Biomedical Engineering Reference
In-Depth Information
Table 4.23
Comparison of selected predictors using highest
R
i
for mutated network
Gene
Correct Predictor
Selected Predictor
Selected Predictor
(linear)
(sigmoid)
x
2
Rb
x
1
,
x
4
,
x
5
,
x
9
x
1
,
x
4
x
1
,
x
4
,
x
9
x
3
E2F
x
2
,
x
5
,
x
9
x
2
,
x
5
,
x
9
x
2
,
x
5
,
x
9
x
4
CycE
x
2
,
x
3
x
2
,
x
3
x
2
,
x
3
x
5
CycA
x
2
,
x
3
,
x
6
,
x
7
,
x
8
x
2
,
x
6
x
2
,
x
3
,
x
6
,
x
8
x
6
Cdc
20
x
9
x
9
x
9
x
7
Cdh
1
x
5
,
x
6
,
x
9
x
6
,
x
9
x
5
,
x
6
,
x
9
x
8
UbcH
10
x
5
,
x
6
,
x
7
,
x
9
x
7
x
7
x
9
CycB
x
6
,
x
7
x
6
,
x
7
x
6
,
x
7
correct
4
5
Table 4.24
Normalized gene expression lines for melanoma network
PIRIN
S
100
P
RET
1
MART
1
HADHB
STC
2
WNT
5A
x
1
x
2
x
3
x
4
x
5
x
6
x
7
0.002
0.020
0.275
0.010
0.227
1.000
1.000
0.056
0.000
0.239
0.011
0.430
0.583
0.318
0.387
0.006
0.440
0.008
0.070
0.055
0.511
0.137
0.147
0.156
0.005
0.227
0.014
0.026
0.222
0.168
0.532
0.141
0.395
0.000
0.000
0.751
0.016
0.349
0.152
0.564
0.197
0.005
0.401
0.411
0.028
0.778
0.663
0.016
0.018
4.3.3
Melanoma Network
Based on the results from the mutated mammal cell-cycle network using synthetic
data, we evaluate our predictor ranking and selection algorithms to the actual data
from a melanoma network study [
17
]. This study identified seven genes
PIRIN
,
S
100
P
,
RET
1,
MART
1,
HADHB
,
STC
2, and
WNT
5
A
, to be closely related with the
metastasis of melanoma. From [
17
], 31 gene expression lines (states) were measured,
and then reduced to seven distinct lines (shown in Table
4.24
).
We apply our ranking algorithm on each gene
x
1
to
x
7
in the melanoma network and
show the results in Table
4.25
through
4.31
respectively. From [
4
] which inferred the
predictor set on the binary valued melanoma gene expression data, we observed that
one of the attractor cycle ordering is common to the majority of predictor selection
results. This attractor cycle ordering (
S
1
→
S
1
,
S
2
→
S
3
→
S
2
,
S
4
→
S
4
,
S
5
→
S
6
→
S
5
,
S
7
→
S
7
) is used along with the normalized gene expression data in our
algorithm.
Due to the limited sample size of 7 states, we limit the algorithm to search pre-
dictors up to 3 input genes. Similar to assumptions made in the mutated mammal
network, this is a reasonable constraint as most genes have been observed to be
regulated by few input genes.
From the predictor rankings, we apply our selection method based on MSE and
resolution ratio and find the best predictors for each gene. The selected predictors
