median of each row ((i,j)-log(t(i)) as the estimate of log(f(j)), until stable estimates of each
value are reached.
This initial estimate of target response (t(i)) is output as signal if the user selects the quick option for PLIER. Note that these estimates
do not minimize the PLIER goodness of fit, and are simply provided for completeness.
Finding a Minimum
One way of finding (a) minimum of such a function is to use Newtons method. Starting with an assignment of values to target responses (t) and feature responses (f), given the rate at which the function is varying, find a good step size to take to find the next assignment
of values to target responses and feature responses. Newtons method turns out to be computationally intensive and requires inversion
of large matrices, and therefore an approximation to the method is used instead.
Identifiability Constraint
Since the method deals with two variables multiplied by each other, it is always the case that feature responses can be multiplied by
some constant and target responses can be divided by some constant and wind up with the same fit to the data. To resolve this ambiguity, it is required that sum(f(j)) = n, the number of features (not counting any used for bias removal). Note that feature responses are relative to the other features in a probe set and cannot be directly compared between probe sets.
Avoiding Local Minima
If good values for target and feature response have been found, so that there are no local improvements possible, it may be that the
method is trapped in a local minimum, and possible improvements to the estimates should be checked for. The natural method for finding an improved estimate is to examine the values a variable can plausibly take on to see if any of them improve the estimate.
In particular, in this case, each feature pair in each experiment provides a natural estimate for alternative values for target and feature
response that might be near to better minima. The reasoning is that a minimum should be near to the value that is perfect for some
feature in some experiment. These two estimates are:
t (i) = PM(i,j) MM(i,j)
f(j)

f (j) = PM(i,j) MM(i,j)

t(j)
These values (if non-negative) are checked systematically to find if there is a possible improvement. If there is possible improvement, the
search is continued from the new value using the Newton-like method. While it is possible to check all possible values, or do a line search,
it is unlikely that a "good" minimum will be far from one of these estimates.
Penalties
It is often useful to modify the likelihood function by putting constraints on the variables, based on the prior knowledge that they are unlikely
to be extremely different from each other. Therefore, a roughness penalty can be incorporated in the likelihood function that penalizes
extremely unusual variable values. This is an advanced option, but included for completeness. The actual functions are the obvious:
GT = differential_target_response_penalty*(sum(log(t(i))-average(log(t))2)
GF= differential_feature_response_penalty*(sum(log(f(j)-average(log(f)))2)
LL = sum(H)+sum(GT)+sum(GF)
This reduces to the standard LL when the penalties are zero [no penalty for roughness].

II
II

AFFYMETRIX, INC.

AFFYMETRIX UK Ltd

AFFYMETRIX JAPAN K.K.

3380 Central Expressway

Santa Clara, CA 95051 USA
Tel: 1-888-DNA-CHIP (1-888-362-2447)
Fax: 1-408-731-5441
[email protected]
[email protected]

Voyager, Mercury Park

Wycombe Lane, Wooburn Green,
High Wycombe HP10 0HH
United Kingdom
UK and Others Tel: +44 (0) 1628 552550
France Tel: 0800919505
Germany Tel: 01803001334
Fax: +44 (0) 1628 552585
[email protected]
[email protected]

Mita NN Bldg., 16 F
4-1-23 Shiba, Minato-ku,
Tokyo 108-0014 Japan
Tel: +81-(0)3-5730-8200
Fax: +81-(0)3-5730-8201
[email protected]
[email protected]

www.affymetrix.com Please visit our web site for international distributor contact information.
For research use only. Not for use in diagnostic procedures.
Part No. 701903 Rev. 1
, GeneChip, HuSNP, GenFlex, Flying Objective, CustomExpress, CustomSeq , NetAffx, Tools To Take You As Far As Your Vision,
2005 Affymetrix, Inc. All rights reserved. Affymetrix,
The Way Ahead, Powered by Affymetrix, and GeneChip-compatible, are trademarks of Affymetrix, Inc. All other trademarks are the property of their respective owners. Software products may
be covered by one or more of the following patents: U.S. Patent Nos. 5,733,729; 5,795,716; 5,974,164; 6,066,454; 6,090,555, 6,185,561 6,188,783, 6,223,127; 6,228,593; 6,229,911; 6,242,180;
6,308,170; 6,361,937; 6,420,108; 6,484,183; 6,505,125; 6510,391; 6,532,462; 6,546,340; 6,687,692; 6,607,887; and other U.S. or foreign patents.