Differential Gene Expression in

1
Differential Gene Expression in Metastatic
Medulloblastoma
Kevin Lai Mandy Tam Sharon Karackattu
2
The Anatomy of the Human Brain
3
Tumor Characteristics

• Embryonal neuroepithelial tumor
• Highly cellular, soft friable tumors
• Little cytoplasm
• Degrees of glial or neuroblastic differentiation
  - primitive cell of origin with bipotential
  capacity
• Desmoplastic - tumors with a large stromal
  component

4
Medulloblastoma (HE x160). Hypercellular tumour
composed of small cells with a dense nucleus and
scant cytoplasm. Embryonal tumour of essentially
neuronal lineage. The patient in this case
presented with Turcot syndrome (an inherited
autosomal dominant disorder characterized by the
association of colorectal tumours and a
glioblastoma or a medulloblastoma).
5
Medulloblastoma demonstrating infiltration into
the choroid plexus
Medullomyoblastoma (MT x400). Cross-striation
can be clearly seen in the bundles of small
muscle cells.
Medullomyoblastoma (MT x160). This typical
medulloblastoma contains small striated muscle
cells, generally arranged in bundles.
6
Desmoplastic medulloblastoma (HE x40). Clear
islands with the characteristics of
medulloblastoma and dark, collagen-rich
areas containing small tumour cells. Collagen
production is generally considered a reaction of
the leptomeninges, which are frequently invaded
by medulloblastoma.
7
Desmoplastic medulloblastoma (silver reticulin
impregnation x40). The reticulin network spares
the clear islands that vary greatly in size.
Neuroblastic medulloblastoma (cerebellar
neuroblastoma) (MT x40). The islands of tumour
cells form both typical rosettes and
long fibrillary bundles, denoting neuro- blastic
differentiation.
8
Large cell medulloblastoma (medulloblastoma with
neuronal differentiation) (cresyl violet x100).
Scattered among the small cells of this
medulloblastoma are clearer areas, composed of
small nerve cells with a vesicular nucleus and a
large nucleolus. Scant cytoplasm, very rich in
Nissl substance.
9
Demography

• Cerebellar tumor - cerebellar vermis
• 10-20 of primary CNS neoplasms in children
• Peak age of incidence is 5 years
• Most tumors occur in first 10 years of life
• 21 male to female ratio

10
Metastatic Potential

• Medulloblastoma has highest tendency for
  extraneural spread of all pediatric CNS neoplasms
• Bone is most common site for metastases (80)
• Bone marrow, lymph nodes, liver and lungs are
  also common organs for metastases

11
Chang Staging System for Medulloblastoma
12
Chang Staging System for Medulloblastoma
13
Gene Expression in Medulloblastoma

• C-myc is amplified or rearranged in 5 of 7
  medulloblastoma cells lines
• In fresh tumor specimens, only 3 of 43 samples
  showed c-myc amplification
• Certain in vitro tumor markers are of limited
  clinical relevance

14
Applications of Gene Expression and Clustering
Analysis

• Medulloblastoma is one of the most chemosensitive
  primary CNS tumors
• Subclassifying and identifying relevant marker
  genes may uncover potential drug targets

15
In vivo selection scheme from Clark et al.,
Nature 406, 532-5
(melanoma cells)
16
Informative features that distinguish metastatic
from non-metastatic melanomas in mice (from Clark
et al.)
Fibronectin RhoC Thymosin ?4 t-PA Angiopoetin
1 IEX-1/Glu96 RTP/NDR1 Fibromodulin Hsp70
IL13 receptor, ?2 subunit Sec61? snRNP,
polypetide C Collagen I?2 UBE21 KIAA0156 TGF?
superfamily Surfactant protein C Matrix Gla
protein
17
Main question
In humans, are the same genes important in
metastasis? i.e., Do metastatic human tumor
cells surrounded by mouse cells express the same
genes as human tumor cells surrounded by human
cells?
18
Medulloblastoma samples (60)
Classic subtype (40)
Desmoplastic subtype (12)
Other subtypes (8)
Affymetrix chip data contributed by Todd Golub
M0/M1 (28)
M3/M4 (8)
Other (4)
Log transform data and normalize using whole chip
signal
19
Treatment of data

• 1) Take only classic samples, split into 2 groups
  by Chang stage
• M0/M1 vs. M3/M4
• (low metastatic potential) (high
  metastatic potential)
• 28 samples 8 samples
• 2) Divide into training set and test set
• each 14 M0/M1 samples and 4 M3/M4 samples
• 3) Use training set to select informative
  features, by likelihood ratio test
• calculate mean and variance of each gene for each
  class
• select genes that are most different
• arbitrarily set threshold for T to get around 10
  genes
• too many genes will introduce unnecessary noise
• too few genes will not provide enough
  information for class distinction
• 4) Cluster the samples using informative genes to
  verify feature selection algorithm worked
• samples should cluster according to their class
  labels

20
Are the genes identified in Clark et al. also
differentially expressed in different stages of
medulloblastoma?

• Take the genes in the paper as informative
  features
• Do the genes cluster M0/M1 vs. M3/M4 samples
  correctly?
• Do the genes correctly predict class M0/M1 vs.
  class M3/M4?

21
Informative genes that distinguish M0/M1 vs.
M3/M4 medulloblastoma
Frizzled gene product mRNA UDP-Galactose 4
epimerase (GALE) gene Serum protein MSE55 GTP
binding protein (ARL3) mRNA RIG-G
mRNA Adenosine triphosphatase DNM1 Dynamin
1 Von Willebrand Factor Precursor KRT8 Keratin
8 GPD2 Glycerol-3-phosphate dehydrogenase 2
(mitochondrial)
22
Hierarchical clustering of M0/M1 vs. M3/M4
samples using informative genes from training set
Chang Stage M3/M4
No dot Chang Stage M0/M1
23
K-means clustering of M0/M1 vs. M3/M4 samples
using informative genes from training set
24
Prediction by weighted voting
Adapted from Golub et al., Science 286,
531-7 Weight P(g,c) (?1-?2)/(?1?2) where g
expression vector, c class vector Decision
boundary b (?1?2)/2 Vote by each gene Vg
P(g,c) (x-b) weight(g) distance(x,b) Sum all
negative and positive votes Winner is the class
receiving the larger total vote
25
Classification of medulloblastoma samples by
Chang stage
Training Set 14 M0/M1, 4 M3/M4 Prediction Us
ing features selected from the training
set Correct predictions using an independent
set 12 of 14 M0/M1 2 of 4 M3/M4 -------------
-------- 14 of 18 Total
26
Informative features that distinguish metastatic
from non-metastatic melanomas in mice (from Clark
et al.)
Fibronectin RhoC Thymosin ?4 t-PA Angiopoetin
1 IEX-1/Glu96 RTP/NDR1 Fibromodulin Hsp70
IL13 receptor, ?2 subunit Sec61? snRNP,
polypetide C Collagen I?2 UBE21 KIAA0156 TGF?
superfamily Surfactant protein C Matrix Gla
protein
27
(No Transcript)
28
(No Transcript)
29
Classification of medulloblastoma samples by
Chang stage
Training Set 14 M0/M1, 4 M3/M4 Prediction Us
ing all 18 genes identified by Clark et
al. Correct predictions using an independent
set 10 of 14 M0/M1 0 of 4 M3/M4 -------------
-------- 10 of 18 Total
30
Genes selected for 4-gene predictor from Clark et
al.
IL-13 receptor, ?2 subunit snRNP, polypeptide
C Collagen I?2 Surfactant protein C
31
(No Transcript)
32
(No Transcript)
33
Classification of medulloblastoma samples by
Chang stage
Training Set 14 M0/M1, 4 M3/M4 Prediction Us
ing 4 of the 18 genes identified in Clark et
al. Correct predictions using an independent
set 11 of 14 M0/M1 2 of 4 M3/M4 -------------
-------- 13 of 18 Total
34
Discussion 1.) Informative genes in
classification of metastatic vs. non-metastatic
medulloblastomas are not the same as those found
by Clark et al. for melanomas. ? Cells
surrounding tumor cells (mouse or human) affect
gene expression ? Genes important for metastasis
are different in different cell types ? Genetic
variation among patients ? Class distinction is
not as clear as ALL vs. AML ? Not enough M4
samples (only 1 of 8 M3/M4 class)- i.e., not
enough highly metastatic samples to select
features correctly
35
Discussion 2.) Prediction by weighted voting was
unsuccessful. ? Too few samples in one class-
i.e., only 8 M3/M4 samples vs. 28 M0/M1 total, so
training set only has 4 M3/M4 samples ? To verify
that the prediction algorithm worked Used all 36
samples to select informative genes. Then,
divided into a training set and a test set.
Correct predictions 13 of 14 M0/M1 samples and 4
of 4 M3/M4 samples.
36

• Future Work
• Cross validation
• Subtype classification and prediction
• classic vs. desmoplastic
• T stage
• Look at different human cancer types

37
Acknowledgements Nicola!! Todd Golub
38
What have we learned?