Information Technology Reference
In-Depth Information
Table 1.
Results of a training session showing the weights attributed to each method
indicating the relative performance of this method in comparison to others. The
three columns show results in predicting helix,ҏsheet or both. (set consisted of 1000
randomly selected proteins).
number
name
E-
sheet
D-
helix
both
1
ChouFasman
0.7
0.9
0.7
2
PREDATOR
2
10.0
9.6
10.0
3
Garnier
0.4
0.6
0.3
4
Simpa96
1.4
1.6
1.2
5
GOR4
0.7
0.6
0.6
6
DSC
0.7
0.6
0.7
7
DSC-l
0.5
0.6
0.5
8
CFpred
0.4
0.6
0.3
9
PHD
6.1
10.0
9.4
10
NNpredict
0.6
0.6
0.5
The results in table 1 also show that there is a difference in performance in
predicting D-helix or E-sheet between different methods. This effect can only be clearly
seen in some of the methods (for instance the method PHD). In the following experiments
the difference between prediction of D-helix or E-sheet is no longer taken into account. In
these experiments the weights are for the prediction of both D-helix and E-sheet.
Next a neural network training session was done on a training set of 6000 proteins.
The weights acquired in this session can be seen in table 2 (column labelled “weight 1”).
Again the methods PREDATOR 2 and PHD are getting far better scores compared to other
methods. To investigate a possible suppressive effect of the two aforementioned methods
on the weights of other methods another experiment was performed. Data of the methods
PREDATOR 2 and PHD was excluded from the training. The test set consisted of 2000
proteins and again a neural network was trained to find the weights for the remaining
methods. The results are shown in table 3. Although the weights seem to be much higher,
they are not. This is because the weights represent the relative weight for a method. The
table shows us that indeed PREDATOR 2 and PHD suppress other methods, but by leaving
them out there is no real change on our views of the other methods. The method, which is
the best now in comparison with the other methods, Simpa96, was also better than these
methods in former experiments (except in comparison with PREDATOR 2 and PHD of
course). Also the methods with bad scores, like Garnier and CFpred, stay at the bottom.
In order to investigate whether the weights from the last experiment have changed
at the same rate for all the remaining methods in comparison with one and another, the ratio
of the weights from the experiments with and without the programs PREDATOR 2 and
PHD is calculated.
When the ratios of the weights of the last two experiments are compared a striking
difference between better valued methods from the last experiment and less valued methods
emerges. Table 4 shows that Simpa96, GOR 4, DSC and DSC-l all have a ratio of
approximately 9. Chou-Fasman, Garnier, Cfpred and NNpredict all have a ratio of about 4
to 5. Assuming that the neural network favours better methods by assigning higher weights,
