Transcript Microarray Image Data Analysis

Gene Discovery from Microarray
Images
陳朝欽、 高成炎、張春梵
ARCNTU, NTU-Hospital
[email protected]
[email protected]
Project#: 93-EC-17-A-19-S1-0016
Motivation and Data Acquisition
• Parts of our current works attempt to
investigate and discover “a subset of
genes” related to some specific diseases
such as Hepatoma and Gastric Cancers
by microarray experiments. Hence, we
collect data from cDNA microarray images
which are “spot signal intensities” via a
sequence of biological experiments
A Paradigm for Microarray Image
Data Analysis
Outline
•
•
•
•
•
•
Microarray Image Data Acquisition
Gridding for Image Segmentation
Normalization from MA-Plot
Finding Differentially Expressed Genes
Finding Discriminative Genes
Performance Evaluation by Dendrogram
and K-means Algorithms
A Look at a Microarray Slide
Examples of Microarray Images
Gridding for Spot Segmentation
Gridding for a Block of 30*9 Spots
Spot Feature Computation
• Cy3 (for Column 1)
639
54879
5980
1984
324
910
2153
236
• Cy5 (for Column 6)
104
52858
567
189
36
1489
5083
407
M-A plot and Piecewise
Normalization
Normalized Ratio from MA-Plot
Pre-Processing / Normalization
• Due to the process of measurements or
some unavoidable factors, “Raw Data”
directly collected from experiments may
contain noise and may have different
scales, or have missing items. Thus, a
pre-processing step for filtering out some
inappropriate data, or normalization may
be done.
Spot Features for Gene
Discovery
Cy3
201
520
28276
4072
14807
1058
572
Cy5
67
153
21747
6324
690
1451
524
M=(log2Cy3 − log2Cy5)
A= (log2Cy3+log2Cy5)/2
Program compustt.c
computes spot features
and pieceline.c does
normalization and
maplot.c does M-A plot
Microarray Pattern Analysis
• Microarrays consisting of 13574 effected
genes from 18564 in a chip with tumor
dyed in Cy3 and normal dyed in Cy5
• 12 HCV, 27 HBV, 1 HCV+HBV, 4 neither
HCV nor HBV patients
• Criterion for Differentially Expressed is
defined as log2(Lowess normalized ratio of
Cy3/Cy5) is greater than T (↑) or less than
-T (↓)
Feature Selection/Extraction (1)
• Given a set of N patterns from K
categories (K=2, a problem of dichotomy)
with Ni , 1≤ i ≤ K, patterns belonging to
category i, each pattern consists of M
redundant features, e.g., a microarray
can be represented as a pattern consisting
of 13574 features corresponding to 13574
effected genes. The goal is to select a
small subset of features for “Recognition”
Feature Selection/Extraction (2)
• Given a set of N patterns from K
categories (K=2, a problem of dichotomy)
with Ni , 1≤ i ≤ K, patterns belonging to
category i. The goal of extraction is to
transform an M-dimensional pattern into
an m-dimensional pattern with m<<M for
classification. A selected feature preserves
the original meaning but an extraction
usually does not preserve the original one.
16 Most Discriminative Genes to
distinguish HCV from HBV [YCT39]
Index
13796
7197
2918
8495
11189
11087
9443
9546
Accession#
U35376
BG259957
BI520001
AJ012159
AB008549
BC006496
CAC51145
X52125
Index
16144
16496
17213
14579
587
113
17215
16760
Accession#
AK024601
Y00083
BC007437
BC011568
AF386492
Y16961
AF195766
AI022747
Next 16 Most Discriminative Genes
to distinguish HCV from HBV
Index
5947
4885
11291
1262
8055
10965
4164
8088
Accession#
BG207354
AK021818
AF155110
BI861005
AJ224741
AAF36120
NM_000423
BC000187
Index
7353
5434
12727
14993
4182
5341
10052
8140
Accession#
AF070641
AB050785
AB062987
AA974308
AI970531
X65882
AB011542
AK026068
32 Discriminative Genes by
Fisher’s Ratios for a Dendrogram
32 Discriminative Genes by
Chuang+Kao’s for a Dendrogram
Dendrogram from Chen’s 32
Most Discriminative Genes [CC39]
Dendrogram from Genasia’s 32
Most Discriminative Genes
K-means Clustering Results by
using 32 Best Discriminative Genes
• G45 from Genasia: distortion 341.26
1222221222 2211111111 111111111111111111
• X47 from C. Chen: distortion 302.33
1222221222 2211111111 112111111111111111
• Y48 by Fisher’s Ratio on YCT39: distortion 307.49
1222221222 2211111111 112111111111111111
• PY50 by Chuang+Kao’s on YCT39: distortion 290.06
2222222222 2211211111 112111111111111111
Leave-one-out errors by 1-nn :
4, 3, 2, 1 (/39)
Leave-one-out errors by Fisher : 15, 7, 8, 9 (/39)
Up (Down) Regulated Genes for
Gastric Cancers
• 5 Advanced and 5 Early Stage of Patients
with Gastric Cancer
• We find the following genes which can
completely discriminate Patients of
“Advanced Stage” from “Early Stage”
under clinical diagnosis
Dengrogram for Gastric Patients
Top 16 Discriminative Genes for
Advanced and Early Stages
Index
15843
12994
18370
2070
1118
9661
2017
1128
Accession#
AF316855
BF868865
BC002996
AK021788
BC000249
AP000350
U53530
AF035281
Index
8728
494
10990
342
10425
6052
170
1016
Accession#
AL591713
AB014526
L77570
BC007848
BG745129
AF073362
AK000278
BF526386
Thank You
• http://www.bioinfo.ntu.edu.tw
• http://www.cs.nthu.edu.tw/~cchen
• Tel: (02) 2312 3456 ~ 5917
• Tel: (02) 2362 5336 ~ 418
• Tel: (03) 573 1078