Identification of novel stem cell markers using gap analysis of gene expression data

gb-2007-8-9-r193 GBJ Method Identification of novel stem cell markers using gap analysis of gene expression data Krzyzanowski M Paul pkrzyzanowski@ohri.ca Andrade-Navarro A Miguel mandrade@ohri.ca

Molecular Medicine, Ottawa Health Research Institute, 501 Smyth Road, Ottawa, Ontario, K1H 8L6, Canada

Faculty of Medicine, University of Ottawa, 451 Smyth Road, Ottawa, Ontario, K1H 8M5, Canada

Genome Biology 1465-6906 2007 8 9 R193 http://genomebiology.com/2007/8/9/R193 17875203 10.1186/gb-2007-8-9-r193 4 5 2007 17 9 2007 17 09 2007 2007 Krzyzanowski and Andrade-Navarro.; licensee 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.

Gap analysis for stem cell markers

A method for the detection of marker genes in large heterogeneous collections of gene expression data is described and applied to DNA microarray data generated from 83 mouse stem cell-related samples.

Abstract

We describe a method for detecting marker genes in large heterogeneous collections of gene expression data. Markers are identified and characterized by the existence of demarcations in their expression values across the whole dataset, which suggest the presence of groupings of samples. We apply this method to DNA microarray data generated from 83 mouse stem cell related samples and describe 426 selected markers associated with differentiation to establish principles of stem cell evolution.

Bioinformatics Development Genome studies

Background

Gene expression microarrays allow thousands of transcripts in a cellular sample to be quantified simultaneously. (For reviews of the technology and applications, see the reports by Heller 1 and Sloughton 2.) Continuing improvements in microarray technology, in terms of transcript density, technical robustness, and cost, have led to widespread usage of arrays in experiments. The size of single studies has grown and can encompass the analysis of up to hundreds of arrays simultaneously 345. This vast explosion of reusable data being generated has resulted in efforts being directed at producing expression data repositories in which the data are curated and presented in an ordered manner 678. The large number of data points makes such resources an exceptional source of biologic information.

Some common uses of gene expression data are the identification of co-regulated genes across many samples 9, identification of differentially expressed genes in samples of interest 10, and, more recently, analysis of alternative splicing 111213 and genome-wide surveillance of transcription 141516. They can also be used to identify marker genes associated with specific sets of samples. As distinguishing features, such markers can be used as diagnostic tests for disease 1718 or for the identification and purification of particular cell types 1920. The identification of multiple markers for a particular phenotype may also reveal biologic mechanisms by which certain genes act in concert.

A simple method to identify marker gene candidates is to identify genes that are differentially expressed between a set of control samples and samples from a condition of interest. A two-state comparison can be made, and genes associated with each type of sample can be identified and used as markers. Current gene expression databases typically contain data from many types of samples, and this heterogeneity provides the potential for more powerful analyses. One can, for example, identify transcripts that are specific to a sample (or samples) of interest, or conduct novel comparisons between different combinations of transcription profiles. The increased size of the databases also increases the number of possible two-state comparisons exponentially, which poses a computational problem. Overcoming this problem requires a computational method.

We have developed a methodology that uses large heterogeneous gene expression datasets to identify genes that can function as markers. In summary, we examine the distribution of expression values of each probe set to identify gaps. These gaps can be used to partition the database into groups of low-expressing and high-expressing samples, which suggest the existence of distinct subpopulations of samples. We then score other probe sets based on their ability to reproduce these database partitions. The characteristics of samples in each database partition identify the context in which genes may act as markers, which aids in the subsequent evaluation of genes in terms of their putative marker roles.

In this study we illustrate our methodology in the analysis of a database of stem-cell related DNA microarray samples that we previously developed (StemBase 7). In particular, we study 83 mouse stem cell related samples analyzed using the Affymetrix MOE430 genechip set (Affymetrix Inc., Santa Clara, CA, USA), which includes approximately 45,000 probe sets. Unbiased application of the method produces a set of 4,449 cell and tissue markers, including 45 out of 71 known stem cell markers (69%). Analysis of the markers that segregate six types of stem cells (hematopoietic, mast, mammospheres, osteoblasts, and two embryonic) from their differentiated counterparts suggests 426 high confidence markers, 206 of which are highly expressed in the stem cell and 222 are highly expressed in the differentiated counterpart (two being highly expressed in stem cells in some cases, and in the differentiated counterpart in others). Of those 426 markers, 17 are involved in multiple distinct lineages that include at least one non-embryonic cell type; nine markers are highly expressed in the stem cells, six are highly expressed in the differentiated cells, and two exhibit opposite variation in different stem-derivative cell pairs. Analysis of the functions of the 222 genes that are highly expressed in the differentiated cells indicates enrichment of extracellular gene products and enzyme inhibitors (12 genes, five of them serpins).

The set of 426 stem cell markers allows us to focus on gene superfamilies that have undergone repeated gene duplication events for a phylogenetic analysis of the evolution of proteins involved in stem cell function. By sequence similarity analysis, we identify four such families (nuclear receptors, cytochrome P450, Rab family GTPases, and early B-cell factors) with multiple members in this set. The study of examples from each reveals multiple events of gene duplication along the vertebrate lineage giving rise to genes with a very high degree of sequence similarity, but very different patterns of expression in stem cells. This leads to a hypothesis that many stem cell related genes expressed in particular tissues arose by duplication and specialization of stem cell related genes originally expressed in other tissues. Superfamilies with large rates of duplication in the vertebrate lineage may have functions related to the development of an increasingly complex organism, including the generation and control of tissue-specific stem cell pools. All results and data presented here are available and can be queried through a web server 21.

Results

We applied our method to a set of DNA microarray data from 83 samples from mouse stem cells and derivatives. Samples included embryonic, hematopoietic, mammosphere, retinal, neurosphere, adipose, and muscle cells (Table 1). All data were obtained using the Affymetrix MOE430 platform (Affymetrix Inc.) and subjected to quality controls as previously described 7.

Table 1

Set of mouse samples selected for our analysis from StemBase

Class

Samples

Replicates

SampleIDs

Adipose derived stem cells

S199

Dermis derived stem cells

S200

Embryonal carcinoma

S129, S130, S131, S132

Embryonic

S219, S220, S164, S165, S166, S167, S168, S169

Embryonic fibroblasts

S180, S286

Embryonic stem cell differentiation

105

S153, S154, S155, S156, S157, S158, S159, S181, S174, S241, S175, S206, S207, S208, S209, S210, S211, S212, S213, S215, S216, S217, S242, S243, S244, S251, S252, S245, S250, S246, S247, S248, S249, S127, S128

Hematopoietic

S233, S234, S235, S236, S237, S147, S291, S292, S293, S294, S295, S296

Mammary

S255, S256, S311, S312, S313

Muscle derived stem cells

S274, S184, S197

Neural

S271, S272, S198

Osteoblast differentiation

S185, S186, S188, S190, S192, S194, S196

Retinal derived stem cells

S232, S240

Various stem cells (SC) are represented in the data set. All original data including sample and experiment descriptions are accessible from the StemBase homepage [64].

At a U-score cutoff of 0.9 (see Materials and methods, below), our method identified 893 different two-state classifications with which to segregate the 83 sample dataset. Figure 1 shows three exemplar distributions of hybridization values for three probe sets that clearly segregate the dataset and are desirable choices for markers. Interpretation of the segregation pattern becomes obvious from the distribution itself as the group of samples with low gene expression values is separated from the group of samples with high expression values by a gap. Simplicity of interrogation allowed us to create a web tool (accessible on the internet 21) to query these classifications so that users can determine whether a given gene is a marker, and find which samples it represents. Also, the tool allows us to find the markers separating two sets of samples of choice.

Figure 1

Distributions of hybridization values for probe sets

Distributions of hybridization values for probe sets. Each histogram depicts the number of replicates (from a total of 241) with a given hybridization value for a given probe set. For illustrative purposes, we display the distribution of hybridization values of three probe sets selected using the gap method as markers corresponding to (a) a known neural stem cell marker (Nestin; probe set 1449022_at), (b) a novel stem cell marker encoding a protein of known function (phospholipase Pla2g7; probe set 1430700_a_at) that we observed to be upregulated in bone marrow mast cell precursors and in undifferentiated mammospheres, and (c) a novel stem cell marker corresponding to an uncharacterized transcript upregulated in undifferentiated V6.5 and J1 murine embryonic stem cells (2410146L05Rik; probe set 1460471_at). Details on these three cases can be obtained through our online webserver [21] by viewing group numbers 73, 497, and 265, respectively. (d) For comparison, we show the distribution for a housekeeping gene (Eef1a1; probe set 1424635_at), a ribosomal translation elongation factor protein, which is always expressed and was not classified as a marker in our analysis. For robustness, the segregation of the samples (indicated by red and blue bars in the three marker distributions for the down and upregulated groups, respectively) is derived by analysis of the global set of patterns (see Materials and methods) and might not correspond perfectly to the distribution observed here. See, for example, a single replicate in the distribution of Nestin, which is left of the gap in the distribution but was (correctly) associated to the upregulated (blue) group.

Properties of the set of potential markers

Classifications contained varying numbers of probe sets. Figure 2a shows that patterns with small numbers of markers are more numerous. The 893 classifications also separated different numbers of samples into groups. Most patterns assigned small numbers of samples to 'upregulated' groups (for example, the gene was highly expressed in three samples versus 80). Over 80% of the patterns separate ten or fewer samples from the remainder of the database as a highly expressed group (Figure 2b). The complete set of putative markers associated with the 893 patterns includes 10,401 probe sets, or approximately 25% of the probes on the microarray, which is intuitively quite a large fraction.

Figure 2

Properties of the set of 893 patterns

Properties of the set of 893 patterns. (a) Number of patterns with a given number of probe sets associated with a score of 90%. (b) Number of patterns with a given number of samples segregated in the high expression group.

We expected that patterns defined by smaller numbers of marker genes would be more likely to include genes important for stem cell related functions. To investigate this, we examined the distribution of known stem cell markers in this set and whether the method was able to select them preferentially within small clusters.

Known stem cell markers in dataset

We investigated whether known stem cell markers were identified in our dataset in order to understand the properties and usefulness of the selected patterns within the context of stem cell research.

By examining the literature, we selected 88 marker genes that represented the variety of stem cell types in our dataset (Table 2). We identified the corresponding entries in the Entrez Gene database for 72 of the 88 marker genes. In seven of the remaining cases we were unable to identify definitively the correct gene because of ambiguity of the provided gene identifier (for example, 'laminin' could indicate one of several possible genes), and in nine cases the identifier could not be related to any entry in the database (for example, 'neuralstemmin'). Of these 72 Entrez Gene IDs, 71 had at least one associated probe set on the Affymetrix MOE430A/B chip set. Our set of 10,401 'potential marker' probe sets, which contains patterns with a maximum of 200 associated probe sets (described under Filtering marker lists by size and number of classified samples, below), contained probe sets for 49 of these 71 marker genes (69%).

Table 2

Stem cell markers

Ref.

Cellular type

Gene name

Entrez Gene

Probe set polled

Pattern

Score

[74]

Differentiated retinal

309L

Rho1D4 (rhodopsin)

D2P4 (rhodopsin)

CHX10

Chx10

1419628_at.A

Not found

PKC

ROM-1

Rom1

1448996_at.A

879

0.942

Photoreceptor specific homeobox

Crx

1418705_at.A

Not found

Muller glia

10E4

[75]

Human central nervous system

CD24-/lo

CD24a

1416034_at.A

573

0.987

CD34-

CD34

1416072_at.A

0.990

CD45-

Ptprc

1422124_a_at.A

427

0.987

Human neuronal lineage

N-CAM

Ncam1

1426864_a_at.A

318

1.000

Neural

CD133

Prom1

1419700_a_at.A

573

0.930

Hematopoietic

CD133

Prom1

1419700_a_at.A

573

0.930

[76]

Proliferating neural

Ki-67

Mki67

1426817_at.A

0.991

[77]

Trophoblast

Cdx2

1422074_at.A

Not found

Ectoderm

Fgf5

1438883_at.B

442

1.000

Neuroectoderm

Isl1

1422720_at.A

0.931

Pluripotent stem cell

Nanog

1429388_at.A

390

0.947

Oct3/4

Pou5f1

1417945_at.A

151

1.000

Rex1

Zfp42

1418362_at.A

547

0.964

Mesoderm

Brachyury

1419304_at.A

232

0.938

[78]

Neural stem cell

Neuralstemmin

ABCG2

Abcg2

1422906_at.A

Not found

LeX/SSEA-1

Fut4

1455843_at.B

771

0.924

Musashi (Msi1)

Msi1

1421409_at.A

Not found

Sox-1

Sox1

1438729_at.B

542

0.958

Sox-2

Sox2

1416967_at.A

0.986

[79]

Muscle specific

MCK (muscle creatine kinase)

Ckm

1417614_at.A

104

1.000

MHC (myosin heavy chain)

Myh4

1427026_at.A

Not found

Myofiber sarcolemma

Dystrophin

Dmd

1417307_at.A

835

0.958

Basal lamina

Laminin

Myogenic lineage

Myf5

1420757_at.A

Not found

MyoD

Myod1

1418420_at.A

1.000

Late myogenic lineage

MRF4

Myf6

1419150_at.A

Not found

Myogenin

Myog

1419391_at.A

104

1.000

Satellite cells

Pax7

1452510_at.A

Not found

[80]

Neural restricted precursors

MAP2

Mtap2

1434194_at.B

108

1.000

Beta3 tubulin

Tubb3

1415978_at.A

Not found

[81]

SKP

Fibronectin

Fn1

1426642_at.A

681

0.984

GAP43

Gap43

1423537_at.A

363

0.959

MAP2

Mtap2

1434194_At.B

108

1.000

Nestin

Nes

1449022_at.A

0.980

p75NTR

Ngfr

1421241_at.A

Not found

Vimentin

Vim

1450641_at.A

770

0.977

[82]

Hematopoietic stem cell

BCRP1

Abcg2

1422906_at.A

Not found

Epidermal side population

BCRP1

Abcg2

1422906_at.A

Not found

alpha6-integrin

Itga6

1422444_at.A

Not found

beta1-integrin

Itgb1

1426918_at.A

0.967

keratin 14

Krt1-14

1460347_at.A

386

0.983

Sca-1

Ly6a

1417185_at.A

637

0.929

CD34-

CD34

1416072_at.A

0.990

E-cadherin

Cdh1

1448261_at.A

0.993

Keratin 19

Krt1-19

1417156_at.A

268

0.989

CD71-

Tfrc

1452661_at.A

845

0.958

[83]

Muller glia; retinal

Gln synthetase

Glul

1426235_a_at.A

573

0.901

Retinal

syntaxin

Pax6

1419271_at.A

374

0.960

rhodopsin

Rho

1425171_at.A

Not found

[84]

Neural

(NeuN) Neuron specific protein

Neuron-specific enolase

Eno2

1418829_a_at.A

Not found

Osteoblasts

Alkaline phosphatase

Akp2

1423611_at.A

Not found

BMP2

Bmp2

1423635_at.A

Not found

BMP4

Bmp4

1422912_at.A

Not found

BMP Receptor 1

Bmpr1a

1425492_at.A

581

0.939

Bmpr1b

1437312_at.B

446

0.917

BMP Receptor 2

Bmpr2

1434310_at.B

602

0.954

PTH receptor

Pthr2

1452129_at.A

Not found

Type 1 collagen

bone sialoprotein

Ibsp

1417484_at.A

Not found

PTH receptor

Pthr1

1417092_at.A

368

0.964

RunX-1

Runx1

1422865_at.A

295

0.982

osteonectin

Sparc

1448392_at.A

581

1.000

osteopontin

spp1

1449254_at.A

737

0.996

General stem cell factor receptor

CD117

Kit

1459588_at.B

243

0.912

Muscle

merosin

Lama2

1426285_at.A

799

0.986

Cartilage related extracellular matrix

aggrecan

Agc1

1449827_at.A

Not found

collagen II

collagen IV

PRELP

Prelp

1416322_at.A

231

0.958

Adipose derived stem cell

CD49d+

Itga4

1421194_at.A

Not found

CD106-

[85]

Mammary stem cell

Bmi-1

Bmi1

Not on array

p21

Cdkn1a

1424638_at.A

400

0.967

CD49f

Itga6

1422444_at.A

Not found

Cytokeratin 19

Krt1-19

1417156_at.A

268

0.989

Sca-1

Ly6a

1417185_at.A

637

0.930

Musashi (Msi1)

Msi1

1421409_at.A

Not found

Cytokeratin 5/6

[86]

Neural enrichment

CD24

CD24a

1416034_at.A

573

0.990

Skin

CD29

Itgb1

1426918_at.A

0.970

[87]

Neural

GFAP

Gfap

1426508_at.A

249

0.924

Neurofilament

Nefl

1426255_at.A

151

0.935

Nestin

Nes

1449022_at.A

0.980

Photoreceptors

Recoverin

Rcvrn

1450215_at.A

Not found

Epithelial lineage

Cytokeratin (904-clone 34betaB4?)

CK18

Krt1-18

1448169_at.A

0.917

AE1

Slc4a1

1416464_at.A

1.000

[88]

Epithelial lineage

AE3

Slc4a3

1418485_at.A

391

0.931

Note that some markers may be included in multiple rows of the table (as many as indicated in the right-most column), because a number of genes have been identified as markers of more than one cell type (for example, Abcg2 for hematopoietic and neural cells [78,82], or Nestin for neural and skin-derived precursors [SKPs] [81,87]). Some of the markers could not be assigned to an Entrez gene either because the marker name was ambiguous (indicated by '^a') or was absent from the database of gene names (indicated by '?'). Patterns can be examined online via the web server [21]. All patterns for which a polled probe set is a marker can also be examined at the web server. Note that the chip containing the probe set is indicated by '.A' or '.B' appended to the identifier, because some probe sets are included in both the MOE430A and the MOE430B chips.

As previously observed, we obtained numerous patterns that segregate small subsets of samples and that are defined by small numbers of genes. To test our hypothesis that these would tend to contain relevant genes (in this case, genes useful for characterizing stem cells), we examined the recall and precision of our method for the 71 known markers as the maximum marker list length was reduced from 200 to 3 (Figure 3). Limiting the list to patterns defined by 63 or fewer markers reduced the total number of probe sets assigned marker roles from 10,401 to 5,848 (44% reduction) while losing only four known stem cell markers, representing a recall rate of 63% (45 out 71; point marked with a circle in Figure 3). This supports our theory that marker genes are more often contained in small clusters.

Figure 3

Precision/Recall curves for genes selected by the gap method and k-means

Precision/Recall curves for genes selected by the gap method and k-means. Precision/recall curves in red are associated with the gap method, and the point marked with a circle denotes the precision/recall associated with patterns associated with 63 probe sets or less. Gray curves show precision/recall for all replicates of k-means clustering, and the average expected precision/recall curve is shown in black. Recall values are based on the 71 stem cell markers defined in Table 2, whereas precision is the fraction of marker genes identified in the total number of predicted marker genes.

The 705 patterns defined by 63 or fewer markers segregated a mean of nine samples (median 5) in the upregulated group and were associated with a mean of 21 markers (median 15). This set associates 5,848 probe sets (4,449 genes) with at least one pattern, accounting for approximately 13% of the probe sets on the MOE430 microarray platform. We propose that many of those can be developed into useful markers. We define this as 'selected marker' set.

We compared the performance of our method with a popular method for analysis of gene expression data, namely k-means, which is a standard clustering algorithm (see Materials and methods, below). Both methods performed similarly in producing groups of genes that are expected to be enriched for stem cell markers (Figure 3). However, our method differs from a clustering algorithm in that we identify markers that segregate sets of samples whereas clustering algorithms group markers with similar expression patterns. Accordingly, the groups of associated markers produced by the gap method were somewhat different from the clusters obtained using k-means (mean overlap of 69.8%).

Overview of the selected marker set

To illustrate the variety of patterns identified, Figure 4 shows the expression patterns of the 49 probe sets that represent previously known stem cell marker genes identified by the algorithm (yellow), together with another 1,252 genes that were assigned a perfect score by the algorithm (blue). All major divisions in the dataset appear clearly defined, with samples related to one of hematopoietic, sphere 22, or embryonic sample types forming major groups. For example, hematopoietic samples define many patterns with probe sets identified uniquely within them.

Figure 4

Heatmap indicating the distribution of patterns for markers

Heatmap indicating the distribution of patterns for markers. The horizontal axis shows 241 mouse samples used in this study. The vertical axis shows patterns for 1,301 markers either predicted to have high reliability (scoring 100%, n = 1,252; blue) or probe sets belonging to genes ascribed marker roles based on evidence in the literature (n = 49; yellow). Rows were clustered and diagonalized. Vertical separators were used to distinguish major sample cell types. Sample identities are grouped as follows: embryonic, blue; P19 embryonal carcinoma, orange; fibroblasts, purple; spheres, yellow; hematopoietic, red. Gene names are indicated only for the 49 stem cell markers.

Gene Ontology statistics of the selected marker set

Since the selection of markers was done without reference to the identity of the samples, we would expect to find not just stem cell markers but also general markers of cell and tissue identity (for example, distinguishing differentiated blood cells from epithelial cells). To investigate in general terms the function of the genes that were selected in the marker set, we collected the 5,848 probe sets defining the 705 selected patterns (with 3 to 63 associated probe sets). These corresponded to 4,449 Entrez Gene IDs, for which we determined the over-representation of Gene Ontology (GO) annotations. Using the 'potential marker' set of 7,478 genes (which defined groups with up to 200 probe sets) as a reference, we found several significantly enriched functional categories (Table 3).

Table 3

Gene Ontology terms of marker sets

All

Selected set

High in differentiated

GOID

7,478

4,449

222

Total

976

727

2.85 × 10^-22

1.60 × 10^-16

Extracellular region