Page 1

Multivariate regression analysis of distance matrices

for testing associations between gene expression

patterns and related variables

Matthew A. Zapala* and Nicholas J. Schork*†‡

*Biomedical Sciences Graduate Program and the Polymorphism Research Laboratory, Department of Psychiatry, and†Division of Biostatistics, Department of

Family and Preventive Medicine, Moores UCSD Cancer Center, Center for Human Genetics and Genomics, and the California Institute of Telecommunications

and Information Technology, University of California at San Diego, 9500 Gilman Drive, La Jolla, CA 92093

Communicated by Dennis A. Carson, University of California at San Diego School of Medicine, La Jolla, CA, October 25, 2006 (received for review

July 25, 2006)

A fundamental step in the analysis of gene expression and other

high-dimensional genomic data is the calculation of the similarity

or distance between pairs of individual samples in a study. If one

has collected N total samples and assayed the expression level of

G genes on those samples, then an N ? N similarity matrix can be

formed that reflects the correlation or similarity of the samples

with respect to the expression values over the G genes. This matrix

can then be examined for patterns via standard data reduction and

cluster analysis techniques. We consider an alternative to conven-

tional data reduction and cluster analyses of similarity matrices

that is rooted in traditional linear models. This analysis method

allows predictor variables collected on the samples to be related to

variation in the pairwise similarity/distance values reflected in the

matrix. The proposed multivariate method avoids the need for

reducing the dimensions of a similarity matrix, can be used to

assess relationships between the genes used to construct the

matrix and additional information collected on the samples under

study, and can be used to analyze individual genes or groups of

genes identified in different ways. The technique can be used with

any high-dimensional assay or data type and is ideally suited for

testing subsets of genes defined by their participation in a bio-

chemical pathway or other a priori grouping. We showcase the

methodology using three published gene expression data sets.

analysis of variance ? high-dimensional data

T

researchers with a set of assays that are unprecedented in their

sophistication. These technologies allow researchers to interro-

gate the expression levels of thousands to tens-of-thousands of

genes or proteins simultaneously (1, 2). Although of tremendous

importance, the use of these technologies is plagued by the fact

that they generate enormous amounts of data, whose signifi-

cance, both statistically and biologically, can be difficult to

fathom in any single experiment (3). In essence, the collection of

expression levels on thousands of genes on relatively few indi-

viduals or other units of observation, such as cells or cell types,

creates enormous potential for false positive results when each

gene is analyzed in isolation (4).

Many clever and useful data analysis strategies for the assess-

ment of gene expression and related high-dimensional genomic

data have been proposed (5). The vast majority of these strat-

egies rely on either some form of data reduction, such as cluster

analysis (6), or eigenstructure analysis (7, 8), which raises a

number of questions about the appropriateness of the cluster

method used, the number of clusters or eigenvalue/eigenvector

pairs seen as ‘‘optimal,’’ appropriate or statistically significant, as

well as the biological meaning of the clusters or eigenvectors that

emerge (9). Despite this fact, one common and appropriate

strategy exploited by a number of data analysis approaches,

which is in fact a precursor and fundamental construct to many

he introduction of high-throughput technologies, such as

DNA microarrays and proteomics platforms, has provided

contemporary gene expression analysis methods, involves the

construction of a similarity or distance matrix, which reflects the

similarity/dissimilarity of each pair of individuals with respect to

the gene expression values obtained on them. This strategy was

outlined in many of the earliest proposed gene expression

analysismethods(6,10–12),hasbecomeastandardtoolforgene

expression data analysis and visualization tools (13, 14), and is,

in fact, even a typical ingredient in cluster and eigenstructure

analyses.

We describe a method for testing the relationship between

variation in a distance matrix and predictor information col-

lected on the samples whose gene expression levels have been

used to construct the matrix. The method provides a formal test

of the organization of a similarity or distance matrix as it relates

to predictor variable information collected on the individual

samples, such as clinical parameters on subjects whose tumors

have been evaluated for gene expression or genotype data of

different inbred mouse strains assayed for gene expression. As a

result, the method is the perfect companion for heat map and

tree-based representations of high-dimensional data organized

by some feature or a priori grouping factor meant to graphically

represent and reveal a relationship between the genes used to

construct the heat map or tree and these features or groupings.

By testing more global hypotheses about the patterns within a

similarity or distance matrix, the procedure avoids the need for

cluster analysis and is very appropriate for situations where the

number of data points collected is much larger than the number

of samples or individuals. We first describe the derivation of the

method, and then showcase its application to three publicly

available data sets. We also want to emphasize that the proce-

dure can be used to study any number of groups of genes,

including single genes or all of the genes in a data set, making

it very flexible and a method that only adds to existing univariate

approaches.

Methods

Basic Model. In describing the proposed analysis methodologies,

we follow the notation in McArdle and Anderson (15). We do

not focus on many of the alternative methodologies for distance-

based analyses developed by Krzanowski (16), Gower and Kr-

zanowski (17), Legendre and Anderson (18), and Gower and

Legendre (19), although many of these techniques may have

some merit in the analysis of genomic data. Note that we used

boldface to indicate matrices or vectors in our notation. Let Y be

Author contributions: M.A.Z. and N.J.S. designed research, performed research, analyzed

data, and wrote the paper.

The authors declare no conflict of interest.

‡To whom correspondence should be addressed. E-mail: nschork@ucsd.edu.

This article contains supporting information online at www.pnas.org/cgi/content/full/

0609333103/DC1.

© 2006 by The National Academy of Sciences of the USA

19430–19435 ?

PNAS ?

December 19, 2006 ?

vol. 103 ?

no. 51www.pnas.org?cgi?doi?10.1073?pnas.0609333103

Page 2

an N ? P matrix harboring gene expression values on N subjects

for P genes. Let X be an N ? M matrix harboring information

on M predictor or regressor variables whose relationship to the

gene expression values is of interest, where the first column

contains a column vector whose every element is 1, and reflects

an intercept term, as in standard regression contexts. These

predictor variables could include the ages of individuals assayed,

clinical diagnoses, strain memberships, cell line types, or geno-

type information. A standard multivariate multiple regression

model for this situation would be (20, 21)

Y ? X? ? ?,

[1]

where ? is an M ? P matrix of regression coefficients and ? is

an error term, often thought be distributed as a (multivariate)

normal vector. The least-squares solution for ? is ?ˆ ?

(X?X)?1X?Y, with the matrix of residual errors for the model

being

R ? Y ? Yˆ? Y ? X?? ?I ? H?Y,

[2]

where H ? (X?X)?1X? and is the traditional ‘‘hat’’ matrix.

Unfortunately, If N ? ? P, as is often the case with gene

expression and other genomic data types, then this model is

problematic. An alternative would consider how the M predictor

variables relate to the similarity or dissimilarity of the subjects

under study with respect to the P gene expression values as a

whole or as a series of unique subsets of the data.

Let D be an N ? N distance matrix, whose elements, dij, reflect

the distance (or dissimilarity) of subjects i and j with respect to

thePgeneexpressionvalues.Forexample,dijcouldbecalculated

as the Euclidean distance or as a function of the correlation

coefficient (see Forming the Distance Matrix below). Let A ?

(aij) ? (?1⁄2dij

A by calculating

G ??I ?1

2). One can form Gower’s centered matrix G from

n11??A?I ?1

n11??,

[3]

where 1 is a N-dimensional column vector whose every element

is 1 and I is an N ? N identity matrix. An appropriate F statistic

for assessing the relationship between the M predictor variables

and variation in the dissimilarities among the N subjects with

respect to the P variables is

F ?

tr?HGH???M ? 1?

tr??I ? H?G?I ? H????N ? M?,

[4]

where H is a hat matrix, G is Gower’s centered matrix, and I is

the identity matrix, formed as above. M is scalar and reflects the

number of predictors and N is the number of subjects. If P ? 1

(i.e., a univariate analysis) and the distance matrix is computed

through the use of the standard Euclidean distance measure,

then F in Eq. 4 is the standard F statistic and possesses the typical

properties associated with F statistics in ANOVA contexts. This

result is due to the fact that the inner product matrix (Y?Y) used

in standard univariate analysis of variance and regression con-

texts contains the same information, in terms of total sums-of-

squares, as the outer product matrix (YY?), which reflects

interpoint squared differences or distances (tr(Y?Y) ? tr(YY?))

(15). When different distance measures are used, the properties

of F are more complicated, suggesting the use of alternative

methods for assessing statistical significance (see Assessing Sta-

tistical Significance below).

Forming the Distance Matrix. The formation of the distance matrix

is an important step in the use of the proposed procedure. There

is a bewildering array of potential distance measures one could

use with the proposed method (22). The correlation coefficient,

r, is often used to assess the similarity between two individuals

based on gene expression values (14). A correlation matrix with

elements rijcan be converted to a distance matrix with elements

dijeasily enough through a simple transformation

dij? ?2?1 ? rij?.

[5]

This transformation leads to a distance matrix with metric

properties, although distance measures with nonmetric proper-

ties can be used in the analysis method described as well (17). We

discuss aspects of the choice of a distance measure in Results, but

more work in this area is needed. One additional aspect of the

formation of a distance matrix that deserves attention involves

handling missing data. Intuitively, if one has collected thousands

of gene expression values to be used to create distance profiles,

then a few missing observations are not likely to have much of

an influence. For example, one could simply not use these genes

in the formation of the distance matrix, ignore the missing values

only when assessing pair-wise distance for a pair of observations

with missing data, or assign individuals with missing data

imputed values that are then used to compute distance. How-

ever, the delineation of a threshold beyond which the number of

missing values creates problems for a distance-based analysis is

important and worthy of further research.

Assessing Statistical Significance. The distribution of the F statistic

defined in Eq. 4 is complicated and its derivation for any

particular distance matrix is unlikely to generalize to other

distance matrices, especially with small sample sizes. Therefore,

one can rely on permutation tests to evaluate the probabilistic

significance of an observed F statistic computed from Eq. 4

(23–25). Permutations can either involve permuting the raw data

or simultaneously permuting the rows and columns of the G

matrix, as is done in Mantel’s matrix correspondence test (15).

In addition, if permutation tests are used, the degrees-of-

freedom terms in the numerator (M ? 1) and the denominator

(N ? M) are not required in the formulation of the statistic

presented in Eq. 4. Finally, given that different predictor vari-

ables, or subsets of variables, can be tested for association with

variation in a distance matrix, one can pursue step-wise or

variable selection procedures with the technique, identical to

univariate standard multiple regression analysis (26). Beyond a

P value, an estimation of the proportion of variation within the

matrix that is explained by a particular set of M predictor

variables can be calculated by dividing tr(HGH) (i.e., the sum of

the diagonal elements of a matrix) by tr(G). In our analyses,

independent variables are tested both individually and in a

forward stepwise manner. The independent variables selected

for the model in a stepwise manner are based on the highest

cumulative proportion of variance that is explained by the

inclusion of an additional variable in the regression model. An

F statistic and P value are calculated for the addition of this

variable to the model.

Assessing Level Accuracy and Power of the Proposed Hypothesis Test.

To examine properties of the proposed analysis procedure, a

series of studies investigating the level accuracy and power of the

proposed test statistics were performed. To examine the level

accuracy of the test, we simulated 30 samples each measured on

100 variables. The variables were assumed to follow a standard

normal distribution; hence, there was no structure to the data.

We assumed that the first 15 samples had a different origin than

the second 15. We then tested the relationship between this

grouping factor (coded as 0 for the first 15 samples and 1 for the

second 15 samples) and the distance between the samples

calculated with different distance measures using the proposed

procedure with 1,000 random permutations of the data. We

Zapala and Schork PNAS ?

December 19, 2006 ?

vol. 103 ?

no. 51 ?

19431

GENETICS

Page 3

repeated this process 1,000 times. Table 1 describes the results

and clearly shows that the nominal level of the test matches

closely with the simulation results, suggesting that the proposed

test procedure is nonbiased.

We also considered the power of the proposed test. We

simulated data for 30 samples and 100 variables in which 15

samples were assigned to a hypothetical control group (inde-

pendent variable ? 0) and 15 samples were assigned to a

hypothetical experimental group (independent variable ? 1).

Data in the control group were generated as standard normal

variables with a mean of 0 and variance 1. Data in the experi-

mental group were generated as standard normal variables with

variance ? 1 and means that took on values of 0 to 1.5 in

increments of 0.001. The power of the proposed permutation-

based statistical test was then investigated in these settings. In

this context we also generated different simulated data sets for

which 100%, 50%, 25%, 10%, or 5% of the variables used in the

construction of the distance matrix had means adjusted from 0

(in the appropriate increments) in the experimental group. Fig.

1 describes the results. Note that the gray line in Fig. 1 represents

the power curve obtained based on a t test with the Bonferroni

correction, corrected for 100 multiple comparisons. Fig. 1 clearly

shows that the proposed procedure can detect ‘‘signals’’ in the

dataaslongasthenumberofvariablescontributingtothatsignal

used in the construction of the distance matrix is moderate.

Results

The proposed method was tested on three different published

data sets to display its utility. We briefly consider some of the

implementation details and properties of the proposed tech-

nique, such as the need for evaluating the distance between the

observations, and the dependence of the test statistic on subsets

of genes among all those used to derive a distance measure. We

note that for the following applications we used the correlation

coefficient to derive the distance measure, as this measure has

been the standard for gene expression data (14). In addition, we

used 1,000 permutations to compute P values.

Embryonic Imprint of the Adult Mouse Brain. The first data set

involved gene expression data from multiple brain regions and

multiple inbred mouse strains (27). The normalized data can be

downloaded from the Gene Expression Omnibus (GEO) using

record number GSE3594. The authors had three hypotheses

about the relationships of the gene expression patterns between

the different brain regions in the adult mouse. The gene expres-

sion patterns of these brain regions could be related to each

other based on adult anatomy, evolutionary relationships, or

embryonic origin. The authors performed hierarchical cluster

analysis and created a Pearson correlation heat map matrix

where they hypothesized that the gene expression patterns of the

adult mouse brain bear an imprint based on the adult tissue’s

embryological origin. The heat map and the hierarchical tree

constructed based on the similarity of gene expression patterns

across all of the genes suggest that adult structures are related

to each other based on the classic five vesicle embryonic neural

regions [telencephalon, diencephalon, mesencephalon, meten-

cephalon, and myelencephalon; see supporting information (SI)

Fig. 3). Using the proposed regression analysis procedure, we

provide statistical evidence that embryonic origin is the most

likely hypothesis of the three because it explains the largest

proportion of variation in the similarity of the overall gene

expression profile of the brain regions (P ? 0.001 and proportion

of variation in pair-wise distances explained by embryological

origin ? 0.33, adult anatomy ? 0.26, and evolutionary relation-

ships ? 0.19). The authors also suggested that anterior-to-

posterior (A/P) position along the neural tube could dictate

expression patterns in the adult neural structures. The position

along the neural tube was tested individually and in combination

with embryological origin. Importantly, A/P position added a

significant proportion of explained variation in brain region gene

expression similarity over and above embryological origin (P ?

0.001 and proportion of variation explained above embryological

origin ? 0.032).

Aging Human Brain. The second data set examined gene expres-

sion patterns in the human frontal cortex among individuals who

died at various ages (26–106 years) (28). The normalized data

can be downloaded from GEO using record number GSE1572.

The authors performed Spearman rank correlations to deter-

mine 463 genes that correlated with age (P ? 0.005) and a

Pearson correlation-based heat map matrix was then calculated

that covered all pair-wise comparisons of individuals. We ana-

lyzedthedistancematrixbasedonthepairwisecorrelations(Eq.

5) using the proposed regression method to quantify the effect

that age and sex may have on the gene expression patterns for

the 463 genes found to be correlated with age (age P ? 0.001 and

proportion of variation explained ? 0.35; sex P ? 0.224).

Although sex was not a significant predictor of the gene expres-

sion patterns in the frontal cortex, age appeared to explain

?35% of the variation in the similarity in the gene expression

patterns among the individuals based on the age-related genes

(see SI Fig. 4A). Moreover, the association with age was not only

apparent in age-related genes (as identified from Spearman rank

correlations), but was also evident in the correlation matrix

created using all of the genes scored as ‘‘Present’’ in at least one

of the samples (age P ? 0.001 and proportion of variation

explained ? 0.16; sex P ? 0.78) (see SI Fig. 4B). Therefore, it

Fig. 1.

as a function of both increasing differences in control vs. experimental set-

tingsaswellasoverallsignal-to-noiseratio-basedsimulateddatasets(seetext

for details).

Power of the proposed distance matrix-based regression procedure

Table 1. Level accuracy of the proposed permutation test

Distance

metric

% of tests

P ? 0.01

% of tests

P ? 0.05

% of tests

P ? 0.25

% of tests

P ? 0.50

Pear.

Spear.

Kend.

Conc.

Eucl.

Cheb.

1.3

1.5

1.3

1.3

1.2

1.3

4.8

4.6

4.9

4.8

6.1

5.9

26.5

27.4

23.9

26.7

24.7

25.4

51.4

52.9

47.9

52

49.1

48.6

P values were calculated using the proposed method based on 1,000

permutations of the data. The simulations were repeated 1,000 times and the

percentage of P values below a certain threshold is reported for each of the

following metrics: Pearson correlation (Pear.), Spearman rank (Spear.), Ken-

dall Tau (Kend.), Lin’s concordance correlation (Conc.), Euclidean distance

(Eucl.), and Chebychev distance (Cheb).

19432 ?

www.pnas.org?cgi?doi?10.1073?pnas.0609333103Zapala and Schork

Page 4

appears that the age effect is pronounced enough to be signif-

icant even when including all detectable genes on the array

(8,507 probe sets).

As emphasized, the proposed regression technique can be

applied to each gene in a univariate manner. Univariate analysis

can be used to identify a set of genes that, when considered in

the construction of a distance matrix, are strongly related to a

specific predictor variable. For example, we wanted to identify a

set of age-associated genes that would explain a larger propor-

tion of variation than the set that Lu et al. (28) identified by using

Spearman rank correlations. Using the matrix regression tech-

nique for each individual gene (using the Euclidean distance),

we calculated an F statistic that, as noted when Euclidean

distances are used with a single variable, is identical to an

ANOVA statistic (15). The resulting F statistics were then used

to rank genes and identify those which demonstrated the largest

age association (Fig. 2 A and B). These ranked genes were then

serially used to construct matrices tested with our proposed

procedure to identify an optimal set of genes that resulted in the

largest proportion of variation explained by age effects (Fig. 2 C

and D). An optimal set was found to occur with the top 100

ranking genes, as age explained 52% of the variation in dissim-

ilarity within the matrix constructed with these genes, which is

much higher than the 35% explained with the age-related genes

chosen by Lu et al. (28). Of the 100 genes identified as an optimal

set, 80 were within the Lu et al. set. Even using an equivalent

number of genes as Lu et al. used, the proportion of variation

explained by our highest ranked 463 genes is 0.42 compared with

the Lu et al. proportion of 0.35. Of our 463 highest ranked genes,

only 256 are also present in the Lu et al. study. Interestingly,

among the 20 genes within the top 100 genes identified by our

analysis that were not identified in the Lu et al. study were two

genes involved specifically with neurological function, mitogen-

activated protein kinase kinase 1 (MAP2K1) and amyloid beta

(A4) precursor protein-binding, family A, member 1 (APBA1).

MAP2K1 is known to control apoptosis signaling specifically in

astrocytes (29) and to regulate MAPK1, which was found in the

original Lu et al. study and is also involved in synaptic transmis-

sion. APBA1 is a neuronal adaptor protein that interacts with,

stabilizes and inhibits the proteolysis of the Alzheimer’s disease

amyloid precursor protein (APP) (30). A large number of genes

involved in neurological disorders are present in the 207 genes

of our expanded gene list which were not identified in the

original 463 aging genes identified by Lu et al. These include

interesting genes implicated in age-specific neurological dis-

eases,suchasCDK5,CDK5R1,BCL2L1,andPLAT(30–33).The

full gene list is found in SI Table 2.

Aging Human Kidney. The third data set considered gene expres-

sion in human kidney tissue across two regions (cortex and

medulla) in multiple patients of different ages (34). The nor-

malized data can be downloaded from the Stanford Microarray

Database. In addition to the gene expression data, there were

numerous clinical parameters available on a majority of the

patients. The clinical parameters included indicators of renal

pathology, such as the degree of glomerular sclerosis or arterial

intimal hyalinosis (AIH). There was also information about

creatininelevels,medicalhistory,andsystolicanddiastolicblood

pressures. All of this information was used as independent

variables to predict gene expression patterns. We restricted our

analysis to patient samples for which there was clinical data

available for all of the parameters. This limited the analysis to 63

samples. First, we analyzed the Pearson correlation derived

distance matrix (Eq. 5) for genes that had to be scored as

‘‘Present’’ in at least one of the samples. Three variables were

found to significantly contribute to the variation in the gene

expression similarity/dissimilarity: tissue type (cortex or medulla

P ? 0.001; proportion ? 0.18), AIH (P ? 0.001, proportion ?

0.05), and past medical history (P ? 0.033, proportion ? 0.03).

Past medical history included a history of hypertension, diabetes

mellitus, chronic renal insufficiency, hepatitis B virus, hepatitis

C virus, or combinations of those diseases. Next, we analyzed the

distance matrix computed from only the age associated genes

that the authors had identified by using a linear regression model

(985 genes). Interestingly, age was not a predictor of the gene

expression pattern in the age associated gene expression corre-

lation matrix. The most significant predictors were AIH (P ?

0.001; proportion ? 0.15), tissue type (P ? 0.002; proportion ?

0.08), and race (P ? 0.001 proportion ? 0.07) (see SI Fig. 5). Age

was insignificant with a P value of 0.63 and a contribution of

Fig. 2.

multivariate matrix regression and hierarchical clustering. (A) Hierarchical

cluster of top 100 genes identified from univariate regression of single genes.

Thedendrogramabovetheheatmapshowsthesamplesorderedfromyoung-

est to oldest (from left to right) with the leftmost branches of the tree

connecting data on individuals with an average age of 43, whereas the

rightmost branches connect data on individuals with an average age of 80 (t

test, P ? 0.000001). Hierarchical clustering was performed by using CLUSTER

with the average linkage metric and displayed with JAVATREEVIEW. (B) Plot

ofpermutedPvaluesofageeffects(purple)andsexeffects(red)aswellasthe

P values obtained from standard ANOVA F statistic for age (pink) and sex

(green)effects.(C)GeneswererankedbytheageFstatisticandthengrouped

together to identify the set of highly significant age-related genes (here

found to be 100 genes) whose expression levels would produce a sample-

based distance matrix such that the variation in its elements could be ex-

plained maximally by age effects. (D) A heat map matrix of the optimal set of

100 highly significant age-related genes for which age explains ?52% of the

variation in dissimilarity across the individuals based on these genes’ expres-

sion values.

Use of univariate regressions to identify optimal set of genes for

Zapala and Schork PNAS ?

December 19, 2006 ?

vol. 103 ?

no. 51 ?

19433

GENETICS

Page 5

?0.01. However, the chronicity index, which was an index

developed by the authors that scores the morphological and

physiological state of the kidney and was designed to give a

physiological age to the kidney, was almost significant with a P

value of 0.055 and proportion of variation explained of 0.02.

Thus,collinearityamongchronicityindexandotherindependent

variables may have prevented it from entering into the final

model as a significant predictor. When we tested the indepen-

dent variables individually, the chronicity index was significant

(P ? 0.001; proportion ? 0.15).

Beyond testing whole sets of genes, the method can test

specific subsets of genes for which it may be hypothesized that

gene expression is specifically altered. For example, one may be

interested in whether genes involved in the Pharm-GKB derived

ACE-inhibitor pathway show altered gene expression patterns

consistent with a specific form of renal pathology. Testing all of

the ACE-inhibitor pathway genes using the proposed procedure,

we discovered that not only are there large tissue differences

between the cortex and medulla of the kidney in the ACE-

inhibitor pathway (P ? 0.001, proportion ? 0.12), but there is a

significant association above tissue differences in regards to the

patient’s level of tubular atrophy/interstitial fibrosis (P ? 0.007,

proportion ? 0.08, cumulative proportion ? 0.20).

Evaluating Different Distance Measures. We considered the effect

of the use of different distance measures on the tests for

association. Although not an exhaustive study, we present this to

showcase the importance of choosing a distance matrix. The

choice of a distance matrix is important in a number of related

contexts, such as the choice of a distance matrix for graphically

representing data in heat maps or tree diagrams or in cluster

analysis settings (35–37). We reevaluated the associations in-

volving the above data sets using the Pearson correlation coef-

ficient, the Spearman correlation, the Kendall Tau correlation,

Lin’s concordance correlation, the Euclidean distance, and the

Chebychev distance to derive the distance matrix (see SI Table

3). This analysis considered the distance matrices constructed

from the same genes used in the analyses above. Lin’s concor-

dance correlation, the Euclidean distance and the Chebychev

distance emphasize the actual proximity of the numerical values

of the genes used to compute the distance matrix, and hence

stand in contrast to the correlation coefficient which merely

considers the linear relationship of the values across the genes

used (38, 39). The choice of a distance measure influences the

proportion of variation in the distance matrix explained but not

necessarily the significance of the relationship between the

predictor variables and the distance matrix entries. A more

thorough evaluation of this issue is required.

Signal Strength and Distance Matrix. Because it is unlikely that all

of the genes considered in a study will be related to a particular

predictor variable, the formation of a distance matrix with all of

the genes may not show a signal, or as strong a signal, with the

predictor variable as a distance matrix constructed with only

those genes that are relevant to the predictor variable. Unfor-

tunately, it will be difficult to know a priori which genes should

go into the construction of the distance matrix. Although our

procedure can be used to test each gene individually, or subsets

of genes, as noted, we have also considered the more ‘‘omnibus’’

hypothesis testing situation in which one is interested in knowing

whether there is any relationship between a predictor variable

and gene expression patterns as a whole or across all genes

assayed in a study. We were therefore interested in determining

how strong the relationship between gene expression similarity

and predictor variables considered in our examples is as a

function of the number of genes considered in the construction

ofthedistancematrix.Thiswouldprovideuswithinsightintothe

amount of ‘‘noise’’ that could be tolerated and still allow the

‘‘signal’’ relating the gene expression values and the predictor

variable to appear. We therefore considered the inclusion of

random, simulated gene expression values in the construction of

the distance matrix, knowing that these random simulated gene

expression values would saturate the signal if enough were

added. SI Fig. 6 shows the relationship between the F statistic,

the proportion of variation in similarity/dissimilarity explained,

and the permutation P value as a function of the number of

extraneous gene expression values that are added in the con-

struction of the distance matrix for all data sets tested. Large

amounts of noise reduce the overall proportion of variation in

similarity/dissimilarity explained as well as the F statistic, as one

would expect; however, the permutation test derived P values

remain significant. Thus, it takes the addition of ?98% noise to

saturate the signal to the point of statistical insignificance.

Discussion

Our proposed method of analysis can easily complement many

traditional and alternative methods of analysis for high-

dimensional data. In fact, because the proposed procedure can

be used to analyze each gene in a univariate manner, it extends

traditional univariate procedures. In addition, unlike other

approaches, the proposed approach does not require a reduction

ofthedataviaprincipalcomponents(40),cluster(6),factor(41),

or multidimensional scaling analysis (42). The proposed analysis

procedure also differs from related procedures, such as GSEA

and globalTest (43, 44), in that it can be used to emphasize the

multivariate nature of the expression values of many genes in the

same pathway and treats the system being interrogated as a

whole and does not consider each individual gene in a univariate

analysiswhichthenconsiderstheresultoftheunivariateanalyses

in aggregate. The exploitation of this fact can have disadvan-

tages, obviously, because one may be interested in knowing

which particular sets of genes are the most perturbed in a

particular setting. However, it is arguable that physiological

perturbations and variations are likely to ‘‘re-set’’ the coordi-

nated expression patterns of many genes to reach biochemical or

physiological homeostasis or equilibrium. Thus, the assessment

of the similarity of global gene expression profiles of multiple

samples with different features or exposures is appropriate.

Depending on the number of data points that are selected for

analysis, it is possible to over fit the regression and identify

significant predictor variables whose effect could be assigned to

a large number of data points, when in fact only a smaller subset

of the data points is truly associated with the predictor variable.

However, because the multivariate regression technique can be

reduced to a univariate analysis that focuses on single data

points, it is possible to identify specific subsets of the data within

a larger group for which the predictor variable is having the

largest significant effect. The method we have proposed is, in

fact, flexible enough to be used in settings for which insight into

the effects of single genes or subsets of genes is the goal.

Alternatively, one could test subsets or groups of genes based on

some (a priori) grouping factor, such as participation in a

biochemical pathway or genetic network. One could also com-

bine the proposed approach with standard non-distance-based

univariate and/or cluster analysis methods to assess the signifi-

cance of groups of genes identified with these methods with

respect to a predictor variable. Finally, beyond testing the

relevance of specific clinical or phenotypic predictor variables,

the method can be used as a quality control measure to identify

potential sources of nonbiological error, such as technician, chip

lot or dissection error. These sources of nonbiological variation

can be included in the multiple regression as additional inde-

pendent variables and thus these factors can be controlled for in

the analysis.

There are some limitations to the proposed method that go

beyond the choice of a distance metric or the manner in which

19434 ?

www.pnas.org?cgi?doi?10.1073?pnas.0609333103Zapala and Schork